diff options
Diffstat (limited to 'src/gallium/drivers/cell')
48 files changed, 7222 insertions, 2272 deletions
diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h index cb0631baf5..1f6860da11 100644 --- a/src/gallium/drivers/cell/common.h +++ b/src/gallium/drivers/cell/common.h @@ -49,6 +49,15 @@ } + +#define JOIN(x, y) JOIN_AGAIN(x, y) +#define JOIN_AGAIN(x, y) x ## y + +#define STATIC_ASSERT(e) \ +{typedef char JOIN(assertion_failed_at_line_, __LINE__) [(e) ? 1 : -1];} + + + /** for sanity checking */ #define ASSERT_ALIGN16(ptr) \ ASSERT((((unsigned long) (ptr)) & 0xf) == 0); @@ -64,9 +73,13 @@ #define ROUNDUP16(k) (((k) + 0xf) & ~0xf) -#define CELL_MAX_SPUS 6 +#define CELL_MAX_SPUS 8 #define CELL_MAX_SAMPLERS 4 +#define CELL_MAX_TEXTURE_LEVELS 12 /* 2k x 2k */ +#define CELL_MAX_CONSTANTS 32 /**< number of float[4] constants */ +#define CELL_MAX_WIDTH 1024 /**< max framebuffer width */ +#define CELL_MAX_HEIGHT 1024 /**< max framebuffer width */ #define TILE_SIZE 32 @@ -94,48 +107,106 @@ #define CELL_CMD_STATE_BIND_VS 18 #define CELL_CMD_STATE_FRAGMENT_PROGRAM 19 #define CELL_CMD_STATE_ATTRIB_FETCH 20 -#define CELL_CMD_VS_EXECUTE 22 -#define CELL_CMD_FLUSH_BUFFER_RANGE 23 +#define CELL_CMD_STATE_FS_CONSTANTS 21 +#define CELL_CMD_STATE_RASTERIZER 22 +#define CELL_CMD_VS_EXECUTE 23 +#define CELL_CMD_FLUSH_BUFFER_RANGE 24 +#define CELL_CMD_FENCE 25 +/** Command/batch buffers */ #define CELL_NUM_BUFFERS 4 #define CELL_BUFFER_SIZE (4*1024) /**< 16KB would be the max */ #define CELL_BUFFER_STATUS_FREE 10 #define CELL_BUFFER_STATUS_USED 20 +/** Debug flags */ +#define CELL_DEBUG_CHECKER (1 << 0) +#define CELL_DEBUG_ASM (1 << 1) +#define CELL_DEBUG_SYNC (1 << 2) +#define CELL_DEBUG_FRAGMENT_OPS (1 << 3) +#define CELL_DEBUG_FRAGMENT_OP_FALLBACK (1 << 4) +#define CELL_DEBUG_CMD (1 << 5) +#define CELL_DEBUG_CACHE (1 << 6) -#define CELL_DEBUG_CHECKER (1 << 0) -#define CELL_DEBUG_SYNC (1 << 1) +#define CELL_FENCE_IDLE 0 +#define CELL_FENCE_EMITTED 1 +#define CELL_FENCE_SIGNALLED 2 +#define CELL_FACING_FRONT 0 +#define CELL_FACING_BACK 1 +struct cell_fence +{ + /** There's a 16-byte status qword per SPU */ + volatile uint status[CELL_MAX_SPUS][4]; +}; -/** Max instructions for doing per-fragment operations */ -#define SPU_MAX_FRAGMENT_OPS_INSTS 64 +#ifdef __SPU__ +typedef vector unsigned int opcode_t; +#else +typedef unsigned int opcode_t[4]; +#endif + +/** + * Fence command sent to SPUs. In response, the SPUs will write + * CELL_FENCE_STATUS_SIGNALLED back to the fence status word in main memory. + */ +struct cell_command_fence +{ + opcode_t opcode; /**< CELL_CMD_FENCE */ + struct cell_fence *fence; + uint32_t pad_[3]; +}; /** * Command to specify per-fragment operations state and generated code. + * Note that this is a variant-length structure, allocated with as + * much memory as needed to hold the generated code; the "code" + * field *must* be the last field in the structure. Also, the entire + * length of the structure (including the variant code field) must be + * a multiple of 8 bytes; we require that this structure itself be + * a multiple of 8 bytes, and that the generated code also be a multiple + * of 8 bytes. + * + * Also note that the dsa, blend, blend_color fields are really only needed + * for the fallback/C per-pixel code. They're not used when we generate + * dynamic SPU fragment code (which is the normal case), and will eventually + * be removed from this structure. */ struct cell_command_fragment_ops { - uint64_t opcode; /**< CELL_CMD_STATE_FRAGMENT_OPS */ + opcode_t opcode; /**< CELL_CMD_STATE_FRAGMENT_OPS */ + + /* Fields for the fallback case */ struct pipe_depth_stencil_alpha_state dsa; struct pipe_blend_state blend; - unsigned code[SPU_MAX_FRAGMENT_OPS_INSTS]; + struct pipe_blend_color blend_color; + + /* Fields for the generated SPU code */ + unsigned total_code_size; + unsigned front_code_index; + unsigned back_code_index; + /* this field has variant length, and must be the last field in + * the structure + */ + unsigned code[0]; }; /** Max instructions for fragment programs */ -#define SPU_MAX_FRAGMENT_PROGRAM_INSTS 128 +#define SPU_MAX_FRAGMENT_PROGRAM_INSTS 512 /** - * Command to send a fragment progra to SPUs. + * Command to send a fragment program to SPUs. */ struct cell_command_fragment_program { - uint64_t opcode; /**< CELL_CMD_STATE_FRAGMENT_PROGRAM */ + opcode_t opcode; /**< CELL_CMD_STATE_FRAGMENT_PROGRAM */ uint num_inst; /**< Number of instructions */ + uint32_t pad[3]; unsigned code[SPU_MAX_FRAGMENT_PROGRAM_INSTS]; }; @@ -145,10 +216,21 @@ struct cell_command_fragment_program */ struct cell_command_framebuffer { - uint64_t opcode; /**< CELL_CMD_FRAMEBUFFER */ + opcode_t opcode; /**< CELL_CMD_STATE_FRAMEBUFFER */ int width, height; void *color_start, *depth_start; enum pipe_format color_format, depth_format; + uint32_t pad_[2]; +}; + + +/** + * Tell SPUs about rasterizer state. + */ +struct cell_command_rasterizer +{ + opcode_t opcode; /**< CELL_CMD_STATE_RASTERIZER */ + struct pipe_rasterizer_state rasterizer; }; @@ -157,9 +239,10 @@ struct cell_command_framebuffer */ struct cell_command_clear_surface { - uint64_t opcode; /**< CELL_CMD_CLEAR_SURFACE */ + opcode_t opcode; /**< CELL_CMD_CLEAR_SURFACE */ uint surface; /**< Temporary: 0=color, 1=Z */ uint value; + uint32_t pad[2]; }; @@ -206,7 +289,7 @@ struct cell_shader_info #define SPU_VERTS_PER_BATCH 64 struct cell_command_vs { - uint64_t opcode; /**< CELL_CMD_VS_EXECUTE */ + opcode_t opcode; /**< CELL_CMD_VS_EXECUTE */ uint64_t vOut[SPU_VERTS_PER_BATCH]; unsigned num_elts; unsigned elts[SPU_VERTS_PER_BATCH]; @@ -218,7 +301,7 @@ struct cell_command_vs struct cell_command_render { - uint64_t opcode; /**< CELL_CMD_RENDER */ + opcode_t opcode; /**< CELL_CMD_RENDER */ uint prim_type; /**< PIPE_PRIM_x */ uint num_verts; uint vertex_size; /**< bytes per vertex */ @@ -227,44 +310,51 @@ struct cell_command_render float xmin, ymin, xmax, ymax; /* XXX another dummy field */ uint min_index; boolean inline_verts; + uint32_t pad_[1]; }; struct cell_command_release_verts { - uint64_t opcode; /**< CELL_CMD_RELEASE_VERTS */ + opcode_t opcode; /**< CELL_CMD_RELEASE_VERTS */ uint vertex_buf; /**< in [0, CELL_NUM_BUFFERS-1] */ + uint32_t pad_[3]; }; struct cell_command_sampler { - uint64_t opcode; /**< CELL_CMD_STATE_SAMPLER */ + opcode_t opcode; /**< CELL_CMD_STATE_SAMPLER */ uint unit; struct pipe_sampler_state state; + uint32_t pad_[1]; }; struct cell_command_texture { - uint64_t opcode; /**< CELL_CMD_STATE_TEXTURE */ + opcode_t opcode; /**< CELL_CMD_STATE_TEXTURE */ + uint target; /**< PIPE_TEXTURE_x */ uint unit; - void *start; /**< Address in main memory */ - ushort width, height; + void *start[CELL_MAX_TEXTURE_LEVELS]; /**< Address in main memory */ + ushort width[CELL_MAX_TEXTURE_LEVELS]; + ushort height[CELL_MAX_TEXTURE_LEVELS]; + ushort depth[CELL_MAX_TEXTURE_LEVELS]; }; -/** XXX unions don't seem to work */ -/* XXX this should go away; all commands should be placed in batch buffers */ -struct cell_command +#define MAX_SPU_FUNCTIONS 12 +/** + * Used to tell the PPU about the address of particular functions in the + * SPU's address space. + */ +struct cell_spu_function_info { -#if 0 - struct cell_command_framebuffer fb; - struct cell_command_clear_surface clear; - struct cell_command_render render; -#endif - struct cell_command_vs vs; -} ALIGN16_ATTRIB; + uint num; + char names[MAX_SPU_FUNCTIONS][16]; + uint addrs[MAX_SPU_FUNCTIONS]; + char pad[12]; /**< Pad struct to multiple of 16 bytes (256 currently) */ +}; /** This is the object passed to spe_create_thread() */ @@ -273,11 +363,13 @@ struct cell_init_info unsigned id; unsigned num_spus; unsigned debug_flags; /**< mask of CELL_DEBUG_x flags */ - struct cell_command *cmd; + float inv_timebase; /**< 1.0/timebase, for perf measurement */ /** Buffers for command batches, vertex/index data */ ubyte *buffers[CELL_NUM_BUFFERS]; uint *buffer_status; /**< points at cell_context->buffer_status */ + + struct cell_spu_function_info *spu_functions; } ALIGN16_ATTRIB; diff --git a/src/gallium/drivers/cell/ppu/Makefile b/src/gallium/drivers/cell/ppu/Makefile index b28f4c5c31..c92f8e5cba 100644 --- a/src/gallium/drivers/cell/ppu/Makefile +++ b/src/gallium/drivers/cell/ppu/Makefile @@ -24,6 +24,7 @@ SOURCES = \ cell_clear.c \ cell_context.c \ cell_draw_arrays.c \ + cell_fence.c \ cell_flush.c \ cell_gen_fragment.c \ cell_gen_fp.c \ @@ -38,8 +39,7 @@ SOURCES = \ cell_texture.c \ cell_vbuf.c \ cell_vertex_fetch.c \ - cell_vertex_shader.c \ - cell_winsys.c + cell_vertex_shader.c OBJECTS = $(SOURCES:.c=.o) \ @@ -54,6 +54,9 @@ INCLUDE_DIRS = \ $(CC) -c $(INCLUDE_DIRS) $(CFLAGS) $< -o $@ +.c.s: + $(CC) -S $(INCLUDE_DIRS) $(CFLAGS) $< -o $@ + default: $(CELL_LIB) diff --git a/src/gallium/drivers/cell/ppu/cell_batch.c b/src/gallium/drivers/cell/ppu/cell_batch.c index 16882c0129..fe144f8b84 100644 --- a/src/gallium/drivers/cell/ppu/cell_batch.c +++ b/src/gallium/drivers/cell/ppu/cell_batch.c @@ -28,6 +28,7 @@ #include "cell_context.h" #include "cell_batch.h" +#include "cell_fence.h" #include "cell_spu.h" @@ -42,7 +43,9 @@ uint cell_get_empty_buffer(struct cell_context *cell) { - uint buf = 0, tries = 0; + static uint prev_buffer = 0; + uint buf = (prev_buffer + 1) % CELL_NUM_BUFFERS; + uint tries = 0; /* Find a buffer that's marked as free by all SPUs */ while (1) { @@ -58,8 +61,13 @@ cell_get_empty_buffer(struct cell_context *cell) cell->buffer_status[spu][buf][0] = CELL_BUFFER_STATUS_USED; } /* - printf("PPU: ALLOC BUFFER %u\n", buf); + printf("PPU: ALLOC BUFFER %u, %u tries\n", buf, tries); */ + prev_buffer = buf; + + /* release tex buffer associated w/ prev use of this batch buf */ + cell_free_fenced_buffers(cell, &cell->fenced_buffers[buf]); + return buf; } } @@ -82,6 +90,38 @@ cell_get_empty_buffer(struct cell_context *cell) /** + * Append a fence command to the current batch buffer. + * Note that we're sure there's always room for this because of the + * adjusted size check in cell_batch_free_space(). + */ +static void +emit_fence(struct cell_context *cell) +{ + const uint batch = cell->cur_batch; + const uint size = cell->buffer_size[batch]; + struct cell_command_fence *fence_cmd; + struct cell_fence *fence = &cell->fenced_buffers[batch].fence; + uint i; + + /* set fence status to emitted, not yet signalled */ + for (i = 0; i < cell->num_spus; i++) { + fence->status[i][0] = CELL_FENCE_EMITTED; + } + + STATIC_ASSERT(sizeof(struct cell_command_fence) % 16 == 0); + ASSERT(size % 16 == 0); + ASSERT(size + sizeof(struct cell_command_fence) <= CELL_BUFFER_SIZE); + + fence_cmd = (struct cell_command_fence *) (cell->buffer[batch] + size); + fence_cmd->opcode[0] = CELL_CMD_FENCE; + fence_cmd->fence = fence; + + /* update batch buffer size */ + cell->buffer_size[batch] = size + sizeof(struct cell_command_fence); +} + + +/** * Flush the current batch buffer to the SPUs. * An empty buffer will be found and set as the new current batch buffer * for subsequent commands/data. @@ -91,7 +131,7 @@ cell_batch_flush(struct cell_context *cell) { static boolean flushing = FALSE; uint batch = cell->cur_batch; - const uint size = cell->buffer_size[batch]; + uint size = cell->buffer_size[batch]; uint spu, cmd_word; assert(!flushing); @@ -99,6 +139,14 @@ cell_batch_flush(struct cell_context *cell) if (size == 0) return; + /* Before we use this batch buffer, make sure any fenced texture buffers + * are released. + */ + if (cell->fenced_buffers[batch].head) { + emit_fence(cell); + size = cell->buffer_size[batch]; + } + flushing = TRUE; assert(batch < CELL_NUM_BUFFERS); @@ -139,74 +187,24 @@ uint cell_batch_free_space(const struct cell_context *cell) { uint free = CELL_BUFFER_SIZE - cell->buffer_size[cell->cur_batch]; + free -= sizeof(struct cell_command_fence); return free; } /** - * Append data to the current batch buffer. - * \param data address of block of bytes to append - * \param bytes size of block of bytes - */ -void -cell_batch_append(struct cell_context *cell, const void *data, uint bytes) -{ - uint size; - - ASSERT(bytes % 8 == 0); - ASSERT(bytes <= CELL_BUFFER_SIZE); - ASSERT(cell->cur_batch >= 0); - -#ifdef ASSERT - { - uint spu; - for (spu = 0; spu < cell->num_spus; spu++) { - ASSERT(cell->buffer_status[spu][cell->cur_batch][0] - == CELL_BUFFER_STATUS_USED); - } - } -#endif - - size = cell->buffer_size[cell->cur_batch]; - - if (size + bytes > CELL_BUFFER_SIZE) { - cell_batch_flush(cell); - size = 0; - } - - ASSERT(size + bytes <= CELL_BUFFER_SIZE); - - memcpy(cell->buffer[cell->cur_batch] + size, data, bytes); - - cell->buffer_size[cell->cur_batch] = size + bytes; -} - - -/** * Allocate space in the current batch buffer for 'bytes' space. + * Bytes must be a multiple of 16 bytes. Allocation will be 16 byte aligned. * \return address in batch buffer to put data */ void * -cell_batch_alloc(struct cell_context *cell, uint bytes) -{ - return cell_batch_alloc_aligned(cell, bytes, 1); -} - - -/** - * Same as \sa cell_batch_alloc, but return an address at a particular - * alignment. - */ -void * -cell_batch_alloc_aligned(struct cell_context *cell, uint bytes, - uint alignment) +cell_batch_alloc16(struct cell_context *cell, uint bytes) { void *pos; - uint size, padbytes; + uint size; - ASSERT(bytes % 8 == 0); + ASSERT(bytes % 16 == 0); ASSERT(bytes <= CELL_BUFFER_SIZE); - ASSERT(alignment > 0); ASSERT(cell->cur_batch >= 0); #ifdef ASSERT @@ -221,17 +219,12 @@ cell_batch_alloc_aligned(struct cell_context *cell, uint bytes, size = cell->buffer_size[cell->cur_batch]; - padbytes = (alignment - (size % alignment)) % alignment; - - if (padbytes + size + bytes > CELL_BUFFER_SIZE) { + if (bytes > cell_batch_free_space(cell)) { cell_batch_flush(cell); size = 0; } - else { - size += padbytes; - } - ASSERT(size % alignment == 0); + ASSERT(size % 16 == 0); ASSERT(size + bytes <= CELL_BUFFER_SIZE); pos = (void *) (cell->buffer[cell->cur_batch] + size); diff --git a/src/gallium/drivers/cell/ppu/cell_batch.h b/src/gallium/drivers/cell/ppu/cell_batch.h index f74dd60079..290136031a 100644 --- a/src/gallium/drivers/cell/ppu/cell_batch.h +++ b/src/gallium/drivers/cell/ppu/cell_batch.h @@ -44,15 +44,8 @@ cell_batch_flush(struct cell_context *cell); extern uint cell_batch_free_space(const struct cell_context *cell); -extern void -cell_batch_append(struct cell_context *cell, const void *data, uint bytes); - -extern void * -cell_batch_alloc(struct cell_context *cell, uint bytes); - extern void * -cell_batch_alloc_aligned(struct cell_context *cell, uint bytes, - uint alignment); +cell_batch_alloc16(struct cell_context *cell, uint bytes); extern void cell_init_batch_buffers(struct cell_context *cell); diff --git a/src/gallium/drivers/cell/ppu/cell_clear.c b/src/gallium/drivers/cell/ppu/cell_clear.c index c9c0c721bb..edc06747ac 100644 --- a/src/gallium/drivers/cell/ppu/cell_clear.c +++ b/src/gallium/drivers/cell/ppu/cell_clear.c @@ -70,18 +70,12 @@ void cell_clear_surface(struct pipe_context *pipe, struct pipe_surface *ps, unsigned clearValue) { - struct pipe_screen *screen = pipe->screen; struct cell_context *cell = cell_context(pipe); uint surfIndex; if (cell->dirty) cell_update_derived(cell); - - if (!cell->cbuf_map[0]) - cell->cbuf_map[0] = screen->surface_map(screen, ps, - PIPE_BUFFER_USAGE_GPU_WRITE); - if (ps == cell->framebuffer.zsbuf) { /* clear z/stencil buffer */ surfIndex = 1; @@ -99,11 +93,25 @@ cell_clear_surface(struct pipe_context *pipe, struct pipe_surface *ps, /* Build a CLEAR command and place it in the current batch buffer */ { + STATIC_ASSERT(sizeof(struct cell_command_clear_surface) % 16 == 0); struct cell_command_clear_surface *clr = (struct cell_command_clear_surface *) - cell_batch_alloc(cell, sizeof(*clr)); - clr->opcode = CELL_CMD_CLEAR_SURFACE; + cell_batch_alloc16(cell, sizeof(*clr)); + clr->opcode[0] = CELL_CMD_CLEAR_SURFACE; clr->surface = surfIndex; clr->value = clearValue; } + + /* Technically, the surface's contents are now known and cleared, + * so we could set the status to PIPE_SURFACE_STATUS_CLEAR. But + * it turns out it's quite painful to recognize when any particular + * surface goes from PIPE_SURFACE_STATUS_CLEAR to + * PIPE_SURFACE_STATUS_DEFINED (i.e. with known contents), because + * the drawing commands could be operating on numerous draw buffers, + * which we'd have to iterate through to set all their stati... + * For now, we cheat a bit and set the surface's status to DEFINED + * right here. Later we should revisit this and set the status to + * CLEAR here, and find a better place to set the status to DEFINED. + */ + ps->status = PIPE_SURFACE_STATUS_DEFINED; } diff --git a/src/gallium/drivers/cell/ppu/cell_context.c b/src/gallium/drivers/cell/ppu/cell_context.c index 71f1a3049d..ae82ded334 100644 --- a/src/gallium/drivers/cell/ppu/cell_context.c +++ b/src/gallium/drivers/cell/ppu/cell_context.c @@ -36,7 +36,7 @@ #include "pipe/p_defines.h" #include "pipe/p_format.h" #include "util/u_memory.h" -#include "pipe/p_winsys.h" +#include "pipe/internal/p_winsys_screen.h" #include "pipe/p_screen.h" #include "draw/draw_context.h" @@ -47,6 +47,7 @@ #include "cell_clear.h" #include "cell_context.h" #include "cell_draw_arrays.h" +#include "cell_fence.h" #include "cell_flush.h" #include "cell_state.h" #include "cell_surface.h" @@ -62,6 +63,8 @@ cell_destroy_context( struct pipe_context *pipe ) { struct cell_context *cell = cell_context(pipe); + util_delete_keymap(cell->fragment_ops_cache, NULL); + cell_spu_exit(cell); align_free(cell); @@ -85,13 +88,16 @@ cell_draw_create(struct cell_context *cell) } -#ifdef DEBUG static const struct debug_named_value cell_debug_flags[] = { {"checker", CELL_DEBUG_CHECKER},/**< modulate tile clear color by SPU ID */ + {"asm", CELL_DEBUG_ASM}, /**< dump SPU asm code */ {"sync", CELL_DEBUG_SYNC}, /**< SPUs do synchronous DMA */ + {"fragops", CELL_DEBUG_FRAGMENT_OPS}, /**< SPUs emit fragment ops debug messages*/ + {"fragopfallback", CELL_DEBUG_FRAGMENT_OP_FALLBACK}, /**< SPUs use reference implementation for fragment ops*/ + {"cmd", CELL_DEBUG_CMD}, /**< SPUs dump command buffer info */ + {"cache", CELL_DEBUG_CACHE}, /**< report texture cache stats on exit */ {NULL, 0} }; -#endif struct pipe_context * @@ -99,6 +105,7 @@ cell_create_context(struct pipe_screen *screen, struct cell_winsys *cws) { struct cell_context *cell; + uint i; /* some fields need to be 16-byte aligned, so align the whole object */ cell = (struct cell_context*) align_malloc(sizeof(struct cell_context), 16); @@ -125,11 +132,14 @@ cell_create_context(struct pipe_screen *screen, cell_init_state_functions(cell); cell_init_shader_functions(cell); cell_init_surface_functions(cell); - cell_init_texture_functions(cell); cell_init_vertex_functions(cell); cell->draw = cell_draw_create(cell); + /* Create cache of fragment ops generated code */ + cell->fragment_ops_cache = + util_new_keymap(sizeof(struct cell_fragment_ops_key), ~0, NULL); + cell_init_vbuf(cell); draw_set_rasterize_stage(cell->draw, cell->vbuf); @@ -143,17 +153,31 @@ cell_create_context(struct pipe_screen *screen, cell_debug_flags, 0 ); + for (i = 0; i < CELL_NUM_BUFFERS; i++) + cell_fence_init(&cell->fenced_buffers[i].fence); + + /* * SPU stuff */ - cell->num_spus = 6; - /* XXX is this in SDK 3.0 only? - cell->num_spus = spe_cpu_info_get(SPE_COUNT_PHYSICAL_SPES, -1); - */ + /* This call only works with SDK 3.0. Anyone still using 2.1??? */ + cell->num_cells = spe_cpu_info_get(SPE_COUNT_PHYSICAL_CPU_NODES, -1); + cell->num_spus = spe_cpu_info_get(SPE_COUNT_USABLE_SPES, -1); + if (cell->debug_flags) { + printf("Cell: found %d Cell(s) with %u SPUs\n", + cell->num_cells, cell->num_spus); + } + if (getenv("CELL_NUM_SPUS")) { + cell->num_spus = atoi(getenv("CELL_NUM_SPUS")); + assert(cell->num_spus > 0); + } cell_start_spus(cell); cell_init_batch_buffers(cell); + /* make sure SPU initializations are done before proceeding */ + cell_flush_int(cell, CELL_FLUSH_WAIT); + return &cell->pipe; } diff --git a/src/gallium/drivers/cell/ppu/cell_context.h b/src/gallium/drivers/cell/ppu/cell_context.h index 14914b9c6f..ca03dc1511 100644 --- a/src/gallium/drivers/cell/ppu/cell_context.h +++ b/src/gallium/drivers/cell/ppu/cell_context.h @@ -38,6 +38,7 @@ #include "cell/common.h" #include "rtasm/rtasm_ppc_spe.h" #include "tgsi/tgsi_scan.h" +#include "util/u_keymap.h" struct cell_vbuf_render; @@ -67,31 +68,29 @@ struct cell_fragment_shader_state /** - * Cell blend state atom, subclass of pipe_blend_state. + * Key for mapping per-fragment state to cached SPU machine code. + * keymap(cell_fragment_ops_key) => cell_command_fragment_ops */ -struct cell_blend_state +struct cell_fragment_ops_key { - struct pipe_blend_state base; - - /** - * Generated code to perform alpha blending - */ - struct spe_function code; + struct pipe_blend_state blend; + struct pipe_blend_color blend_color; + struct pipe_depth_stencil_alpha_state dsa; + enum pipe_format color_format; + enum pipe_format zs_format; }; +struct cell_buffer_node; + /** - * Cell depth/stencil/alpha state atom, subclass of - * pipe_depth_stencil_alpha_state. + * Fenced buffer list. List of buffers which can be unreferenced after + * the fence has been executed/signalled. */ -struct cell_depth_stencil_alpha_state +struct cell_buffer_list { - struct pipe_depth_stencil_alpha_state base; - - /** - * Generated code to perform alpha, stencil, and depth testing on the SPE - */ - struct spe_function code; + struct cell_fence fence ALIGN16_ATTRIB; + struct cell_buffer_node *head; }; @@ -104,10 +103,10 @@ struct cell_context struct cell_winsys *winsys; - const struct cell_blend_state *blend; + const struct pipe_blend_state *blend; const struct pipe_sampler_state *sampler[PIPE_MAX_SAMPLERS]; uint num_samplers; - const struct cell_depth_stencil_alpha_state *depth_stencil; + const struct pipe_depth_stencil_alpha_state *depth_stencil; const struct pipe_rasterizer_state *rasterizer; const struct cell_vertex_shader_state *vs; const struct cell_fragment_shader_state *fs; @@ -128,6 +127,9 @@ struct cell_context struct pipe_vertex_element vertex_element[PIPE_MAX_ATTRIBS]; uint num_vertex_elements; + struct pipe_transfer *cbuf_transfer[PIPE_MAX_COLOR_BUFS]; + struct pipe_transfer *zsbuf_transfer; + ubyte *cbuf_map[PIPE_MAX_COLOR_BUFS]; ubyte *zsbuf_map; @@ -135,6 +137,11 @@ struct cell_context uint *tex_map; uint dirty; + uint dirty_textures; /* bitmask of texture units */ + uint dirty_samplers; /* bitmask of sampler units */ + + /** Cache of code generated for per-fragment ops */ + struct keymap *fragment_ops_cache; /** The primitive drawing context */ struct draw_context *draw; @@ -149,8 +156,9 @@ struct cell_context /** Mapped constant buffers */ void *mapped_constants[PIPE_SHADER_TYPES]; + struct cell_spu_function_info spu_functions ALIGN16_ATTRIB; - uint num_spus; + uint num_cells, num_spus; /** Buffers for command batches, vertex/index data */ uint buffer_size[CELL_NUM_BUFFERS]; @@ -162,6 +170,14 @@ struct cell_context uint buffer_status[CELL_MAX_SPUS][CELL_NUM_BUFFERS][4] ALIGN16_ATTRIB; + /** Associated with each command/batch buffer is a list of pipe_buffers + * that are fenced. When the last command in a buffer is executed, the + * fence will be signalled, indicating that any pipe_buffers preceeding + * that fence can be unreferenced (and probably freed). + */ + struct cell_buffer_list fenced_buffers[CELL_NUM_BUFFERS]; + + struct spe_function attrib_fetch; unsigned attrib_fetch_offsets[PIPE_MAX_ATTRIBS]; diff --git a/src/gallium/drivers/cell/ppu/cell_draw_arrays.c b/src/gallium/drivers/cell/ppu/cell_draw_arrays.c index 880d535320..644496db40 100644 --- a/src/gallium/drivers/cell/ppu/cell_draw_arrays.c +++ b/src/gallium/drivers/cell/ppu/cell_draw_arrays.c @@ -33,7 +33,7 @@ #include "pipe/p_defines.h" #include "pipe/p_context.h" -#include "pipe/p_winsys.h" +#include "pipe/internal/p_winsys_screen.h" #include "pipe/p_inlines.h" #include "cell_context.h" @@ -51,9 +51,9 @@ cell_map_constant_buffers(struct cell_context *sp) struct pipe_winsys *ws = sp->pipe.winsys; uint i; for (i = 0; i < 2; i++) { - if (sp->constants[i].size) { + if (sp->constants[i].buffer && sp->constants[i].buffer->size) { sp->mapped_constants[i] = ws->buffer_map(ws, sp->constants[i].buffer, - PIPE_BUFFER_USAGE_CPU_READ); + PIPE_BUFFER_USAGE_CPU_READ); cell_flush_buffer_range(sp, sp->mapped_constants[i], sp->constants[i].buffer->size); } @@ -61,7 +61,7 @@ cell_map_constant_buffers(struct cell_context *sp) draw_set_mapped_constant_buffer(sp->draw, sp->mapped_constants[PIPE_SHADER_VERTEX], - sp->constants[PIPE_SHADER_VERTEX].size); + sp->constants[PIPE_SHADER_VERTEX].buffer->size); } static void @@ -70,7 +70,7 @@ cell_unmap_constant_buffers(struct cell_context *sp) struct pipe_winsys *ws = sp->pipe.winsys; uint i; for (i = 0; i < 2; i++) { - if (sp->constants[i].size) + if (sp->constants[i].buffer && sp->constants[i].buffer->size) ws->buffer_unmap(ws, sp->constants[i].buffer); sp->mapped_constants[i] = NULL; } diff --git a/src/gallium/drivers/cell/ppu/cell_fence.c b/src/gallium/drivers/cell/ppu/cell_fence.c new file mode 100644 index 0000000000..867b5dcaa0 --- /dev/null +++ b/src/gallium/drivers/cell/ppu/cell_fence.c @@ -0,0 +1,168 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#include <unistd.h> +#include "util/u_memory.h" +#include "pipe/p_inlines.h" +#include "cell_context.h" +#include "cell_batch.h" +#include "cell_fence.h" +#include "cell_texture.h" + + +void +cell_fence_init(struct cell_fence *fence) +{ + uint i; + ASSERT_ALIGN16(fence->status); + for (i = 0; i < CELL_MAX_SPUS; i++) { + fence->status[i][0] = CELL_FENCE_IDLE; + } +} + + +boolean +cell_fence_signalled(const struct cell_context *cell, + const struct cell_fence *fence) +{ + uint i; + for (i = 0; i < cell->num_spus; i++) { + if (fence->status[i][0] != CELL_FENCE_SIGNALLED) + return FALSE; + /*assert(fence->status[i][0] == CELL_FENCE_EMITTED);*/ + } + return TRUE; +} + + +void +cell_fence_finish(const struct cell_context *cell, + const struct cell_fence *fence) +{ + while (!cell_fence_signalled(cell, fence)) { + usleep(10); + } + +#ifdef DEBUG + { + uint i; + for (i = 0; i < cell->num_spus; i++) { + assert(fence->status[i][0] == CELL_FENCE_SIGNALLED); + } + } +#endif +} + + + + +struct cell_buffer_node +{ + struct pipe_buffer *buffer; + struct cell_buffer_node *next; +}; + + +static void +cell_add_buffer_to_list(struct cell_context *cell, + struct cell_buffer_list *list, + struct pipe_buffer *buffer) +{ + struct pipe_screen *ps = cell->pipe.screen; + struct cell_buffer_node *node = CALLOC_STRUCT(cell_buffer_node); + /* create new list node which references the buffer, insert at head */ + if (node) { + pipe_buffer_reference(ps, &node->buffer, buffer); + node->next = list->head; + list->head = node; + } +} + + +/** + * Wait for completion of the given fence, then unreference any buffers + * on the list. + * This typically unrefs/frees texture buffers after any rendering which uses + * them has completed. + */ +void +cell_free_fenced_buffers(struct cell_context *cell, + struct cell_buffer_list *list) +{ + if (list->head) { + struct pipe_screen *ps = cell->pipe.screen; + struct cell_buffer_node *node; + + cell_fence_finish(cell, &list->fence); + + /* traverse the list, unreferencing buffers, freeing nodes */ + node = list->head; + while (node) { + struct cell_buffer_node *next = node->next; + assert(node->buffer); + pipe_buffer_unmap(ps, node->buffer); +#if 0 + printf("Unref buffer %p\n", node->buffer); + if (node->buffer->refcount == 1) + printf(" Delete!\n"); +#endif + pipe_buffer_reference(ps, &node->buffer, NULL); + FREE(node); + node = next; + } + list->head = NULL; + } +} + + +/** + * This should be called for each render command. + * Any texture buffers that are current bound will be added to a fenced + * list to be freed later when the fence is executed/signalled. + */ +void +cell_add_fenced_textures(struct cell_context *cell) +{ + struct cell_buffer_list *list = &cell->fenced_buffers[cell->cur_batch]; + uint i; + + for (i = 0; i < cell->num_textures; i++) { + struct cell_texture *ct = cell->texture[i]; + if (ct) { + uint level; + for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) { + if (ct->tiled_buffer[level]) { +#if 0 + printf("Adding texture %p buffer %p to list\n", + ct, ct->tiled_buffer[level]); +#endif + cell_add_buffer_to_list(cell, list, ct->tiled_buffer[level]); + } + } + } + } +} diff --git a/src/gallium/drivers/cell/ppu/cell_fence.h b/src/gallium/drivers/cell/ppu/cell_fence.h new file mode 100644 index 0000000000..536b4ba411 --- /dev/null +++ b/src/gallium/drivers/cell/ppu/cell_fence.h @@ -0,0 +1,57 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#ifndef CELL_FENCE_H +#define CELL_FENCE_H + + +extern void +cell_fence_init(struct cell_fence *fence); + + +extern boolean +cell_fence_signalled(const struct cell_context *cell, + const struct cell_fence *fence); + + +extern void +cell_fence_finish(const struct cell_context *cell, + const struct cell_fence *fence); + + + +extern void +cell_free_fenced_buffers(struct cell_context *cell, + struct cell_buffer_list *list); + + +extern void +cell_add_fenced_textures(struct cell_context *cell); + + +#endif /* CELL_FENCE_H */ diff --git a/src/gallium/drivers/cell/ppu/cell_flush.c b/src/gallium/drivers/cell/ppu/cell_flush.c index 6596b72010..8275c9dc9c 100644 --- a/src/gallium/drivers/cell/ppu/cell_flush.c +++ b/src/gallium/drivers/cell/ppu/cell_flush.c @@ -49,7 +49,7 @@ cell_flush(struct pipe_context *pipe, unsigned flags, flags |= CELL_FLUSH_WAIT; } - if (flags & PIPE_FLUSH_SWAPBUFFERS) + if (flags & (PIPE_FLUSH_SWAPBUFFERS | PIPE_FLUSH_RENDER_CACHE)) flags |= CELL_FLUSH_WAIT; draw_flush( cell->draw ); @@ -72,8 +72,9 @@ cell_flush_int(struct cell_context *cell, unsigned flags) flushing = TRUE; if (flags & CELL_FLUSH_WAIT) { - uint64_t *cmd = (uint64_t *) cell_batch_alloc(cell, sizeof(uint64_t)); - *cmd = CELL_CMD_FINISH; + STATIC_ASSERT(sizeof(opcode_t) % 16 == 0); + opcode_t *cmd = (opcode_t*) cell_batch_alloc16(cell, sizeof(opcode_t)); + *cmd[0] = CELL_CMD_FINISH; } cell_batch_flush(cell); @@ -101,11 +102,11 @@ void cell_flush_buffer_range(struct cell_context *cell, void *ptr, unsigned size) { - uint64_t batch[1 + (ROUNDUP8(sizeof(struct cell_buffer_range)) / 8)]; - struct cell_buffer_range *br = (struct cell_buffer_range *) & batch[1]; - + STATIC_ASSERT((sizeof(opcode_t) + sizeof(struct cell_buffer_range)) % 16 == 0); + uint32_t *batch = (uint32_t*)cell_batch_alloc16(cell, + sizeof(opcode_t) + sizeof(struct cell_buffer_range)); + struct cell_buffer_range *br = (struct cell_buffer_range *) &batch[4]; batch[0] = CELL_CMD_FLUSH_BUFFER_RANGE; br->base = (uintptr_t) ptr; br->size = size; - cell_batch_append(cell, batch, sizeof(batch)); } diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fp.c b/src/gallium/drivers/cell/ppu/cell_gen_fp.c index 6ffe94eb14..5a889a6119 100644 --- a/src/gallium/drivers/cell/ppu/cell_gen_fp.c +++ b/src/gallium/drivers/cell/ppu/cell_gen_fp.c @@ -2,6 +2,7 @@ * * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. * All Rights Reserved. + * Copyright 2009 VMware, Inc. All rights reserved. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the @@ -37,7 +38,7 @@ * \author Brian Paul */ - +#include <math.h> #include "pipe/p_defines.h" #include "pipe/p_state.h" #include "pipe/p_shader_tokens.h" @@ -51,25 +52,55 @@ #include "cell_gen_fp.h" -/** Set to 1 to enable debug/disassembly printfs */ -#define DISASSEM 01 +#define MAX_TEMPS 16 +#define MAX_IMMED 8 +#define CHAN_X 0 +#define CHAN_Y 1 +#define CHAN_Z 2 +#define CHAN_W 3 /** * Context needed during code generation. */ struct codegen { + struct cell_context *cell; int inputs_reg; /**< 1st function parameter */ int outputs_reg; /**< 2nd function parameter */ int constants_reg; /**< 3rd function parameter */ - int temp_regs[8][4]; /**< maps TGSI temps to SPE registers */ + int temp_regs[MAX_TEMPS][4]; /**< maps TGSI temps to SPE registers */ + int imm_regs[MAX_IMMED][4]; /**< maps TGSI immediates to SPE registers */ + + int num_imm; /**< number of immediates */ int one_reg; /**< register containing {1.0, 1.0, 1.0, 1.0} */ + int addr_reg; /**< address register, integer values */ + /** Per-instruction temps / intermediate temps */ int num_itemps; - int itemps[3]; + int itemps[12]; + + /** Current IF/ELSE/ENDIF nesting level */ + int if_nesting; + /** Current BGNLOOP/ENDLOOP nesting level */ + int loop_nesting; + /** Location of start of current loop */ + int loop_start; + + /** Index of if/conditional mask register */ + int cond_mask_reg; + /** Index of loop mask register */ + int loop_mask_reg; + + /** Index of master execution mask register */ + int exec_mask_reg; + + /** KIL mask: indicates which fragments have been killed */ + int kill_mask_reg; + + int frame_size; /**< Stack frame size, in words */ struct spe_function *f; boolean error; @@ -112,19 +143,122 @@ get_const_one_reg(struct codegen *gen) { if (gen->one_reg <= 0) { gen->one_reg = spe_allocate_available_register(gen->f); - } - /* one = {1.0, 1.0, 1.0, 1.0} */ - spe_load_float(gen->f, gen->one_reg, 1.0f); -#if DISASSEM - printf("il\tr%d, 1.0f\n", gen->one_reg); -#endif + spe_indent(gen->f, 4); + spe_comment(gen->f, -4, "init constant reg = 1.0:"); + + /* one = {1.0, 1.0, 1.0, 1.0} */ + spe_load_float(gen->f, gen->one_reg, 1.0f); + + spe_indent(gen->f, -4); + } return gen->one_reg; } /** + * Return index of the address register. + * Used for indirect register loads/stores. + */ +static int +get_address_reg(struct codegen *gen) +{ + if (gen->addr_reg <= 0) { + gen->addr_reg = spe_allocate_available_register(gen->f); + + spe_indent(gen->f, 4); + spe_comment(gen->f, -4, "init address reg = 0:"); + + /* init addr = {0, 0, 0, 0} */ + spe_zero(gen->f, gen->addr_reg); + + spe_indent(gen->f, -4); + } + + return gen->addr_reg; +} + + +/** + * Return index of the master execution mask. + * The register is allocated an initialized upon the first call. + * + * The master execution mask controls which pixels in a quad are + * modified, according to surrounding conditionals, loops, etc. + */ +static int +get_exec_mask_reg(struct codegen *gen) +{ + if (gen->exec_mask_reg <= 0) { + gen->exec_mask_reg = spe_allocate_available_register(gen->f); + + /* XXX this may not be needed */ + spe_comment(gen->f, 0*-4, "initialize master execution mask = ~0"); + spe_load_int(gen->f, gen->exec_mask_reg, ~0); + } + + return gen->exec_mask_reg; +} + + +/** Return index of the conditional (if/else) execution mask register */ +static int +get_cond_mask_reg(struct codegen *gen) +{ + if (gen->cond_mask_reg <= 0) { + gen->cond_mask_reg = spe_allocate_available_register(gen->f); + } + + return gen->cond_mask_reg; +} + + +/** Return index of the loop execution mask register */ +static int +get_loop_mask_reg(struct codegen *gen) +{ + if (gen->loop_mask_reg <= 0) { + gen->loop_mask_reg = spe_allocate_available_register(gen->f); + } + + return gen->loop_mask_reg; +} + + + +static boolean +is_register_src(struct codegen *gen, int channel, + const struct tgsi_full_src_register *src) +{ + int swizzle = tgsi_util_get_full_src_register_extswizzle(src, channel); + int sign_op = tgsi_util_get_full_src_register_sign_mode(src, channel); + + if (swizzle > TGSI_SWIZZLE_W || sign_op != TGSI_UTIL_SIGN_KEEP) { + return FALSE; + } + if (src->SrcRegister.File == TGSI_FILE_TEMPORARY || + src->SrcRegister.File == TGSI_FILE_IMMEDIATE) { + return TRUE; + } + return FALSE; +} + + +static boolean +is_memory_dst(struct codegen *gen, int channel, + const struct tgsi_full_dst_register *dst) +{ + if (dst->DstRegister.File == TGSI_FILE_OUTPUT) { + return TRUE; + } + else { + return FALSE; + } +} + + +/** * Return the index of the SPU temporary containing the named TGSI * source register. If the TGSI register is a TGSI_FILE_TEMPORARY we * just return the corresponding SPE register. If the TGIS register @@ -136,35 +270,99 @@ get_src_reg(struct codegen *gen, int channel, const struct tgsi_full_src_register *src) { - int reg; + int reg = -1; + int swizzle = tgsi_util_get_full_src_register_extswizzle(src, channel); + boolean reg_is_itemp = FALSE; + uint sign_op; + + assert(swizzle >= TGSI_SWIZZLE_X); + assert(swizzle <= TGSI_EXTSWIZZLE_ONE); + + if (swizzle == TGSI_EXTSWIZZLE_ONE) { + /* Load const one float and early out */ + reg = get_const_one_reg(gen); + } + else if (swizzle == TGSI_EXTSWIZZLE_ZERO) { + /* Load const zero float and early out */ + reg = get_itemp(gen); + spe_xor(gen->f, reg, reg, reg); + } + else { + int index = src->SrcRegister.Index; + + assert(swizzle < 4); + + if (src->SrcRegister.Indirect) { + /* XXX unfinished */ + } + + switch (src->SrcRegister.File) { + case TGSI_FILE_TEMPORARY: + reg = gen->temp_regs[index][swizzle]; + break; + case TGSI_FILE_INPUT: + { + /* offset is measured in quadwords, not bytes */ + int offset = index * 4 + swizzle; + reg = get_itemp(gen); + reg_is_itemp = TRUE; + /* Load: reg = memory[(machine_reg) + offset] */ + spe_lqd(gen->f, reg, gen->inputs_reg, offset * 16); + } + break; + case TGSI_FILE_IMMEDIATE: + reg = gen->imm_regs[index][swizzle]; + break; + case TGSI_FILE_CONSTANT: + { + /* offset is measured in quadwords, not bytes */ + int offset = index * 4 + swizzle; + reg = get_itemp(gen); + reg_is_itemp = TRUE; + /* Load: reg = memory[(machine_reg) + offset] */ + spe_lqd(gen->f, reg, gen->constants_reg, offset * 16); + } + break; + default: + assert(0); + } + } - /* XXX need to examine src swizzle info here. - * That will involve changing the channel var... + /* + * Handle absolute value, negate or set-negative of src register. */ + sign_op = tgsi_util_get_full_src_register_sign_mode(src, channel); + if (sign_op != TGSI_UTIL_SIGN_KEEP) { + /* + * All sign ops are done by manipulating bit 31, the IEEE float sign bit. + */ + const int bit31mask_reg = get_itemp(gen); + int result_reg; + + if (reg_is_itemp) { + /* re-use 'reg' for the result */ + result_reg = reg; + } + else { + /* alloc a new reg for the result */ + result_reg = get_itemp(gen); + } + /* mask with bit 31 set, the rest cleared */ + spe_load_uint(gen->f, bit31mask_reg, (1 << 31)); - switch (src->SrcRegister.File) { - case TGSI_FILE_TEMPORARY: - reg = gen->temp_regs[src->SrcRegister.Index][channel]; - break; - case TGSI_FILE_INPUT: - { - /* offset is measured in quadwords, not bytes */ - int offset = src->SrcRegister.Index * 4 + channel; - reg = get_itemp(gen); - /* Load: reg = memory[(machine_reg) + offset] */ - spe_lqd(gen->f, reg, gen->inputs_reg, offset); -#if DISASSEM - printf("lqd\tr%d, r%d + %d\n", reg, gen->inputs_reg, offset); -#endif + if (sign_op == TGSI_UTIL_SIGN_CLEAR) { + spe_andc(gen->f, result_reg, reg, bit31mask_reg); } - break; - case TGSI_FILE_IMMEDIATE: - /* xxx fall-through for now / fix */ - case TGSI_FILE_CONSTANT: - /* xxx fall-through for now / fix */ - default: - assert(0); + else if (sign_op == TGSI_UTIL_SIGN_SET) { + spe_and(gen->f, result_reg, reg, bit31mask_reg); + } + else { + assert(sign_op == TGSI_UTIL_SIGN_TOGGLE); + spe_xor(gen->f, result_reg, reg, bit31mask_reg); + } + + reg = result_reg; } return reg; @@ -183,11 +381,14 @@ get_dst_reg(struct codegen *gen, int channel, const struct tgsi_full_dst_register *dest) { - int reg; + int reg = -1; switch (dest->DstRegister.File) { case TGSI_FILE_TEMPORARY: - reg = gen->temp_regs[dest->DstRegister.Index][channel]; + if (gen->if_nesting > 0 || gen->loop_nesting > 0) + reg = get_itemp(gen); + else + reg = gen->temp_regs[dest->DstRegister.Index][channel]; break; case TGSI_FILE_OUTPUT: reg = get_itemp(gen); @@ -211,19 +412,59 @@ store_dest_reg(struct codegen *gen, int value_reg, int channel, const struct tgsi_full_dst_register *dest) { + /* + * XXX need to implement dst reg clamping/saturation + */ +#if 0 + switch (inst->Instruction.Saturate) { + case TGSI_SAT_NONE: + break; + case TGSI_SAT_ZERO_ONE: + break; + case TGSI_SAT_MINUS_PLUS_ONE: + break; + default: + assert( 0 ); + } +#endif + switch (dest->DstRegister.File) { case TGSI_FILE_TEMPORARY: - /* no-op */ + if (gen->if_nesting > 0 || gen->loop_nesting > 0) { + int d_reg = gen->temp_regs[dest->DstRegister.Index][channel]; + int exec_reg = get_exec_mask_reg(gen); + /* Mix d with new value according to exec mask: + * d[i] = mask_reg[i] ? value_reg : d_reg + */ + spe_selb(gen->f, d_reg, d_reg, value_reg, exec_reg); + } + else { + /* we're not inside a condition or loop: do nothing special */ + + } break; case TGSI_FILE_OUTPUT: { /* offset is measured in quadwords, not bytes */ int offset = dest->DstRegister.Index * 4 + channel; - /* Store: memory[(machine_reg) + offset] = reg */ - spe_stqd(gen->f, value_reg, gen->outputs_reg, offset); -#if DISASSEM - printf("stqd\tr%d, r%d + %d\n", value_reg, gen->outputs_reg, offset); -#endif + if (gen->if_nesting > 0 || gen->loop_nesting > 0) { + int exec_reg = get_exec_mask_reg(gen); + int curval_reg = get_itemp(gen); + /* First read the current value from memory: + * Load: curval = memory[(machine_reg) + offset] + */ + spe_lqd(gen->f, curval_reg, gen->outputs_reg, offset * 16); + /* Mix curval with newvalue according to exec mask: + * d[i] = mask_reg[i] ? value_reg : d_reg + */ + spe_selb(gen->f, curval_reg, curval_reg, value_reg, exec_reg); + /* Store: memory[(machine_reg) + offset] = curval */ + spe_stqd(gen->f, curval_reg, gen->outputs_reg, offset * 16); + } + else { + /* Store: memory[(machine_reg) + offset] = reg */ + spe_stqd(gen->f, value_reg, gen->outputs_reg, offset * 16); + } } break; default: @@ -232,125 +473,1265 @@ store_dest_reg(struct codegen *gen, } + +static void +emit_prologue(struct codegen *gen) +{ + gen->frame_size = 1024; /* XXX temporary, should be dynamic */ + + spe_comment(gen->f, 0, "Function prologue:"); + + /* save $lr on stack # stqd $lr,16($sp) */ + spe_stqd(gen->f, SPE_REG_RA, SPE_REG_SP, 16); + + if (gen->frame_size >= 512) { + /* offset is too large for ai instruction */ + int offset_reg = spe_allocate_available_register(gen->f); + int sp_reg = spe_allocate_available_register(gen->f); + /* offset = -framesize */ + spe_load_int(gen->f, offset_reg, -gen->frame_size); + /* sp = $sp */ + spe_move(gen->f, sp_reg, SPE_REG_SP); + /* $sp = $sp + offset_reg */ + spe_a(gen->f, SPE_REG_SP, SPE_REG_SP, offset_reg); + /* save $sp in stack frame */ + spe_stqd(gen->f, sp_reg, SPE_REG_SP, 0); + /* clean up */ + spe_release_register(gen->f, offset_reg); + spe_release_register(gen->f, sp_reg); + } + else { + /* save stack pointer # stqd $sp,-frameSize($sp) */ + spe_stqd(gen->f, SPE_REG_SP, SPE_REG_SP, -gen->frame_size); + + /* adjust stack pointer # ai $sp,$sp,-frameSize */ + spe_ai(gen->f, SPE_REG_SP, SPE_REG_SP, -gen->frame_size); + } +} + + +static void +emit_epilogue(struct codegen *gen) +{ + const int return_reg = 3; + + spe_comment(gen->f, 0, "Function epilogue:"); + + spe_comment(gen->f, 0, "return the killed mask"); + if (gen->kill_mask_reg > 0) { + /* shader called KIL, return the "alive" mask */ + spe_move(gen->f, return_reg, gen->kill_mask_reg); + } + else { + /* return {0,0,0,0} */ + spe_load_uint(gen->f, return_reg, 0); + } + + spe_comment(gen->f, 0, "restore stack and return"); + if (gen->frame_size >= 512) { + /* offset is too large for ai instruction */ + int offset_reg = spe_allocate_available_register(gen->f); + /* offset = framesize */ + spe_load_int(gen->f, offset_reg, gen->frame_size); + /* $sp = $sp + offset */ + spe_a(gen->f, SPE_REG_SP, SPE_REG_SP, offset_reg); + /* clean up */ + spe_release_register(gen->f, offset_reg); + } + else { + /* restore stack pointer # ai $sp,$sp,frameSize */ + spe_ai(gen->f, SPE_REG_SP, SPE_REG_SP, gen->frame_size); + } + + /* restore $lr # lqd $lr,16($sp) */ + spe_lqd(gen->f, SPE_REG_RA, SPE_REG_SP, 16); + + /* return from function call */ + spe_bi(gen->f, SPE_REG_RA, 0, 0); +} + + +#define FOR_EACH_ENABLED_CHANNEL(inst, ch) \ + for (ch = 0; ch < 4; ch++) \ + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) + + +static boolean +emit_ARL(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch = 0, src_reg, addr_reg; + + src_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + addr_reg = get_address_reg(gen); + + /* convert float to int */ + spe_cflts(gen->f, addr_reg, src_reg, 0); + + free_itemps(gen); + + return TRUE; +} + + static boolean emit_MOV(struct codegen *gen, const struct tgsi_full_instruction *inst) { + int ch, src_reg[4], dst_reg[4]; + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + src_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + dst_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + if (is_register_src(gen, ch, &inst->FullSrcRegisters[0]) && + is_memory_dst(gen, ch, &inst->FullDstRegisters[0])) { + /* special-case: register to memory store */ + store_dest_reg(gen, src_reg[ch], ch, &inst->FullDstRegisters[0]); + } + else { + spe_move(gen->f, dst_reg[ch], src_reg[ch]); + store_dest_reg(gen, dst_reg[ch], ch, &inst->FullDstRegisters[0]); + } + } + + free_itemps(gen); + + return TRUE; +} + +/** + * Emit binary operation + */ +static boolean +emit_binop(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], s2_reg[4], d_reg[4]; + + /* Loop over Red/Green/Blue/Alpha channels, fetch src operands */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + s2_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + } + + /* Loop over Red/Green/Blue/Alpha channels, do the op, store results */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + /* Emit actual SPE instruction: d = s1 + s2 */ + switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_ADD: + spe_fa(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); + break; + case TGSI_OPCODE_SUB: + spe_fs(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); + break; + case TGSI_OPCODE_MUL: + spe_fm(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); + break; + default: + ; + } + } + + /* Store the result (a no-op for TGSI_FILE_TEMPORARY dests) */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + + /* Free any intermediate temps we allocated */ + free_itemps(gen); + + return TRUE; +} + + +/** + * Emit multiply add. See emit_ADD for comments. + */ +static boolean +emit_MAD(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], s2_reg[4], s3_reg[4], d_reg[4]; + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + s2_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); + s3_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[2]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + } + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_fma(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch], s3_reg[ch]); + } + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + free_itemps(gen); + return TRUE; +} + + +/** + * Emit linear interpolate. See emit_ADD for comments. + */ +static boolean +emit_LERP(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], s2_reg[4], s3_reg[4], d_reg[4], tmp_reg[4]; + + /* setup/get src/dst/temp regs */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + s2_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); + s3_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[2]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + tmp_reg[ch] = get_itemp(gen); + } + + /* d = s3 + s1(s2 - s3) */ + /* do all subtracts, then all fma, then all stores to better pipeline */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_fs(gen->f, tmp_reg[ch], s2_reg[ch], s3_reg[ch]); + } + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_fma(gen->f, d_reg[ch], tmp_reg[ch], s1_reg[ch], s3_reg[ch]); + } + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + free_itemps(gen); + return TRUE; +} + + + +/** + * Emit reciprocal or recip sqrt. + */ +static boolean +emit_RCP_RSQ(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], d_reg[4], tmp_reg[4]; + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + tmp_reg[ch] = get_itemp(gen); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + if (inst->Instruction.Opcode == TGSI_OPCODE_RCP) { + /* tmp = 1/s1 */ + spe_frest(gen->f, tmp_reg[ch], s1_reg[ch]); + } + else { + /* tmp = 1/sqrt(s1) */ + spe_frsqest(gen->f, tmp_reg[ch], s1_reg[ch]); + } + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + /* d = float_interp(s1, tmp) */ + spe_fi(gen->f, d_reg[ch], s1_reg[ch], tmp_reg[ch]); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + + +/** + * Emit absolute value. See emit_ADD for comments. + */ +static boolean +emit_ABS(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], d_reg[4]; + const int bit31mask_reg = get_itemp(gen); + + /* mask with bit 31 set, the rest cleared */ + spe_load_uint(gen->f, bit31mask_reg, (1 << 31)); + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + } + + /* d = sign bit cleared in s1 */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_andc(gen->f, d_reg[ch], s1_reg[ch], bit31mask_reg); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + +/** + * Emit 3 component dot product. See emit_ADD for comments. + */ +static boolean +emit_DP3(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ int ch; - for (ch = 0; ch < 4; ch++) { - if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { - int src_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); - int dst_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); - /* XXX we don't always need to actually emit a mov instruction here */ - spe_move(gen->f, dst_reg, src_reg); -#if DISASSEM - printf("mov\tr%d, r%d\n", dst_reg, src_reg); -#endif - store_dest_reg(gen, dst_reg, ch, &inst->FullDstRegisters[0]); - free_itemps(gen); + int s1x_reg, s1y_reg, s1z_reg; + int s2x_reg, s2y_reg, s2z_reg; + int t0_reg = get_itemp(gen), t1_reg = get_itemp(gen); + + s1x_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[0]); + s2x_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[1]); + s1y_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[0]); + s2y_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[1]); + s1z_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[0]); + s2z_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[1]); + + /* t0 = x0 * x1 */ + spe_fm(gen->f, t0_reg, s1x_reg, s2x_reg); + + /* t1 = y0 * y1 */ + spe_fm(gen->f, t1_reg, s1y_reg, s2y_reg); + + /* t0 = z0 * z1 + t0 */ + spe_fma(gen->f, t0_reg, s1z_reg, s2z_reg, t0_reg); + + /* t0 = t0 + t1 */ + spe_fa(gen->f, t0_reg, t0_reg, t1_reg); + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + spe_move(gen->f, d_reg, t0_reg); + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + +/** + * Emit 4 component dot product. See emit_ADD for comments. + */ +static boolean +emit_DP4(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + int s0x_reg, s0y_reg, s0z_reg, s0w_reg; + int s1x_reg, s1y_reg, s1z_reg, s1w_reg; + int t0_reg = get_itemp(gen), t1_reg = get_itemp(gen); + + s0x_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[0]); + s1x_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[1]); + s0y_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[0]); + s1y_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[1]); + s0z_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[0]); + s1z_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[1]); + s0w_reg = get_src_reg(gen, CHAN_W, &inst->FullSrcRegisters[0]); + s1w_reg = get_src_reg(gen, CHAN_W, &inst->FullSrcRegisters[1]); + + /* t0 = x0 * x1 */ + spe_fm(gen->f, t0_reg, s0x_reg, s1x_reg); + + /* t1 = y0 * y1 */ + spe_fm(gen->f, t1_reg, s0y_reg, s1y_reg); + + /* t0 = z0 * z1 + t0 */ + spe_fma(gen->f, t0_reg, s0z_reg, s1z_reg, t0_reg); + + /* t1 = w0 * w1 + t1 */ + spe_fma(gen->f, t1_reg, s0w_reg, s1w_reg, t1_reg); + + /* t0 = t0 + t1 */ + spe_fa(gen->f, t0_reg, t0_reg, t1_reg); + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + spe_move(gen->f, d_reg, t0_reg); + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + +/** + * Emit homogeneous dot product. See emit_ADD for comments. + */ +static boolean +emit_DPH(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + /* XXX rewrite this function to look more like DP3/DP4 */ + int ch; + int s1_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[0]); + int s2_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[1]); + int tmp_reg = get_itemp(gen); + + /* t = x0 * x1 */ + spe_fm(gen->f, tmp_reg, s1_reg, s2_reg); + + s1_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[0]); + s2_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[1]); + /* t = y0 * y1 + t */ + spe_fma(gen->f, tmp_reg, s1_reg, s2_reg, tmp_reg); + + s1_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[0]); + s2_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[1]); + /* t = z0 * z1 + t */ + spe_fma(gen->f, tmp_reg, s1_reg, s2_reg, tmp_reg); + + s2_reg = get_src_reg(gen, CHAN_W, &inst->FullSrcRegisters[1]); + /* t = w1 + t */ + spe_fa(gen->f, tmp_reg, s2_reg, tmp_reg); + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + spe_move(gen->f, d_reg, tmp_reg); + store_dest_reg(gen, tmp_reg, ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + +/** + * Emit 3-component vector normalize. + */ +static boolean +emit_NRM3(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + int src_reg[3]; + int t0_reg = get_itemp(gen), t1_reg = get_itemp(gen); + + src_reg[0] = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[0]); + src_reg[1] = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[0]); + src_reg[2] = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[0]); + + /* t0 = x * x */ + spe_fm(gen->f, t0_reg, src_reg[0], src_reg[0]); + + /* t1 = y * y */ + spe_fm(gen->f, t1_reg, src_reg[1], src_reg[1]); + + /* t0 = z * z + t0 */ + spe_fma(gen->f, t0_reg, src_reg[2], src_reg[2], t0_reg); + + /* t0 = t0 + t1 */ + spe_fa(gen->f, t0_reg, t0_reg, t1_reg); + + /* t1 = 1.0 / sqrt(t0) */ + spe_frsqest(gen->f, t1_reg, t0_reg); + spe_fi(gen->f, t1_reg, t0_reg, t1_reg); + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + /* dst = src[ch] * t1 */ + spe_fm(gen->f, d_reg, src_reg[ch], t1_reg); + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + + +/** + * Emit cross product. See emit_ADD for comments. + */ +static boolean +emit_XPD(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int s1_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[0]); + int s2_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[1]); + int tmp_reg = get_itemp(gen); + + /* t = z0 * y1 */ + spe_fm(gen->f, tmp_reg, s1_reg, s2_reg); + + s1_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[0]); + s2_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[1]); + /* t = y0 * z1 - t */ + spe_fms(gen->f, tmp_reg, s1_reg, s2_reg, tmp_reg); + + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << CHAN_X)) { + store_dest_reg(gen, tmp_reg, CHAN_X, &inst->FullDstRegisters[0]); + } + + s1_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[0]); + s2_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[1]); + /* t = x0 * z1 */ + spe_fm(gen->f, tmp_reg, s1_reg, s2_reg); + + s1_reg = get_src_reg(gen, CHAN_Z, &inst->FullSrcRegisters[0]); + s2_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[1]); + /* t = z0 * x1 - t */ + spe_fms(gen->f, tmp_reg, s1_reg, s2_reg, tmp_reg); + + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << CHAN_Y)) { + store_dest_reg(gen, tmp_reg, CHAN_Y, &inst->FullDstRegisters[0]); + } + + s1_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[0]); + s2_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[1]); + /* t = y0 * x1 */ + spe_fm(gen->f, tmp_reg, s1_reg, s2_reg); + + s1_reg = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[0]); + s2_reg = get_src_reg(gen, CHAN_Y, &inst->FullSrcRegisters[1]); + /* t = x0 * y1 - t */ + spe_fms(gen->f, tmp_reg, s1_reg, s2_reg, tmp_reg); + + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << CHAN_Z)) { + store_dest_reg(gen, tmp_reg, CHAN_Z, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + + +/** + * Emit inequality instruction. + * Note that the SPE fcgt instruction produces 0x0 and 0xffffffff as + * the result but OpenGL/TGSI needs 0.0 and 1.0 results. + * We can easily convert 0x0/0xffffffff to 0.0/1.0 with a bitwise AND. + */ +static boolean +emit_inequality(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], s2_reg[4], d_reg[4], one_reg; + bool complement = FALSE; + + one_reg = get_const_one_reg(gen); + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + s2_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_SGT: + spe_fcgt(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); + break; + case TGSI_OPCODE_SLT: + spe_fcgt(gen->f, d_reg[ch], s2_reg[ch], s1_reg[ch]); + break; + case TGSI_OPCODE_SGE: + spe_fcgt(gen->f, d_reg[ch], s2_reg[ch], s1_reg[ch]); + complement = TRUE; + break; + case TGSI_OPCODE_SLE: + spe_fcgt(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); + complement = TRUE; + break; + case TGSI_OPCODE_SEQ: + spe_fceq(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); + break; + case TGSI_OPCODE_SNE: + spe_fceq(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); + complement = TRUE; + break; + default: + assert(0); } } - return true; + + /* convert d from 0x0/0xffffffff to 0.0/1.0 */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + /* d = d & one_reg */ + if (complement) + spe_andc(gen->f, d_reg[ch], one_reg, d_reg[ch]); + else + spe_and(gen->f, d_reg[ch], one_reg, d_reg[ch]); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; } /** - * Emit addition instructions. Recall that a single TGSI_OPCODE_ADD - * becomes (up to) four SPU "fa" instructions because we're doing SOA - * processing. + * Emit compare. */ static boolean -emit_ADD(struct codegen *gen, const struct tgsi_full_instruction *inst) +emit_CMP(struct codegen *gen, const struct tgsi_full_instruction *inst) { int ch; - /* Loop over Red/Green/Blue/Alpha channels */ - for (ch = 0; ch < 4; ch++) { - /* If the dest R, G, B or A writemask is enabled... */ - if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { - /* get indexes of the two src, one dest SPE registers */ - int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); - int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); - int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); - - /* Emit actual SPE instruction: d = s1 + s2 */ - spe_fa(gen->f, d_reg, s1_reg, s2_reg); -#if DISASSEM - printf("fa\tr%d, r%d, r%d\n", d_reg, s1_reg, s2_reg); + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); + int s3_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[2]); + int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + int zero_reg = get_itemp(gen); + + spe_zero(gen->f, zero_reg); + + /* d = (s1 < 0) ? s2 : s3 */ + spe_fcgt(gen->f, d_reg, zero_reg, s1_reg); + spe_selb(gen->f, d_reg, s3_reg, s2_reg, d_reg); + + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + + return TRUE; +} + +/** + * Emit trunc. + * Convert float to signed int + * Convert signed int to float + */ +static boolean +emit_TRUNC(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], d_reg[4]; + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + } + + /* Convert float to int */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_cflts(gen->f, d_reg[ch], s1_reg[ch], 0); + } + + /* Convert int to float */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_csflt(gen->f, d_reg[ch], d_reg[ch], 0); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + + +/** + * Emit floor. + * If negative int subtract one + * Convert float to signed int + * Convert signed int to float + */ +static boolean +emit_FLR(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], d_reg[4], tmp_reg[4], zero_reg, one_reg; + + zero_reg = get_itemp(gen); + spe_zero(gen->f, zero_reg); + one_reg = get_const_one_reg(gen); + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + tmp_reg[ch] = get_itemp(gen); + } + + /* If negative, subtract 1.0 */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_fcgt(gen->f, tmp_reg[ch], zero_reg, s1_reg[ch]); + } + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_selb(gen->f, tmp_reg[ch], zero_reg, one_reg, tmp_reg[ch]); + } + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_fs(gen->f, tmp_reg[ch], s1_reg[ch], tmp_reg[ch]); + } + + /* Convert float to int */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_cflts(gen->f, tmp_reg[ch], tmp_reg[ch], 0); + } + + /* Convert int to float */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_csflt(gen->f, d_reg[ch], tmp_reg[ch], 0); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + + +/** + * Compute frac = Input - FLR(Input) + */ +static boolean +emit_FRC(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], d_reg[4], tmp_reg[4], zero_reg, one_reg; + + zero_reg = get_itemp(gen); + spe_zero(gen->f, zero_reg); + one_reg = get_const_one_reg(gen); + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + tmp_reg[ch] = get_itemp(gen); + } + + /* If negative, subtract 1.0 */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_fcgt(gen->f, tmp_reg[ch], zero_reg, s1_reg[ch]); + } + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_selb(gen->f, tmp_reg[ch], zero_reg, one_reg, tmp_reg[ch]); + } + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_fs(gen->f, tmp_reg[ch], s1_reg[ch], tmp_reg[ch]); + } + + /* Convert float to int */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_cflts(gen->f, tmp_reg[ch], tmp_reg[ch], 0); + } + + /* Convert int to float */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_csflt(gen->f, tmp_reg[ch], tmp_reg[ch], 0); + } + + /* d = s1 - FLR(s1) */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_fs(gen->f, d_reg[ch], s1_reg[ch], tmp_reg[ch]); + } + + /* store result */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + + +#if 0 +static void +print_functions(struct cell_context *cell) +{ + struct cell_spu_function_info *funcs = &cell->spu_functions; + uint i; + for (i = 0; i < funcs->num; i++) { + printf("SPU func %u: %s at %u\n", + i, funcs->names[i], funcs->addrs[i]); + } +} #endif - /* Store the result (a no-op for TGSI_FILE_TEMPORARY dests) */ - store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); - /* Free any intermediate temps we allocated */ - free_itemps(gen); + +static uint +lookup_function(struct cell_context *cell, const char *funcname) +{ + const struct cell_spu_function_info *funcs = &cell->spu_functions; + uint i, addr = 0; + for (i = 0; i < funcs->num; i++) { + if (strcmp(funcs->names[i], funcname) == 0) { + addr = funcs->addrs[i]; } } - return true; + assert(addr && "spu function not found"); + return addr / 4; /* discard 2 least significant bits */ } /** - * Emit multiply. See emit_ADD for comments. + * Emit code to call a SPU function. + * Used to implement instructions like SIN/COS/POW/TEX/etc. + * If scalar, only the X components of the src regs are used, and the + * result is replicated across the dest register's XYZW components. */ static boolean -emit_MUL(struct codegen *gen, const struct tgsi_full_instruction *inst) +emit_function_call(struct codegen *gen, + const struct tgsi_full_instruction *inst, + char *funcname, uint num_args, boolean scalar) +{ + const uint addr = lookup_function(gen->cell, funcname); + char comment[100]; + int s_regs[3]; + int func_called = FALSE; + uint a, ch; + int retval_reg = -1; + + assert(num_args <= 3); + + snprintf(comment, sizeof(comment), "CALL %s:", funcname); + spe_comment(gen->f, -4, comment); + + if (scalar) { + for (a = 0; a < num_args; a++) { + s_regs[a] = get_src_reg(gen, CHAN_X, &inst->FullSrcRegisters[a]); + } + /* we'll call the function, put the return value in this register, + * then replicate it across all write-enabled components in d_reg. + */ + retval_reg = spe_allocate_available_register(gen->f); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + int d_reg; + ubyte usedRegs[SPE_NUM_REGS]; + uint i, numUsed; + + if (!scalar) { + for (a = 0; a < num_args; a++) { + s_regs[a] = get_src_reg(gen, ch, &inst->FullSrcRegisters[a]); + } + } + + d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + + if (!scalar || !func_called) { + /* for a scalar function, we'll really only call the function once */ + + numUsed = spe_get_registers_used(gen->f, usedRegs); + assert(numUsed < gen->frame_size / 16 - 2); + + /* save registers to stack */ + for (i = 0; i < numUsed; i++) { + uint reg = usedRegs[i]; + int offset = 2 + i; + spe_stqd(gen->f, reg, SPE_REG_SP, 16 * offset); + } + + /* setup function arguments */ + for (a = 0; a < num_args; a++) { + spe_move(gen->f, 3 + a, s_regs[a]); + } + + /* branch to function, save return addr */ + spe_brasl(gen->f, SPE_REG_RA, addr); + + /* save function's return value */ + if (scalar) + spe_move(gen->f, retval_reg, 3); + else + spe_move(gen->f, d_reg, 3); + + /* restore registers from stack */ + for (i = 0; i < numUsed; i++) { + uint reg = usedRegs[i]; + if (reg != d_reg && reg != retval_reg) { + int offset = 2 + i; + spe_lqd(gen->f, reg, SPE_REG_SP, 16 * offset); + } + } + + func_called = TRUE; + } + + if (scalar) { + spe_move(gen->f, d_reg, retval_reg); + } + + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + + if (scalar) { + spe_release_register(gen->f, retval_reg); + } + + return TRUE; +} + + +static boolean +emit_TEX(struct codegen *gen, const struct tgsi_full_instruction *inst) { + const uint target = inst->InstructionExtTexture.Texture; + const uint unit = inst->FullSrcRegisters[1].SrcRegister.Index; + uint addr; int ch; + int coord_regs[4], d_regs[4]; + + switch (target) { + case TGSI_TEXTURE_1D: + case TGSI_TEXTURE_2D: + addr = lookup_function(gen->cell, "spu_tex_2d"); + break; + case TGSI_TEXTURE_3D: + addr = lookup_function(gen->cell, "spu_tex_3d"); + break; + case TGSI_TEXTURE_CUBE: + addr = lookup_function(gen->cell, "spu_tex_cube"); + break; + default: + ASSERT(0 && "unsupported texture target"); + return FALSE; + } + + assert(inst->FullSrcRegisters[1].SrcRegister.File == TGSI_FILE_SAMPLER); + + spe_comment(gen->f, -4, "CALL tex:"); + + /* get src/dst reg info */ for (ch = 0; ch < 4; ch++) { - if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { - int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); - int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); - int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); - /* d = s1 * s2 */ - spe_fm(gen->f, d_reg, s1_reg, s2_reg); -#if DISASSEM - printf("fm\tr%d, r%d, r%d\n", d_reg, s1_reg, s2_reg); -#endif - store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); - free_itemps(gen); + coord_regs[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + d_regs[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + } + + { + ubyte usedRegs[SPE_NUM_REGS]; + uint i, numUsed; + + numUsed = spe_get_registers_used(gen->f, usedRegs); + assert(numUsed < gen->frame_size / 16 - 2); + + /* save registers to stack */ + for (i = 0; i < numUsed; i++) { + uint reg = usedRegs[i]; + int offset = 2 + i; + spe_stqd(gen->f, reg, SPE_REG_SP, 16 * offset); + } + + /* setup function arguments (XXX depends on target) */ + for (i = 0; i < 4; i++) { + spe_move(gen->f, 3 + i, coord_regs[i]); } + spe_load_uint(gen->f, 7, unit); /* sampler unit */ + + /* branch to function, save return addr */ + spe_brasl(gen->f, SPE_REG_RA, addr); + + /* save function's return values (four pixel's colors) */ + for (i = 0; i < 4; i++) { + spe_move(gen->f, d_regs[i], 3 + i); + } + + /* restore registers from stack */ + for (i = 0; i < numUsed; i++) { + uint reg = usedRegs[i]; + if (reg != d_regs[0] && + reg != d_regs[1] && + reg != d_regs[2] && + reg != d_regs[3]) { + int offset = 2 + i; + spe_lqd(gen->f, reg, SPE_REG_SP, 16 * offset); + } + } + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_regs[ch], ch, &inst->FullDstRegisters[0]); + free_itemps(gen); } - return true; + + return TRUE; } /** - * Emit set-if-greater-than. - * Note that the SPE fcgt instruction produces 0x0 and 0xffffffff as - * the result but OpenGL/TGSI needs 0.0 and 1.0 results. - * We can easily convert 0x0/0xffffffff to 0.0/1.0 with a bitwise AND. + * KILL if any of src reg values are less than zero. */ static boolean -emit_SGT(struct codegen *gen, const struct tgsi_full_instruction *inst) +emit_KIL(struct codegen *gen, const struct tgsi_full_instruction *inst) { int ch; + int s_regs[4], kil_reg = -1, cmp_reg, zero_reg; - for (ch = 0; ch < 4; ch++) { - if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { - int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); - int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); - int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); - - /* d = (s1 > s2) */ - spe_fcgt(gen->f, d_reg, s1_reg, s2_reg); -#if DISASSEM - printf("fcgt\tr%d, r%d, r%d\n", d_reg, s1_reg, s2_reg); -#endif + spe_comment(gen->f, -4, "CALL kil:"); - /* convert d from 0x0/0xffffffff to 0.0/1.0 */ - /* d = d & one_reg */ - spe_and(gen->f, d_reg, d_reg, get_const_one_reg(gen)); -#if DISASSEM - printf("and\tr%d, r%d, r%d\n", d_reg, d_reg, get_const_one_reg(gen)); -#endif + /* zero = {0,0,0,0} */ + zero_reg = get_itemp(gen); + spe_zero(gen->f, zero_reg); + + cmp_reg = get_itemp(gen); + + /* get src regs */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s_regs[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + } + + /* test if any src regs are < 0 */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + if (kil_reg >= 0) { + /* cmp = 0 > src ? : ~0 : 0 */ + spe_fcgt(gen->f, cmp_reg, zero_reg, s_regs[ch]); + /* kil = kil | cmp */ + spe_or(gen->f, kil_reg, kil_reg, cmp_reg); + } + else { + kil_reg = get_itemp(gen); + /* kil = 0 > src ? : ~0 : 0 */ + spe_fcgt(gen->f, kil_reg, zero_reg, s_regs[ch]); + } + } + + if (gen->if_nesting || gen->loop_nesting) { + /* may have been a conditional kil */ + spe_and(gen->f, kil_reg, kil_reg, gen->exec_mask_reg); + } - store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); - free_itemps(gen); + /* allocate the kill mask reg if needed */ + if (gen->kill_mask_reg <= 0) { + gen->kill_mask_reg = spe_allocate_available_register(gen->f); + spe_move(gen->f, gen->kill_mask_reg, kil_reg); + } + else { + spe_or(gen->f, gen->kill_mask_reg, gen->kill_mask_reg, kil_reg); + } + + free_itemps(gen); + + return TRUE; +} + + + +/** + * Emit min or max. + */ +static boolean +emit_MIN_MAX(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s0_reg[4], s1_reg[4], d_reg[4], tmp_reg[4]; + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + s0_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + s1_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); + d_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + tmp_reg[ch] = get_itemp(gen); + } + + /* d = (s0 > s1) ? s0 : s1 */ + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + if (inst->Instruction.Opcode == TGSI_OPCODE_MAX) + spe_fcgt(gen->f, tmp_reg[ch], s0_reg[ch], s1_reg[ch]); + else + spe_fcgt(gen->f, tmp_reg[ch], s1_reg[ch], s0_reg[ch]); + } + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + spe_selb(gen->f, d_reg[ch], s1_reg[ch], s0_reg[ch], tmp_reg[ch]); + } + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + + free_itemps(gen); + return TRUE; +} + + +/** + * Emit code to update the execution mask. + * This needs to be done whenever the execution status of a conditional + * or loop is changed. + */ +static void +emit_update_exec_mask(struct codegen *gen) +{ + const int exec_reg = get_exec_mask_reg(gen); + const int cond_reg = gen->cond_mask_reg; + const int loop_reg = gen->loop_mask_reg; + + spe_comment(gen->f, 0, "Update master execution mask"); + + if (gen->if_nesting > 0 && gen->loop_nesting > 0) { + /* exec_mask = cond_mask & loop_mask */ + assert(cond_reg > 0); + assert(loop_reg > 0); + spe_and(gen->f, exec_reg, cond_reg, loop_reg); + } + else if (gen->if_nesting > 0) { + assert(cond_reg > 0); + spe_move(gen->f, exec_reg, cond_reg); + } + else if (gen->loop_nesting > 0) { + assert(loop_reg > 0); + spe_move(gen->f, exec_reg, loop_reg); + } + else { + spe_load_int(gen->f, exec_reg, ~0x0); + } +} + + +static boolean +emit_IF(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + const int channel = 0; + int cond_reg; + + cond_reg = get_cond_mask_reg(gen); + + /* XXX push cond exec mask */ + + spe_comment(gen->f, 0, "init conditional exec mask = ~0:"); + spe_load_int(gen->f, cond_reg, ~0); + + /* update conditional execution mask with the predicate register */ + int tmp_reg = get_itemp(gen); + int s1_reg = get_src_reg(gen, channel, &inst->FullSrcRegisters[0]); + + /* tmp = (s1_reg == 0) */ + spe_ceqi(gen->f, tmp_reg, s1_reg, 0); + /* tmp = !tmp */ + spe_complement(gen->f, tmp_reg, tmp_reg); + /* cond_mask = cond_mask & tmp */ + spe_and(gen->f, cond_reg, cond_reg, tmp_reg); + + gen->if_nesting++; + + /* update the master execution mask */ + emit_update_exec_mask(gen); + + free_itemps(gen); + + return TRUE; +} + + +static boolean +emit_ELSE(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + const int cond_reg = get_cond_mask_reg(gen); + + spe_comment(gen->f, 0, "cond exec mask = !cond exec mask"); + spe_complement(gen->f, cond_reg, cond_reg); + emit_update_exec_mask(gen); + + return TRUE; +} + + +static boolean +emit_ENDIF(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + /* XXX todo: pop cond exec mask */ + + gen->if_nesting--; + + emit_update_exec_mask(gen); + + return TRUE; +} + + +static boolean +emit_BGNLOOP(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int exec_reg, loop_reg; + + exec_reg = get_exec_mask_reg(gen); + loop_reg = get_loop_mask_reg(gen); + + /* XXX push loop_exec mask */ + + spe_comment(gen->f, 0*-4, "initialize loop exec mask = ~0"); + spe_load_int(gen->f, loop_reg, ~0x0); + + gen->loop_nesting++; + gen->loop_start = spe_code_size(gen->f); /* in bytes */ + + return TRUE; +} + + +static boolean +emit_ENDLOOP(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + const int loop_reg = get_loop_mask_reg(gen); + const int tmp_reg = get_itemp(gen); + int offset; + + /* tmp_reg = exec[0] | exec[1] | exec[2] | exec[3] */ + spe_orx(gen->f, tmp_reg, loop_reg); + + offset = gen->loop_start - spe_code_size(gen->f); /* in bytes */ + + /* branch back to top of loop if tmp_reg != 0 */ + spe_brnz(gen->f, tmp_reg, offset / 4); + + /* XXX pop loop_exec mask */ + + gen->loop_nesting--; + + emit_update_exec_mask(gen); + + return TRUE; +} + + +static boolean +emit_BRK(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + const int exec_reg = get_exec_mask_reg(gen); + const int loop_reg = get_loop_mask_reg(gen); + + assert(gen->loop_nesting > 0); + + spe_comment(gen->f, 0, "loop exec mask &= ~master exec mask"); + spe_andc(gen->f, loop_reg, loop_reg, exec_reg); + + emit_update_exec_mask(gen); + + return TRUE; +} + + +static boolean +emit_CONT(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + assert(gen->loop_nesting > 0); + + return TRUE; +} + + +static boolean +emit_DDX_DDY(struct codegen *gen, const struct tgsi_full_instruction *inst, + boolean ddx) +{ + int ch; + + FOR_EACH_ENABLED_CHANNEL(inst, ch) { + int s_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + + int t1_reg = get_itemp(gen); + int t2_reg = get_itemp(gen); + + spe_splat_word(gen->f, t1_reg, s_reg, 0); /* upper-left pixel */ + if (ddx) { + spe_splat_word(gen->f, t2_reg, s_reg, 1); /* upper-right pixel */ + } + else { + spe_splat_word(gen->f, t2_reg, s_reg, 2); /* lower-left pixel */ } + spe_fs(gen->f, d_reg, t2_reg, t1_reg); + + free_itemps(gen); } - return true; + return TRUE; } + + /** * Emit END instruction. * We just return from the shader function at this point. @@ -361,12 +1742,8 @@ emit_SGT(struct codegen *gen, const struct tgsi_full_instruction *inst) static boolean emit_END(struct codegen *gen) { - /* return from function call */ - spe_bi(gen->f, SPE_REG_RA, 0, 0); -#if DISASSEM - printf("bi\trRA\n"); -#endif - return true; + emit_epilogue(gen); + return TRUE; } @@ -378,24 +1755,153 @@ emit_instruction(struct codegen *gen, const struct tgsi_full_instruction *inst) { switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_ARL: + return emit_ARL(gen, inst); case TGSI_OPCODE_MOV: + case TGSI_OPCODE_SWZ: return emit_MOV(gen, inst); - case TGSI_OPCODE_MUL: - return emit_MUL(gen, inst); case TGSI_OPCODE_ADD: - return emit_ADD(gen, inst); + case TGSI_OPCODE_SUB: + case TGSI_OPCODE_MUL: + return emit_binop(gen, inst); + case TGSI_OPCODE_MAD: + return emit_MAD(gen, inst); + case TGSI_OPCODE_LERP: + return emit_LERP(gen, inst); + case TGSI_OPCODE_DP3: + return emit_DP3(gen, inst); + case TGSI_OPCODE_DP4: + return emit_DP4(gen, inst); + case TGSI_OPCODE_DPH: + return emit_DPH(gen, inst); + case TGSI_OPCODE_NRM: + return emit_NRM3(gen, inst); + case TGSI_OPCODE_XPD: + return emit_XPD(gen, inst); + case TGSI_OPCODE_RCP: + case TGSI_OPCODE_RSQ: + return emit_RCP_RSQ(gen, inst); + case TGSI_OPCODE_ABS: + return emit_ABS(gen, inst); case TGSI_OPCODE_SGT: - return emit_SGT(gen, inst); + case TGSI_OPCODE_SLT: + case TGSI_OPCODE_SGE: + case TGSI_OPCODE_SLE: + case TGSI_OPCODE_SEQ: + case TGSI_OPCODE_SNE: + return emit_inequality(gen, inst); + case TGSI_OPCODE_CMP: + return emit_CMP(gen, inst); + case TGSI_OPCODE_MIN: + case TGSI_OPCODE_MAX: + return emit_MIN_MAX(gen, inst); + case TGSI_OPCODE_TRUNC: + return emit_TRUNC(gen, inst); + case TGSI_OPCODE_FLR: + return emit_FLR(gen, inst); + case TGSI_OPCODE_FRC: + return emit_FRC(gen, inst); case TGSI_OPCODE_END: return emit_END(gen); + case TGSI_OPCODE_COS: + return emit_function_call(gen, inst, "spu_cos", 1, TRUE); + case TGSI_OPCODE_SIN: + return emit_function_call(gen, inst, "spu_sin", 1, TRUE); + case TGSI_OPCODE_POW: + return emit_function_call(gen, inst, "spu_pow", 2, TRUE); + case TGSI_OPCODE_EXPBASE2: + return emit_function_call(gen, inst, "spu_exp2", 1, TRUE); + case TGSI_OPCODE_LOGBASE2: + return emit_function_call(gen, inst, "spu_log2", 1, TRUE); + case TGSI_OPCODE_TEX: + /* fall-through for now */ + case TGSI_OPCODE_TXD: + /* fall-through for now */ + case TGSI_OPCODE_TXB: + /* fall-through for now */ + case TGSI_OPCODE_TXL: + /* fall-through for now */ + case TGSI_OPCODE_TXP: + return emit_TEX(gen, inst); + case TGSI_OPCODE_KIL: + return emit_KIL(gen, inst); + + case TGSI_OPCODE_IF: + return emit_IF(gen, inst); + case TGSI_OPCODE_ELSE: + return emit_ELSE(gen, inst); + case TGSI_OPCODE_ENDIF: + return emit_ENDIF(gen, inst); + + case TGSI_OPCODE_BGNLOOP2: + return emit_BGNLOOP(gen, inst); + case TGSI_OPCODE_ENDLOOP2: + return emit_ENDLOOP(gen, inst); + case TGSI_OPCODE_BRK: + return emit_BRK(gen, inst); + case TGSI_OPCODE_CONT: + return emit_CONT(gen, inst); + + case TGSI_OPCODE_DDX: + return emit_DDX_DDY(gen, inst, TRUE); + case TGSI_OPCODE_DDY: + return emit_DDX_DDY(gen, inst, FALSE); + /* XXX lots more cases to do... */ default: - return false; + fprintf(stderr, "Cell: unimplemented TGSI instruction %d!\n", + inst->Instruction.Opcode); + return FALSE; } - return true; + return TRUE; +} + + + +/** + * Emit code for a TGSI immediate value (vector of four floats). + * This involves register allocation and initialization. + * XXX the initialization should be done by a "prepare" stage, not + * per quad execution! + */ +static boolean +emit_immediate(struct codegen *gen, const struct tgsi_full_immediate *immed) +{ + int ch; + + assert(gen->num_imm < MAX_TEMPS); + + for (ch = 0; ch < 4; ch++) { + float val = immed->u.ImmediateFloat32[ch].Float; + + if (ch > 0 && val == immed->u.ImmediateFloat32[ch - 1].Float) { + /* re-use previous register */ + gen->imm_regs[gen->num_imm][ch] = gen->imm_regs[gen->num_imm][ch - 1]; + } + else { + char str[100]; + int reg = spe_allocate_available_register(gen->f); + + if (reg < 0) + return FALSE; + + sprintf(str, "init $%d = %f", reg, val); + spe_comment(gen->f, 0, str); + + /* update immediate map */ + gen->imm_regs[gen->num_imm][ch] = reg; + + /* emit initializer instruction */ + spe_load_float(gen->f, reg, val); + } + } + + gen->num_imm++; + + return TRUE; } @@ -405,44 +1911,46 @@ emit_instruction(struct codegen *gen, * We only care about TGSI TEMPORARY register declarations at this time. * For each TGSI TEMPORARY we allocate four SPE registers. */ -static void -emit_declaration(struct codegen *gen, const struct tgsi_full_declaration *decl) +static boolean +emit_declaration(struct cell_context *cell, + struct codegen *gen, const struct tgsi_full_declaration *decl) { int i, ch; switch (decl->Declaration.File) { case TGSI_FILE_TEMPORARY: -#if DISASSEM - printf("Declare temp reg %d .. %d\n", - decl->DeclarationRange.First, - decl->DeclarationRange.Last); -#endif for (i = decl->DeclarationRange.First; i <= decl->DeclarationRange.Last; i++) { + assert(i < MAX_TEMPS); for (ch = 0; ch < 4; ch++) { gen->temp_regs[i][ch] = spe_allocate_available_register(gen->f); + if (gen->temp_regs[i][ch] < 0) + return FALSE; /* out of regs */ } /* XXX if we run out of SPE registers, we need to spill * to SPU memory. someday... */ -#if DISASSEM - printf(" SPE regs: %d %d %d %d\n", - gen->temp_regs[i][0], - gen->temp_regs[i][1], - gen->temp_regs[i][2], - gen->temp_regs[i][3]); -#endif + { + char buf[100]; + sprintf(buf, "TGSI temp[%d] maps to SPU regs [$%d $%d $%d $%d]", i, + gen->temp_regs[i][0], gen->temp_regs[i][1], + gen->temp_regs[i][2], gen->temp_regs[i][3]); + spe_comment(gen->f, 0, buf); + } } break; default: ; /* ignore */ } + + return TRUE; } + /** * Translate TGSI shader code to SPE instructions. This is done when * the state tracker gives us a new shader (via pipe->create_fs_state()). @@ -458,8 +1966,10 @@ cell_gen_fragment_program(struct cell_context *cell, { struct tgsi_parse_context parse; struct codegen gen; + uint ic = 0; memset(&gen, 0, sizeof(gen)); + gen.cell = cell; gen.f = f; /* For SPE function calls: reg $3 = first param, $4 = second param, etc. */ @@ -472,50 +1982,63 @@ cell_gen_fragment_program(struct cell_context *cell, spe_allocate_register(f, gen.outputs_reg); spe_allocate_register(f, gen.constants_reg); -#if DISASSEM - printf("Begin %s\n", __FUNCTION__); - tgsi_dump(tokens, 0); -#endif + if (cell->debug_flags & CELL_DEBUG_ASM) { + spe_print_code(f, TRUE); + spe_indent(f, 2*8); + printf("Begin %s\n", __FUNCTION__); + tgsi_dump(tokens, 0); + } tgsi_parse_init(&parse, tokens); + emit_prologue(&gen); + while (!tgsi_parse_end_of_tokens(&parse) && !gen.error) { tgsi_parse_token(&parse); switch (parse.FullToken.Token.Type) { case TGSI_TOKEN_TYPE_IMMEDIATE: -#if 0 - if (!note_immediate(&gen, &parse.FullToken.FullImmediate )) - goto fail; -#endif + if (f->print) { + _debug_printf(" # "); + tgsi_dump_immediate(&parse.FullToken.FullImmediate); + } + if (!emit_immediate(&gen, &parse.FullToken.FullImmediate)) + gen.error = TRUE; break; case TGSI_TOKEN_TYPE_DECLARATION: - emit_declaration(&gen, &parse.FullToken.FullDeclaration); + if (f->print) { + _debug_printf(" # "); + tgsi_dump_declaration(&parse.FullToken.FullDeclaration); + } + if (!emit_declaration(cell, &gen, &parse.FullToken.FullDeclaration)) + gen.error = TRUE; break; case TGSI_TOKEN_TYPE_INSTRUCTION: - if (!emit_instruction(&gen, &parse.FullToken.FullInstruction )) { - gen.error = true; + if (f->print) { + _debug_printf(" # "); + ic++; + tgsi_dump_instruction(&parse.FullToken.FullInstruction, ic); } + if (!emit_instruction(&gen, &parse.FullToken.FullInstruction)) + gen.error = TRUE; break; default: assert(0); - } } - if (gen.error) { /* terminate the SPE code */ return emit_END(&gen); } -#if DISASSEM - printf("cell_gen_fragment_program nr instructions: %d\n", f->num_inst); - printf("End %s\n", __FUNCTION__); -#endif + if (cell->debug_flags & CELL_DEBUG_ASM) { + printf("cell_gen_fragment_program nr instructions: %d\n", f->num_inst); + printf("End %s\n", __FUNCTION__); + } tgsi_parse_free( &parse ); diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c index 06219d4e98..66d4b3b6a3 100644 --- a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c @@ -2,6 +2,7 @@ * * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. * All Rights Reserved. + * Copyright 2009 VMware, Inc. All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the @@ -25,11 +26,10 @@ * **************************************************************************/ - - /** * Generate SPU per-fragment code (actually per-quad code). * \author Brian Paul + * \author Bob Ellison */ @@ -54,12 +54,17 @@ * \param ifragZ_reg register containing integer fragment Z values (in) * \param ifbZ_reg register containing integer frame buffer Z values (in/out) * \param zmask_reg register containing result of Z test/comparison (out) + * + * Returns TRUE if the Z-buffer needs to be updated. */ -static void -gen_depth_test(const struct pipe_depth_stencil_alpha_state *dsa, - struct spe_function *f, +static boolean +gen_depth_test(struct spe_function *f, + const struct pipe_depth_stencil_alpha_state *dsa, int mask_reg, int ifragZ_reg, int ifbZ_reg, int zmask_reg) { + /* NOTE: we use clgt below, not cgt, because we want to compare _unsigned_ + * quantities. This only makes a difference for 32-bit Z values though. + */ ASSERT(dsa->depth.enabled); switch (dsa->depth.func) { @@ -79,28 +84,28 @@ gen_depth_test(const struct pipe_depth_stencil_alpha_state *dsa, case PIPE_FUNC_GREATER: /* zmask = (ifragZ > ref) */ - spe_cgt(f, zmask_reg, ifragZ_reg, ifbZ_reg); + spe_clgt(f, zmask_reg, ifragZ_reg, ifbZ_reg); /* mask = (mask & zmask) */ spe_and(f, mask_reg, mask_reg, zmask_reg); break; case PIPE_FUNC_LESS: /* zmask = (ref > ifragZ) */ - spe_cgt(f, zmask_reg, ifbZ_reg, ifragZ_reg); + spe_clgt(f, zmask_reg, ifbZ_reg, ifragZ_reg); /* mask = (mask & zmask) */ spe_and(f, mask_reg, mask_reg, zmask_reg); break; case PIPE_FUNC_LEQUAL: /* zmask = (ifragZ > ref) */ - spe_cgt(f, zmask_reg, ifragZ_reg, ifbZ_reg); + spe_clgt(f, zmask_reg, ifragZ_reg, ifbZ_reg); /* mask = (mask & ~zmask) */ spe_andc(f, mask_reg, mask_reg, zmask_reg); break; case PIPE_FUNC_GEQUAL: /* zmask = (ref > ifragZ) */ - spe_cgt(f, zmask_reg, ifbZ_reg, ifragZ_reg); + spe_clgt(f, zmask_reg, ifbZ_reg, ifragZ_reg); /* mask = (mask & ~zmask) */ spe_andc(f, mask_reg, mask_reg, zmask_reg); break; @@ -129,7 +134,10 @@ gen_depth_test(const struct pipe_depth_stencil_alpha_state *dsa, * framebufferZ = (ztest_passed ? fragmentZ : framebufferZ; */ spe_selb(f, ifbZ_reg, ifbZ_reg, ifragZ_reg, mask_reg); + return TRUE; } + + return FALSE; } @@ -153,7 +161,7 @@ gen_alpha_test(const struct pipe_depth_stencil_alpha_state *dsa, if ((dsa->alpha.func != PIPE_FUNC_NEVER) && (dsa->alpha.func != PIPE_FUNC_ALWAYS)) { /* load/splat the alpha reference float value */ - spe_load_float(f, ref_reg, dsa->alpha.ref); + spe_load_float(f, ref_reg, dsa->alpha.ref_value); } /* emit code to do the alpha comparison, updating 'mask' */ @@ -230,6 +238,134 @@ gen_alpha_test(const struct pipe_depth_stencil_alpha_state *dsa, } +/** + * This pair of functions is used inline to allocate and deallocate + * optional constant registers. Once a constant is discovered to be + * needed, we will likely need it again, so we don't want to deallocate + * it and have to allocate and load it again unnecessarily. + */ +static INLINE void +setup_optional_register(struct spe_function *f, + int *r) +{ + if (*r < 0) + *r = spe_allocate_available_register(f); +} + +static INLINE void +release_optional_register(struct spe_function *f, + int r) +{ + if (r >= 0) + spe_release_register(f, r); +} + +static INLINE void +setup_const_register(struct spe_function *f, + int *r, + float value) +{ + if (*r >= 0) + return; + setup_optional_register(f, r); + spe_load_float(f, *r, value); +} + +static INLINE void +release_const_register(struct spe_function *f, + int r) +{ + release_optional_register(f, r); +} + + + +/** + * Unpack/convert framebuffer colors from four 32-bit packed colors + * (fbRGBA) to four float RGBA vectors (fbR, fbG, fbB, fbA). + * Each 8-bit color component is expanded into a float in [0.0, 1.0]. + */ +static void +unpack_colors(struct spe_function *f, + enum pipe_format color_format, + int fbRGBA_reg, + int fbR_reg, int fbG_reg, int fbB_reg, int fbA_reg) +{ + int mask0_reg = spe_allocate_available_register(f); + int mask1_reg = spe_allocate_available_register(f); + int mask2_reg = spe_allocate_available_register(f); + int mask3_reg = spe_allocate_available_register(f); + + spe_load_int(f, mask0_reg, 0xff); + spe_load_int(f, mask1_reg, 0xff00); + spe_load_int(f, mask2_reg, 0xff0000); + spe_load_int(f, mask3_reg, 0xff000000); + + spe_comment(f, 0, "Unpack framebuffer colors, convert to floats"); + + switch (color_format) { + case PIPE_FORMAT_A8R8G8B8_UNORM: + /* fbB = fbRGBA & mask */ + spe_and(f, fbB_reg, fbRGBA_reg, mask0_reg); + + /* fbG = fbRGBA & mask */ + spe_and(f, fbG_reg, fbRGBA_reg, mask1_reg); + + /* fbR = fbRGBA & mask */ + spe_and(f, fbR_reg, fbRGBA_reg, mask2_reg); + + /* fbA = fbRGBA & mask */ + spe_and(f, fbA_reg, fbRGBA_reg, mask3_reg); + + /* fbG = fbG >> 8 */ + spe_roti(f, fbG_reg, fbG_reg, -8); + + /* fbR = fbR >> 16 */ + spe_roti(f, fbR_reg, fbR_reg, -16); + + /* fbA = fbA >> 24 */ + spe_roti(f, fbA_reg, fbA_reg, -24); + break; + + case PIPE_FORMAT_B8G8R8A8_UNORM: + /* fbA = fbRGBA & mask */ + spe_and(f, fbA_reg, fbRGBA_reg, mask0_reg); + + /* fbR = fbRGBA & mask */ + spe_and(f, fbR_reg, fbRGBA_reg, mask1_reg); + + /* fbG = fbRGBA & mask */ + spe_and(f, fbG_reg, fbRGBA_reg, mask2_reg); + + /* fbB = fbRGBA & mask */ + spe_and(f, fbB_reg, fbRGBA_reg, mask3_reg); + + /* fbR = fbR >> 8 */ + spe_roti(f, fbR_reg, fbR_reg, -8); + + /* fbG = fbG >> 16 */ + spe_roti(f, fbG_reg, fbG_reg, -16); + + /* fbB = fbB >> 24 */ + spe_roti(f, fbB_reg, fbB_reg, -24); + break; + + default: + ASSERT(0); + } + + /* convert int[4] in [0,255] to float[4] in [0.0, 1.0] */ + spe_cuflt(f, fbR_reg, fbR_reg, 8); + spe_cuflt(f, fbG_reg, fbG_reg, 8); + spe_cuflt(f, fbB_reg, fbB_reg, 8); + spe_cuflt(f, fbA_reg, fbA_reg, 8); + + spe_release_register(f, mask0_reg); + spe_release_register(f, mask1_reg); + spe_release_register(f, mask2_reg); + spe_release_register(f, mask3_reg); +} + /** * Generate SPE code to implement the given blend mode for a quad of pixels. @@ -242,6 +378,7 @@ gen_alpha_test(const struct pipe_depth_stencil_alpha_state *dsa, */ static void gen_blend(const struct pipe_blend_state *blend, + const struct pipe_blend_color *blend_color, struct spe_function *f, enum pipe_format color_format, int fragR_reg, int fragG_reg, int fragB_reg, int fragA_reg, @@ -262,211 +399,464 @@ gen_blend(const struct pipe_blend_state *blend, int fbB_reg = spe_allocate_available_register(f); int fbA_reg = spe_allocate_available_register(f); - int one_reg = spe_allocate_available_register(f); int tmp_reg = spe_allocate_available_register(f); - boolean one_reg_set = false; /* avoid setting one_reg more than once */ - - ASSERT(blend->blend_enable); - - /* Unpack/convert framebuffer colors from four 32-bit packed colors - * (fbRGBA) to four float RGBA vectors (fbR, fbG, fbB, fbA). - * Each 8-bit color component is expanded into a float in [0.0, 1.0]. + /* Optional constant registers we might or might not end up using; + * if we do use them, make sure we only allocate them once by + * keeping a flag on each one. */ - { - int mask_reg = spe_allocate_available_register(f); - - /* mask = {0x000000ff, 0x000000ff, 0x000000ff, 0x000000ff} */ - spe_load_int(f, mask_reg, 0xff); - - /* XXX there may be more clever ways to implement the following code */ - switch (color_format) { - case PIPE_FORMAT_A8R8G8B8_UNORM: - /* fbB = fbB & mask */ - spe_and(f, fbB_reg, fbRGBA_reg, mask_reg); - /* mask = mask << 8 */ - spe_roti(f, mask_reg, mask_reg, 8); - - /* fbG = fbRGBA & mask */ - spe_and(f, fbG_reg, fbRGBA_reg, mask_reg); - /* fbG = fbG >> 8 */ - spe_roti(f, fbG_reg, fbG_reg, -8); - /* mask = mask << 8 */ - spe_roti(f, mask_reg, mask_reg, 8); - - /* fbR = fbRGBA & mask */ - spe_and(f, fbR_reg, fbRGBA_reg, mask_reg); - /* fbR = fbR >> 16 */ - spe_roti(f, fbR_reg, fbR_reg, -16); - /* mask = mask << 8 */ - spe_roti(f, mask_reg, mask_reg, 8); - - /* fbA = fbRGBA & mask */ - spe_and(f, fbA_reg, fbRGBA_reg, mask_reg); - /* fbA = fbA >> 24 */ - spe_roti(f, fbA_reg, fbA_reg, -24); - break; - - case PIPE_FORMAT_B8G8R8A8_UNORM: - /* fbA = fbA & mask */ - spe_and(f, fbA_reg, fbRGBA_reg, mask_reg); - /* mask = mask << 8 */ - spe_roti(f, mask_reg, mask_reg, 8); - - /* fbR = fbRGBA & mask */ - spe_and(f, fbR_reg, fbRGBA_reg, mask_reg); - /* fbR = fbR >> 8 */ - spe_roti(f, fbR_reg, fbR_reg, -8); - /* mask = mask << 8 */ - spe_roti(f, mask_reg, mask_reg, 8); - - /* fbG = fbRGBA & mask */ - spe_and(f, fbG_reg, fbRGBA_reg, mask_reg); - /* fbG = fbG >> 16 */ - spe_roti(f, fbG_reg, fbG_reg, -16); - /* mask = mask << 8 */ - spe_roti(f, mask_reg, mask_reg, 8); - - /* fbB = fbRGBA & mask */ - spe_and(f, fbB_reg, fbRGBA_reg, mask_reg); - /* fbB = fbB >> 24 */ - spe_roti(f, fbB_reg, fbB_reg, -24); - break; + int one_reg = -1; + int constR_reg = -1, constG_reg = -1, constB_reg = -1, constA_reg = -1; - default: - ASSERT(0); - } - - /* convert int[4] in [0,255] to float[4] in [0.0, 1.0] */ - spe_cuflt(f, fbR_reg, fbR_reg, 8); - spe_cuflt(f, fbG_reg, fbG_reg, 8); - spe_cuflt(f, fbB_reg, fbB_reg, 8); - spe_cuflt(f, fbA_reg, fbA_reg, 8); - - spe_release_register(f, mask_reg); - } + ASSERT(blend->blend_enable); + /* packed RGBA -> float colors */ + unpack_colors(f, color_format, fbRGBA_reg, + fbR_reg, fbG_reg, fbB_reg, fbA_reg); /* - * Compute Src RGB terms + * Compute Src RGB terms. We're actually looking for the value + * of (the appropriate RGB factors) * (the incoming source RGB color), + * because in some cases (like PIPE_BLENDFACTOR_ONE and + * PIPE_BLENDFACTOR_ZERO) we can avoid doing unnecessary math. */ switch (blend->rgb_src_factor) { case PIPE_BLENDFACTOR_ONE: + /* factors = (1,1,1), so term = (R,G,B) */ spe_move(f, term1R_reg, fragR_reg); spe_move(f, term1G_reg, fragG_reg); spe_move(f, term1B_reg, fragB_reg); break; case PIPE_BLENDFACTOR_ZERO: - spe_zero(f, term1R_reg); - spe_zero(f, term1G_reg); - spe_zero(f, term1B_reg); + /* factors = (0,0,0), so term = (0,0,0) */ + spe_load_float(f, term1R_reg, 0.0f); + spe_load_float(f, term1G_reg, 0.0f); + spe_load_float(f, term1B_reg, 0.0f); break; case PIPE_BLENDFACTOR_SRC_COLOR: + /* factors = (R,G,B), so term = (R*R, G*G, B*B) */ spe_fm(f, term1R_reg, fragR_reg, fragR_reg); spe_fm(f, term1G_reg, fragG_reg, fragG_reg); spe_fm(f, term1B_reg, fragB_reg, fragB_reg); break; case PIPE_BLENDFACTOR_SRC_ALPHA: + /* factors = (A,A,A), so term = (R*A, G*A, B*A) */ spe_fm(f, term1R_reg, fragR_reg, fragA_reg); spe_fm(f, term1G_reg, fragG_reg, fragA_reg); spe_fm(f, term1B_reg, fragB_reg, fragA_reg); break; - /* XXX more cases */ + case PIPE_BLENDFACTOR_INV_SRC_COLOR: + /* factors = (1-R,1-G,1-B), so term = (R*(1-R), G*(1-G), B*(1-B)) + * or in other words term = (R-R*R, G-G*G, B-B*B) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term1R_reg, fragR_reg, fragR_reg, fragR_reg); + spe_fnms(f, term1G_reg, fragG_reg, fragG_reg, fragG_reg); + spe_fnms(f, term1B_reg, fragB_reg, fragB_reg, fragB_reg); + break; + case PIPE_BLENDFACTOR_DST_COLOR: + /* factors = (Rfb,Gfb,Bfb), so term = (R*Rfb, G*Gfb, B*Bfb) */ + spe_fm(f, term1R_reg, fragR_reg, fbR_reg); + spe_fm(f, term1G_reg, fragG_reg, fbG_reg); + spe_fm(f, term1B_reg, fragB_reg, fbB_reg); + break; + case PIPE_BLENDFACTOR_INV_DST_COLOR: + /* factors = (1-Rfb,1-Gfb,1-Bfb), so term = (R*(1-Rfb),G*(1-Gfb),B*(1-Bfb)) + * or term = (R-R*Rfb, G-G*Gfb, B-B*Bfb) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term1R_reg, fragR_reg, fbR_reg, fragR_reg); + spe_fnms(f, term1G_reg, fragG_reg, fbG_reg, fragG_reg); + spe_fnms(f, term1B_reg, fragB_reg, fbB_reg, fragB_reg); + break; + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: + /* factors = (1-A,1-A,1-A), so term = (R*(1-A),G*(1-A),B*(1-A)) + * or term = (R-R*A,G-G*A,B-B*A) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term1R_reg, fragR_reg, fragA_reg, fragR_reg); + spe_fnms(f, term1G_reg, fragG_reg, fragA_reg, fragG_reg); + spe_fnms(f, term1B_reg, fragB_reg, fragA_reg, fragB_reg); + break; + case PIPE_BLENDFACTOR_DST_ALPHA: + /* factors = (Afb, Afb, Afb), so term = (R*Afb, G*Afb, B*Afb) */ + spe_fm(f, term1R_reg, fragR_reg, fbA_reg); + spe_fm(f, term1G_reg, fragG_reg, fbA_reg); + spe_fm(f, term1B_reg, fragB_reg, fbA_reg); + break; + case PIPE_BLENDFACTOR_INV_DST_ALPHA: + /* factors = (1-Afb, 1-Afb, 1-Afb), so term = (R*(1-Afb),G*(1-Afb),B*(1-Afb)) + * or term = (R-R*Afb,G-G*Afb,b-B*Afb) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term1R_reg, fragR_reg, fbA_reg, fragR_reg); + spe_fnms(f, term1G_reg, fragG_reg, fbA_reg, fragG_reg); + spe_fnms(f, term1B_reg, fragB_reg, fbA_reg, fragB_reg); + break; + case PIPE_BLENDFACTOR_CONST_COLOR: + /* We need the optional constant color registers */ + setup_const_register(f, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg, blend_color->color[2]); + /* now, factor = (Rc,Gc,Bc), so term = (R*Rc,G*Gc,B*Bc) */ + spe_fm(f, term1R_reg, fragR_reg, constR_reg); + spe_fm(f, term1G_reg, fragG_reg, constG_reg); + spe_fm(f, term1B_reg, fragB_reg, constB_reg); + break; + case PIPE_BLENDFACTOR_CONST_ALPHA: + /* we'll need the optional constant alpha register */ + setup_const_register(f, &constA_reg, blend_color->color[3]); + /* factor = (Ac,Ac,Ac), so term = (R*Ac,G*Ac,B*Ac) */ + spe_fm(f, term1R_reg, fragR_reg, constA_reg); + spe_fm(f, term1G_reg, fragG_reg, constA_reg); + spe_fm(f, term1B_reg, fragB_reg, constA_reg); + break; + case PIPE_BLENDFACTOR_INV_CONST_COLOR: + /* We need the optional constant color registers */ + setup_const_register(f, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg, blend_color->color[2]); + /* factor = (1-Rc,1-Gc,1-Bc), so term = (R*(1-Rc),G*(1-Gc),B*(1-Bc)) + * or term = (R-R*Rc, G-G*Gc, B-B*Bc) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term1R_reg, fragR_reg, constR_reg, fragR_reg); + spe_fnms(f, term1G_reg, fragG_reg, constG_reg, fragG_reg); + spe_fnms(f, term1B_reg, fragB_reg, constB_reg, fragB_reg); + break; + case PIPE_BLENDFACTOR_INV_CONST_ALPHA: + /* We need the optional constant color registers */ + setup_const_register(f, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg, blend_color->color[2]); + /* factor = (1-Ac,1-Ac,1-Ac), so term = (R*(1-Ac),G*(1-Ac),B*(1-Ac)) + * or term = (R-R*Ac,G-G*Ac,B-B*Ac) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term1R_reg, fragR_reg, constA_reg, fragR_reg); + spe_fnms(f, term1G_reg, fragG_reg, constA_reg, fragG_reg); + spe_fnms(f, term1B_reg, fragB_reg, constA_reg, fragB_reg); + break; + case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: + /* We'll need the optional {1,1,1,1} register */ + setup_const_register(f, &one_reg, 1.0f); + /* factor = (min(A,1-Afb),min(A,1-Afb),min(A,1-Afb)), so + * term = (R*min(A,1-Afb), G*min(A,1-Afb), B*min(A,1-Afb)) + * We could expand the term (as a*min(b,c) == min(a*b,a*c) + * as long as a is positive), but then we'd have to do three + * spe_float_min() functions instead of one, so this is simpler. + */ + /* tmp = 1 - Afb */ + spe_fs(f, tmp_reg, one_reg, fbA_reg); + /* tmp = min(A,tmp) */ + spe_float_min(f, tmp_reg, fragA_reg, tmp_reg); + /* term = R*tmp */ + spe_fm(f, term1R_reg, fragR_reg, tmp_reg); + spe_fm(f, term1G_reg, fragG_reg, tmp_reg); + spe_fm(f, term1B_reg, fragB_reg, tmp_reg); + break; + + /* These are special D3D cases involving a second color output + * from the fragment shader. I'm not sure we can support them + * yet... XXX + */ + case PIPE_BLENDFACTOR_SRC1_COLOR: + case PIPE_BLENDFACTOR_SRC1_ALPHA: + case PIPE_BLENDFACTOR_INV_SRC1_COLOR: + case PIPE_BLENDFACTOR_INV_SRC1_ALPHA: + default: ASSERT(0); } /* - * Compute Src Alpha term + * Compute Src Alpha term. Like the above, we're looking for + * the full term A*factor, not just the factor itself, because + * in many cases we can avoid doing unnecessary multiplies. */ switch (blend->alpha_src_factor) { + case PIPE_BLENDFACTOR_ZERO: + /* factor = 0, so term = 0 */ + spe_load_float(f, term1A_reg, 0.0f); + break; + + case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: /* fall through */ case PIPE_BLENDFACTOR_ONE: + /* factor = 1, so term = A */ spe_move(f, term1A_reg, fragA_reg); break; + case PIPE_BLENDFACTOR_SRC_COLOR: + /* factor = A, so term = A*A */ spe_fm(f, term1A_reg, fragA_reg, fragA_reg); break; case PIPE_BLENDFACTOR_SRC_ALPHA: spe_fm(f, term1A_reg, fragA_reg, fragA_reg); break; - /* XXX more cases */ + + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_INV_SRC_COLOR: + /* factor = 1-A, so term = A*(1-A) = A-A*A */ + /* fnms(a,b,c,d) computes a = d - b*c */ + spe_fnms(f, term1A_reg, fragA_reg, fragA_reg, fragA_reg); + break; + + case PIPE_BLENDFACTOR_DST_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_DST_COLOR: + /* factor = Afb, so term = A*Afb */ + spe_fm(f, term1A_reg, fragA_reg, fbA_reg); + break; + + case PIPE_BLENDFACTOR_INV_DST_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_INV_DST_COLOR: + /* factor = 1-Afb, so term = A*(1-Afb) = A - A*Afb */ + /* fnms(a,b,c,d) computes a = d - b*c */ + spe_fnms(f, term1A_reg, fragA_reg, fbA_reg, fragA_reg); + break; + + case PIPE_BLENDFACTOR_CONST_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_CONST_COLOR: + /* We need the optional constA_reg register */ + setup_const_register(f, &constA_reg, blend_color->color[3]); + /* factor = Ac, so term = A*Ac */ + spe_fm(f, term1A_reg, fragA_reg, constA_reg); + break; + + case PIPE_BLENDFACTOR_INV_CONST_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_INV_CONST_COLOR: + /* We need the optional constA_reg register */ + setup_const_register(f, &constA_reg, blend_color->color[3]); + /* factor = 1-Ac, so term = A*(1-Ac) = A-A*Ac */ + /* fnms(a,b,c,d) computes a = d - b*c */ + spe_fnms(f, term1A_reg, fragA_reg, constA_reg, fragA_reg); + break; + + /* These are special D3D cases involving a second color output + * from the fragment shader. I'm not sure we can support them + * yet... XXX + */ + case PIPE_BLENDFACTOR_SRC1_COLOR: + case PIPE_BLENDFACTOR_SRC1_ALPHA: + case PIPE_BLENDFACTOR_INV_SRC1_COLOR: + case PIPE_BLENDFACTOR_INV_SRC1_ALPHA: default: ASSERT(0); } /* - * Compute Dest RGB terms + * Compute Dest RGB term. Like the above, we're looking for + * the full term (Rfb,Gfb,Bfb)*(factor), not just the factor itself, because + * in many cases we can avoid doing unnecessary multiplies. */ switch (blend->rgb_dst_factor) { case PIPE_BLENDFACTOR_ONE: + /* factors = (1,1,1), so term = (Rfb,Gfb,Bfb) */ spe_move(f, term2R_reg, fbR_reg); spe_move(f, term2G_reg, fbG_reg); spe_move(f, term2B_reg, fbB_reg); break; case PIPE_BLENDFACTOR_ZERO: - spe_zero(f, term2R_reg); - spe_zero(f, term2G_reg); - spe_zero(f, term2B_reg); + /* factor s= (0,0,0), so term = (0,0,0) */ + spe_load_float(f, term2R_reg, 0.0f); + spe_load_float(f, term2G_reg, 0.0f); + spe_load_float(f, term2B_reg, 0.0f); break; case PIPE_BLENDFACTOR_SRC_COLOR: + /* factors = (R,G,B), so term = (R*Rfb, G*Gfb, B*Bfb) */ spe_fm(f, term2R_reg, fbR_reg, fragR_reg); spe_fm(f, term2G_reg, fbG_reg, fragG_reg); spe_fm(f, term2B_reg, fbB_reg, fragB_reg); break; + case PIPE_BLENDFACTOR_INV_SRC_COLOR: + /* factors = (1-R,1-G,1-B), so term = (Rfb*(1-R), Gfb*(1-G), Bfb*(1-B)) + * or in other words term = (Rfb-Rfb*R, Gfb-Gfb*G, Bfb-Bfb*B) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term2R_reg, fragR_reg, fbR_reg, fbR_reg); + spe_fnms(f, term2G_reg, fragG_reg, fbG_reg, fbG_reg); + spe_fnms(f, term2B_reg, fragB_reg, fbB_reg, fbB_reg); + break; case PIPE_BLENDFACTOR_SRC_ALPHA: + /* factors = (A,A,A), so term = (Rfb*A, Gfb*A, Bfb*A) */ spe_fm(f, term2R_reg, fbR_reg, fragA_reg); spe_fm(f, term2G_reg, fbG_reg, fragA_reg); spe_fm(f, term2B_reg, fbB_reg, fragA_reg); break; case PIPE_BLENDFACTOR_INV_SRC_ALPHA: - /* one = {1.0, 1.0, 1.0, 1.0} */ - if (!one_reg_set) { - spe_load_float(f, one_reg, 1.0f); - one_reg_set = true; - } - /* tmp = one - fragA */ - spe_fs(f, tmp_reg, one_reg, fragA_reg); - /* term = fb * tmp */ - spe_fm(f, term2R_reg, fbR_reg, tmp_reg); - spe_fm(f, term2G_reg, fbG_reg, tmp_reg); - spe_fm(f, term2B_reg, fbB_reg, tmp_reg); - break; - /* XXX more cases */ + /* factors = (1-A,1-A,1-A) so term = (Rfb-Rfb*A,Gfb-Gfb*A,Bfb-Bfb*A) */ + /* fnms(a,b,c,d) computes a = d - b*c */ + spe_fnms(f, term2R_reg, fbR_reg, fragA_reg, fbR_reg); + spe_fnms(f, term2G_reg, fbG_reg, fragA_reg, fbG_reg); + spe_fnms(f, term2B_reg, fbB_reg, fragA_reg, fbB_reg); + break; + case PIPE_BLENDFACTOR_DST_COLOR: + /* factors = (Rfb,Gfb,Bfb), so term = (Rfb*Rfb, Gfb*Gfb, Bfb*Bfb) */ + spe_fm(f, term2R_reg, fbR_reg, fbR_reg); + spe_fm(f, term2G_reg, fbG_reg, fbG_reg); + spe_fm(f, term2B_reg, fbB_reg, fbB_reg); + break; + case PIPE_BLENDFACTOR_INV_DST_COLOR: + /* factors = (1-Rfb,1-Gfb,1-Bfb), so term = (Rfb*(1-Rfb),Gfb*(1-Gfb),Bfb*(1-Bfb)) + * or term = (Rfb-Rfb*Rfb, Gfb-Gfb*Gfb, Bfb-Bfb*Bfb) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term2R_reg, fbR_reg, fbR_reg, fbR_reg); + spe_fnms(f, term2G_reg, fbG_reg, fbG_reg, fbG_reg); + spe_fnms(f, term2B_reg, fbB_reg, fbB_reg, fbB_reg); + break; + + case PIPE_BLENDFACTOR_DST_ALPHA: + /* factors = (Afb, Afb, Afb), so term = (Rfb*Afb, Gfb*Afb, Bfb*Afb) */ + spe_fm(f, term2R_reg, fbR_reg, fbA_reg); + spe_fm(f, term2G_reg, fbG_reg, fbA_reg); + spe_fm(f, term2B_reg, fbB_reg, fbA_reg); + break; + case PIPE_BLENDFACTOR_INV_DST_ALPHA: + /* factors = (1-Afb, 1-Afb, 1-Afb), so term = (Rfb*(1-Afb),Gfb*(1-Afb),Bfb*(1-Afb)) + * or term = (Rfb-Rfb*Afb,Gfb-Gfb*Afb,Bfb-Bfb*Afb) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term2R_reg, fbR_reg, fbA_reg, fbR_reg); + spe_fnms(f, term2G_reg, fbG_reg, fbA_reg, fbG_reg); + spe_fnms(f, term2B_reg, fbB_reg, fbA_reg, fbB_reg); + break; + case PIPE_BLENDFACTOR_CONST_COLOR: + /* We need the optional constant color registers */ + setup_const_register(f, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg, blend_color->color[2]); + /* now, factor = (Rc,Gc,Bc), so term = (Rfb*Rc,Gfb*Gc,Bfb*Bc) */ + spe_fm(f, term2R_reg, fbR_reg, constR_reg); + spe_fm(f, term2G_reg, fbG_reg, constG_reg); + spe_fm(f, term2B_reg, fbB_reg, constB_reg); + break; + case PIPE_BLENDFACTOR_CONST_ALPHA: + /* we'll need the optional constant alpha register */ + setup_const_register(f, &constA_reg, blend_color->color[3]); + /* factor = (Ac,Ac,Ac), so term = (Rfb*Ac,Gfb*Ac,Bfb*Ac) */ + spe_fm(f, term2R_reg, fbR_reg, constA_reg); + spe_fm(f, term2G_reg, fbG_reg, constA_reg); + spe_fm(f, term2B_reg, fbB_reg, constA_reg); + break; + case PIPE_BLENDFACTOR_INV_CONST_COLOR: + /* We need the optional constant color registers */ + setup_const_register(f, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg, blend_color->color[2]); + /* factor = (1-Rc,1-Gc,1-Bc), so term = (Rfb*(1-Rc),Gfb*(1-Gc),Bfb*(1-Bc)) + * or term = (Rfb-Rfb*Rc, Gfb-Gfb*Gc, Bfb-Bfb*Bc) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term2R_reg, fbR_reg, constR_reg, fbR_reg); + spe_fnms(f, term2G_reg, fbG_reg, constG_reg, fbG_reg); + spe_fnms(f, term2B_reg, fbB_reg, constB_reg, fbB_reg); + break; + case PIPE_BLENDFACTOR_INV_CONST_ALPHA: + /* We need the optional constant color registers */ + setup_const_register(f, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg, blend_color->color[2]); + /* factor = (1-Ac,1-Ac,1-Ac), so term = (Rfb*(1-Ac),Gfb*(1-Ac),Bfb*(1-Ac)) + * or term = (Rfb-Rfb*Ac,Gfb-Gfb*Ac,Bfb-Bfb*Ac) + * fnms(a,b,c,d) computes a = d - b*c + */ + spe_fnms(f, term2R_reg, fbR_reg, constA_reg, fbR_reg); + spe_fnms(f, term2G_reg, fbG_reg, constA_reg, fbG_reg); + spe_fnms(f, term2B_reg, fbB_reg, constA_reg, fbB_reg); + break; + case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: /* not supported for dest RGB */ + ASSERT(0); + break; + + /* These are special D3D cases involving a second color output + * from the fragment shader. I'm not sure we can support them + * yet... XXX + */ + case PIPE_BLENDFACTOR_SRC1_COLOR: + case PIPE_BLENDFACTOR_SRC1_ALPHA: + case PIPE_BLENDFACTOR_INV_SRC1_COLOR: + case PIPE_BLENDFACTOR_INV_SRC1_ALPHA: + default: ASSERT(0); } /* - * Compute Dest Alpha term + * Compute Dest Alpha term. Like the above, we're looking for + * the full term Afb*factor, not just the factor itself, because + * in many cases we can avoid doing unnecessary multiplies. */ switch (blend->alpha_dst_factor) { case PIPE_BLENDFACTOR_ONE: + /* factor = 1, so term = Afb */ spe_move(f, term2A_reg, fbA_reg); break; case PIPE_BLENDFACTOR_ZERO: - spe_zero(f, term2A_reg); + /* factor = 0, so term = 0 */ + spe_load_float(f, term2A_reg, 0.0f); break; - case PIPE_BLENDFACTOR_SRC_ALPHA: + + case PIPE_BLENDFACTOR_SRC_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_SRC_COLOR: + /* factor = A, so term = Afb*A */ spe_fm(f, term2A_reg, fbA_reg, fragA_reg); break; - case PIPE_BLENDFACTOR_INV_SRC_ALPHA: - /* one = {1.0, 1.0, 1.0, 1.0} */ - if (!one_reg_set) { - spe_load_float(f, one_reg, 1.0f); - one_reg_set = true; - } - /* tmp = one - fragA */ - spe_fs(f, tmp_reg, one_reg, fragA_reg); - /* termA = fbA * tmp */ - spe_fm(f, term2A_reg, fbA_reg, tmp_reg); + + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_INV_SRC_COLOR: + /* factor = 1-A, so term = Afb*(1-A) = Afb-Afb*A */ + /* fnms(a,b,c,d) computes a = d - b*c */ + spe_fnms(f, term2A_reg, fbA_reg, fragA_reg, fbA_reg); + break; + + case PIPE_BLENDFACTOR_DST_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_DST_COLOR: + /* factor = Afb, so term = Afb*Afb */ + spe_fm(f, term2A_reg, fbA_reg, fbA_reg); + break; + + case PIPE_BLENDFACTOR_INV_DST_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_INV_DST_COLOR: + /* factor = 1-Afb, so term = Afb*(1-Afb) = Afb - Afb*Afb */ + /* fnms(a,b,c,d) computes a = d - b*c */ + spe_fnms(f, term2A_reg, fbA_reg, fbA_reg, fbA_reg); + break; + + case PIPE_BLENDFACTOR_CONST_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_CONST_COLOR: + /* We need the optional constA_reg register */ + setup_const_register(f, &constA_reg, blend_color->color[3]); + /* factor = Ac, so term = Afb*Ac */ + spe_fm(f, term2A_reg, fbA_reg, constA_reg); break; - /* XXX more cases */ + + case PIPE_BLENDFACTOR_INV_CONST_ALPHA: /* fall through */ + case PIPE_BLENDFACTOR_INV_CONST_COLOR: + /* We need the optional constA_reg register */ + setup_const_register(f, &constA_reg, blend_color->color[3]); + /* factor = 1-Ac, so term = Afb*(1-Ac) = Afb-Afb*Ac */ + /* fnms(a,b,c,d) computes a = d - b*c */ + spe_fnms(f, term2A_reg, fbA_reg, constA_reg, fbA_reg); + break; + + case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: /* not supported for dest alpha */ + ASSERT(0); + break; + + /* These are special D3D cases involving a second color output + * from the fragment shader. I'm not sure we can support them + * yet... XXX + */ + case PIPE_BLENDFACTOR_SRC1_COLOR: + case PIPE_BLENDFACTOR_SRC1_ALPHA: + case PIPE_BLENDFACTOR_INV_SRC1_COLOR: + case PIPE_BLENDFACTOR_INV_SRC1_ALPHA: default: ASSERT(0); } /* - * Combine Src/Dest RGB terms + * Combine Src/Dest RGB terms as per the blend equation. */ switch (blend->rgb_func) { case PIPE_BLEND_ADD: @@ -479,7 +869,21 @@ gen_blend(const struct pipe_blend_state *blend, spe_fs(f, fragG_reg, term1G_reg, term2G_reg); spe_fs(f, fragB_reg, term1B_reg, term2B_reg); break; - /* XXX more cases */ + case PIPE_BLEND_REVERSE_SUBTRACT: + spe_fs(f, fragR_reg, term2R_reg, term1R_reg); + spe_fs(f, fragG_reg, term2G_reg, term1G_reg); + spe_fs(f, fragB_reg, term2B_reg, term1B_reg); + break; + case PIPE_BLEND_MIN: + spe_float_min(f, fragR_reg, term1R_reg, term2R_reg); + spe_float_min(f, fragG_reg, term1G_reg, term2G_reg); + spe_float_min(f, fragB_reg, term1B_reg, term2B_reg); + break; + case PIPE_BLEND_MAX: + spe_float_max(f, fragR_reg, term1R_reg, term2R_reg); + spe_float_max(f, fragG_reg, term1G_reg, term2G_reg); + spe_float_max(f, fragB_reg, term1B_reg, term2B_reg); + break; default: ASSERT(0); } @@ -494,7 +898,15 @@ gen_blend(const struct pipe_blend_state *blend, case PIPE_BLEND_SUBTRACT: spe_fs(f, fragA_reg, term1A_reg, term2A_reg); break; - /* XXX more cases */ + case PIPE_BLEND_REVERSE_SUBTRACT: + spe_fs(f, fragA_reg, term2A_reg, term1A_reg); + break; + case PIPE_BLEND_MIN: + spe_float_min(f, fragA_reg, term1A_reg, term2A_reg); + break; + case PIPE_BLEND_MAX: + spe_float_max(f, fragA_reg, term1A_reg, term2A_reg); + break; default: ASSERT(0); } @@ -514,8 +926,14 @@ gen_blend(const struct pipe_blend_state *blend, spe_release_register(f, fbB_reg); spe_release_register(f, fbA_reg); - spe_release_register(f, one_reg); spe_release_register(f, tmp_reg); + + /* Free any optional registers that actually got used */ + release_const_register(f, one_reg); + release_const_register(f, constR_reg); + release_const_register(f, constG_reg); + release_const_register(f, constB_reg); + release_const_register(f, constA_reg); } @@ -524,24 +942,74 @@ gen_logicop(const struct pipe_blend_state *blend, struct spe_function *f, int fragRGBA_reg, int fbRGBA_reg) { - /* XXX to-do */ - /* operate on 32-bit packed pixels, not float colors */ -} - - -static void -gen_colormask(uint colormask, - struct spe_function *f, - int fragRGBA_reg, int fbRGBA_reg) -{ - /* XXX to-do */ - /* operate on 32-bit packed pixels, not float colors */ + /* We've got four 32-bit RGBA packed pixels in each of + * fragRGBA_reg and fbRGBA_reg, not sets of floating-point + * reds, greens, blues, and alphas. + * */ + ASSERT(blend->logicop_enable); + + switch(blend->logicop_func) { + case PIPE_LOGICOP_CLEAR: /* 0 */ + spe_zero(f, fragRGBA_reg); + break; + case PIPE_LOGICOP_NOR: /* ~(s | d) */ + spe_nor(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg); + break; + case PIPE_LOGICOP_AND_INVERTED: /* ~s & d */ + /* andc R, A, B computes R = A & ~B */ + spe_andc(f, fragRGBA_reg, fbRGBA_reg, fragRGBA_reg); + break; + case PIPE_LOGICOP_COPY_INVERTED: /* ~s */ + spe_complement(f, fragRGBA_reg, fragRGBA_reg); + break; + case PIPE_LOGICOP_AND_REVERSE: /* s & ~d */ + /* andc R, A, B computes R = A & ~B */ + spe_andc(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg); + break; + case PIPE_LOGICOP_INVERT: /* ~d */ + /* Note that (A nor A) == ~(A|A) == ~A */ + spe_nor(f, fragRGBA_reg, fbRGBA_reg, fbRGBA_reg); + break; + case PIPE_LOGICOP_XOR: /* s ^ d */ + spe_xor(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg); + break; + case PIPE_LOGICOP_NAND: /* ~(s & d) */ + spe_nand(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg); + break; + case PIPE_LOGICOP_AND: /* s & d */ + spe_and(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg); + break; + case PIPE_LOGICOP_EQUIV: /* ~(s ^ d) */ + spe_xor(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg); + spe_complement(f, fragRGBA_reg, fragRGBA_reg); + break; + case PIPE_LOGICOP_NOOP: /* d */ + spe_move(f, fragRGBA_reg, fbRGBA_reg); + break; + case PIPE_LOGICOP_OR_INVERTED: /* ~s | d */ + /* orc R, A, B computes R = A | ~B */ + spe_orc(f, fragRGBA_reg, fbRGBA_reg, fragRGBA_reg); + break; + case PIPE_LOGICOP_COPY: /* s */ + break; + case PIPE_LOGICOP_OR_REVERSE: /* s | ~d */ + /* orc R, A, B computes R = A | ~B */ + spe_orc(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg); + break; + case PIPE_LOGICOP_OR: /* s | d */ + spe_or(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg); + break; + case PIPE_LOGICOP_SET: /* 1 */ + spe_load_int(f, fragRGBA_reg, 0xffffffff); + break; + default: + ASSERT(0); + } } - /** - * Generate code to pack a quad of float colors into a four 32-bit integers. + * Generate code to pack a quad of float colors into four 32-bit integers. * * \param f SPE function to append instruction onto. * \param color_format the dest color packing format @@ -557,13 +1025,16 @@ gen_pack_colors(struct spe_function *f, int r_reg, int g_reg, int b_reg, int a_reg, int rgba_reg) { + int rg_reg = spe_allocate_available_register(f); + int ba_reg = spe_allocate_available_register(f); + /* Convert float[4] in [0.0,1.0] to int[4] in [0,~0], with clamping */ spe_cfltu(f, r_reg, r_reg, 32); spe_cfltu(f, g_reg, g_reg, 32); spe_cfltu(f, b_reg, b_reg, 32); spe_cfltu(f, a_reg, a_reg, 32); - /* Shift the most significant bytes to least the significant positions. + /* Shift the most significant bytes to the least significant positions. * I.e.: reg = reg >> 24 */ spe_rotmi(f, r_reg, r_reg, -24); @@ -595,12 +1066,936 @@ gen_pack_colors(struct spe_function *f, * OR-ing all those together gives us four packed colors: * RGBA = {0xffffffff, 0xaa114477, 0xbb225588, 0xcc336699} */ - spe_or(f, rgba_reg, r_reg, g_reg); - spe_or(f, rgba_reg, rgba_reg, b_reg); - spe_or(f, rgba_reg, rgba_reg, a_reg); + spe_or(f, rg_reg, r_reg, g_reg); + spe_or(f, ba_reg, a_reg, b_reg); + spe_or(f, rgba_reg, rg_reg, ba_reg); + + spe_release_register(f, rg_reg); + spe_release_register(f, ba_reg); } +static void +gen_colormask(struct spe_function *f, + uint colormask, + enum pipe_format color_format, + int fragRGBA_reg, int fbRGBA_reg) +{ + /* We've got four 32-bit RGBA packed pixels in each of + * fragRGBA_reg and fbRGBA_reg, not sets of floating-point + * reds, greens, blues, and alphas. Further, the pixels + * are packed according to the given color format, not + * necessarily RGBA... + */ + uint r_mask; + uint g_mask; + uint b_mask; + uint a_mask; + + /* Calculate exactly where the bits for any particular color + * end up, so we can mask them correctly. + */ + switch(color_format) { + case PIPE_FORMAT_A8R8G8B8_UNORM: + /* ARGB */ + a_mask = 0xff000000; + r_mask = 0x00ff0000; + g_mask = 0x0000ff00; + b_mask = 0x000000ff; + break; + case PIPE_FORMAT_B8G8R8A8_UNORM: + /* BGRA */ + b_mask = 0xff000000; + g_mask = 0x00ff0000; + r_mask = 0x0000ff00; + a_mask = 0x000000ff; + break; + default: + ASSERT(0); + } + + /* For each R, G, B, and A component we're supposed to mask out, + * clear its bits. Then our mask operation later will work + * as expected. + */ + if (!(colormask & PIPE_MASK_R)) { + r_mask = 0; + } + if (!(colormask & PIPE_MASK_G)) { + g_mask = 0; + } + if (!(colormask & PIPE_MASK_B)) { + b_mask = 0; + } + if (!(colormask & PIPE_MASK_A)) { + a_mask = 0; + } + + /* Get a temporary register to hold the mask that will be applied + * to the fragment + */ + int colormask_reg = spe_allocate_available_register(f); + + /* The actual mask we're going to use is an OR of the remaining R, G, B, + * and A masks. Load the result value into our temporary register. + */ + spe_load_uint(f, colormask_reg, r_mask | g_mask | b_mask | a_mask); + + /* Use the mask register to select between the fragment color + * values and the frame buffer color values. Wherever the + * mask has a 0 bit, the current frame buffer color should override + * the fragment color. Wherever the mask has a 1 bit, the + * fragment color should persevere. The Select Bits (selb rt, rA, rB, rM) + * instruction will select bits from its first operand rA wherever the + * the mask bits rM are 0, and from its second operand rB wherever the + * mask bits rM are 1. That means that the frame buffer color is the + * first operand, and the fragment color the second. + */ + spe_selb(f, fragRGBA_reg, fbRGBA_reg, fragRGBA_reg, colormask_reg); + + /* Release the temporary register and we're done */ + spe_release_register(f, colormask_reg); +} + + +/** + * This function is annoyingly similar to gen_depth_test(), above, except + * that instead of comparing two varying values (i.e. fragment and buffer), + * we're comparing a varying value with a static value. As such, we have + * access to the Compare Immediate instructions where we don't in + * gen_depth_test(), which is what makes us very different. + * + * There's some added complexity if there's a non-trivial state->mask + * value; then stencil and reference both must be masked + * + * The return value in the stencil_pass_reg is a bitmask of valid + * fragments that also passed the stencil test. The bitmask of valid + * fragments that failed would be found in + * (fragment_mask_reg & ~stencil_pass_reg). + */ +static void +gen_stencil_test(struct spe_function *f, + const struct pipe_stencil_state *state, + uint stencil_max_value, + int fragment_mask_reg, + int fbS_reg, + int stencil_pass_reg) +{ + /* Generate code that puts the set of passing fragments into the + * stencil_pass_reg register, taking into account whether each fragment + * was active to begin with. + */ + switch (state->func) { + case PIPE_FUNC_EQUAL: + if (state->valuemask == stencil_max_value) { + /* stencil_pass = fragment_mask & (s == reference) */ + spe_compare_equal_uint(f, stencil_pass_reg, fbS_reg, state->ref_value); + spe_and(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + } + else { + /* stencil_pass = fragment_mask & ((s&mask) == (reference&mask)) */ + uint tmp_masked_stencil = spe_allocate_available_register(f); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->valuemask); + spe_compare_equal_uint(f, stencil_pass_reg, tmp_masked_stencil, + state->valuemask & state->ref_value); + spe_and(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_masked_stencil); + } + break; + + case PIPE_FUNC_NOTEQUAL: + if (state->valuemask == stencil_max_value) { + /* stencil_pass = fragment_mask & ~(s == reference) */ + spe_compare_equal_uint(f, stencil_pass_reg, fbS_reg, state->ref_value); + spe_andc(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + } + else { + /* stencil_pass = fragment_mask & ~((s&mask) == (reference&mask)) */ + int tmp_masked_stencil = spe_allocate_available_register(f); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->valuemask); + spe_compare_equal_uint(f, stencil_pass_reg, tmp_masked_stencil, + state->valuemask & state->ref_value); + spe_andc(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_masked_stencil); + } + break; + + case PIPE_FUNC_LESS: + if (state->valuemask == stencil_max_value) { + /* stencil_pass = fragment_mask & (reference < s) */ + spe_compare_greater_uint(f, stencil_pass_reg, fbS_reg, state->ref_value); + spe_and(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + } + else { + /* stencil_pass = fragment_mask & ((reference&mask) < (s & mask)) */ + int tmp_masked_stencil = spe_allocate_available_register(f); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->valuemask); + spe_compare_greater_uint(f, stencil_pass_reg, tmp_masked_stencil, + state->valuemask & state->ref_value); + spe_and(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_masked_stencil); + } + break; + + case PIPE_FUNC_GREATER: + if (state->valuemask == stencil_max_value) { + /* stencil_pass = fragment_mask & (reference > s) */ + /* There's no convenient Compare Less Than Immediate instruction, so + * we'll have to do this one the harder way, by loading a register and + * comparing directly. Compare Logical Greater Than Word (clgt) + * treats its operands as unsigned - no sign extension. + */ + int tmp_reg = spe_allocate_available_register(f); + spe_load_uint(f, tmp_reg, state->ref_value); + spe_clgt(f, stencil_pass_reg, tmp_reg, fbS_reg); + spe_and(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_reg); + } + else { + /* stencil_pass = fragment_mask & ((reference&mask) > (s&mask)) */ + int tmp_reg = spe_allocate_available_register(f); + int tmp_masked_stencil = spe_allocate_available_register(f); + spe_load_uint(f, tmp_reg, state->valuemask & state->ref_value); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->valuemask); + spe_clgt(f, stencil_pass_reg, tmp_reg, tmp_masked_stencil); + spe_and(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_reg); + spe_release_register(f, tmp_masked_stencil); + } + break; + + case PIPE_FUNC_GEQUAL: + if (state->valuemask == stencil_max_value) { + /* stencil_pass = fragment_mask & (reference >= s) + * = fragment_mask & ~(s > reference) */ + spe_compare_greater_uint(f, stencil_pass_reg, fbS_reg, + state->ref_value); + spe_andc(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + } + else { + /* stencil_pass = fragment_mask & ~((s&mask) > (reference&mask)) */ + int tmp_masked_stencil = spe_allocate_available_register(f); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->valuemask); + spe_compare_greater_uint(f, stencil_pass_reg, tmp_masked_stencil, + state->valuemask & state->ref_value); + spe_andc(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_masked_stencil); + } + break; + + case PIPE_FUNC_LEQUAL: + if (state->valuemask == stencil_max_value) { + /* stencil_pass = fragment_mask & (reference <= s) ] + * = fragment_mask & ~(reference > s) */ + /* As above, we have to do this by loading a register */ + int tmp_reg = spe_allocate_available_register(f); + spe_load_uint(f, tmp_reg, state->ref_value); + spe_clgt(f, stencil_pass_reg, tmp_reg, fbS_reg); + spe_andc(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_reg); + } + else { + /* stencil_pass = fragment_mask & ~((reference&mask) > (s&mask)) */ + int tmp_reg = spe_allocate_available_register(f); + int tmp_masked_stencil = spe_allocate_available_register(f); + spe_load_uint(f, tmp_reg, state->ref_value & state->valuemask); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->valuemask); + spe_clgt(f, stencil_pass_reg, tmp_reg, tmp_masked_stencil); + spe_andc(f, stencil_pass_reg, fragment_mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_reg); + spe_release_register(f, tmp_masked_stencil); + } + break; + + case PIPE_FUNC_NEVER: + /* stencil_pass = fragment_mask & 0 = 0 */ + spe_load_uint(f, stencil_pass_reg, 0); + break; + + case PIPE_FUNC_ALWAYS: + /* stencil_pass = fragment_mask & 1 = fragment_mask */ + spe_move(f, stencil_pass_reg, fragment_mask_reg); + break; + } + + /* The fragments that passed the stencil test are now in stencil_pass_reg. + * The fragments that failed would be (fragment_mask_reg & ~stencil_pass_reg). + */ +} + + +/** + * This function generates code that calculates a set of new stencil values + * given the earlier values and the operation to apply. It does not + * apply any tests. It is intended to be called up to 3 times + * (for the stencil fail operation, for the stencil pass-z fail operation, + * and for the stencil pass-z pass operation) to collect up to three + * possible sets of values, and for the caller to combine them based + * on the result of the tests. + * + * stencil_max_value should be (2^n - 1) where n is the number of bits + * in the stencil buffer - in other words, it should be usable as a mask. + */ +static void +gen_stencil_values(struct spe_function *f, + uint stencil_op, + uint stencil_ref_value, + uint stencil_max_value, + int fbS_reg, + int newS_reg) +{ + /* The code below assumes that newS_reg and fbS_reg are not the same + * register; if they can be, the calculations below will have to use + * an additional temporary register. For now, mark the assumption + * with an assertion that will fail if they are the same. + */ + ASSERT(fbS_reg != newS_reg); + + /* The code also assumes the the stencil_max_value is of the form + * 2^n-1 and can therefore be used as a mask for the valid bits in + * addition to a maximum. Make sure this is the case as well. + * The clever math below exploits the fact that incrementing a + * binary number serves to flip all the bits of a number starting at + * the LSB and continuing to (and including) the first zero bit + * found. That means that a number and its increment will always + * have at least one bit in common (the high order bit, if nothing + * else) *unless* the number is zero, *or* the number is of a form + * consisting of some number of 1s in the low-order bits followed + * by nothing but 0s in the high-order bits. The latter case + * implies it's of the form 2^n-1. + */ + ASSERT(stencil_max_value > 0 && ((stencil_max_value + 1) & stencil_max_value) == 0); + + switch(stencil_op) { + case PIPE_STENCIL_OP_KEEP: + /* newS = S */ + spe_move(f, newS_reg, fbS_reg); + break; + + case PIPE_STENCIL_OP_ZERO: + /* newS = 0 */ + spe_zero(f, newS_reg); + break; + + case PIPE_STENCIL_OP_REPLACE: + /* newS = stencil reference value */ + spe_load_uint(f, newS_reg, stencil_ref_value); + break; + + case PIPE_STENCIL_OP_INCR: { + /* newS = (s == max ? max : s + 1) */ + int equals_reg = spe_allocate_available_register(f); + + spe_compare_equal_uint(f, equals_reg, fbS_reg, stencil_max_value); + /* Add Word Immediate computes rT = rA + 10-bit signed immediate */ + spe_ai(f, newS_reg, fbS_reg, 1); + /* Select from the current value or the new value based on the equality test */ + spe_selb(f, newS_reg, newS_reg, fbS_reg, equals_reg); + + spe_release_register(f, equals_reg); + break; + } + case PIPE_STENCIL_OP_DECR: { + /* newS = (s == 0 ? 0 : s - 1) */ + int equals_reg = spe_allocate_available_register(f); + + spe_compare_equal_uint(f, equals_reg, fbS_reg, 0); + /* Add Word Immediate with a (-1) value works */ + spe_ai(f, newS_reg, fbS_reg, -1); + /* Select from the current value or the new value based on the equality test */ + spe_selb(f, newS_reg, newS_reg, fbS_reg, equals_reg); + + spe_release_register(f, equals_reg); + break; + } + case PIPE_STENCIL_OP_INCR_WRAP: + /* newS = (s == max ? 0 : s + 1), but since max is 2^n-1, we can + * do a normal add and mask off the correct bits + */ + spe_ai(f, newS_reg, fbS_reg, 1); + spe_and_uint(f, newS_reg, newS_reg, stencil_max_value); + break; + + case PIPE_STENCIL_OP_DECR_WRAP: + /* newS = (s == 0 ? max : s - 1), but we'll pull the same mask trick as above */ + spe_ai(f, newS_reg, fbS_reg, -1); + spe_and_uint(f, newS_reg, newS_reg, stencil_max_value); + break; + + case PIPE_STENCIL_OP_INVERT: + /* newS = ~s. We take advantage of the mask/max value to invert only + * the valid bits for the field so we don't have to do an extra "and". + */ + spe_xor_uint(f, newS_reg, fbS_reg, stencil_max_value); + break; + + default: + ASSERT(0); + } +} + + +/** + * This function generates code to get all the necessary possible + * stencil values. For each of the output registers (fail_reg, + * zfail_reg, and zpass_reg), it either allocates a new register + * and calculates a new set of values based on the stencil operation, + * or it reuses a register allocation and calculation done for an + * earlier (matching) operation, or it reuses the fbS_reg register + * (if the stencil operation is KEEP, which doesn't change the + * stencil buffer). + * + * Since this function allocates a variable number of registers, + * to avoid incurring complex logic to free them, they should + * be allocated after a spe_allocate_register_set() call + * and released by the corresponding spe_release_register_set() call. + */ +static void +gen_get_stencil_values(struct spe_function *f, + const struct pipe_stencil_state *stencil, + const uint depth_enabled, + int fbS_reg, + int *fail_reg, + int *zfail_reg, + int *zpass_reg) +{ + uint zfail_op; + + /* Stenciling had better be enabled here */ + ASSERT(stencil->enabled); + + /* If the depth test is not enabled, it is treated as though it always + * passes, which means that the zfail_op is not considered - a + * failing stencil test triggers the fail_op, and a passing one + * triggers the zpass_op + * + * As an optimization, override calculation of the zfail_op values + * if they aren't going to be used. By setting the value of + * the operation to PIPE_STENCIL_OP_KEEP, its value will be assumed + * to match the incoming stencil values, and no calculation will + * be done. + */ + if (depth_enabled) { + zfail_op = stencil->zfail_op; + } + else { + zfail_op = PIPE_STENCIL_OP_KEEP; + } + + /* One-sided or front-facing stencil */ + if (stencil->fail_op == PIPE_STENCIL_OP_KEEP) { + *fail_reg = fbS_reg; + } + else { + *fail_reg = spe_allocate_available_register(f); + gen_stencil_values(f, stencil->fail_op, stencil->ref_value, + 0xff, fbS_reg, *fail_reg); + } + + /* Check the possibly overridden value, not the structure value */ + if (zfail_op == PIPE_STENCIL_OP_KEEP) { + *zfail_reg = fbS_reg; + } + else if (zfail_op == stencil->fail_op) { + *zfail_reg = *fail_reg; + } + else { + *zfail_reg = spe_allocate_available_register(f); + gen_stencil_values(f, stencil->zfail_op, stencil->ref_value, + 0xff, fbS_reg, *zfail_reg); + } + + if (stencil->zpass_op == PIPE_STENCIL_OP_KEEP) { + *zpass_reg = fbS_reg; + } + else if (stencil->zpass_op == stencil->fail_op) { + *zpass_reg = *fail_reg; + } + else if (stencil->zpass_op == zfail_op) { + *zpass_reg = *zfail_reg; + } + else { + *zpass_reg = spe_allocate_available_register(f); + gen_stencil_values(f, stencil->zpass_op, stencil->ref_value, + 0xff, fbS_reg, *zpass_reg); + } +} + +/** + * Note that fbZ_reg may *not* be set on entry, if in fact + * the depth test is not enabled. This function must not use + * the register if depth is not enabled. + */ +static boolean +gen_stencil_depth_test(struct spe_function *f, + const struct pipe_depth_stencil_alpha_state *dsa, + const uint facing, + const int mask_reg, const int fragZ_reg, + const int fbZ_reg, const int fbS_reg) +{ + /* True if we've generated code that could require writeback to the + * depth and/or stencil buffers + */ + boolean modified_buffers = FALSE; + + boolean need_to_calculate_stencil_values; + boolean need_to_writemask_stencil_values; + + struct pipe_stencil_state *stencil; + + /* Registers. We may or may not actually allocate these, depending + * on whether the state values indicate that we need them. + */ + int stencil_pass_reg, stencil_fail_reg; + int stencil_fail_values, stencil_pass_depth_fail_values, stencil_pass_depth_pass_values; + int stencil_writemask_reg; + int zmask_reg; + int newS_reg; + + /* Stenciling is quite complex: up to six different configurable stencil + * operations/calculations can be required (three each for front-facing + * and back-facing fragments). Many of those operations will likely + * be identical, so there's good reason to try to avoid calculating + * the same values more than once (which unfortunately makes the code less + * straightforward). + * + * To make register management easier, we start a new + * register set; we can release all the registers in the set at + * once, and avoid having to keep track of exactly which registers + * we allocate. We can still allocate and free registers as + * desired (if we know we no longer need a register), but we don't + * have to spend the complexity to track the more difficult variant + * register usage scenarios. + */ + spe_comment(f, 0, "Allocating stencil register set"); + spe_allocate_register_set(f); + + /* The facing we're given is the fragment facing; it doesn't + * exactly match the stencil facing. If stencil is enabled, + * but two-sided stencil is *not* enabled, we use the same + * stencil settings for both front- and back-facing fragments. + * We only use the "back-facing" stencil for backfacing fragments + * if two-sided stenciling is enabled. + */ + if (facing == CELL_FACING_BACK && dsa->stencil[1].enabled) { + stencil = &dsa->stencil[1]; + } + else { + stencil = &dsa->stencil[0]; + } + + /* Calculate the writemask. If the writemask is trivial (either + * all 0s, meaning that we don't need to calculate any stencil values + * because they're not going to change the stencil anyway, or all 1s, + * meaning that we have to calculate the stencil values but do not + * need to mask them), we can avoid generating code. Don't forget + * that we need to consider backfacing stencil, if enabled. + * + * Note that if the backface stencil is *not* enabled, the backface + * stencil will have the same values as the frontface stencil. + */ + if (stencil->fail_op == PIPE_STENCIL_OP_KEEP && + stencil->zfail_op == PIPE_STENCIL_OP_KEEP && + stencil->zpass_op == PIPE_STENCIL_OP_KEEP) { + need_to_calculate_stencil_values = FALSE; + need_to_writemask_stencil_values = FALSE; + } + else if (stencil->writemask == 0x0) { + /* All changes are writemasked out, so no need to calculate + * what those changes might be, and no need to write anything back. + */ + need_to_calculate_stencil_values = FALSE; + need_to_writemask_stencil_values = FALSE; + } + else if (stencil->writemask == 0xff) { + /* Still trivial, but a little less so. We need to write the stencil + * values, but we don't need to mask them. + */ + need_to_calculate_stencil_values = TRUE; + need_to_writemask_stencil_values = FALSE; + } + else { + /* The general case: calculate, mask, and write */ + need_to_calculate_stencil_values = TRUE; + need_to_writemask_stencil_values = TRUE; + + /* While we're here, generate code that calculates what the + * writemask should be. If backface stenciling is enabled, + * and the backface writemask is not the same as the frontface + * writemask, we'll have to generate code that merges the + * two masks into a single effective mask based on fragment facing. + */ + spe_comment(f, 0, "Computing stencil writemask"); + stencil_writemask_reg = spe_allocate_available_register(f); + spe_load_uint(f, stencil_writemask_reg, dsa->stencil[facing].writemask); + } + + /* At least one-sided stenciling must be on. Generate code that + * runs the stencil test on the basic/front-facing stencil, leaving + * the mask of passing stencil bits in stencil_pass_reg. This mask will + * be used both to mask the set of active pixels, and also to + * determine how the stencil buffer changes. + * + * This test will *not* change the value in mask_reg (because we don't + * yet know whether to apply the two-sided stencil or one-sided stencil). + */ + spe_comment(f, 0, "Running basic stencil test"); + stencil_pass_reg = spe_allocate_available_register(f); + gen_stencil_test(f, stencil, 0xff, mask_reg, fbS_reg, stencil_pass_reg); + + /* Generate code that, given the mask of valid fragments and the + * mask of valid fragments that passed the stencil test, computes + * the mask of valid fragments that failed the stencil test. We + * have to do this before we run a depth test (because the + * depth test should not be performed on fragments that failed the + * stencil test, and because the depth test will update the + * mask of valid fragments based on the results of the depth test). + */ + spe_comment(f, 0, "Computing stencil fail mask and updating fragment mask"); + stencil_fail_reg = spe_allocate_available_register(f); + spe_andc(f, stencil_fail_reg, mask_reg, stencil_pass_reg); + /* Now remove the stenciled-out pixels from the valid fragment mask, + * so we can later use the valid fragment mask in the depth test. + */ + spe_and(f, mask_reg, mask_reg, stencil_pass_reg); + + /* We may not need to calculate stencil values, if the writemask is off */ + if (need_to_calculate_stencil_values) { + /* Generate code that calculates exactly which stencil values we need, + * without calculating the same value twice (say, if two different + * stencil ops have the same value). This code will work for one-sided + * and two-sided stenciling (so that we take into account that operations + * may match between front and back stencils), and will also take into + * account whether the depth test is enabled (if the depth test is off, + * we don't need any of the zfail results, because the depth test always + * is considered to pass if it is disabled). Any register value that + * does not need to be calculated will come back with the same value + * that's in fbS_reg. + * + * This function will allocate a variant number of registers that + * will be released as part of the register set. + */ + spe_comment(f, 0, facing == CELL_FACING_FRONT + ? "Computing front-facing stencil values" + : "Computing back-facing stencil values"); + gen_get_stencil_values(f, stencil, dsa->depth.enabled, fbS_reg, + &stencil_fail_values, &stencil_pass_depth_fail_values, + &stencil_pass_depth_pass_values); + } + + /* We now have all the stencil values we need. We also need + * the results of the depth test to figure out which + * stencil values will become the new stencil values. (Even if + * we aren't actually calculating stencil values, we need to apply + * the depth test if it's enabled.) + * + * The code generated by gen_depth_test() returns the results of the + * test in the given register, but also alters the mask_reg based + * on the results of the test. + */ + if (dsa->depth.enabled) { + spe_comment(f, 0, "Running stencil depth test"); + zmask_reg = spe_allocate_available_register(f); + modified_buffers |= gen_depth_test(f, dsa, mask_reg, fragZ_reg, + fbZ_reg, zmask_reg); + } + + if (need_to_calculate_stencil_values) { + + /* If we need to writemask the stencil values before going into + * the stencil buffer, we'll have to use a new register to + * hold the new values. If not, we can just keep using the + * current register. + */ + if (need_to_writemask_stencil_values) { + newS_reg = spe_allocate_available_register(f); + spe_comment(f, 0, "Saving current stencil values for writemasking"); + spe_move(f, newS_reg, fbS_reg); + } + else { + newS_reg = fbS_reg; + } + + /* Merge in the selected stencil fail values */ + if (stencil_fail_values != fbS_reg) { + spe_comment(f, 0, "Loading stencil fail values"); + spe_selb(f, newS_reg, newS_reg, stencil_fail_values, stencil_fail_reg); + modified_buffers = TRUE; + } + + /* Same for the stencil pass/depth fail values. If this calculation + * is not needed (say, if depth test is off), then the + * stencil_pass_depth_fail_values register will be equal to fbS_reg + * and we'll skip the calculation. + */ + if (stencil_pass_depth_fail_values != fbS_reg) { + /* We don't actually have a stencil pass/depth fail mask yet. + * Calculate it here from the stencil passing mask and the + * depth passing mask. Note that zmask_reg *must* have been + * set above if we're here. + */ + uint stencil_pass_depth_fail_mask = + spe_allocate_available_register(f); + + spe_comment(f, 0, "Loading stencil pass/depth fail values"); + spe_andc(f, stencil_pass_depth_fail_mask, stencil_pass_reg, zmask_reg); + + spe_selb(f, newS_reg, newS_reg, stencil_pass_depth_fail_values, + stencil_pass_depth_fail_mask); + + spe_release_register(f, stencil_pass_depth_fail_mask); + modified_buffers = TRUE; + } + + /* Same for the stencil pass/depth pass mask. Note that we + * *can* get here with zmask_reg being unset (if the depth + * test is off but the stencil test is on). In this case, + * we assume the depth test passes, and don't need to mask + * the stencil pass mask with the Z mask. + */ + if (stencil_pass_depth_pass_values != fbS_reg) { + if (dsa->depth.enabled) { + uint stencil_pass_depth_pass_mask = spe_allocate_available_register(f); + /* We'll need a separate register */ + spe_comment(f, 0, "Loading stencil pass/depth pass values"); + spe_and(f, stencil_pass_depth_pass_mask, stencil_pass_reg, zmask_reg); + spe_selb(f, newS_reg, newS_reg, stencil_pass_depth_pass_values, stencil_pass_depth_pass_mask); + spe_release_register(f, stencil_pass_depth_pass_mask); + } + else { + /* We can use the same stencil-pass register */ + spe_comment(f, 0, "Loading stencil pass values"); + spe_selb(f, newS_reg, newS_reg, stencil_pass_depth_pass_values, stencil_pass_reg); + } + modified_buffers = TRUE; + } + + /* Almost done. If we need to writemask, do it now, leaving the + * results in the fbS_reg register passed in. If we don't need + * to writemask, then the results are *already* in the fbS_reg, + * so there's nothing more to do. + */ + + if (need_to_writemask_stencil_values && modified_buffers) { + /* The Select Bytes command makes a fine writemask. Where + * the mask is 0, the first (original) values are retained, + * effectively masking out changes. Where the mask is 1, the + * second (new) values are retained, incorporating changes. + */ + spe_comment(f, 0, "Writemasking new stencil values"); + spe_selb(f, fbS_reg, fbS_reg, newS_reg, stencil_writemask_reg); + } + + } /* done calculating stencil values */ + + /* The stencil and/or depth values have been applied, and the + * mask_reg, fbS_reg, and fbZ_reg values have been updated. + * We're all done, except that we've allocated a fair number + * of registers that we didn't bother tracking. Release all + * those registers as part of the register set, and go home. + */ + spe_comment(f, 0, "Releasing stencil register set"); + spe_release_register_set(f); + + /* Return TRUE if we could have modified the stencil and/or + * depth buffers. + */ + return modified_buffers; +} + + +/** + * Generate depth and/or stencil test code. + * \param cell context + * \param dsa depth/stencil/alpha state + * \param f spe function to emit + * \param facing either CELL_FACING_FRONT or CELL_FACING_BACK + * \param mask_reg register containing the pixel alive/dead mask + * \param depth_tile_reg register containing address of z/stencil tile + * \param quad_offset_reg offset to quad from start of tile + * \param fragZ_reg register containg fragment Z values + */ +static void +gen_depth_stencil(struct cell_context *cell, + const struct pipe_depth_stencil_alpha_state *dsa, + struct spe_function *f, + uint facing, + int mask_reg, + int depth_tile_reg, + int quad_offset_reg, + int fragZ_reg) + +{ + const enum pipe_format zs_format = cell->framebuffer.zsbuf->format; + boolean write_depth_stencil; + + /* framebuffer's combined z/stencil values register */ + int fbZS_reg = spe_allocate_available_register(f); + + /* Framebufer Z values register */ + int fbZ_reg = spe_allocate_available_register(f); + + /* Framebuffer stencil values register (may not be used) */ + int fbS_reg = spe_allocate_available_register(f); + + /* 24-bit mask register (may not be used) */ + int zmask_reg = spe_allocate_available_register(f); + + /** + * The following code: + * 1. fetch quad of packed Z/S values from the framebuffer tile. + * 2. extract the separate the Z and S values from packed values + * 3. convert fragment Z values from float in [0,1] to 32/24/16-bit ints + * + * The instructions for doing this are interleaved for better performance. + */ + spe_comment(f, 0, "Fetch Z/stencil quad from tile"); + + switch(zs_format) { + case PIPE_FORMAT_S8Z24_UNORM: /* fall through */ + case PIPE_FORMAT_X8Z24_UNORM: + /* prepare mask to extract Z vals from ZS vals */ + spe_load_uint(f, zmask_reg, 0x00ffffff); + + /* convert fragment Z from [0,1] to 32-bit ints */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + + /* Load: fbZS_reg = memory[depth_tile_reg + offset_reg] */ + spe_lqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); + + /* right shift 32-bit fragment Z to 24 bits */ + spe_rotmi(f, fragZ_reg, fragZ_reg, -8); + + /* extract 24-bit Z values from ZS values by masking */ + spe_and(f, fbZ_reg, fbZS_reg, zmask_reg); + + /* extract 8-bit stencil values by shifting */ + spe_rotmi(f, fbS_reg, fbZS_reg, -24); + break; + + case PIPE_FORMAT_Z24S8_UNORM: /* fall through */ + case PIPE_FORMAT_Z24X8_UNORM: + /* convert fragment Z from [0,1] to 32-bit ints */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + + /* Load: fbZS_reg = memory[depth_tile_reg + offset_reg] */ + spe_lqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); + + /* right shift 32-bit fragment Z to 24 bits */ + spe_rotmi(f, fragZ_reg, fragZ_reg, -8); + + /* extract 24-bit Z values from ZS values by shifting */ + spe_rotmi(f, fbZ_reg, fbZS_reg, -8); + + /* extract 8-bit stencil values by masking */ + spe_and_uint(f, fbS_reg, fbZS_reg, 0x000000ff); + break; + + case PIPE_FORMAT_Z32_UNORM: + /* Load: fbZ_reg = memory[depth_tile_reg + offset_reg] */ + spe_lqx(f, fbZ_reg, depth_tile_reg, quad_offset_reg); + + /* convert fragment Z from [0,1] to 32-bit ints */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + + /* No stencil, so can't do anything there */ + break; + + case PIPE_FORMAT_Z16_UNORM: + /* XXX This code for 16bpp Z is broken! */ + + /* Load: fbZS_reg = memory[depth_tile_reg + offset_reg] */ + spe_lqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); + + /* Copy over 4 32-bit values */ + spe_move(f, fbZ_reg, fbZS_reg); + + /* convert Z from [0,1] to 16-bit ints */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + spe_rotmi(f, fragZ_reg, fragZ_reg, -16); + /* No stencil */ + break; + + default: + ASSERT(0); /* invalid format */ + } + + /* If stencil is enabled, use the stencil-specific code + * generator to generate both the stencil and depth (if needed) + * tests. Otherwise, if only depth is enabled, generate + * a quick depth test. The test generators themselves will + * report back whether the depth/stencil buffer has to be + * written back. + */ + if (dsa->stencil[0].enabled) { + /* This will perform the stencil and depth tests, and update + * the mask_reg, fbZ_reg, and fbS_reg as required by the + * tests. + */ + ASSERT(fbS_reg >= 0); + spe_comment(f, 0, "Perform stencil test"); + + /* Note that fbZ_reg may not be set on entry, if stenciling + * is enabled but there's no Z-buffer. The + * gen_stencil_depth_test() function must ignore the + * fbZ_reg register if depth is not enabled. + */ + write_depth_stencil = gen_stencil_depth_test(f, dsa, facing, + mask_reg, fragZ_reg, + fbZ_reg, fbS_reg); + } + else if (dsa->depth.enabled) { + int zmask_reg = spe_allocate_available_register(f); + ASSERT(fbZ_reg >= 0); + spe_comment(f, 0, "Perform depth test"); + write_depth_stencil = gen_depth_test(f, dsa, mask_reg, fragZ_reg, + fbZ_reg, zmask_reg); + spe_release_register(f, zmask_reg); + } + else { + write_depth_stencil = FALSE; + } + + if (write_depth_stencil) { + /* Merge latest Z and Stencil values into fbZS_reg. + * fbZ_reg has four Z vals in bits [23..0] or bits [15..0]. + * fbS_reg has four 8-bit Z values in bits [7..0]. + */ + spe_comment(f, 0, "Store quad's depth/stencil values in tile"); + if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM) { + spe_shli(f, fbS_reg, fbS_reg, 24); /* fbS = fbS << 24 */ + spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */ + } + else if (zs_format == PIPE_FORMAT_Z24S8_UNORM || + zs_format == PIPE_FORMAT_Z24X8_UNORM) { + spe_shli(f, fbZ_reg, fbZ_reg, 8); /* fbZ = fbZ << 8 */ + spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */ + } + else if (zs_format == PIPE_FORMAT_Z32_UNORM) { + spe_move(f, fbZS_reg, fbZ_reg); /* fbZS = fbZ */ + } + else if (zs_format == PIPE_FORMAT_Z16_UNORM) { + spe_move(f, fbZS_reg, fbZ_reg); /* fbZS = fbZ */ + } + else if (zs_format == PIPE_FORMAT_S8_UNORM) { + ASSERT(0); /* XXX to do */ + } + else { + ASSERT(0); /* bad zs_format */ + } + + /* Store: memory[depth_tile_reg + quad_offset_reg] = fbZS */ + spe_stqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); + } + + /* Don't need these any more */ + spe_release_register(f, fbZS_reg); + spe_release_register(f, fbZ_reg); + spe_release_register(f, fbS_reg); + spe_release_register(f, zmask_reg); +} + /** @@ -621,14 +2016,21 @@ gen_pack_colors(struct spe_function *f, * should be much faster. * * \param cell the rendering context (in) - * \param f the generated function (out) + * \param facing whether the generated code is for front-facing or + * back-facing fragments + * \param f the generated function (in/out); on input, the function + * must already have been initialized. On exit, whatever + * instructions within the generated function have had + * the fragment ops appended. */ void -cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) +cell_gen_fragment_function(struct cell_context *cell, + const uint facing, + struct spe_function *f) { - const struct pipe_depth_stencil_alpha_state *dsa = - &cell->depth_stencil->base; - const struct pipe_blend_state *blend = &cell->blend->base; + const struct pipe_depth_stencil_alpha_state *dsa = cell->depth_stencil; + const struct pipe_blend_state *blend = cell->blend; + const struct pipe_blend_color *blend_color = &cell->blend_color; const enum pipe_format color_format = cell->framebuffer.cbufs[0]->format; /* For SPE function calls: reg $3 = first param, $4 = second param, etc. */ @@ -643,15 +2045,23 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) const int fragA_reg = 11; /* vector float */ const int mask_reg = 12; /* vector uint */ + ASSERT(facing == CELL_FACING_FRONT || facing == CELL_FACING_BACK); + /* offset of quad from start of tile * XXX assuming 4-byte pixels for color AND Z/stencil!!!! */ int quad_offset_reg; int fbRGBA_reg; /**< framebuffer's RGBA colors for quad */ - int fbZS_reg; /**< framebuffer's combined z/stencil values for quad */ - spe_init_func(f, SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); + if (cell->debug_flags & CELL_DEBUG_ASM) { + spe_print_code(f, TRUE); + spe_indent(f, 8); + spe_comment(f, -4, facing == CELL_FACING_FRONT + ? "Begin front-facing per-fragment ops" + : "Begin back-facing per-fragment ops"); + } + spe_allocate_register(f, x_reg); spe_allocate_register(f, y_reg); spe_allocate_register(f, color_tile_reg); @@ -665,7 +2075,6 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) quad_offset_reg = spe_allocate_available_register(f); fbRGBA_reg = spe_allocate_available_register(f); - fbZS_reg = spe_allocate_available_register(f); /* compute offset of quad from start of tile, in bytes */ { @@ -674,8 +2083,9 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) ASSERT(TILE_SIZE == 32); - spe_rotmi(f, x2_reg, x_reg, -1); /* x2 = x / 2 */ + spe_comment(f, 0, "Compute quad offset within tile"); spe_rotmi(f, y2_reg, y_reg, -1); /* y2 = y / 2 */ + spe_rotmi(f, x2_reg, x_reg, -1); /* x2 = x / 2 */ spe_shli(f, y2_reg, y2_reg, 4); /* y2 *= 16 */ spe_a(f, quad_offset_reg, y2_reg, x2_reg); /* offset = y2 + x2 */ spe_shli(f, quad_offset_reg, quad_offset_reg, 4); /* offset *= 16 */ @@ -684,139 +2094,33 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) spe_release_register(f, y2_reg); } - + /* Generate the alpha test, if needed. */ if (dsa->alpha.enabled) { gen_alpha_test(dsa, f, mask_reg, fragA_reg); } + /* generate depth and/or stencil test code */ if (dsa->depth.enabled || dsa->stencil[0].enabled) { - const enum pipe_format zs_format = cell->framebuffer.zsbuf->format; - boolean write_depth_stencil; - - int fbZ_reg = spe_allocate_available_register(f); /* Z values */ - int fbS_reg = spe_allocate_available_register(f); /* Stencil values */ - - /* fetch quad of depth/stencil values from tile at (x,y) */ - /* Load: fbZS_reg = memory[depth_tile_reg + offset_reg] */ - spe_lqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); - - if (dsa->depth.enabled) { - /* Extract Z bits from fbZS_reg into fbZ_reg */ - if (zs_format == PIPE_FORMAT_S8Z24_UNORM || - zs_format == PIPE_FORMAT_X8Z24_UNORM) { - int mask_reg = spe_allocate_available_register(f); - spe_fsmbi(f, mask_reg, 0x7777); /* mask[0,1,2,3] = 0x00ffffff */ - spe_and(f, fbZ_reg, fbZS_reg, mask_reg); /* fbZ = fbZS & mask */ - spe_release_register(f, mask_reg); - /* OK, fbZ_reg has four 24-bit Z values now */ - } - else { - /* XXX handle other z/stencil formats */ - ASSERT(0); - } - - /* Convert fragZ values from float[4] to uint[4] */ - if (zs_format == PIPE_FORMAT_S8Z24_UNORM || - zs_format == PIPE_FORMAT_X8Z24_UNORM || - zs_format == PIPE_FORMAT_Z24S8_UNORM || - zs_format == PIPE_FORMAT_Z24X8_UNORM) { - /* 24-bit Z values */ - int scale_reg = spe_allocate_available_register(f); - - /* scale_reg[0,1,2,3] = float(2^24-1) */ - spe_load_float(f, scale_reg, (float) 0xffffff); - - /* XXX these two instructions might be combined */ - spe_fm(f, fragZ_reg, fragZ_reg, scale_reg); /* fragZ *= scale */ - spe_cfltu(f, fragZ_reg, fragZ_reg, 0); /* fragZ = (int) fragZ */ - - spe_release_register(f, scale_reg); - } - else { - /* XXX handle 16-bit Z format */ - ASSERT(0); - } - } - - if (dsa->stencil[0].enabled) { - /* Extract Stencil bit sfrom fbZS_reg into fbS_reg */ - if (zs_format == PIPE_FORMAT_S8Z24_UNORM || - zs_format == PIPE_FORMAT_X8Z24_UNORM) { - /* XXX extract with a shift */ - ASSERT(0); - } - else if (zs_format == PIPE_FORMAT_Z24S8_UNORM || - zs_format == PIPE_FORMAT_Z24X8_UNORM) { - /* XXX extract with a mask */ - ASSERT(0); - } - } - - - if (dsa->stencil[0].enabled) { - /* XXX this may involve depth testing too */ - // gen_stencil_test(dsa, f, ... ); - ASSERT(0); - } - else if (dsa->depth.enabled) { - int zmask_reg = spe_allocate_available_register(f); - gen_depth_test(dsa, f, mask_reg, fragZ_reg, fbZ_reg, zmask_reg); - spe_release_register(f, zmask_reg); - } - - /* do we need to write Z and/or Stencil back into framebuffer? */ - write_depth_stencil = (dsa->depth.writemask | - dsa->stencil[0].write_mask | - dsa->stencil[1].write_mask); - - if (write_depth_stencil) { - /* Merge latest Z and Stencil values into fbZS_reg. - * fbZ_reg has four Z vals in bits [23..0] or bits [15..0]. - * fbS_reg has four 8-bit Z values in bits [7..0]. - */ - if (zs_format == PIPE_FORMAT_S8Z24_UNORM || - zs_format == PIPE_FORMAT_X8Z24_UNORM) { - spe_shli(f, fbS_reg, fbS_reg, 24); /* fbS = fbS << 24 */ - spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */ - } - else if (zs_format == PIPE_FORMAT_S8Z24_UNORM || - zs_format == PIPE_FORMAT_X8Z24_UNORM) { - /* XXX to do */ - ASSERT(0); - } - else if (zs_format == PIPE_FORMAT_Z16_UNORM) { - /* XXX to do */ - ASSERT(0); - } - else if (zs_format == PIPE_FORMAT_S8_UNORM) { - /* XXX to do */ - ASSERT(0); - } - else { - /* bad zs_format */ - ASSERT(0); - } - - /* Store: memory[depth_tile_reg + quad_offset_reg] = fbZS */ - spe_stqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); - } - - spe_release_register(f, fbZ_reg); - spe_release_register(f, fbS_reg); + gen_depth_stencil(cell, dsa, f, + facing, + mask_reg, + depth_tile_reg, + quad_offset_reg, + fragZ_reg); } - /* Get framebuffer quad/colors. We'll need these for blending, * color masking, and to obey the quad/pixel mask. * Load: fbRGBA_reg = memory[color_tile + quad_offset] * Note: if mask={~0,~0,~0,~0} and we're not blending or colormasking * we could skip this load. */ + spe_comment(f, 0, "Fetch quad colors from tile"); spe_lqx(f, fbRGBA_reg, color_tile_reg, quad_offset_reg); - if (blend->blend_enable) { - gen_blend(blend, f, color_format, + spe_comment(f, 0, "Perform blending"); + gen_blend(blend, blend_color, f, color_format, fragR_reg, fragG_reg, fragB_reg, fragA_reg, fbRGBA_reg); } @@ -829,19 +2133,21 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) int rgba_reg = spe_allocate_available_register(f); /* Pack four float colors as four 32-bit int colors */ + spe_comment(f, 0, "Convert float quad colors to packed int framebuffer colors"); gen_pack_colors(f, color_format, fragR_reg, fragG_reg, fragB_reg, fragA_reg, rgba_reg); if (blend->logicop_enable) { + spe_comment(f, 0, "Compute logic op"); gen_logicop(blend, f, rgba_reg, fbRGBA_reg); } - if (blend->colormask != 0xf) { - gen_colormask(blend->colormask, f, rgba_reg, fbRGBA_reg); + if (blend->colormask != PIPE_MASK_RGBA) { + spe_comment(f, 0, "Compute color mask"); + gen_colormask(f, blend->colormask, color_format, rgba_reg, fbRGBA_reg); } - /* Mix fragment colors with framebuffer colors using the quad/pixel mask: * if (mask[i]) * rgba[i] = rgba[i]; @@ -853,6 +2159,7 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) /* Store updated quad in tile: * memory[color_tile + quad_offset] = rgba_reg; */ + spe_comment(f, 0, "Store quad colors into color tile"); spe_stqx(f, rgba_reg, color_tile_reg, quad_offset_reg); spe_release_register(f, rgba_reg); @@ -862,9 +2169,13 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) spe_bi(f, SPE_REG_RA, 0, 0); /* return from function call */ - spe_release_register(f, fbRGBA_reg); - spe_release_register(f, fbZS_reg); spe_release_register(f, quad_offset_reg); -} + if (cell->debug_flags & CELL_DEBUG_ASM) { + char buffer[1024]; + sprintf(buffer, "End %s-facing per-fragment ops: %d instructions", + facing == CELL_FACING_FRONT ? "front" : "back", f->num_inst); + spe_comment(f, -4, buffer); + } +} diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.h b/src/gallium/drivers/cell/ppu/cell_gen_fragment.h index b59de198dc..21b35d1faf 100644 --- a/src/gallium/drivers/cell/ppu/cell_gen_fragment.h +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.h @@ -31,7 +31,7 @@ extern void -cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f); +cell_gen_fragment_function(struct cell_context *cell, const uint facing, struct spe_function *f); #endif /* CELL_GEN_FRAGMENT_H */ diff --git a/src/gallium/drivers/cell/ppu/cell_pipe_state.c b/src/gallium/drivers/cell/ppu/cell_pipe_state.c index 475c6ef0ce..facd9551fe 100644 --- a/src/gallium/drivers/cell/ppu/cell_pipe_state.c +++ b/src/gallium/drivers/cell/ppu/cell_pipe_state.c @@ -35,9 +35,9 @@ #include "draw/draw_context.h" #include "cell_context.h" #include "cell_flush.h" +#include "cell_pipe_state.h" #include "cell_state.h" #include "cell_texture.h" -#include "cell_state_per_fragment.h" @@ -45,24 +45,18 @@ static void * cell_create_blend_state(struct pipe_context *pipe, const struct pipe_blend_state *blend) { - struct cell_blend_state *cb = MALLOC(sizeof(struct cell_blend_state)); - - (void) memcpy(cb, blend, sizeof(*blend)); -#if 0 - cell_generate_alpha_blend(cb); -#endif - return cb; + return mem_dup(blend, sizeof(*blend)); } static void -cell_bind_blend_state(struct pipe_context *pipe, void *state) +cell_bind_blend_state(struct pipe_context *pipe, void *blend) { struct cell_context *cell = cell_context(pipe); draw_flush(cell->draw); - cell->blend = (struct cell_blend_state *) state; + cell->blend = (struct pipe_blend_state *) blend; cell->dirty |= CELL_NEW_BLEND; } @@ -70,10 +64,7 @@ cell_bind_blend_state(struct pipe_context *pipe, void *state) static void cell_delete_blend_state(struct pipe_context *pipe, void *blend) { - struct cell_blend_state *cb = (struct cell_blend_state *) blend; - - spe_release_func(& cb->code); - FREE(cb); + FREE(blend); } @@ -95,41 +86,29 @@ cell_set_blend_color(struct pipe_context *pipe, static void * cell_create_depth_stencil_alpha_state(struct pipe_context *pipe, - const struct pipe_depth_stencil_alpha_state *depth_stencil) + const struct pipe_depth_stencil_alpha_state *dsa) { - struct cell_depth_stencil_alpha_state *cdsa = - MALLOC(sizeof(struct cell_depth_stencil_alpha_state)); - - (void) memcpy(cdsa, depth_stencil, sizeof(*depth_stencil)); -#if 0 - cell_generate_depth_stencil_test(cdsa); -#endif - return cdsa; + return mem_dup(dsa, sizeof(*dsa)); } static void cell_bind_depth_stencil_alpha_state(struct pipe_context *pipe, - void *depth_stencil) + void *dsa) { struct cell_context *cell = cell_context(pipe); draw_flush(cell->draw); - cell->depth_stencil = - (struct cell_depth_stencil_alpha_state *) depth_stencil; + cell->depth_stencil = (struct pipe_depth_stencil_alpha_state *) dsa; cell->dirty |= CELL_NEW_DEPTH_STENCIL; } static void -cell_delete_depth_stencil_alpha_state(struct pipe_context *pipe, void *depth) +cell_delete_depth_stencil_alpha_state(struct pipe_context *pipe, void *dsa) { - struct cell_depth_stencil_alpha_state *cdsa = - (struct cell_depth_stencil_alpha_state *) depth; - - spe_release_func(& cdsa->code); - FREE(cdsa); + FREE(dsa); } @@ -191,24 +170,23 @@ cell_set_polygon_stipple( struct pipe_context *pipe, static void * cell_create_rasterizer_state(struct pipe_context *pipe, - const struct pipe_rasterizer_state *setup) + const struct pipe_rasterizer_state *rasterizer) { - struct pipe_rasterizer_state *state - = MALLOC(sizeof(struct pipe_rasterizer_state)); - memcpy(state, setup, sizeof(struct pipe_rasterizer_state)); - return state; + return mem_dup(rasterizer, sizeof(*rasterizer)); } static void -cell_bind_rasterizer_state(struct pipe_context *pipe, void *setup) +cell_bind_rasterizer_state(struct pipe_context *pipe, void *rast) { + struct pipe_rasterizer_state *rasterizer = + (struct pipe_rasterizer_state *) rast; struct cell_context *cell = cell_context(pipe); /* pass-through to draw module */ - draw_set_rasterizer_state(cell->draw, setup); + draw_set_rasterizer_state(cell->draw, rasterizer); - cell->rasterizer = (struct pipe_rasterizer_state *)setup; + cell->rasterizer = rasterizer; cell->dirty |= CELL_NEW_RASTERIZER; } @@ -235,17 +213,24 @@ cell_bind_sampler_states(struct pipe_context *pipe, unsigned num, void **samplers) { struct cell_context *cell = cell_context(pipe); + uint i, changed = 0x0; assert(num <= CELL_MAX_SAMPLERS); draw_flush(cell->draw); - memcpy(cell->sampler, samplers, num * sizeof(void *)); - memset(&cell->sampler[num], 0, (CELL_MAX_SAMPLERS - num) * - sizeof(void *)); - cell->num_samplers = num; + for (i = 0; i < CELL_MAX_SAMPLERS; i++) { + struct pipe_sampler_state *new_samp = i < num ? samplers[i] : NULL; + if (cell->sampler[i] != new_samp) { + cell->sampler[i] = new_samp; + changed |= (1 << i); + } + } - cell->dirty |= CELL_NEW_SAMPLER; + if (changed) { + cell->dirty |= CELL_NEW_SAMPLER; + cell->dirty_samplers |= changed; + } } @@ -263,30 +248,101 @@ cell_set_sampler_textures(struct pipe_context *pipe, unsigned num, struct pipe_texture **texture) { struct cell_context *cell = cell_context(pipe); - uint i; + uint i, changed = 0x0; assert(num <= CELL_MAX_SAMPLERS); - /* Check for no-op */ - if (num == cell->num_textures && - !memcmp(cell->texture, texture, num * sizeof(struct pipe_texture *))) - return; - - draw_flush(cell->draw); - for (i = 0; i < CELL_MAX_SAMPLERS; i++) { - struct pipe_texture *tex = i < num ? texture[i] : NULL; + struct cell_texture *new_tex = cell_texture(i < num ? texture[i] : NULL); + struct cell_texture *old_tex = cell->texture[i]; + if (old_tex != new_tex) { + + pipe_texture_reference((struct pipe_texture **) &cell->texture[i], + (struct pipe_texture *) new_tex); - pipe_texture_reference((struct pipe_texture **) &cell->texture[i], tex); + changed |= (1 << i); + } } + cell->num_textures = num; - cell_update_texture_mapping(cell); + if (changed) { + cell->dirty |= CELL_NEW_TEXTURE; + cell->dirty_textures |= changed; + } +} + - cell->dirty |= CELL_NEW_TEXTURE; +/** + * Map color and z/stencil framebuffer surfaces. + */ +static void +cell_map_surfaces(struct cell_context *cell) +{ + struct pipe_screen *screen = cell->pipe.screen; + uint i; + + for (i = 0; i < 1; i++) { + struct pipe_surface *ps = cell->framebuffer.cbufs[i]; + if (ps) { + cell->cbuf_transfer[i] = + screen->get_tex_transfer(screen, ps->texture, ps->face, + ps->level, ps->zslice, + PIPE_TRANSFER_READ_WRITE, + 0, 0, ps->width, ps->height); + + cell->cbuf_map[i] = + screen->transfer_map(screen, cell->cbuf_transfer[i]); + } + } + + { + struct pipe_surface *ps = cell->framebuffer.zsbuf; + if (ps) { + cell->zsbuf_transfer = + screen->get_tex_transfer(screen, ps->texture, ps->face, + ps->level, ps->zslice, + PIPE_TRANSFER_READ_WRITE, + 0, 0, ps->width, ps->height); + + cell->zsbuf_map = + screen->transfer_map(screen, cell->zsbuf_transfer); + } + } } +/** + * Unmap color and z/stencil framebuffer surfaces. + */ +static void +cell_unmap_surfaces(struct cell_context *cell) +{ + struct pipe_screen *screen = cell->pipe.screen; + uint i; + + for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) { + if (cell->cbuf_transfer[i] && cell->cbuf_map[i]) { + /* unmap color buffer/surface [i] */ + screen->transfer_unmap(screen, cell->cbuf_transfer[i]); + cell->cbuf_map[i] = NULL; + + /* get rid of transfer object [i] */ + screen->tex_transfer_release(screen, &cell->cbuf_transfer[i]); + assert(cell->cbuf_transfer[i] == NULL); + } + } + + if (cell->zsbuf_transfer && cell->zsbuf_map) { + screen->transfer_unmap(screen, cell->zsbuf_transfer); + cell->zsbuf_map = NULL; + + /* get rid of transfer object */ + screen->tex_transfer_release(screen, &cell->zsbuf_transfer); + assert(cell->zsbuf_transfer == NULL); + } +} + static void cell_set_framebuffer_state(struct pipe_context *pipe, @@ -295,24 +351,10 @@ cell_set_framebuffer_state(struct pipe_context *pipe, struct cell_context *cell = cell_context(pipe); if (1 /*memcmp(&cell->framebuffer, fb, sizeof(*fb))*/) { - struct pipe_surface *csurf = fb->cbufs[0]; - struct pipe_surface *zsurf = fb->zsbuf; uint i; - uint flags = (PIPE_BUFFER_USAGE_GPU_WRITE | - PIPE_BUFFER_USAGE_GPU_READ); /* unmap old surfaces */ - for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) { - if (cell->framebuffer.cbufs[i] && cell->cbuf_map[i]) { - pipe_surface_unmap(cell->framebuffer.cbufs[i]); - cell->cbuf_map[i] = NULL; - } - } - - if (cell->framebuffer.zsbuf && cell->zsbuf_map) { - pipe_surface_unmap(cell->framebuffer.zsbuf); - cell->zsbuf_map = NULL; - } + cell_unmap_surfaces(cell); /* Finish any pending rendering to the current surface before * installing a new surface! @@ -324,18 +366,14 @@ cell_set_framebuffer_state(struct pipe_context *pipe, */ cell->framebuffer.width = fb->width; cell->framebuffer.height = fb->height; - cell->framebuffer.num_cbufs = fb->num_cbufs; + cell->framebuffer.nr_cbufs = fb->nr_cbufs; for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) { pipe_surface_reference(&cell->framebuffer.cbufs[i], fb->cbufs[i]); } pipe_surface_reference(&cell->framebuffer.zsbuf, fb->zsbuf); /* map new surfaces */ - if (csurf) - cell->cbuf_map[0] = pipe_surface_map(csurf, flags); - - if (zsurf) - cell->zsbuf_map = pipe_surface_map(zsurf, flags); + cell_map_surfaces(cell); cell->dirty |= CELL_NEW_FRAMEBUFFER; } diff --git a/src/gallium/drivers/cell/ppu/cell_render.c b/src/gallium/drivers/cell/ppu/cell_render.c index dd25ae880e..79cb8df82f 100644 --- a/src/gallium/drivers/cell/ppu/cell_render.c +++ b/src/gallium/drivers/cell/ppu/cell_render.c @@ -152,6 +152,7 @@ cell_flush_prim_buffer(struct cell_context *cell) struct cell_command_render *render = &cell_global.command[i].render; render->prim_type = PIPE_PRIM_TRIANGLES; render->num_verts = cell->prim_buffer.num_verts; + render->front_winding = cell->rasterizer->front_winding; render->vertex_size = cell->vertex_info->size * 4; render->xmin = cell->prim_buffer.xmin; render->ymin = cell->prim_buffer.ymin; diff --git a/src/gallium/drivers/cell/ppu/cell_screen.c b/src/gallium/drivers/cell/ppu/cell_screen.c index 139b3719b6..512d85d352 100644 --- a/src/gallium/drivers/cell/ppu/cell_screen.c +++ b/src/gallium/drivers/cell/ppu/cell_screen.c @@ -27,7 +27,8 @@ #include "util/u_memory.h" -#include "pipe/p_winsys.h" +#include "util/u_simple_screen.h" +#include "pipe/internal/p_winsys_screen.h" #include "pipe/p_defines.h" #include "pipe/p_screen.h" @@ -58,9 +59,9 @@ cell_get_param(struct pipe_screen *screen, int param) case PIPE_CAP_MAX_TEXTURE_IMAGE_UNITS: return CELL_MAX_SAMPLERS; case PIPE_CAP_NPOT_TEXTURES: - return 0; + return 1; case PIPE_CAP_TWO_SIDED_STENCIL: - return 0; + return 1; case PIPE_CAP_GLSL: return 1; case PIPE_CAP_S3TC: @@ -68,19 +69,23 @@ cell_get_param(struct pipe_screen *screen, int param) case PIPE_CAP_ANISOTROPIC_FILTER: return 0; case PIPE_CAP_POINT_SPRITE: - return 0; + return 1; case PIPE_CAP_MAX_RENDER_TARGETS: return 1; case PIPE_CAP_OCCLUSION_QUERY: - return 0; + return 1; case PIPE_CAP_TEXTURE_SHADOW_MAP: - return 0; + return 10; case PIPE_CAP_MAX_TEXTURE_2D_LEVELS: - return 12; /* max 2Kx2K */ + return CELL_MAX_TEXTURE_LEVELS; case PIPE_CAP_MAX_TEXTURE_3D_LEVELS: return 8; /* max 128x128x128 */ case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS: - return 12; /* max 2Kx2K */ + return CELL_MAX_TEXTURE_LEVELS; + case PIPE_CAP_TEXTURE_MIRROR_REPEAT: + return 1; /* XXX not really true */ + case PIPE_CAP_TEXTURE_MIRROR_CLAMP: + return 0; /* XXX to do */ default: return 0; } @@ -165,6 +170,7 @@ cell_create_screen(struct pipe_winsys *winsys) screen->is_format_supported = cell_is_format_supported; cell_init_screen_texture_funcs(screen); + u_simple_screen_init(screen); return screen; } diff --git a/src/gallium/drivers/cell/ppu/cell_spu.c b/src/gallium/drivers/cell/ppu/cell_spu.c index 9508227e29..28e5e6d706 100644 --- a/src/gallium/drivers/cell/ppu/cell_spu.c +++ b/src/gallium/drivers/cell/ppu/cell_spu.c @@ -36,6 +36,7 @@ #include "cell_spu.h" #include "pipe/p_format.h" #include "pipe/p_state.h" +#include "util/u_memory.h" #include "cell/common.h" @@ -52,6 +53,35 @@ struct cell_global_info cell_global; /** + * Scan /proc/cpuinfo to determine the timebase for the system. + * This is used by the SPUs to convert 'decrementer' ticks to seconds. + * There may be a better way to get this value... + */ +static unsigned +get_timebase(void) +{ + FILE *f = fopen("/proc/cpuinfo", "r"); + unsigned timebase; + + assert(f); + while (!feof(f)) { + char line[80]; + fgets(line, sizeof(line), f); + if (strncmp(line, "timebase", 8) == 0) { + char *colon = strchr(line, ':'); + if (colon) { + timebase = atoi(colon + 2); + break; + } + } + } + fclose(f); + + return timebase; +} + + +/** * Write a 1-word message to the given SPE mailbox. */ void @@ -114,6 +144,7 @@ cell_start_spus(struct cell_context *cell) { static boolean one_time_init = FALSE; uint i, j; + uint timebase = get_timebase(); if (one_time_init) { fprintf(stderr, "PPU: Multiple rendering contexts not yet supported " @@ -123,24 +154,29 @@ cell_start_spus(struct cell_context *cell) one_time_init = TRUE; - assert(cell->num_spus <= MAX_SPUS); - - ASSERT_ALIGN16(&cell_global.command[0]); - ASSERT_ALIGN16(&cell_global.command[1]); + assert(cell->num_spus <= CELL_MAX_SPUS); ASSERT_ALIGN16(&cell_global.inits[0]); ASSERT_ALIGN16(&cell_global.inits[1]); + /* + * Initialize the global 'inits' structure for each SPU. + * A pointer to the init struct will be passed to each SPU. + * The SPUs will then each grab their init info with mfc_get(). + */ for (i = 0; i < cell->num_spus; i++) { cell_global.inits[i].id = i; cell_global.inits[i].num_spus = cell->num_spus; cell_global.inits[i].debug_flags = cell->debug_flags; - cell_global.inits[i].cmd = &cell_global.command[i]; + cell_global.inits[i].inv_timebase = 1000.0f / timebase; + for (j = 0; j < CELL_NUM_BUFFERS; j++) { cell_global.inits[i].buffers[j] = cell->buffer[j]; } cell_global.inits[i].buffer_status = &cell->buffer_status[0][0][0]; + cell_global.inits[i].spu_functions = &cell->spu_functions; + cell_global.spe_contexts[i] = spe_context_create(0, NULL); if (!cell_global.spe_contexts[i]) { fprintf(stderr, "spe_context_create() failed\n"); diff --git a/src/gallium/drivers/cell/ppu/cell_spu.h b/src/gallium/drivers/cell/ppu/cell_spu.h index 137f26612e..c93958a9ed 100644 --- a/src/gallium/drivers/cell/ppu/cell_spu.h +++ b/src/gallium/drivers/cell/ppu/cell_spu.h @@ -30,14 +30,12 @@ #include <libspe2.h> -#include <libmisc.h> +#include <pthread.h> #include "cell/common.h" #include "cell_context.h" -#define MAX_SPUS 8 - /** * Global vars, for now anyway. */ @@ -46,14 +44,13 @@ struct cell_global_info /** * SPU/SPE handles, etc */ - spe_context_ptr_t spe_contexts[MAX_SPUS]; - pthread_t spe_threads[MAX_SPUS]; + spe_context_ptr_t spe_contexts[CELL_MAX_SPUS]; + pthread_t spe_threads[CELL_MAX_SPUS]; /** - * Data sent to SPUs + * Data sent to SPUs at start-up */ - struct cell_init_info inits[MAX_SPUS]; - struct cell_command command[MAX_SPUS]; + struct cell_init_info inits[CELL_MAX_SPUS]; }; diff --git a/src/gallium/drivers/cell/ppu/cell_state.h b/src/gallium/drivers/cell/ppu/cell_state.h index a7771a55a3..b193170f9c 100644 --- a/src/gallium/drivers/cell/ppu/cell_state.h +++ b/src/gallium/drivers/cell/ppu/cell_state.h @@ -44,8 +44,9 @@ #define CELL_NEW_TEXTURE 0x800 #define CELL_NEW_VERTEX 0x1000 #define CELL_NEW_VS 0x2000 -#define CELL_NEW_CONSTANTS 0x4000 -#define CELL_NEW_VERTEX_INFO 0x8000 +#define CELL_NEW_VS_CONSTANTS 0x4000 +#define CELL_NEW_FS_CONSTANTS 0x8000 +#define CELL_NEW_VERTEX_INFO 0x10000 extern void diff --git a/src/gallium/drivers/cell/ppu/cell_state_emit.c b/src/gallium/drivers/cell/ppu/cell_state_emit.c index 2da3097983..ff529fe22c 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_emit.c +++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c @@ -25,26 +25,163 @@ * **************************************************************************/ +#include "pipe/p_inlines.h" #include "util/u_memory.h" #include "cell_context.h" #include "cell_gen_fragment.h" #include "cell_state.h" #include "cell_state_emit.h" -#include "cell_state_per_fragment.h" #include "cell_batch.h" #include "cell_texture.h" #include "draw/draw_context.h" #include "draw/draw_private.h" +/** + * Find/create a cell_command_fragment_ops object corresponding to the + * current blend/stencil/z/colormask/etc. state. + */ +static struct cell_command_fragment_ops * +lookup_fragment_ops(struct cell_context *cell) +{ + struct cell_fragment_ops_key key; + struct cell_command_fragment_ops *ops; + + /* + * Build key + */ + memset(&key, 0, sizeof(key)); + key.blend = *cell->blend; + key.blend_color = cell->blend_color; + key.dsa = *cell->depth_stencil; + + if (cell->framebuffer.cbufs[0]) + key.color_format = cell->framebuffer.cbufs[0]->format; + else + key.color_format = PIPE_FORMAT_NONE; + + if (cell->framebuffer.zsbuf) + key.zs_format = cell->framebuffer.zsbuf->format; + else + key.zs_format = PIPE_FORMAT_NONE; + + /* + * Look up key in cache. + */ + ops = (struct cell_command_fragment_ops *) + util_keymap_lookup(cell->fragment_ops_cache, &key); + + /* + * If not found, create/save new fragment ops command. + */ + if (!ops) { + struct spe_function spe_code_front, spe_code_back; + unsigned int facing_dependent, total_code_size; + + if (0) + debug_printf("**** Create New Fragment Ops\n"); + + /* Prepare the buffer that will hold the generated code. The + * "0" passed in for the size means that the SPE code will + * use a default size. + */ + spe_init_func(&spe_code_front, 0); + spe_init_func(&spe_code_back, 0); + + /* Generate new code. Always generate new code for both front-facing + * and back-facing fragments, even if it's the same code in both + * cases. + */ + cell_gen_fragment_function(cell, CELL_FACING_FRONT, &spe_code_front); + cell_gen_fragment_function(cell, CELL_FACING_BACK, &spe_code_back); + + /* Make sure the code is a multiple of 8 bytes long; this is + * required to ensure that the dual pipe instruction alignment + * is correct. It's also important for the SPU unpacking, + * which assumes 8-byte boundaries. + */ + unsigned int front_code_size = spe_code_size(&spe_code_front); + while (front_code_size % 8 != 0) { + spe_lnop(&spe_code_front); + front_code_size = spe_code_size(&spe_code_front); + } + unsigned int back_code_size = spe_code_size(&spe_code_back); + while (back_code_size % 8 != 0) { + spe_lnop(&spe_code_back); + back_code_size = spe_code_size(&spe_code_back); + } + + /* Determine whether the code we generated is facing-dependent, by + * determining whether the generated code is different for the front- + * and back-facing fragments. + */ + if (front_code_size == back_code_size && memcmp(spe_code_front.store, spe_code_back.store, front_code_size) == 0) { + /* Code is identical; only need one copy. */ + facing_dependent = 0; + total_code_size = front_code_size; + } + else { + /* Code is different for front-facing and back-facing fragments. + * Need to send both copies. + */ + facing_dependent = 1; + total_code_size = front_code_size + back_code_size; + } + + /* alloc new fragment ops command. Note that this structure + * has variant length based on the total code size required. + */ + ops = CALLOC_VARIANT_LENGTH_STRUCT(cell_command_fragment_ops, total_code_size); + /* populate the new cell_command_fragment_ops object */ + ops->opcode[0] = CELL_CMD_STATE_FRAGMENT_OPS; + ops->total_code_size = total_code_size; + ops->front_code_index = 0; + memcpy(ops->code, spe_code_front.store, front_code_size); + if (facing_dependent) { + /* We have separate front- and back-facing code. Append the + * back-facing code to the buffer. Be careful because the code + * size is in bytes, but the buffer is of unsigned elements. + */ + ops->back_code_index = front_code_size / sizeof(spe_code_front.store[0]); + memcpy(ops->code + ops->back_code_index, spe_code_back.store, back_code_size); + } + else { + /* Use the same code for front- and back-facing fragments */ + ops->back_code_index = ops->front_code_index; + } + + /* Set the fields for the fallback case. Note that these fields + * (and the whole fallback case) will eventually go away. + */ + ops->dsa = *cell->depth_stencil; + ops->blend = *cell->blend; + ops->blend_color = cell->blend_color; + + /* insert cell_command_fragment_ops object into keymap/cache */ + util_keymap_insert(cell->fragment_ops_cache, &key, ops, NULL); + + /* release rtasm buffer */ + spe_release_func(&spe_code_front); + spe_release_func(&spe_code_back); + } + else { + if (0) + debug_printf("**** Re-use Fragment Ops\n"); + } + + return ops; +} + + + static void emit_state_cmd(struct cell_context *cell, uint cmd, const void *state, uint state_size) { - uint64_t *dst = (uint64_t *) - cell_batch_alloc(cell, ROUNDUP8(sizeof(uint64_t) + state_size)); + uint32_t *dst = (uint32_t *) + cell_batch_alloc16(cell, ROUNDUP16(sizeof(opcode_t) + state_size)); *dst = cmd; - memcpy(dst + 1, state, state_size); + memcpy(dst + 4, state, state_size); } @@ -58,9 +195,10 @@ cell_emit_state(struct cell_context *cell) if (cell->dirty & CELL_NEW_FRAMEBUFFER) { struct pipe_surface *cbuf = cell->framebuffer.cbufs[0]; struct pipe_surface *zbuf = cell->framebuffer.zsbuf; + STATIC_ASSERT(sizeof(struct cell_command_framebuffer) % 16 == 0); struct cell_command_framebuffer *fb - = cell_batch_alloc(cell, sizeof(*fb)); - fb->opcode = CELL_CMD_STATE_FRAMEBUFFER; + = cell_batch_alloc16(cell, sizeof(*fb)); + fb->opcode[0] = CELL_CMD_STATE_FRAMEBUFFER; fb->color_start = cell->cbuf_map[0]; fb->color_format = cbuf->format; fb->depth_start = cell->zsbuf_map; @@ -73,11 +211,20 @@ cell_emit_state(struct cell_context *cell) #endif } + if (cell->dirty & (CELL_NEW_RASTERIZER)) { + STATIC_ASSERT(sizeof(struct cell_command_rasterizer) % 16 == 0); + struct cell_command_rasterizer *rast = + cell_batch_alloc16(cell, sizeof(*rast)); + rast->opcode[0] = CELL_CMD_STATE_RASTERIZER; + rast->rasterizer = *cell->rasterizer; + } + if (cell->dirty & (CELL_NEW_FS)) { /* Send new fragment program to SPUs */ + STATIC_ASSERT(sizeof(struct cell_command_fragment_program) % 16 == 0); struct cell_command_fragment_program *fp - = cell_batch_alloc(cell, sizeof(*fp)); - fp->opcode = CELL_CMD_STATE_FRAGMENT_PROGRAM; + = cell_batch_alloc16(cell, sizeof(*fp)); + fp->opcode[0] = CELL_CMD_STATE_FRAGMENT_PROGRAM; fp->num_inst = cell->fs->code.num_inst; memcpy(&fp->code, cell->fs->code.store, SPU_MAX_FRAGMENT_PROGRAM_INSTS * SPE_INST_SIZE); @@ -90,59 +237,83 @@ cell_emit_state(struct cell_context *cell) } } + if (cell->dirty & (CELL_NEW_FS_CONSTANTS)) { + const uint shader = PIPE_SHADER_FRAGMENT; + const uint num_const = cell->constants[shader].buffer->size / sizeof(float); + uint i, j; + float *buf = cell_batch_alloc16(cell, ROUNDUP16(32 + num_const * sizeof(float))); + uint32_t *ibuf = (uint32_t *) buf; + const float *constants = pipe_buffer_map(cell->pipe.screen, + cell->constants[shader].buffer, + PIPE_BUFFER_USAGE_CPU_READ); + ibuf[0] = CELL_CMD_STATE_FS_CONSTANTS; + ibuf[4] = num_const; + j = 8; + for (i = 0; i < num_const; i++) { + buf[j++] = constants[i]; + } + pipe_buffer_unmap(cell->pipe.screen, cell->constants[shader].buffer); + } + if (cell->dirty & (CELL_NEW_FRAMEBUFFER | CELL_NEW_DEPTH_STENCIL | CELL_NEW_BLEND)) { - /* XXX we don't want to always do codegen here. We should have - * a hash/lookup table to cache previous results... - */ - struct cell_command_fragment_ops *fops - = cell_batch_alloc(cell, sizeof(*fops)); - struct spe_function spe_code; - - /* generate new code */ - cell_gen_fragment_function(cell, &spe_code); - /* put the new code into the batch buffer */ - fops->opcode = CELL_CMD_STATE_FRAGMENT_OPS; - memcpy(&fops->code, spe_code.store, - SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); - fops->dsa = cell->depth_stencil->base; - fops->blend = cell->blend->base; - /* free codegen buffer */ - spe_release_func(&spe_code); + struct cell_command_fragment_ops *fops, *fops_cmd; + /* Note that cell_command_fragment_ops is a variant-sized record */ + fops = lookup_fragment_ops(cell); + fops_cmd = cell_batch_alloc16(cell, ROUNDUP16(sizeof(*fops_cmd) + fops->total_code_size)); + memcpy(fops_cmd, fops, sizeof(*fops) + fops->total_code_size); } if (cell->dirty & CELL_NEW_SAMPLER) { uint i; for (i = 0; i < CELL_MAX_SAMPLERS; i++) { - if (cell->sampler[i]) { - struct cell_command_sampler *sampler - = cell_batch_alloc(cell, sizeof(*sampler)); - sampler->opcode = CELL_CMD_STATE_SAMPLER; - sampler->unit = i; - sampler->state = *cell->sampler[i]; + if (cell->dirty_samplers & (1 << i)) { + if (cell->sampler[i]) { + STATIC_ASSERT(sizeof(struct cell_command_sampler) % 16 == 0); + struct cell_command_sampler *sampler + = cell_batch_alloc16(cell, sizeof(*sampler)); + sampler->opcode[0] = CELL_CMD_STATE_SAMPLER; + sampler->unit = i; + sampler->state = *cell->sampler[i]; + } } } + cell->dirty_samplers = 0x0; } if (cell->dirty & CELL_NEW_TEXTURE) { uint i; for (i = 0;i < CELL_MAX_SAMPLERS; i++) { - struct cell_command_texture *texture - = cell_batch_alloc(cell, sizeof(*texture)); - texture->opcode = CELL_CMD_STATE_TEXTURE; - texture->unit = i; - if (cell->texture[i]) { - texture->start = cell->texture[i]->tiled_data; - texture->width = cell->texture[i]->base.width[0]; - texture->height = cell->texture[i]->base.height[0]; - } - else { - texture->start = NULL; - texture->width = 1; - texture->height = 1; + if (cell->dirty_textures & (1 << i)) { + STATIC_ASSERT(sizeof(struct cell_command_texture) % 16 == 0); + struct cell_command_texture *texture + = (struct cell_command_texture *)cell_batch_alloc16(cell, sizeof(*texture)); + texture->opcode[0] = CELL_CMD_STATE_TEXTURE; + texture->unit = i; + if (cell->texture[i]) { + uint level; + for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) { + texture->start[level] = cell->texture[i]->tiled_mapped[level]; + texture->width[level] = cell->texture[i]->base.width[level]; + texture->height[level] = cell->texture[i]->base.height[level]; + texture->depth[level] = cell->texture[i]->base.depth[level]; + } + texture->target = cell->texture[i]->base.target; + } + else { + uint level; + for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) { + texture->start[level] = NULL; + texture->width[level] = 0; + texture->height[level] = 0; + texture->depth[level] = 0; + } + texture->target = 0; + } } } + cell->dirty_textures = 0x0; } if (cell->dirty & CELL_NEW_VERTEX_INFO) { diff --git a/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c b/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c index 78cb446c14..d97c22b2ef 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c +++ b/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c @@ -297,7 +297,7 @@ emit_stencil_test(struct pipe_depth_stencil_alpha_state *dsa, int face_stencil = spe_allocate_available_register(f); int stencil_src = stencil; const unsigned ref = (dsa->stencil[face].ref_value - & dsa->stencil[face].value_mask); + & dsa->stencil[face].valuemask); boolean complement = FALSE; int stored; int tmp = spe_allocate_available_register(f); @@ -305,9 +305,9 @@ emit_stencil_test(struct pipe_depth_stencil_alpha_state *dsa, if ((dsa->stencil[face].func != PIPE_FUNC_NEVER) && (dsa->stencil[face].func != PIPE_FUNC_ALWAYS) - && (dsa->stencil[face].value_mask != 0x0ff)) { + && (dsa->stencil[face].valuemask != 0x0ff)) { stored = spe_allocate_available_register(f); - spe_andi(f, stored, stencil, dsa->stencil[face].value_mask); + spe_andi(f, stored, stencil, dsa->stencil[face].valuemask); } else { stored = stencil; } @@ -395,7 +395,7 @@ emit_stencil_test(struct pipe_depth_stencil_alpha_state *dsa, * - For depth-pass if the stencil test is NEVER * - Any of the 3 conditions if the operation is KEEP */ - if (dsa->stencil[face].write_mask != 0) { + if (dsa->stencil[face].writemask != 0) { if ((dsa->stencil[face].func != PIPE_FUNC_ALWAYS) && (dsa->stencil[face].fail_op != PIPE_STENCIL_OP_KEEP)) { if (complement) { @@ -449,10 +449,10 @@ emit_stencil_test(struct pipe_depth_stencil_alpha_state *dsa, */ if (stencil_src == stencil) { spe_release_register(f, face_stencil); - } else if (dsa->stencil[face].write_mask != 0x0ff) { + } else if (dsa->stencil[face].writemask != 0x0ff) { int tmp = spe_allocate_available_register(f); - spe_il(f, tmp, dsa->stencil[face].write_mask); + spe_il(f, tmp, dsa->stencil[face].writemask); spe_selb(f, stencil_src, stencil, stencil_src, tmp); spe_release_register(f, tmp); @@ -580,8 +580,8 @@ cell_generate_depth_stencil_test(struct cell_depth_stencil_alpha_state *cdsa) dsa->stencil[i].zpass_op); printf("# ref value / value mask / write mask: %02x %02x %02x\n", dsa->stencil[i].ref_value, - dsa->stencil[i].value_mask, - dsa->stencil[i].write_mask); + dsa->stencil[i].valuemask, + dsa->stencil[i].writemask); } printf("\t.text\n"); diff --git a/src/gallium/drivers/cell/ppu/cell_state_shader.c b/src/gallium/drivers/cell/ppu/cell_state_shader.c index 3a0d066da2..bf517ea563 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_shader.c +++ b/src/gallium/drivers/cell/ppu/cell_state_shader.c @@ -28,7 +28,7 @@ #include "pipe/p_defines.h" #include "util/u_memory.h" #include "pipe/p_inlines.h" -#include "pipe/p_winsys.h" +#include "pipe/internal/p_winsys_screen.h" #include "draw/draw_context.h" #include "tgsi/tgsi_parse.h" @@ -186,18 +186,21 @@ cell_set_constant_buffer(struct pipe_context *pipe, const struct pipe_constant_buffer *buf) { struct cell_context *cell = cell_context(pipe); - struct pipe_winsys *ws = pipe->winsys; assert(shader < PIPE_SHADER_TYPES); assert(index == 0); - /* note: reference counting */ - winsys_buffer_reference(ws, - &cell->constants[shader].buffer, - buf->buffer); - cell->constants[shader].size = buf->size; + draw_flush(cell->draw); - cell->dirty |= CELL_NEW_CONSTANTS; + /* note: reference counting */ + pipe_buffer_reference(pipe->screen, + &cell->constants[shader].buffer, + buf->buffer); + + if (shader == PIPE_SHADER_VERTEX) + cell->dirty |= CELL_NEW_VS_CONSTANTS; + else if (shader == PIPE_SHADER_FRAGMENT) + cell->dirty |= CELL_NEW_FS_CONSTANTS; } diff --git a/src/gallium/drivers/cell/ppu/cell_surface.c b/src/gallium/drivers/cell/ppu/cell_surface.c index 732c64082e..c9203fee08 100644 --- a/src/gallium/drivers/cell/ppu/cell_surface.c +++ b/src/gallium/drivers/cell/ppu/cell_surface.c @@ -27,6 +27,7 @@ #include "util/u_rect.h" #include "cell_context.h" +#include "cell_surface.h" void diff --git a/src/gallium/drivers/cell/ppu/cell_texture.c b/src/gallium/drivers/cell/ppu/cell_texture.c index b6590dfb86..fa52e2cbea 100644 --- a/src/gallium/drivers/cell/ppu/cell_texture.c +++ b/src/gallium/drivers/cell/ppu/cell_texture.c @@ -28,12 +28,13 @@ * Authors: * Keith Whitwell <keith@tungstengraphics.com> * Michel Dänzer <michel@tungstengraphics.com> + * Brian Paul */ #include "pipe/p_context.h" #include "pipe/p_defines.h" #include "pipe/p_inlines.h" -#include "pipe/p_winsys.h" +#include "pipe/internal/p_winsys_screen.h" #include "util/u_math.h" #include "util/u_memory.h" @@ -42,30 +43,31 @@ #include "cell_texture.h" -/* Simple, maximally packed layout. - */ -static unsigned minify( unsigned d ) +static unsigned +minify(unsigned d) { return MAX2(1, d>>1); } static void -cell_texture_layout(struct cell_texture * spt) +cell_texture_layout(struct cell_texture *ct) { - struct pipe_texture *pt = &spt->base; + struct pipe_texture *pt = &ct->base; unsigned level; unsigned width = pt->width[0]; unsigned height = pt->height[0]; unsigned depth = pt->depth[0]; - spt->buffer_size = 0; + ct->buffer_size = 0; for ( level = 0 ; level <= pt->last_level ; level++ ) { unsigned size; unsigned w_tile, h_tile; + assert(level < CELL_MAX_TEXTURE_LEVELS); + /* width, height, rounded up to tile size */ w_tile = align(width, TILE_SIZE); h_tile = align(height, TILE_SIZE); @@ -76,9 +78,9 @@ cell_texture_layout(struct cell_texture * spt) pt->nblocksx[level] = pf_get_nblocksx(&pt->block, w_tile); pt->nblocksy[level] = pf_get_nblocksy(&pt->block, h_tile); - spt->stride[level] = pt->nblocksx[level] * pt->block.size; + ct->stride[level] = pt->nblocksx[level] * pt->block.size; - spt->level_offset[level] = spt->buffer_size; + ct->level_offset[level] = ct->buffer_size; size = pt->nblocksx[level] * pt->nblocksy[level] * pt->block.size; if (pt->target == PIPE_TEXTURE_CUBE) @@ -86,7 +88,7 @@ cell_texture_layout(struct cell_texture * spt) else size *= depth; - spt->buffer_size += size; + ct->buffer_size += size; width = minify(width); height = minify(height); @@ -100,26 +102,25 @@ cell_texture_create(struct pipe_screen *screen, const struct pipe_texture *templat) { struct pipe_winsys *ws = screen->winsys; - struct cell_texture *spt = CALLOC_STRUCT(cell_texture); - if (!spt) + struct cell_texture *ct = CALLOC_STRUCT(cell_texture); + if (!ct) return NULL; - spt->base = *templat; - spt->base.refcount = 1; - spt->base.screen = screen; + ct->base = *templat; + ct->base.refcount = 1; + ct->base.screen = screen; - cell_texture_layout(spt); + cell_texture_layout(ct); - spt->buffer = ws->buffer_create(ws, 32, - PIPE_BUFFER_USAGE_PIXEL, - spt->buffer_size); + ct->buffer = ws->buffer_create(ws, 32, PIPE_BUFFER_USAGE_PIXEL, + ct->buffer_size); - if (!spt->buffer) { - FREE(spt); + if (!ct->buffer) { + FREE(ct); return NULL; } - return &spt->base; + return &ct->base; } @@ -135,244 +136,514 @@ cell_texture_release(struct pipe_screen *screen, __FUNCTION__, (void *) *pt, (*pt)->refcount - 1); */ if (--(*pt)->refcount <= 0) { - struct cell_texture *spt = cell_texture(*pt); + /* Delete this texture now. + * But note that the underlying pipe_buffer may linger... + */ + struct cell_texture *ct = cell_texture(*pt); + uint i; /* - DBG("%s deleting %p\n", __FUNCTION__, (void *) spt); + DBG("%s deleting %p\n", __FUNCTION__, (void *) ct); */ - pipe_buffer_reference(screen, &spt->buffer, NULL); + pipe_buffer_reference(screen, &ct->buffer, NULL); + + for (i = 0; i < CELL_MAX_TEXTURE_LEVELS; i++) { + /* Unreference the tiled image buffer. + * It may not actually be deleted until a fence is hit. + */ + if (ct->tiled_buffer[i]) { + ct->tiled_mapped[i] = NULL; + pipe_buffer_reference(screen, &ct->tiled_buffer[i], NULL); + } + } - FREE(spt); + FREE(ct); } *pt = NULL; } -#if 0 + +/** + * Convert image from linear layout to tiled layout. 4-byte pixels. + */ static void -cell_texture_update(struct pipe_context *pipe, struct pipe_texture *texture, - uint face, uint levelsMask) +twiddle_image_uint(uint w, uint h, uint tile_size, uint *dst, + uint src_stride, const uint *src) { - /* XXX TO DO: re-tile the texture data ... */ + const uint tile_size2 = tile_size * tile_size; + const uint h_t = (h + tile_size - 1) / tile_size; + const uint w_t = (w + tile_size - 1) / tile_size; -} -#endif + uint it, jt; /* tile counters */ + uint i, j; /* intra-tile counters */ + src_stride /= 4; /* convert from bytes to pixels */ -static struct pipe_surface * -cell_get_tex_surface(struct pipe_screen *screen, - struct pipe_texture *pt, - unsigned face, unsigned level, unsigned zslice, - unsigned usage) -{ - struct pipe_winsys *ws = screen->winsys; - struct cell_texture *spt = cell_texture(pt); - struct pipe_surface *ps; + /* loop over dest tiles */ + for (it = 0; it < h_t; it++) { + for (jt = 0; jt < w_t; jt++) { + /* start of dest tile: */ + uint *tdst = dst + (it * w_t + jt) * tile_size2; - ps = ws->surface_alloc(ws); - if (ps) { - assert(ps->refcount); - assert(ps->winsys); - winsys_buffer_reference(ws, &ps->buffer, spt->buffer); - ps->format = pt->format; - ps->block = pt->block; - ps->width = pt->width[level]; - ps->height = pt->height[level]; - ps->nblocksx = pt->nblocksx[level]; - ps->nblocksy = pt->nblocksy[level]; - ps->stride = spt->stride[level]; - ps->offset = spt->level_offset[level]; - ps->usage = usage; + /* compute size of this tile (may be smaller than tile_size) */ + /* XXX note: a compiler bug was found here. That's why the code + * looks as it does. + */ + uint tile_width = w - jt * tile_size; + tile_width = MIN2(tile_width, tile_size); + uint tile_height = h - it * tile_size; + tile_height = MIN2(tile_height, tile_size); - /* XXX may need to override usage flags (see sp_texture.c) */ + /* loop over texels in the tile */ + for (i = 0; i < tile_height; i++) { + for (j = 0; j < tile_width; j++) { + const uint srci = it * tile_size + i; + const uint srcj = jt * tile_size + j; + ASSERT(srci < h); + ASSERT(srcj < w); + tdst[i * tile_size + j] = src[srci * src_stride + srcj]; + } + } + } + } +} - pipe_texture_reference(&ps->texture, pt); - ps->face = face; - ps->level = level; - ps->zslice = zslice; - if (pt->target == PIPE_TEXTURE_CUBE || pt->target == PIPE_TEXTURE_3D) { - ps->offset += ((pt->target == PIPE_TEXTURE_CUBE) ? face : zslice) * - ps->nblocksy * - ps->stride; - } - else { - assert(face == 0); - assert(zslice == 0); +/** + * For Cell. Basically, rearrange the pixels/quads from this layout: + * +--+--+--+--+ + * |p0|p1|p2|p3|.... + * +--+--+--+--+ + * + * to this layout: + * +--+--+ + * |p0|p1|.... + * +--+--+ + * |p2|p3| + * +--+--+ + */ +static void +twiddle_tile(const uint *tileIn, uint *tileOut) +{ + int y, x; + + for (y = 0; y < TILE_SIZE; y+=2) { + for (x = 0; x < TILE_SIZE; x+=2) { + int k = 4 * (y/2 * TILE_SIZE/2 + x/2); + tileOut[y * TILE_SIZE + (x + 0)] = tileIn[k]; + tileOut[y * TILE_SIZE + (x + 1)] = tileIn[k+1]; + tileOut[(y + 1) * TILE_SIZE + (x + 0)] = tileIn[k+2]; + tileOut[(y + 1) * TILE_SIZE + (x + 1)] = tileIn[k+3]; } } - return ps; } - /** - * Copy tile data from linear layout to tiled layout. - * XXX this should be rolled into the future surface-creation code. - * XXX also need "untile" code... + * Convert image from tiled layout to linear layout. 4-byte pixels. */ static void -tile_copy_data(uint w, uint h, uint tile_size, uint *dst, const uint *src) +untwiddle_image_uint(uint w, uint h, uint tile_size, uint *dst, + uint dst_stride, const uint *src) { const uint tile_size2 = tile_size * tile_size; - const uint h_t = h / tile_size, w_t = w / tile_size; - + const uint h_t = (h + tile_size - 1) / tile_size; + const uint w_t = (w + tile_size - 1) / tile_size; + uint *tile_buf; uint it, jt; /* tile counters */ uint i, j; /* intra-tile counters */ - /* loop over dest tiles */ + dst_stride /= 4; /* convert from bytes to pixels */ + + tile_buf = align_malloc(tile_size * tile_size * 4, 16); + + /* loop over src tiles */ for (it = 0; it < h_t; it++) { for (jt = 0; jt < w_t; jt++) { - /* start of dest tile: */ - uint *tdst = dst + (it * w_t + jt) * tile_size2; + /* start of src tile: */ + const uint *tsrc = src + (it * w_t + jt) * tile_size2; + + twiddle_tile(tsrc, tile_buf); + tsrc = tile_buf; + + /* compute size of this tile (may be smaller than tile_size) */ + /* XXX note: a compiler bug was found here. That's why the code + * looks as it does. + */ + uint tile_width = w - jt * tile_size; + tile_width = MIN2(tile_width, tile_size); + uint tile_height = h - it * tile_size; + tile_height = MIN2(tile_height, tile_size); + /* loop over texels in the tile */ - for (i = 0; i < tile_size; i++) { - for (j = 0; j < tile_size; j++) { - const uint srci = it * tile_size + i; - const uint srcj = jt * tile_size + j; - *tdst++ = src[srci * w + srcj]; + for (i = 0; i < tile_height; i++) { + for (j = 0; j < tile_width; j++) { + uint dsti = it * tile_size + i; + uint dstj = jt * tile_size + j; + ASSERT(dsti < h); + ASSERT(dstj < w); + dst[dsti * dst_stride + dstj] = tsrc[i * tile_size + j]; } } } } -} + align_free(tile_buf); +} /** * Convert linear texture image data to tiled format for SPU usage. - * XXX recast this in terms of pipe_surfaces (aka texture views). */ static void -cell_tile_texture(struct cell_context *cell, - struct cell_texture *texture) +cell_twiddle_texture(struct pipe_screen *screen, + struct pipe_surface *surface) { - struct pipe_screen *screen = cell->pipe.screen; - uint face = 0, level = 0, zslice = 0; - struct pipe_surface *surf; - const uint w = texture->base.width[0], h = texture->base.height[0]; - const uint *src; - - /* temporary restrictions: */ - assert(w >= TILE_SIZE); - assert(h >= TILE_SIZE); - assert(w % TILE_SIZE == 0); - assert(h % TILE_SIZE == 0); - - surf = screen->get_tex_surface(screen, &texture->base, face, level, zslice, - PIPE_BUFFER_USAGE_CPU_WRITE); - ASSERT(surf); - - src = (const uint *) pipe_surface_map(surf, PIPE_BUFFER_USAGE_CPU_WRITE); - - if (texture->tiled_data) { - align_free(texture->tiled_data); +#if 0 // XXX fix me + struct cell_texture *ct = cell_texture(surface->texture); + const uint level = surface->level; + const uint texWidth = ct->base.width[level]; + const uint texHeight = ct->base.height[level]; + const uint bufWidth = align(texWidth, TILE_SIZE); + const uint bufHeight = align(texHeight, TILE_SIZE); + const void *map = screen->surface_map(screen, surface, PIPE_BUFFER_USAGE_CPU_READ); + const uint *src = (const uint *) map; + + switch (ct->base.format) { + case PIPE_FORMAT_A8R8G8B8_UNORM: + case PIPE_FORMAT_B8G8R8A8_UNORM: + case PIPE_FORMAT_S8Z24_UNORM: + { + int numFaces = ct->base.target == PIPE_TEXTURE_CUBE ? 6 : 1; + int offset = bufWidth * bufHeight * 4 * surface->face; + uint *dst; + + if (!ct->tiled_buffer[level]) { + /* allocate buffer for tiled data now */ + struct pipe_winsys *ws = screen->winsys; + uint bytes = bufWidth * bufHeight * 4 * numFaces; + ct->tiled_buffer[level] = + ws->buffer_create(ws, 16, PIPE_BUFFER_USAGE_PIXEL, bytes); + /* and map it */ + ct->tiled_mapped[level] = + ws->buffer_map(ws, ct->tiled_buffer[level], + PIPE_BUFFER_USAGE_GPU_READ); + } + dst = (uint *) ((ubyte *) ct->tiled_mapped[level] + offset); + + twiddle_image_uint(texWidth, texHeight, TILE_SIZE, dst, + surface->stride, src); + } + break; + default: + printf("Cell: twiddle unsupported texture format %s\n", + pf_name(ct->base.format)); } - texture->tiled_data = align_malloc(w * h * 4, 16); - tile_copy_data(w, h, TILE_SIZE, texture->tiled_data, src); + screen->surface_unmap(screen, surface); +#endif +} + - pipe_surface_unmap(surf); +/** + * Convert SPU tiled texture image data to linear format for app usage. + */ +static void +cell_untwiddle_texture(struct pipe_screen *screen, + struct pipe_surface *surface) +{ +#if 0 // XXX fix me + struct cell_texture *ct = cell_texture(surface->texture); + const uint level = surface->level; + const uint texWidth = ct->base.width[level]; + const uint texHeight = ct->base.height[level]; + const void *map = screen->surface_map(screen, surface, PIPE_BUFFER_USAGE_CPU_READ); + const uint *src = (const uint *) ((const ubyte *) map + surface->offset); + + switch (ct->base.format) { + case PIPE_FORMAT_A8R8G8B8_UNORM: + case PIPE_FORMAT_B8G8R8A8_UNORM: + case PIPE_FORMAT_S8Z24_UNORM: + { + int numFaces = ct->base.target == PIPE_TEXTURE_CUBE ? 6 : 1; + int offset = surface->stride * texHeight * 4 * surface->face; + uint *dst; + + if (!ct->untiled_data[level]) { + ct->untiled_data[level] = + align_malloc(surface->stride * texHeight * 4 * numFaces, 16); + } - pipe_surface_reference(&surf, NULL); + dst = (uint *) ((ubyte *) ct->untiled_data[level] + offset); + + untwiddle_image_uint(texWidth, texHeight, TILE_SIZE, dst, + surface->stride, src); + } + break; + default: + { + ct->untiled_data[level] = NULL; + printf("Cell: untwiddle unsupported texture format %s\n", + pf_name(ct->base.format)); + } + } + + screen->surface_unmap(screen, surface); +#endif } -void -cell_update_texture_mapping(struct cell_context *cell) +static struct pipe_surface * +cell_get_tex_surface(struct pipe_screen *screen, + struct pipe_texture *pt, + unsigned face, unsigned level, unsigned zslice, + unsigned usage) { -#if 0 - uint face = 0, level = 0, zslice = 0; + struct cell_texture *ct = cell_texture(pt); + struct pipe_surface *ps; + + ps = CALLOC_STRUCT(pipe_surface); + if (ps) { + ps->refcount = 1; + pipe_texture_reference(&ps->texture, pt); + ps->format = pt->format; + //ps->block = pt->block; + ps->width = pt->width[level]; + ps->height = pt->height[level]; + //ps->nblocksx = pt->nblocksx[level]; + //ps->nblocksy = pt->nblocksy[level]; + //ps->stride = ct->stride[level]; + ps->offset = ct->level_offset[level]; + ps->usage = usage; + + /* XXX may need to override usage flags (see sp_texture.c) */ + + pipe_texture_reference(&ps->texture, pt); + ps->face = face; + ps->level = level; + ps->zslice = zslice; + + if (pt->target == PIPE_TEXTURE_CUBE || pt->target == PIPE_TEXTURE_3D) { +#if 0 // XXX fix me + ps->offset += ((pt->target == PIPE_TEXTURE_CUBE) ? face : zslice) * + ps->nblocksy * + ps->stride; #endif - uint i; + } + else { + assert(face == 0); + assert(zslice == 0); + } - for (i = 0; i < CELL_MAX_SAMPLERS; i++) { - if (cell->texture[i]) - cell_tile_texture(cell, cell->texture[i]); + if (ps->usage & PIPE_BUFFER_USAGE_CPU_READ) { + /* convert from tiled to linear layout */ + cell_untwiddle_texture(screen, ps); + } } + return ps; +} -#if 0 - if (cell->tex_surf && cell->tex_map) { - pipe_surface_unmap(cell->tex_surf); - cell->tex_map = NULL; + +static void +cell_tex_surface_release(struct pipe_screen *screen, + struct pipe_surface **s) +{ + struct cell_texture *ct = cell_texture((*s)->texture); + const uint level = (*s)->level; + struct pipe_surface *surf = *s; + + if ((surf->usage & PIPE_BUFFER_USAGE_CPU_READ) && (ct->untiled_data[level])) + { + align_free(ct->untiled_data[level]); + ct->untiled_data[level] = NULL; } - /* XXX free old surface */ + if ((ct->base.tex_usage & PIPE_TEXTURE_USAGE_SAMPLER) && + (surf->usage & PIPE_BUFFER_USAGE_CPU_WRITE)) { + /* convert from linear to tiled layout */ + cell_twiddle_texture(screen, surf); + } - cell->tex_surf = cell_get_tex_surface(&cell->pipe, - &cell->texture[0]->base, - face, level, zslice); + /* XXX if done rendering to teximage, re-tile */ - cell->tex_map = pipe_surface_map(cell->tex_surf); -#endif + if (--surf->refcount == 0) { + pipe_texture_reference(&surf->texture, NULL); + FREE(surf); + } + *s = NULL; +} + + +static struct pipe_transfer * +cell_get_tex_transfer(struct pipe_screen *screen, + struct pipe_texture *texture, + unsigned face, unsigned level, unsigned zslice, + enum pipe_transfer_usage usage, + unsigned x, unsigned y, unsigned w, unsigned h) +{ + struct cell_texture *ct = cell_texture(texture); + struct cell_transfer *ctrans; + struct pipe_transfer *pt; + + assert(texture); + assert(level <= texture->last_level); + + ctrans = CALLOC_STRUCT(cell_transfer); + pt = &ctrans->base; + if (ctrans) { + pt->refcount = 1; + pipe_texture_reference(&pt->texture, texture); + pt->format = texture->format; + pt->block = texture->block; + pt->x = x; + pt->y = y; + pt->width = w; + pt->height = h; + pt->nblocksx = texture->nblocksx[level]; + pt->nblocksy = texture->nblocksy[level]; + pt->stride = ct->stride[level]; + ctrans->offset = ct->level_offset[level]; + pt->usage = usage; + pt->face = face; + pt->level = level; + pt->zslice = zslice; + + if (texture->target == PIPE_TEXTURE_CUBE || + texture->target == PIPE_TEXTURE_3D) { + ctrans->offset += ((texture->target == PIPE_TEXTURE_CUBE) ? face : + zslice) * pt->nblocksy * pt->stride; + } + else { + assert(face == 0); + assert(zslice == 0); + } + } + return pt; } static void -cell_tex_surface_release(struct pipe_screen *screen, - struct pipe_surface **s) +cell_tex_transfer_release(struct pipe_screen *screen, + struct pipe_transfer **t) { + struct cell_transfer *transfer = cell_transfer(*t); /* Effectively do the texture_update work here - if texture images * needed post-processing to put them into hardware layout, this is - * where it would happen. For softpipe, nothing to do. + * where it would happen. For cell, nothing to do. */ - assert ((*s)->texture); - pipe_texture_reference(&(*s)->texture, NULL); - - screen->winsys->surface_release(screen->winsys, s); + assert (transfer->base.texture); + if (--transfer->base.refcount == 0) { + pipe_texture_reference(&transfer->base.texture, NULL); + FREE(transfer); + } + *t = NULL; } static void * -cell_surface_map( struct pipe_screen *screen, - struct pipe_surface *surface, - unsigned flags ) +cell_transfer_map( struct pipe_screen *screen, + struct pipe_transfer *transfer ) { ubyte *map; + struct cell_texture *spt; + unsigned flags = 0; - if (flags & ~surface->usage) { - assert(0); - return NULL; + assert(transfer->texture); + spt = cell_texture(transfer->texture); + + if (transfer->usage != PIPE_TRANSFER_READ) { + flags |= PIPE_BUFFER_USAGE_CPU_WRITE; } - map = pipe_buffer_map( screen, surface->buffer, flags ); + if (transfer->usage != PIPE_TRANSFER_WRITE) { + flags |= PIPE_BUFFER_USAGE_CPU_READ; + } + + map = pipe_buffer_map(screen, spt->buffer, flags); if (map == NULL) return NULL; /* May want to different things here depending on read/write nature * of the map: */ - if (surface->texture && - (flags & PIPE_BUFFER_USAGE_CPU_WRITE)) + if (transfer->texture && transfer->usage != PIPE_TRANSFER_READ) { /* Do something to notify sharing contexts of a texture change. - * In softpipe, that would mean flushing the texture cache. + * In cell, that would mean flushing the texture cache. */ -#if 0 +#if 00 cell_screen(screen)->timestamp++; #endif } - return map + surface->offset; + return map + cell_transfer(transfer)->offset + + transfer->y / transfer->block.height * transfer->stride + + transfer->x / transfer->block.width * transfer->block.size; } static void -cell_surface_unmap(struct pipe_screen *screen, - struct pipe_surface *surface) +cell_transfer_unmap(struct pipe_screen *screen, + struct pipe_transfer *transfer) { - pipe_buffer_unmap( screen, surface->buffer ); + struct cell_texture *spt; + + assert(transfer->texture); + spt = cell_texture(transfer->texture); + + pipe_buffer_unmap( screen, spt->buffer ); } -void -cell_init_texture_functions(struct cell_context *cell) +static void * +cell_surface_map(struct pipe_screen *screen, + struct pipe_surface *surface, + unsigned flags) { - /*cell->pipe.texture_update = cell_texture_update;*/ + ubyte *map; + struct cell_texture *ct = cell_texture(surface->texture); + const uint level = surface->level; + + assert(ct); + +#if 0 + if (flags & ~surface->usage) { + assert(0); + return NULL; + } +#endif + + map = pipe_buffer_map( screen, ct->buffer, flags ); + if (map == NULL) { + return NULL; + } + else { + if ((surface->usage & PIPE_BUFFER_USAGE_CPU_READ) && + (ct->untiled_data[level])) { + return (void *) ((ubyte *) ct->untiled_data[level] + surface->offset); + } + else { + return (void *) (map + surface->offset); + } + } } +static void +cell_surface_unmap(struct pipe_screen *screen, + struct pipe_surface *surface) +{ + struct cell_texture *ct = cell_texture(surface->texture); + + assert(ct); + + pipe_buffer_unmap( screen, ct->buffer ); +} + + + void cell_init_screen_texture_funcs(struct pipe_screen *screen) { @@ -382,6 +653,8 @@ cell_init_screen_texture_funcs(struct pipe_screen *screen) screen->get_tex_surface = cell_get_tex_surface; screen->tex_surface_release = cell_tex_surface_release; - screen->surface_map = cell_surface_map; - screen->surface_unmap = cell_surface_unmap; + screen->get_tex_transfer = cell_get_tex_transfer; + screen->tex_transfer_release = cell_tex_transfer_release; + screen->transfer_map = cell_transfer_map; + screen->transfer_unmap = cell_transfer_unmap; } diff --git a/src/gallium/drivers/cell/ppu/cell_texture.h b/src/gallium/drivers/cell/ppu/cell_texture.h index 6d37e95ebc..fc6486adbe 100644 --- a/src/gallium/drivers/cell/ppu/cell_texture.h +++ b/src/gallium/drivers/cell/ppu/cell_texture.h @@ -40,15 +40,31 @@ struct cell_texture { struct pipe_texture base; - unsigned long level_offset[PIPE_MAX_TEXTURE_LEVELS]; - unsigned long stride[PIPE_MAX_TEXTURE_LEVELS]; + unsigned long level_offset[CELL_MAX_TEXTURE_LEVELS]; + unsigned long stride[CELL_MAX_TEXTURE_LEVELS]; /* The data is held here: */ struct pipe_buffer *buffer; unsigned long buffer_size; - void *tiled_data; /* XXX this may be temporary */ /*ALIGN16*/ + /** Texture data in tiled layout is held here */ + struct pipe_buffer *tiled_buffer[CELL_MAX_TEXTURE_LEVELS]; + /** Mapped, tiled texture data */ + void *tiled_mapped[CELL_MAX_TEXTURE_LEVELS]; + + struct pipe_transfer *transfer; + + /** The original, linear texture data */ + void *untiled_data[CELL_MAX_TEXTURE_LEVELS]; +}; + + +struct cell_transfer +{ + struct pipe_transfer base; + + unsigned long offset; }; @@ -60,13 +76,12 @@ cell_texture(struct pipe_texture *pt) } - -extern void -cell_update_texture_mapping(struct cell_context *cell); - - -extern void -cell_init_texture_functions(struct cell_context *cell); +/** cast wrapper */ +static INLINE struct cell_transfer * +cell_transfer(struct pipe_transfer *pt) +{ + return (struct cell_transfer *) pt; +} extern void diff --git a/src/gallium/drivers/cell/ppu/cell_vbuf.c b/src/gallium/drivers/cell/ppu/cell_vbuf.c index aa63435b93..cfaffb52a8 100644 --- a/src/gallium/drivers/cell/ppu/cell_vbuf.c +++ b/src/gallium/drivers/cell/ppu/cell_vbuf.c @@ -38,6 +38,7 @@ #include "cell_batch.h" #include "cell_context.h" +#include "cell_fence.h" #include "cell_flush.h" #include "cell_spu.h" #include "cell_vbuf.h" @@ -61,6 +62,7 @@ struct cell_vbuf_render uint vertex_size; /**< in bytes */ void *vertex_buffer; /**< just for debug, really */ uint vertex_buf; /**< in [0, CELL_NUM_BUFFERS-1] */ + uint vertex_buffer_size; /**< size in bytes */ }; @@ -81,24 +83,26 @@ cell_vbuf_get_vertex_info(struct vbuf_render *vbr) } -static void * +static boolean cell_vbuf_allocate_vertices(struct vbuf_render *vbr, ushort vertex_size, ushort nr_vertices) { struct cell_vbuf_render *cvbr = cell_vbuf_render(vbr); + unsigned size = vertex_size * nr_vertices; /*printf("Alloc verts %u * %u\n", vertex_size, nr_vertices);*/ assert(cvbr->vertex_buf == ~0); cvbr->vertex_buf = cell_get_empty_buffer(cvbr->cell); cvbr->vertex_buffer = cvbr->cell->buffer[cvbr->vertex_buf]; + cvbr->vertex_buffer_size = size; cvbr->vertex_size = vertex_size; - return cvbr->vertex_buffer; + + return cvbr->vertex_buffer != NULL; } static void -cell_vbuf_release_vertices(struct vbuf_render *vbr, void *vertices, - unsigned vertex_size, unsigned vertices_used) +cell_vbuf_release_vertices(struct vbuf_render *vbr) { struct cell_vbuf_render *cvbr = cell_vbuf_render(vbr); struct cell_context *cell = cvbr->cell; @@ -108,23 +112,47 @@ cell_vbuf_release_vertices(struct vbuf_render *vbr, void *vertices, __FUNCTION__, cvbr->vertex_buf, vertices_used); */ + /* Make sure texture buffers aren't released until we're done rendering + * with them. + */ + cell_add_fenced_textures(cell); + /* Tell SPUs they can release the vert buf */ if (cvbr->vertex_buf != ~0U) { + STATIC_ASSERT(sizeof(struct cell_command_release_verts) % 16 == 0); struct cell_command_release_verts *release = (struct cell_command_release_verts *) - cell_batch_alloc(cell, sizeof(struct cell_command_release_verts)); - release->opcode = CELL_CMD_RELEASE_VERTS; + cell_batch_alloc16(cell, sizeof(struct cell_command_release_verts)); + release->opcode[0] = CELL_CMD_RELEASE_VERTS; release->vertex_buf = cvbr->vertex_buf; } cvbr->vertex_buf = ~0; cell_flush_int(cell, 0x0); - assert(vertices == cvbr->vertex_buffer); cvbr->vertex_buffer = NULL; } +static void * +cell_vbuf_map_vertices(struct vbuf_render *vbr) +{ + struct cell_vbuf_render *cvbr = cell_vbuf_render(vbr); + return cvbr->vertex_buffer; +} + + +static void +cell_vbuf_unmap_vertices(struct vbuf_render *vbr, + ushort min_index, + ushort max_index ) +{ + struct cell_vbuf_render *cvbr = cell_vbuf_render(vbr); + assert( cvbr->vertex_buffer_size >= (max_index+1) * cvbr->vertex_size ); + /* do nothing */ +} + + static boolean cell_vbuf_set_primitive(struct vbuf_render *vbr, unsigned prim) @@ -204,15 +232,16 @@ cell_vbuf_draw(struct vbuf_render *vbr, /* build/insert batch RENDER command */ { - const uint index_bytes = ROUNDUP8(nr_indices * 2); - const uint vertex_bytes = nr_vertices * 4 * cell->vertex_info.size; + const uint index_bytes = ROUNDUP16(nr_indices * 2); + const uint vertex_bytes = ROUNDUP16(nr_vertices * 4 * cell->vertex_info.size); + STATIC_ASSERT(sizeof(struct cell_command_render) % 16 == 0); const uint batch_size = sizeof(struct cell_command_render) + index_bytes; struct cell_command_render *render = (struct cell_command_render *) - cell_batch_alloc(cell, batch_size); + cell_batch_alloc16(cell, batch_size); - render->opcode = CELL_CMD_RENDER; + render->opcode[0] = CELL_CMD_RENDER; render->prim_type = cvbr->prim; render->num_indexes = nr_indices; @@ -230,7 +259,7 @@ cell_vbuf_draw(struct vbuf_render *vbr, min_index == 0 && vertex_bytes + 16 <= cell_batch_free_space(cell)) { /* vertex data inlined, after indices, at 16-byte boundary */ - void *dst = cell_batch_alloc_aligned(cell, vertex_bytes, 16); + void *dst = cell_batch_alloc16(cell, vertex_bytes); memcpy(dst, vertices, vertex_bytes); render->inline_verts = TRUE; render->vertex_buf = ~0; @@ -287,6 +316,8 @@ cell_init_vbuf(struct cell_context *cell) cell->vbuf_render->base.get_vertex_info = cell_vbuf_get_vertex_info; cell->vbuf_render->base.allocate_vertices = cell_vbuf_allocate_vertices; + cell->vbuf_render->base.map_vertices = cell_vbuf_map_vertices; + cell->vbuf_render->base.unmap_vertices = cell_vbuf_unmap_vertices; cell->vbuf_render->base.set_primitive = cell_vbuf_set_primitive; cell->vbuf_render->base.draw = cell_vbuf_draw; cell->vbuf_render->base.release_vertices = cell_vbuf_release_vertices; diff --git a/src/gallium/drivers/cell/ppu/cell_vertex_fetch.c b/src/gallium/drivers/cell/ppu/cell_vertex_fetch.c index 566df7f59e..9cba537d9e 100644 --- a/src/gallium/drivers/cell/ppu/cell_vertex_fetch.c +++ b/src/gallium/drivers/cell/ppu/cell_vertex_fetch.c @@ -73,8 +73,8 @@ emit_matrix_transpose(struct spe_function *p, int col3; - spe_lqd(p, shuf_hi, shuf_ptr, 3); - spe_lqd(p, shuf_lo, shuf_ptr, 4); + spe_lqd(p, shuf_hi, shuf_ptr, 3*16); + spe_lqd(p, shuf_lo, shuf_ptr, 4*16); spe_shufb(p, t1, row0, row2, shuf_hi); spe_shufb(p, t2, row0, row2, shuf_lo); @@ -122,13 +122,13 @@ emit_matrix_transpose(struct spe_function *p, */ switch (count) { case 4: - spe_stqd(p, col3, dest_ptr, 3); + spe_stqd(p, col3, dest_ptr, 3 * 16); case 3: - spe_stqd(p, col2, dest_ptr, 2); + spe_stqd(p, col2, dest_ptr, 2 * 16); case 2: - spe_stqd(p, col1, dest_ptr, 1); + spe_stqd(p, col1, dest_ptr, 1 * 16); case 1: - spe_stqd(p, col0, dest_ptr, 0); + spe_stqd(p, col0, dest_ptr, 0 * 16); } @@ -145,6 +145,8 @@ emit_matrix_transpose(struct spe_function *p, } +#if 0 +/* This appears to not be used currently */ static void emit_fetch(struct spe_function *p, unsigned in_ptr, unsigned *offset, @@ -166,17 +168,17 @@ emit_fetch(struct spe_function *p, float scale_signed = 0.0; float scale_unsigned = 0.0; - spe_lqd(p, v0, in_ptr, 0 + offset[0]); - spe_lqd(p, v1, in_ptr, 1 + offset[0]); - spe_lqd(p, v2, in_ptr, 2 + offset[0]); - spe_lqd(p, v3, in_ptr, 3 + offset[0]); + spe_lqd(p, v0, in_ptr, (0 + offset[0]) * 16); + spe_lqd(p, v1, in_ptr, (1 + offset[0]) * 16); + spe_lqd(p, v2, in_ptr, (2 + offset[0]) * 16); + spe_lqd(p, v3, in_ptr, (3 + offset[0]) * 16); offset[0] += 4; switch (bytes) { case 1: scale_signed = 1.0f / 127.0f; scale_unsigned = 1.0f / 255.0f; - spe_lqd(p, tmp, shuf_ptr, 1); + spe_lqd(p, tmp, shuf_ptr, 1 * 16); spe_shufb(p, v0, v0, v0, tmp); spe_shufb(p, v1, v1, v1, tmp); spe_shufb(p, v2, v2, v2, tmp); @@ -185,7 +187,7 @@ emit_fetch(struct spe_function *p, case 2: scale_signed = 1.0f / 32767.0f; scale_unsigned = 1.0f / 65535.0f; - spe_lqd(p, tmp, shuf_ptr, 2); + spe_lqd(p, tmp, shuf_ptr, 2 * 16); spe_shufb(p, v0, v0, v0, tmp); spe_shufb(p, v1, v1, v1, tmp); spe_shufb(p, v2, v2, v2, tmp); @@ -241,11 +243,11 @@ emit_fetch(struct spe_function *p, switch (count) { case 1: - spe_stqd(p, float_zero, out_ptr, 1); + spe_stqd(p, float_zero, out_ptr, 1 * 16); case 2: - spe_stqd(p, float_zero, out_ptr, 2); + spe_stqd(p, float_zero, out_ptr, 2 * 16); case 3: - spe_stqd(p, float_one, out_ptr, 3); + spe_stqd(p, float_one, out_ptr, 3 * 16); } if (float_zero != -1) { @@ -256,6 +258,7 @@ emit_fetch(struct spe_function *p, spe_release_register(p, float_one); } } +#endif void cell_update_vertex_fetch(struct draw_context *draw) diff --git a/src/gallium/drivers/cell/ppu/cell_vertex_shader.c b/src/gallium/drivers/cell/ppu/cell_vertex_shader.c index 2b10c116fa..403cf6d50f 100644 --- a/src/gallium/drivers/cell/ppu/cell_vertex_shader.c +++ b/src/gallium/drivers/cell/ppu/cell_vertex_shader.c @@ -31,7 +31,7 @@ #include "pipe/p_defines.h" #include "pipe/p_context.h" -#include "pipe/p_winsys.h" +#include "pipe/internal/p_winsys_screen.h" #include "util/u_math.h" #include "cell_context.h" diff --git a/src/gallium/drivers/cell/spu/.gitignore b/src/gallium/drivers/cell/spu/.gitignore new file mode 100644 index 0000000000..2be9a2d324 --- /dev/null +++ b/src/gallium/drivers/cell/spu/.gitignore @@ -0,0 +1 @@ +g3d_spu diff --git a/src/gallium/drivers/cell/spu/Makefile b/src/gallium/drivers/cell/spu/Makefile index 1ae0dfb8c1..3cc52301da 100644 --- a/src/gallium/drivers/cell/spu/Makefile +++ b/src/gallium/drivers/cell/spu/Makefile @@ -16,8 +16,10 @@ PROG_SPU_EMBED_O = $(PROG)_spu-embed.o SOURCES = \ - spu_main.c \ + spu_command.c \ spu_dcache.c \ + spu_funcs.c \ + spu_main.c \ spu_per_fragment_op.c \ spu_render.c \ spu_texture.c \ @@ -31,9 +33,10 @@ OLD_SOURCES = \ spu_vertex_shader.c -SPU_OBJECTS = $(SOURCES:.c=.o) \ +SPU_OBJECTS = $(SOURCES:.c=.o) + +SPU_ASM_OUT = $(SOURCES:.c=.s) -SPU_ASM_OUT = $(SOURCES:.c=.s) \ INCLUDE_DIRS = \ -I$(TOP)/src/mesa \ diff --git a/src/gallium/drivers/cell/spu/spu_colorpack.h b/src/gallium/drivers/cell/spu/spu_colorpack.h index fd8dc6ded3..d7ce005524 100644 --- a/src/gallium/drivers/cell/spu/spu_colorpack.h +++ b/src/gallium/drivers/cell/spu/spu_colorpack.h @@ -31,6 +31,7 @@ #define SPU_COLORPACK_H +#include <transpose_matrix4x4.h> #include <spu_intrinsics.h> @@ -84,10 +85,10 @@ spu_unpack_B8G8R8A8(uint color) vector unsigned int color_u4 = spu_splats(color); color_u4 = spu_shuffle(color_u4, color_u4, ((vector unsigned char) { - 10, 10, 10, 10, - 5, 5, 5, 5, + 2, 2, 2, 2, + 1, 1, 1, 1, 0, 0, 0, 0, - 15, 15, 15, 15}) ); + 3, 3, 3, 3}) ); return spu_convtf(color_u4, 32); } @@ -98,13 +99,47 @@ spu_unpack_A8R8G8B8(uint color) vector unsigned int color_u4 = spu_splats(color); color_u4 = spu_shuffle(color_u4, color_u4, ((vector unsigned char) { - 5, 5, 5, 5, - 10, 10, 10, 10, - 15, 15, 15, 15, + 1, 1, 1, 1, + 2, 2, 2, 2, + 3, 3, 3, 3, 0, 0, 0, 0}) ); - return spu_convtf(color_u4, 32); } +/** + * \param color_in - array of 32-bit packed ARGB colors + * \param color_out - returns float colors in RRRR, GGGG, BBBB, AAAA order + */ +static INLINE void +spu_unpack_A8R8G8B8_transpose4(const vector unsigned int color_in[4], + vector float color_out[4]) +{ + vector unsigned int c0; + + c0 = spu_shuffle(color_in[0], color_in[0], + ((vector unsigned char) { + 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 0, 0}) ); + color_out[0] = spu_convtf(c0, 32); + + c0 = spu_shuffle(color_in[1], color_in[1], + ((vector unsigned char) { + 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 0, 0}) ); + color_out[1] = spu_convtf(c0, 32); + + c0 = spu_shuffle(color_in[2], color_in[2], + ((vector unsigned char) { + 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 0, 0}) ); + color_out[2] = spu_convtf(c0, 32); + + c0 = spu_shuffle(color_in[3], color_in[3], + ((vector unsigned char) { + 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 0, 0}) ); + color_out[3] = spu_convtf(c0, 32); + + _transpose_matrix4x4(color_out, color_out); +} + + + #endif /* SPU_COLORPACK_H */ diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c new file mode 100644 index 0000000000..5c0179d954 --- /dev/null +++ b/src/gallium/drivers/cell/spu/spu_command.c @@ -0,0 +1,815 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * SPU command processing code + */ + + +#include <stdio.h> +#include <libmisc.h> + +#include "pipe/p_defines.h" + +#include "spu_command.h" +#include "spu_main.h" +#include "spu_render.h" +#include "spu_per_fragment_op.h" +#include "spu_texture.h" +#include "spu_tile.h" +#include "spu_vertex_shader.h" +#include "spu_dcache.h" +#include "cell/common.h" + + +struct spu_vs_context draw; + + +/** + * Buffers containing dynamically generated SPU code: + */ +static unsigned char attribute_fetch_code_buffer[136 * PIPE_MAX_ATTRIBS] + ALIGN16_ATTRIB; + + + +static INLINE int +align(int value, int alignment) +{ + return (value + alignment - 1) & ~(alignment - 1); +} + + + +/** + * Tell the PPU that this SPU has finished copying a buffer to + * local store and that it may be reused by the PPU. + * This is done by writting a 16-byte batch-buffer-status block back into + * main memory (in cell_context->buffer_status[]). + */ +static void +release_buffer(uint buffer) +{ + /* Evidently, using less than a 16-byte status doesn't work reliably */ + static const vector unsigned int status = {CELL_BUFFER_STATUS_FREE, + CELL_BUFFER_STATUS_FREE, + CELL_BUFFER_STATUS_FREE, + CELL_BUFFER_STATUS_FREE}; + const uint index = 4 * (spu.init.id * CELL_NUM_BUFFERS + buffer); + uint *dst = spu.init.buffer_status + index; + + ASSERT(buffer < CELL_NUM_BUFFERS); + + mfc_put((void *) &status, /* src in local memory */ + (unsigned int) dst, /* dst in main memory */ + sizeof(status), /* size */ + TAG_MISC, /* tag is unimportant */ + 0, /* tid */ + 0 /* rid */); +} + + +/** + * Write CELL_FENCE_SIGNALLED back to the fence status qword in main memory. + * There's a qword of status per SPU. + */ +static void +cmd_fence(struct cell_command_fence *fence_cmd) +{ + static const vector unsigned int status = {CELL_FENCE_SIGNALLED, + CELL_FENCE_SIGNALLED, + CELL_FENCE_SIGNALLED, + CELL_FENCE_SIGNALLED}; + uint *dst = (uint *) fence_cmd->fence; + dst += 4 * spu.init.id; /* main store/memory address, not local store */ + ASSERT_ALIGN16(dst); + mfc_put((void *) &status, /* src in local memory */ + (unsigned int) dst, /* dst in main memory */ + sizeof(status), /* size */ + TAG_FENCE, /* tag */ + 0, /* tid */ + 0 /* rid */); +} + + +static void +cmd_clear_surface(const struct cell_command_clear_surface *clear) +{ + D_PRINTF(CELL_DEBUG_CMD, "CLEAR SURF %u to 0x%08x\n", clear->surface, clear->value); + + if (clear->surface == 0) { + spu.fb.color_clear_value = clear->value; + if (spu.init.debug_flags & CELL_DEBUG_CHECKER) { + uint x = (spu.init.id << 4) | (spu.init.id << 12) | + (spu.init.id << 20) | (spu.init.id << 28); + spu.fb.color_clear_value ^= x; + } + } + else { + spu.fb.depth_clear_value = clear->value; + } + +#define CLEAR_OPT 1 +#if CLEAR_OPT + + /* Simply set all tiles' status to CLEAR. + * When we actually begin rendering into a tile, we'll initialize it to + * the clear value. If any tiles go untouched during the frame, + * really_clear_tiles() will set them to the clear value. + */ + if (clear->surface == 0) { + memset(spu.ctile_status, TILE_STATUS_CLEAR, sizeof(spu.ctile_status)); + } + else { + memset(spu.ztile_status, TILE_STATUS_CLEAR, sizeof(spu.ztile_status)); + } + +#else + + /* + * This path clears the whole framebuffer to the clear color right now. + */ + + /* + printf("SPU: %s num=%d w=%d h=%d\n", + __FUNCTION__, num_tiles, spu.fb.width_tiles, spu.fb.height_tiles); + */ + + /* init a single tile to the clear value */ + if (clear->surface == 0) { + clear_c_tile(&spu.ctile); + } + else { + clear_z_tile(&spu.ztile); + } + + /* walk over my tiles, writing the 'clear' tile's data */ + { + const uint num_tiles = spu.fb.width_tiles * spu.fb.height_tiles; + uint i; + for (i = spu.init.id; i < num_tiles; i += spu.init.num_spus) { + uint tx = i % spu.fb.width_tiles; + uint ty = i / spu.fb.width_tiles; + if (clear->surface == 0) + put_tile(tx, ty, &spu.ctile, TAG_SURFACE_CLEAR, 0); + else + put_tile(tx, ty, &spu.ztile, TAG_SURFACE_CLEAR, 1); + } + } + + if (spu.init.debug_flags & CELL_DEBUG_SYNC) { + wait_on_mask(1 << TAG_SURFACE_CLEAR); + } + +#endif /* CLEAR_OPT */ + + D_PRINTF(CELL_DEBUG_CMD, "CLEAR SURF done\n"); +} + + +static void +cmd_release_verts(const struct cell_command_release_verts *release) +{ + D_PRINTF(CELL_DEBUG_CMD, "RELEASE VERTS %u\n", release->vertex_buf); + ASSERT(release->vertex_buf != ~0U); + release_buffer(release->vertex_buf); +} + + +/** + * Process a CELL_CMD_STATE_FRAGMENT_OPS command. + * This involves installing new fragment ops SPU code. + * If this function is never called, we'll use a regular C fallback function + * for fragment processing. + */ +static void +cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops) +{ + D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_OPS\n"); + + /* Copy state info (for fallback case only - this will eventually + * go away when the fallback case goes away) + */ + memcpy(&spu.depth_stencil_alpha, &fops->dsa, sizeof(fops->dsa)); + memcpy(&spu.blend, &fops->blend, sizeof(fops->blend)); + memcpy(&spu.blend_color, &fops->blend_color, sizeof(fops->blend_color)); + + /* Make sure the SPU knows which buffers it's expected to read when + * it's told to pull tiles. + */ + spu.read_depth_stencil = (spu.depth_stencil_alpha.depth.enabled || spu.depth_stencil_alpha.stencil[0].enabled); + + /* If we're forcing the fallback code to be used (for debug purposes), + * install that. Otherwise install the incoming SPU code. + */ + if ((spu.init.debug_flags & CELL_DEBUG_FRAGMENT_OP_FALLBACK) != 0) { + static unsigned int warned = 0; + if (!warned) { + fprintf(stderr, "Cell Warning: using fallback per-fragment code\n"); + warned = 1; + } + /* The following two lines aren't really necessary if you + * know the debug flags won't change during a run, and if you + * know that the function pointers are initialized correctly. + * We set them here to allow a person to change the debug + * flags during a run (from inside a debugger). + */ + spu.fragment_ops[CELL_FACING_FRONT] = spu_fallback_fragment_ops; + spu.fragment_ops[CELL_FACING_BACK] = spu_fallback_fragment_ops; + return; + } + + /* Make sure the SPU code buffer is large enough to hold the incoming code. + * Note that we *don't* use align_malloc() and align_free(), because + * those utility functions are *not* available in SPU code. + * */ + if (spu.fragment_ops_code_size < fops->total_code_size) { + if (spu.fragment_ops_code != NULL) { + free(spu.fragment_ops_code); + } + spu.fragment_ops_code_size = fops->total_code_size; + spu.fragment_ops_code = malloc(fops->total_code_size); + if (spu.fragment_ops_code == NULL) { + /* Whoops. */ + fprintf(stderr, "CELL Warning: failed to allocate fragment ops code (%d bytes) - using fallback\n", fops->total_code_size); + spu.fragment_ops_code = NULL; + spu.fragment_ops_code_size = 0; + spu.fragment_ops[CELL_FACING_FRONT] = spu_fallback_fragment_ops; + spu.fragment_ops[CELL_FACING_BACK] = spu_fallback_fragment_ops; + return; + } + } + + /* Copy the SPU code from the command buffer to the spu buffer */ + memcpy(spu.fragment_ops_code, fops->code, fops->total_code_size); + + /* Set the pointers for the front-facing and back-facing fragments + * to the specified offsets within the code. Note that if the + * front-facing and back-facing code are the same, they'll have + * the same offset. + */ + spu.fragment_ops[CELL_FACING_FRONT] = (spu_fragment_ops_func) &spu.fragment_ops_code[fops->front_code_index]; + spu.fragment_ops[CELL_FACING_BACK] = (spu_fragment_ops_func) &spu.fragment_ops_code[fops->back_code_index]; +} + +static void +cmd_state_fragment_program(const struct cell_command_fragment_program *fp) +{ + D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_PROGRAM\n"); + /* Copy SPU code from batch buffer to spu buffer */ + memcpy(spu.fragment_program_code, fp->code, + SPU_MAX_FRAGMENT_PROGRAM_INSTS * 4); +#if 01 + /* Point function pointer at new code */ + spu.fragment_program = (spu_fragment_program_func)spu.fragment_program_code; +#endif +} + + +static uint +cmd_state_fs_constants(const qword *buffer, uint pos) +{ + const uint num_const = spu_extract((vector unsigned int)buffer[pos+1], 0); + const float *constants = (const float *) &buffer[pos+2]; + uint i; + + D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FS_CONSTANTS (%u)\n", num_const); + + /* Expand each float to float[4] for SOA execution */ + for (i = 0; i < num_const; i++) { + D_PRINTF(CELL_DEBUG_CMD, " const[%u] = %f\n", i, constants[i]); + spu.constants[i] = spu_splats(constants[i]); + } + + /* return new buffer pos (in 16-byte words) */ + return pos + 2 + (ROUNDUP16(num_const * sizeof(float)) / 16); +} + + +static void +cmd_state_framebuffer(const struct cell_command_framebuffer *cmd) +{ + D_PRINTF(CELL_DEBUG_CMD, "FRAMEBUFFER: %d x %d at %p, cformat 0x%x zformat 0x%x\n", + cmd->width, + cmd->height, + cmd->color_start, + cmd->color_format, + cmd->depth_format); + + ASSERT_ALIGN16(cmd->color_start); + ASSERT_ALIGN16(cmd->depth_start); + + spu.fb.color_start = cmd->color_start; + spu.fb.depth_start = cmd->depth_start; + spu.fb.color_format = cmd->color_format; + spu.fb.depth_format = cmd->depth_format; + spu.fb.width = cmd->width; + spu.fb.height = cmd->height; + spu.fb.width_tiles = (spu.fb.width + TILE_SIZE - 1) / TILE_SIZE; + spu.fb.height_tiles = (spu.fb.height + TILE_SIZE - 1) / TILE_SIZE; + + switch (spu.fb.depth_format) { + case PIPE_FORMAT_Z32_UNORM: + spu.fb.zsize = 4; + spu.fb.zscale = (float) 0xffffffffu; + break; + case PIPE_FORMAT_Z24S8_UNORM: + case PIPE_FORMAT_S8Z24_UNORM: + case PIPE_FORMAT_Z24X8_UNORM: + case PIPE_FORMAT_X8Z24_UNORM: + spu.fb.zsize = 4; + spu.fb.zscale = (float) 0x00ffffffu; + break; + case PIPE_FORMAT_Z16_UNORM: + spu.fb.zsize = 2; + spu.fb.zscale = (float) 0xffffu; + break; + default: + spu.fb.zsize = 0; + break; + } +} + + +/** + * Tex texture mask_s/t and scale_s/t fields depend on the texture size and + * sampler wrap modes. + */ +static void +update_tex_masks(struct spu_texture *texture, + const struct pipe_sampler_state *sampler) +{ + uint i; + + for (i = 0; i < CELL_MAX_TEXTURE_LEVELS; i++) { + int width = texture->level[i].width; + int height = texture->level[i].height; + + if (sampler->wrap_s == PIPE_TEX_WRAP_REPEAT) + texture->level[i].mask_s = spu_splats(width - 1); + else + texture->level[i].mask_s = spu_splats(~0); + + if (sampler->wrap_t == PIPE_TEX_WRAP_REPEAT) + texture->level[i].mask_t = spu_splats(height - 1); + else + texture->level[i].mask_t = spu_splats(~0); + + if (sampler->normalized_coords) { + texture->level[i].scale_s = spu_splats((float) width); + texture->level[i].scale_t = spu_splats((float) height); + } + else { + texture->level[i].scale_s = spu_splats(1.0f); + texture->level[i].scale_t = spu_splats(1.0f); + } + } +} + + +static void +cmd_state_sampler(const struct cell_command_sampler *sampler) +{ + uint unit = sampler->unit; + + D_PRINTF(CELL_DEBUG_CMD, "SAMPLER [%u]\n", unit); + + spu.sampler[unit] = sampler->state; + + switch (spu.sampler[unit].min_img_filter) { + case PIPE_TEX_FILTER_LINEAR: + spu.min_sample_texture_2d[unit] = sample_texture_2d_bilinear; + break; + case PIPE_TEX_FILTER_ANISO: + /* fall-through, for now */ + case PIPE_TEX_FILTER_NEAREST: + spu.min_sample_texture_2d[unit] = sample_texture_2d_nearest; + break; + default: + ASSERT(0); + } + + switch (spu.sampler[sampler->unit].mag_img_filter) { + case PIPE_TEX_FILTER_LINEAR: + spu.mag_sample_texture_2d[unit] = sample_texture_2d_bilinear; + break; + case PIPE_TEX_FILTER_ANISO: + /* fall-through, for now */ + case PIPE_TEX_FILTER_NEAREST: + spu.mag_sample_texture_2d[unit] = sample_texture_2d_nearest; + break; + default: + ASSERT(0); + } + + switch (spu.sampler[sampler->unit].min_mip_filter) { + case PIPE_TEX_MIPFILTER_NEAREST: + case PIPE_TEX_MIPFILTER_LINEAR: + spu.sample_texture_2d[unit] = sample_texture_2d_lod; + break; + case PIPE_TEX_MIPFILTER_NONE: + spu.sample_texture_2d[unit] = spu.mag_sample_texture_2d[unit]; + break; + default: + ASSERT(0); + } + + update_tex_masks(&spu.texture[unit], &spu.sampler[unit]); +} + + +static void +cmd_state_texture(const struct cell_command_texture *texture) +{ + const uint unit = texture->unit; + uint i; + + D_PRINTF(CELL_DEBUG_CMD, "TEXTURE [%u]\n", texture->unit); + + spu.texture[unit].max_level = 0; + spu.texture[unit].target = texture->target; + + for (i = 0; i < CELL_MAX_TEXTURE_LEVELS; i++) { + uint width = texture->width[i]; + uint height = texture->height[i]; + uint depth = texture->depth[i]; + + D_PRINTF(CELL_DEBUG_CMD, " LEVEL %u: at %p size[0] %u x %u\n", i, + texture->start[i], texture->width[i], texture->height[i]); + + spu.texture[unit].level[i].start = texture->start[i]; + spu.texture[unit].level[i].width = width; + spu.texture[unit].level[i].height = height; + spu.texture[unit].level[i].depth = depth; + + spu.texture[unit].level[i].tiles_per_row = + (width + TILE_SIZE - 1) / TILE_SIZE; + + spu.texture[unit].level[i].bytes_per_image = + 4 * align(width, TILE_SIZE) * align(height, TILE_SIZE) * depth; + + spu.texture[unit].level[i].max_s = spu_splats((int) width - 1); + spu.texture[unit].level[i].max_t = spu_splats((int) height - 1); + + if (texture->start[i]) + spu.texture[unit].max_level = i; + } + + update_tex_masks(&spu.texture[unit], &spu.sampler[unit]); +} + + +static void +cmd_state_vertex_info(const struct vertex_info *vinfo) +{ + D_PRINTF(CELL_DEBUG_CMD, "VERTEX_INFO num_attribs=%u\n", vinfo->num_attribs); + ASSERT(vinfo->num_attribs >= 1); + ASSERT(vinfo->num_attribs <= 8); + memcpy(&spu.vertex_info, vinfo, sizeof(*vinfo)); +} + + +static void +cmd_state_vs_array_info(const struct cell_array_info *vs_info) +{ + const unsigned attr = vs_info->attr; + + ASSERT(attr < PIPE_MAX_ATTRIBS); + draw.vertex_fetch.src_ptr[attr] = vs_info->base; + draw.vertex_fetch.pitch[attr] = vs_info->pitch; + draw.vertex_fetch.size[attr] = vs_info->size; + draw.vertex_fetch.code_offset[attr] = vs_info->function_offset; + draw.vertex_fetch.dirty = 1; +} + + +static void +cmd_state_attrib_fetch(const struct cell_attribute_fetch_code *code) +{ + mfc_get(attribute_fetch_code_buffer, + (unsigned int) code->base, /* src */ + code->size, + TAG_BATCH_BUFFER, + 0, /* tid */ + 0 /* rid */); + wait_on_mask(1 << TAG_BATCH_BUFFER); + + draw.vertex_fetch.code = attribute_fetch_code_buffer; +} + + +static void +cmd_finish(void) +{ + D_PRINTF(CELL_DEBUG_CMD, "FINISH\n"); + really_clear_tiles(0); + /* wait for all outstanding DMAs to finish */ + mfc_write_tag_mask(~0); + mfc_read_tag_status_all(); + /* send mbox message to PPU */ + spu_write_out_mbox(CELL_CMD_FINISH); +} + + +/** + * Execute a batch of commands which was sent to us by the PPU. + * See the cell_emit_state.c code to see where the commands come from. + * + * The opcode param encodes the location of the buffer and its size. + */ +static void +cmd_batch(uint opcode) +{ + const uint buf = (opcode >> 8) & 0xff; + uint size = (opcode >> 16); + qword buffer[CELL_BUFFER_SIZE / 16] ALIGN16_ATTRIB; + const unsigned usize = ROUNDUP16(size) / sizeof(buffer[0]); + uint pos; + + D_PRINTF(CELL_DEBUG_CMD, "BATCH buffer %u, len %u, from %p\n", + buf, size, spu.init.buffers[buf]); + + ASSERT((opcode & CELL_CMD_OPCODE_MASK) == CELL_CMD_BATCH); + + ASSERT_ALIGN16(spu.init.buffers[buf]); + + size = ROUNDUP16(size); + + ASSERT_ALIGN16(spu.init.buffers[buf]); + + mfc_get(buffer, /* dest */ + (unsigned int) spu.init.buffers[buf], /* src */ + size, + TAG_BATCH_BUFFER, + 0, /* tid */ + 0 /* rid */); + wait_on_mask(1 << TAG_BATCH_BUFFER); + + /* Tell PPU we're done copying the buffer to local store */ + D_PRINTF(CELL_DEBUG_CMD, "release batch buf %u\n", buf); + release_buffer(buf); + + /* + * Loop over commands in the batch buffer + */ + for (pos = 0; pos < usize; /* no incr */) { + switch (si_to_uint(buffer[pos])) { + /* + * rendering commands + */ + case CELL_CMD_CLEAR_SURFACE: + { + struct cell_command_clear_surface *clr + = (struct cell_command_clear_surface *) &buffer[pos]; + cmd_clear_surface(clr); + pos += sizeof(*clr) / 16; + } + break; + case CELL_CMD_RENDER: + { + struct cell_command_render *render + = (struct cell_command_render *) &buffer[pos]; + uint pos_incr; + cmd_render(render, &pos_incr); + pos += ((pos_incr+1)&~1) / 2; // should 'fix' cmd_render return + } + break; + /* + * state-update commands + */ + case CELL_CMD_STATE_FRAMEBUFFER: + { + struct cell_command_framebuffer *fb + = (struct cell_command_framebuffer *) &buffer[pos]; + cmd_state_framebuffer(fb); + pos += sizeof(*fb) / 16; + } + break; + case CELL_CMD_STATE_FRAGMENT_OPS: + { + struct cell_command_fragment_ops *fops + = (struct cell_command_fragment_ops *) &buffer[pos]; + cmd_state_fragment_ops(fops); + /* This is a variant-sized command */ + pos += ROUNDUP16(sizeof(*fops) + fops->total_code_size) / 16; + } + break; + case CELL_CMD_STATE_FRAGMENT_PROGRAM: + { + struct cell_command_fragment_program *fp + = (struct cell_command_fragment_program *) &buffer[pos]; + cmd_state_fragment_program(fp); + pos += sizeof(*fp) / 16; + } + break; + case CELL_CMD_STATE_FS_CONSTANTS: + pos = cmd_state_fs_constants(buffer, pos); + break; + case CELL_CMD_STATE_RASTERIZER: + { + struct cell_command_rasterizer *rast = + (struct cell_command_rasterizer *) &buffer[pos]; + spu.rasterizer = rast->rasterizer; + pos += sizeof(*rast) / 16; + } + break; + case CELL_CMD_STATE_SAMPLER: + { + struct cell_command_sampler *sampler + = (struct cell_command_sampler *) &buffer[pos]; + cmd_state_sampler(sampler); + pos += sizeof(*sampler) / 16; + } + break; + case CELL_CMD_STATE_TEXTURE: + { + struct cell_command_texture *texture + = (struct cell_command_texture *) &buffer[pos]; + cmd_state_texture(texture); + pos += sizeof(*texture) / 16; + } + break; + case CELL_CMD_STATE_VERTEX_INFO: + cmd_state_vertex_info((struct vertex_info *) &buffer[pos+1]); + pos += 1 + ROUNDUP16(sizeof(struct vertex_info)) / 16; + break; + case CELL_CMD_STATE_VIEWPORT: + (void) memcpy(& draw.viewport, &buffer[pos+1], + sizeof(struct pipe_viewport_state)); + pos += 1 + ROUNDUP16(sizeof(struct pipe_viewport_state)) / 16; + break; + case CELL_CMD_STATE_UNIFORMS: + draw.constants = (const float (*)[4]) (uintptr_t)spu_extract((vector unsigned int)buffer[pos+1],0); + pos += 2; + break; + case CELL_CMD_STATE_VS_ARRAY_INFO: + cmd_state_vs_array_info((struct cell_array_info *) &buffer[pos+1]); + pos += 1 + ROUNDUP16(sizeof(struct cell_array_info)) / 16; + break; + case CELL_CMD_STATE_BIND_VS: +#if 0 + spu_bind_vertex_shader(&draw, + (struct cell_shader_info *) &buffer[pos+1]); +#endif + pos += 1 + ROUNDUP16(sizeof(struct cell_shader_info)) / 16; + break; + case CELL_CMD_STATE_ATTRIB_FETCH: + cmd_state_attrib_fetch((struct cell_attribute_fetch_code *) + &buffer[pos+1]); + pos += 1 + ROUNDUP16(sizeof(struct cell_attribute_fetch_code)) / 16; + break; + /* + * misc commands + */ + case CELL_CMD_FINISH: + cmd_finish(); + pos += 1; + break; + case CELL_CMD_FENCE: + { + struct cell_command_fence *fence_cmd = + (struct cell_command_fence *) &buffer[pos]; + cmd_fence(fence_cmd); + pos += sizeof(*fence_cmd) / 16; + } + break; + case CELL_CMD_RELEASE_VERTS: + { + struct cell_command_release_verts *release + = (struct cell_command_release_verts *) &buffer[pos]; + cmd_release_verts(release); + pos += sizeof(*release) / 16; + } + break; + case CELL_CMD_FLUSH_BUFFER_RANGE: { + struct cell_buffer_range *br = (struct cell_buffer_range *) + &buffer[pos+1]; + + spu_dcache_mark_dirty((unsigned) br->base, br->size); + pos += 1 + ROUNDUP16(sizeof(struct cell_buffer_range)) / 16; + break; + } + default: + printf("SPU %u: bad opcode: 0x%x\n", spu.init.id, si_to_uint(buffer[pos])); + ASSERT(0); + break; + } + } + + D_PRINTF(CELL_DEBUG_CMD, "BATCH complete\n"); +} + + +#define PERF 0 + + +/** + * Main loop for SPEs: Get a command, execute it, repeat. + */ +void +command_loop(void) +{ + int exitFlag = 0; + uint t0, t1; + + D_PRINTF(CELL_DEBUG_CMD, "Enter command loop\n"); + + while (!exitFlag) { + unsigned opcode; + + D_PRINTF(CELL_DEBUG_CMD, "Wait for cmd...\n"); + + if (PERF) + spu_write_decrementer(~0); + + /* read/wait from mailbox */ + opcode = (unsigned int) spu_read_in_mbox(); + D_PRINTF(CELL_DEBUG_CMD, "got cmd 0x%x\n", opcode); + + if (PERF) + t0 = spu_read_decrementer(); + + switch (opcode & CELL_CMD_OPCODE_MASK) { + case CELL_CMD_EXIT: + D_PRINTF(CELL_DEBUG_CMD, "EXIT\n"); + exitFlag = 1; + break; + case CELL_CMD_VS_EXECUTE: +#if 0 + spu_execute_vertex_shader(&draw, &cmd.vs); +#endif + break; + case CELL_CMD_BATCH: + cmd_batch(opcode); + break; + default: + printf("Bad opcode 0x%x!\n", opcode & CELL_CMD_OPCODE_MASK); + } + + if (PERF) { + t1 = spu_read_decrementer(); + printf("wait mbox time: %gms batch time: %gms\n", + (~0u - t0) * spu.init.inv_timebase, + (t0 - t1) * spu.init.inv_timebase); + } + } + + D_PRINTF(CELL_DEBUG_CMD, "Exit command loop\n"); + + if (spu.init.debug_flags & CELL_DEBUG_CACHE) + spu_dcache_report(); +} + +/* Initialize this module; we manage the fragment ops buffer here. */ +void +spu_command_init(void) +{ + /* Install default/fallback fragment processing function. + * This will normally be overriden by a code-gen'd function + * unless CELL_FORCE_FRAGMENT_OPS_FALLBACK is set. + */ + spu.fragment_ops[CELL_FACING_FRONT] = spu_fallback_fragment_ops; + spu.fragment_ops[CELL_FACING_BACK] = spu_fallback_fragment_ops; + + /* Set up the basic empty buffer for code-gen'ed fragment ops */ + spu.fragment_ops_code = NULL; + spu.fragment_ops_code_size = 0; +} + +void +spu_command_close(void) +{ + /* Deallocate the code-gen buffer for fragment ops, and reset the + * fragment ops functions to their initial setting (just to leave + * things in a good state). + */ + if (spu.fragment_ops_code != NULL) { + free(spu.fragment_ops_code); + } + spu_command_init(); +} diff --git a/src/gallium/drivers/cell/spu/spu_command.h b/src/gallium/drivers/cell/spu/spu_command.h new file mode 100644 index 0000000000..83dcdade28 --- /dev/null +++ b/src/gallium/drivers/cell/spu/spu_command.h @@ -0,0 +1,35 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +extern void +command_loop(void); + +extern void +spu_command_init(void); + +extern void +spu_command_close(void); diff --git a/src/gallium/drivers/cell/spu/spu_dcache.c b/src/gallium/drivers/cell/spu/spu_dcache.c index 167404cdc5..a6d67634fd 100644 --- a/src/gallium/drivers/cell/spu/spu_dcache.c +++ b/src/gallium/drivers/cell/spu/spu_dcache.c @@ -36,7 +36,9 @@ #define CACHE_SET_TAGID(set) (((set) & 0x03) + TAG_DCACHE0) #define CACHE_LOG2NNWAY 2 #define CACHE_LOG2NSETS 6 -/*#define CACHE_STATS 1*/ +#ifdef DEBUG +#define CACHE_STATS 1 +#endif #include <cache-api.h> /* Yes folks, this is ugly. diff --git a/src/gallium/drivers/cell/spu/spu_funcs.c b/src/gallium/drivers/cell/spu/spu_funcs.c new file mode 100644 index 0000000000..ff3d609d25 --- /dev/null +++ b/src/gallium/drivers/cell/spu/spu_funcs.c @@ -0,0 +1,173 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +/** + * SPU functions accessed by shaders. + * + * Authors: Brian Paul + */ + + +#include <string.h> +#include <libmisc.h> +#include <math.h> +#include <cos14_v.h> +#include <sin14_v.h> +#include <simdmath/exp2f4.h> +#include <simdmath/log2f4.h> +#include <simdmath/powf4.h> + +#include "cell/common.h" +#include "spu_main.h" +#include "spu_funcs.h" +#include "spu_texture.h" + + +/** For "return"-ing four vectors */ +struct vec_4x4 +{ + vector float v[4]; +}; + + +static vector float +spu_cos(vector float x) +{ + return _cos14_v(x); +} + +static vector float +spu_sin(vector float x) +{ + return _sin14_v(x); +} + +static vector float +spu_pow(vector float x, vector float y) +{ + return _powf4(x, y); +} + +static vector float +spu_exp2(vector float x) +{ + return _exp2f4(x); +} + +static vector float +spu_log2(vector float x) +{ + return _log2f4(x); +} + + +static struct vec_4x4 +spu_tex_2d(vector float s, vector float t, vector float r, vector float q, + unsigned unit) +{ + struct vec_4x4 colors; + (void) r; + (void) q; + spu.sample_texture_2d[unit](s, t, unit, 0, 0, colors.v); + return colors; +} + +static struct vec_4x4 +spu_tex_3d(vector float s, vector float t, vector float r, vector float q, + unsigned unit) +{ + struct vec_4x4 colors; + (void) r; + (void) q; + spu.sample_texture_2d[unit](s, t, unit, 0, 0, colors.v); + return colors; +} + +static struct vec_4x4 +spu_tex_cube(vector float s, vector float t, vector float r, vector float q, + unsigned unit) +{ + struct vec_4x4 colors; + (void) q; + sample_texture_cube(s, t, r, unit, colors.v); + return colors; +} + + +/** + * Add named function to list of "exported" functions that will be + * made available to the PPU-hosted code generator. + */ +static void +export_func(struct cell_spu_function_info *spu_functions, + const char *name, void *addr) +{ + uint n = spu_functions->num; + ASSERT(strlen(name) < 16); + strcpy(spu_functions->names[n], name); + spu_functions->addrs[n] = (uint) addr; + spu_functions->num++; + ASSERT(spu_functions->num <= 16); +} + + +/** + * Return info about the SPU's function to the PPU / main memory. + * The PPU needs to know the address of some SPU-side functions so + * that we can generate shader code with function calls. + */ +void +return_function_info(void) +{ + struct cell_spu_function_info funcs ALIGN16_ATTRIB; + int tag = TAG_MISC; + + ASSERT(sizeof(funcs) == 256); /* must be multiple of 16 bytes */ + + funcs.num = 0; + export_func(&funcs, "spu_cos", &spu_cos); + export_func(&funcs, "spu_sin", &spu_sin); + export_func(&funcs, "spu_pow", &spu_pow); + export_func(&funcs, "spu_exp2", &spu_exp2); + export_func(&funcs, "spu_log2", &spu_log2); + export_func(&funcs, "spu_tex_2d", &spu_tex_2d); + export_func(&funcs, "spu_tex_3d", &spu_tex_3d); + export_func(&funcs, "spu_tex_cube", &spu_tex_cube); + + /* Send the function info back to the PPU / main memory */ + mfc_put((void *) &funcs, /* src in local store */ + (unsigned int) spu.init.spu_functions, /* dst in main memory */ + sizeof(funcs), /* bytes */ + tag, + 0, /* tid */ + 0 /* rid */); + wait_on_mask(1 << tag); +} + + + diff --git a/src/gallium/drivers/cell/ppu/cell_winsys.c b/src/gallium/drivers/cell/spu/spu_funcs.h index d570bbd2f9..3adb6ae99f 100644 --- a/src/gallium/drivers/cell/ppu/cell_winsys.c +++ b/src/gallium/drivers/cell/spu/spu_funcs.h @@ -1,6 +1,6 @@ /************************************************************************** * - * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. * All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a @@ -25,16 +25,11 @@ * **************************************************************************/ +#ifndef SPU_FUNCS_H +#define SPU_FUNCS_H -#include "util/u_memory.h" -#include "cell_winsys.h" +extern void +return_function_info(void); +#endif -struct cell_winsys * -cell_get_winsys(uint format) -{ - struct cell_winsys *cws = CALLOC_STRUCT(cell_winsys); - if (cws) - cws->preferredFormat = format; - return cws; -} diff --git a/src/gallium/drivers/cell/spu/spu_main.c b/src/gallium/drivers/cell/spu/spu_main.c index 78260c4259..97c86d194d 100644 --- a/src/gallium/drivers/cell/spu/spu_main.c +++ b/src/gallium/drivers/cell/spu/spu_main.c @@ -32,16 +32,15 @@ #include <stdio.h> #include <libmisc.h> +#include "pipe/p_defines.h" + +#include "spu_funcs.h" +#include "spu_command.h" #include "spu_main.h" -#include "spu_render.h" #include "spu_per_fragment_op.h" #include "spu_texture.h" -#include "spu_tile.h" //#include "spu_test.h" -#include "spu_vertex_shader.h" -#include "spu_dcache.h" #include "cell/common.h" -#include "pipe/p_defines.h" /* @@ -50,600 +49,8 @@ helpful headers: /opt/cell/sdk/usr/include/libmisc.h */ -boolean Debug = FALSE; - struct spu_global spu; -struct spu_vs_context draw; - - -/** - * Buffers containing dynamically generated SPU code: - */ -static unsigned char attribute_fetch_code_buffer[136 * PIPE_MAX_ATTRIBS] - ALIGN16_ATTRIB; - - - -/** - * Tell the PPU that this SPU has finished copying a buffer to - * local store and that it may be reused by the PPU. - * This is done by writting a 16-byte batch-buffer-status block back into - * main memory (in cell_context->buffer_status[]). - */ -static void -release_buffer(uint buffer) -{ - /* Evidently, using less than a 16-byte status doesn't work reliably */ - static const uint status[4] ALIGN16_ATTRIB - = {CELL_BUFFER_STATUS_FREE, 0, 0, 0}; - - const uint index = 4 * (spu.init.id * CELL_NUM_BUFFERS + buffer); - uint *dst = spu.init.buffer_status + index; - - ASSERT(buffer < CELL_NUM_BUFFERS); - - mfc_put((void *) &status, /* src in local memory */ - (unsigned int) dst, /* dst in main memory */ - sizeof(status), /* size */ - TAG_MISC, /* tag is unimportant */ - 0, /* tid */ - 0 /* rid */); -} - - -/** - * For tiles whose status is TILE_STATUS_CLEAR, write solid-filled - * tiles back to the main framebuffer. - */ -static void -really_clear_tiles(uint surfaceIndex) -{ - const uint num_tiles = spu.fb.width_tiles * spu.fb.height_tiles; - uint i; - - if (surfaceIndex == 0) { - clear_c_tile(&spu.ctile); - - for (i = spu.init.id; i < num_tiles; i += spu.init.num_spus) { - uint tx = i % spu.fb.width_tiles; - uint ty = i / spu.fb.width_tiles; - if (spu.ctile_status[ty][tx] == TILE_STATUS_CLEAR) { - put_tile(tx, ty, &spu.ctile, TAG_SURFACE_CLEAR, 0); - } - } - } - else { - clear_z_tile(&spu.ztile); - - for (i = spu.init.id; i < num_tiles; i += spu.init.num_spus) { - uint tx = i % spu.fb.width_tiles; - uint ty = i / spu.fb.width_tiles; - if (spu.ztile_status[ty][tx] == TILE_STATUS_CLEAR) - put_tile(tx, ty, &spu.ctile, TAG_SURFACE_CLEAR, 1); - } - } - -#if 0 - wait_on_mask(1 << TAG_SURFACE_CLEAR); -#endif -} - - -static void -cmd_clear_surface(const struct cell_command_clear_surface *clear) -{ - if (Debug) - printf("SPU %u: CLEAR SURF %u to 0x%08x\n", spu.init.id, - clear->surface, clear->value); - - if (clear->surface == 0) { - spu.fb.color_clear_value = clear->value; - if (spu.init.debug_flags & CELL_DEBUG_CHECKER) { - uint x = (spu.init.id << 4) | (spu.init.id << 12) | - (spu.init.id << 20) | (spu.init.id << 28); - spu.fb.color_clear_value ^= x; - } - } - else { - spu.fb.depth_clear_value = clear->value; - } - -#define CLEAR_OPT 1 -#if CLEAR_OPT - - /* Simply set all tiles' status to CLEAR. - * When we actually begin rendering into a tile, we'll initialize it to - * the clear value. If any tiles go untouched during the frame, - * really_clear_tiles() will set them to the clear value. - */ - if (clear->surface == 0) { - memset(spu.ctile_status, TILE_STATUS_CLEAR, sizeof(spu.ctile_status)); - } - else { - memset(spu.ztile_status, TILE_STATUS_CLEAR, sizeof(spu.ztile_status)); - } - -#else - - /* - * This path clears the whole framebuffer to the clear color right now. - */ - - /* - printf("SPU: %s num=%d w=%d h=%d\n", - __FUNCTION__, num_tiles, spu.fb.width_tiles, spu.fb.height_tiles); - */ - - /* init a single tile to the clear value */ - if (clear->surface == 0) { - clear_c_tile(&spu.ctile); - } - else { - clear_z_tile(&spu.ztile); - } - - /* walk over my tiles, writing the 'clear' tile's data */ - { - const uint num_tiles = spu.fb.width_tiles * spu.fb.height_tiles; - uint i; - for (i = spu.init.id; i < num_tiles; i += spu.init.num_spus) { - uint tx = i % spu.fb.width_tiles; - uint ty = i / spu.fb.width_tiles; - if (clear->surface == 0) - put_tile(tx, ty, &spu.ctile, TAG_SURFACE_CLEAR, 0); - else - put_tile(tx, ty, &spu.ztile, TAG_SURFACE_CLEAR, 1); - } - } - - if (spu.init.debug_flags & CELL_DEBUG_SYNC) { - wait_on_mask(1 << TAG_SURFACE_CLEAR); - } - -#endif /* CLEAR_OPT */ - - if (Debug) - printf("SPU %u: CLEAR SURF done\n", spu.init.id); -} - - -static void -cmd_release_verts(const struct cell_command_release_verts *release) -{ - if (Debug) - printf("SPU %u: RELEASE VERTS %u\n", - spu.init.id, release->vertex_buf); - ASSERT(release->vertex_buf != ~0U); - release_buffer(release->vertex_buf); -} - - -/** - * Process a CELL_CMD_STATE_FRAGMENT_OPS command. - * This involves installing new fragment ops SPU code. - * If this function is never called, we'll use a regular C fallback function - * for fragment processing. - */ -static void -cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops) -{ - if (Debug) - printf("SPU %u: CMD_STATE_FRAGMENT_OPS\n", spu.init.id); - /* Copy SPU code from batch buffer to spu buffer */ - memcpy(spu.fragment_ops_code, fops->code, SPU_MAX_FRAGMENT_OPS_INSTS * 4); - /* Copy state info (for fallback case only) */ - memcpy(&spu.depth_stencil_alpha, &fops->dsa, sizeof(fops->dsa)); - memcpy(&spu.blend, &fops->blend, sizeof(fops->blend)); - - /* Point function pointer at new code */ - spu.fragment_ops = (spu_fragment_ops_func) spu.fragment_ops_code; - - spu.read_depth = spu.depth_stencil_alpha.depth.enabled; - spu.read_stencil = spu.depth_stencil_alpha.stencil[0].enabled; -} - - -static void -cmd_state_fragment_program(const struct cell_command_fragment_program *fp) -{ - if (Debug) - printf("SPU %u: CMD_STATE_FRAGMENT_PROGRAM\n", spu.init.id); - /* Copy SPU code from batch buffer to spu buffer */ - memcpy(spu.fragment_program_code, fp->code, - SPU_MAX_FRAGMENT_PROGRAM_INSTS * 4); -#if 01 - /* Point function pointer at new code */ - spu.fragment_program = (spu_fragment_program_func)spu.fragment_program_code; -#endif -} - - -static void -cmd_state_framebuffer(const struct cell_command_framebuffer *cmd) -{ - if (Debug) - printf("SPU %u: FRAMEBUFFER: %d x %d at %p, cformat 0x%x zformat 0x%x\n", - spu.init.id, - cmd->width, - cmd->height, - cmd->color_start, - cmd->color_format, - cmd->depth_format); - - ASSERT_ALIGN16(cmd->color_start); - ASSERT_ALIGN16(cmd->depth_start); - - spu.fb.color_start = cmd->color_start; - spu.fb.depth_start = cmd->depth_start; - spu.fb.color_format = cmd->color_format; - spu.fb.depth_format = cmd->depth_format; - spu.fb.width = cmd->width; - spu.fb.height = cmd->height; - spu.fb.width_tiles = (spu.fb.width + TILE_SIZE - 1) / TILE_SIZE; - spu.fb.height_tiles = (spu.fb.height + TILE_SIZE - 1) / TILE_SIZE; - - switch (spu.fb.depth_format) { - case PIPE_FORMAT_Z32_UNORM: - spu.fb.zsize = 4; - spu.fb.zscale = (float) 0xffffffffu; - break; - case PIPE_FORMAT_Z24S8_UNORM: - case PIPE_FORMAT_S8Z24_UNORM: - case PIPE_FORMAT_Z24X8_UNORM: - case PIPE_FORMAT_X8Z24_UNORM: - spu.fb.zsize = 4; - spu.fb.zscale = (float) 0x00ffffffu; - break; - case PIPE_FORMAT_Z16_UNORM: - spu.fb.zsize = 2; - spu.fb.zscale = (float) 0xffffu; - break; - default: - spu.fb.zsize = 0; - break; - } -} - - -static void -cmd_state_sampler(const struct cell_command_sampler *sampler) -{ - if (Debug) - printf("SPU %u: SAMPLER [%u]\n", - spu.init.id, sampler->unit); - - spu.sampler[sampler->unit] = sampler->state; - if (spu.sampler[sampler->unit].min_img_filter == PIPE_TEX_FILTER_LINEAR) - spu.sample_texture[sampler->unit] = sample_texture_bilinear; - else - spu.sample_texture[sampler->unit] = sample_texture_nearest; -} - - -static void -cmd_state_texture(const struct cell_command_texture *texture) -{ - const uint unit = texture->unit; - const uint width = texture->width; - const uint height = texture->height; - - if (Debug) { - printf("SPU %u: TEXTURE [%u] at %p size %u x %u\n", spu.init.id, - texture->unit, texture->start, - texture->width, texture->height); - } - - spu.texture[unit].start = texture->start; - spu.texture[unit].width = width; - spu.texture[unit].height = height; - - spu.texture[unit].tiles_per_row = width / TILE_SIZE; - - spu.texture[unit].tex_size = (vector float) { width, height, 0.0, 0.0}; - spu.texture[unit].tex_size_mask = (vector unsigned int) - { width - 1, height - 1, 0, 0 }; - spu.texture[unit].tex_size_x_mask = spu_splats(width - 1); - spu.texture[unit].tex_size_y_mask = spu_splats(height - 1); -} - - -static void -cmd_state_vertex_info(const struct vertex_info *vinfo) -{ - if (Debug) { - printf("SPU %u: VERTEX_INFO num_attribs=%u\n", spu.init.id, - vinfo->num_attribs); - } - ASSERT(vinfo->num_attribs >= 1); - ASSERT(vinfo->num_attribs <= 8); - memcpy(&spu.vertex_info, vinfo, sizeof(*vinfo)); -} - - -static void -cmd_state_vs_array_info(const struct cell_array_info *vs_info) -{ - const unsigned attr = vs_info->attr; - - ASSERT(attr < PIPE_MAX_ATTRIBS); - draw.vertex_fetch.src_ptr[attr] = vs_info->base; - draw.vertex_fetch.pitch[attr] = vs_info->pitch; - draw.vertex_fetch.size[attr] = vs_info->size; - draw.vertex_fetch.code_offset[attr] = vs_info->function_offset; - draw.vertex_fetch.dirty = 1; -} - - -static void -cmd_state_attrib_fetch(const struct cell_attribute_fetch_code *code) -{ - mfc_get(attribute_fetch_code_buffer, - (unsigned int) code->base, /* src */ - code->size, - TAG_BATCH_BUFFER, - 0, /* tid */ - 0 /* rid */); - wait_on_mask(1 << TAG_BATCH_BUFFER); - - draw.vertex_fetch.code = attribute_fetch_code_buffer; -} - - -static void -cmd_finish(void) -{ - if (Debug) - printf("SPU %u: FINISH\n", spu.init.id); - really_clear_tiles(0); - /* wait for all outstanding DMAs to finish */ - mfc_write_tag_mask(~0); - mfc_read_tag_status_all(); - /* send mbox message to PPU */ - spu_write_out_mbox(CELL_CMD_FINISH); -} - - -/** - * Execute a batch of commands which was sent to us by the PPU. - * See the cell_emit_state.c code to see where the commands come from. - * - * The opcode param encodes the location of the buffer and its size. - */ -static void -cmd_batch(uint opcode) -{ - const uint buf = (opcode >> 8) & 0xff; - uint size = (opcode >> 16); - uint64_t buffer[CELL_BUFFER_SIZE / 8] ALIGN16_ATTRIB; - const unsigned usize = size / sizeof(buffer[0]); - uint pos; - - if (Debug) - printf("SPU %u: BATCH buffer %u, len %u, from %p\n", - spu.init.id, buf, size, spu.init.buffers[buf]); - - ASSERT((opcode & CELL_CMD_OPCODE_MASK) == CELL_CMD_BATCH); - - ASSERT_ALIGN16(spu.init.buffers[buf]); - - size = ROUNDUP16(size); - - ASSERT_ALIGN16(spu.init.buffers[buf]); - - mfc_get(buffer, /* dest */ - (unsigned int) spu.init.buffers[buf], /* src */ - size, - TAG_BATCH_BUFFER, - 0, /* tid */ - 0 /* rid */); - wait_on_mask(1 << TAG_BATCH_BUFFER); - - /* Tell PPU we're done copying the buffer to local store */ - if (Debug) - printf("SPU %u: release batch buf %u\n", spu.init.id, buf); - release_buffer(buf); - - /* - * Loop over commands in the batch buffer - */ - for (pos = 0; pos < usize; /* no incr */) { - switch (buffer[pos]) { - /* - * rendering commands - */ - case CELL_CMD_CLEAR_SURFACE: - { - struct cell_command_clear_surface *clr - = (struct cell_command_clear_surface *) &buffer[pos]; - cmd_clear_surface(clr); - pos += sizeof(*clr) / 8; - } - break; - case CELL_CMD_RENDER: - { - struct cell_command_render *render - = (struct cell_command_render *) &buffer[pos]; - uint pos_incr; - cmd_render(render, &pos_incr); - pos += pos_incr; - } - break; - /* - * state-update commands - */ - case CELL_CMD_STATE_FRAMEBUFFER: - { - struct cell_command_framebuffer *fb - = (struct cell_command_framebuffer *) &buffer[pos]; - cmd_state_framebuffer(fb); - pos += sizeof(*fb) / 8; - } - break; - case CELL_CMD_STATE_FRAGMENT_OPS: - { - struct cell_command_fragment_ops *fops - = (struct cell_command_fragment_ops *) &buffer[pos]; - cmd_state_fragment_ops(fops); - pos += sizeof(*fops) / 8; - } - break; - case CELL_CMD_STATE_FRAGMENT_PROGRAM: - { - struct cell_command_fragment_program *fp - = (struct cell_command_fragment_program *) &buffer[pos]; - cmd_state_fragment_program(fp); - pos += sizeof(*fp) / 8; - } - break; - case CELL_CMD_STATE_SAMPLER: - { - struct cell_command_sampler *sampler - = (struct cell_command_sampler *) &buffer[pos]; - cmd_state_sampler(sampler); - pos += sizeof(*sampler) / 8; - } - break; - case CELL_CMD_STATE_TEXTURE: - { - struct cell_command_texture *texture - = (struct cell_command_texture *) &buffer[pos]; - cmd_state_texture(texture); - pos += sizeof(*texture) / 8; - } - break; - case CELL_CMD_STATE_VERTEX_INFO: - cmd_state_vertex_info((struct vertex_info *) &buffer[pos+1]); - pos += (1 + ROUNDUP8(sizeof(struct vertex_info)) / 8); - break; - case CELL_CMD_STATE_VIEWPORT: - (void) memcpy(& draw.viewport, &buffer[pos+1], - sizeof(struct pipe_viewport_state)); - pos += (1 + ROUNDUP8(sizeof(struct pipe_viewport_state)) / 8); - break; - case CELL_CMD_STATE_UNIFORMS: - draw.constants = (const float (*)[4]) (uintptr_t) buffer[pos + 1]; - pos += 2; - break; - case CELL_CMD_STATE_VS_ARRAY_INFO: - cmd_state_vs_array_info((struct cell_array_info *) &buffer[pos+1]); - pos += (1 + ROUNDUP8(sizeof(struct cell_array_info)) / 8); - break; - case CELL_CMD_STATE_BIND_VS: -#if 0 - spu_bind_vertex_shader(&draw, - (struct cell_shader_info *) &buffer[pos+1]); -#endif - pos += (1 + ROUNDUP8(sizeof(struct cell_shader_info)) / 8); - break; - case CELL_CMD_STATE_ATTRIB_FETCH: - cmd_state_attrib_fetch((struct cell_attribute_fetch_code *) - &buffer[pos+1]); - pos += (1 + ROUNDUP8(sizeof(struct cell_attribute_fetch_code)) / 8); - break; - /* - * misc commands - */ - case CELL_CMD_FINISH: - cmd_finish(); - pos += 1; - break; - case CELL_CMD_RELEASE_VERTS: - { - struct cell_command_release_verts *release - = (struct cell_command_release_verts *) &buffer[pos]; - cmd_release_verts(release); - pos += sizeof(*release) / 8; - } - break; - case CELL_CMD_FLUSH_BUFFER_RANGE: { - struct cell_buffer_range *br = (struct cell_buffer_range *) - &buffer[pos+1]; - - spu_dcache_mark_dirty((unsigned) br->base, br->size); - pos += (1 + ROUNDUP8(sizeof(struct cell_buffer_range)) / 8); - break; - } - default: - printf("SPU %u: bad opcode: 0x%llx\n", spu.init.id, buffer[pos]); - ASSERT(0); - break; - } - } - - if (Debug) - printf("SPU %u: BATCH complete\n", spu.init.id); -} - - -/** - * Temporary/simple main loop for SPEs: Get a command, execute it, repeat. - */ -static void -main_loop(void) -{ - struct cell_command cmd; - int exitFlag = 0; - - if (Debug) - printf("SPU %u: Enter main loop\n", spu.init.id); - - ASSERT((sizeof(struct cell_command) & 0xf) == 0); - ASSERT_ALIGN16(&cmd); - - while (!exitFlag) { - unsigned opcode; - int tag = 0; - - if (Debug) - printf("SPU %u: Wait for cmd...\n", spu.init.id); - - /* read/wait from mailbox */ - opcode = (unsigned int) spu_read_in_mbox(); - - if (Debug) - printf("SPU %u: got cmd 0x%x\n", spu.init.id, opcode); - - /* command payload */ - mfc_get(&cmd, /* dest */ - (unsigned int) spu.init.cmd, /* src */ - sizeof(struct cell_command), /* bytes */ - tag, - 0, /* tid */ - 0 /* rid */); - wait_on_mask( 1 << tag ); - - /* - * NOTE: most commands should be contained in a batch buffer - */ - - switch (opcode & CELL_CMD_OPCODE_MASK) { - case CELL_CMD_EXIT: - if (Debug) - printf("SPU %u: EXIT\n", spu.init.id); - exitFlag = 1; - break; - case CELL_CMD_VS_EXECUTE: -#if 0 - spu_execute_vertex_shader(&draw, &cmd.vs); -#endif - break; - case CELL_CMD_BATCH: - cmd_batch(opcode); - break; - default: - printf("Bad opcode!\n"); - } - - } - - if (Debug) - printf("SPU %u: Exit main loop\n", spu.init.id); - - spu_dcache_report(); -} - - static void one_time_init(void) @@ -651,15 +58,8 @@ one_time_init(void) memset(spu.ctile_status, TILE_STATUS_DEFINED, sizeof(spu.ctile_status)); memset(spu.ztile_status, TILE_STATUS_DEFINED, sizeof(spu.ztile_status)); invalidate_tex_cache(); - - /* Install default/fallback fragment processing function. - * This will normally be overriden by a code-gen'd function. - */ - spu.fragment_ops = spu_fallback_fragment_ops; } - - /* In some versions of the SDK the SPE main takes 'unsigned long' as a * parameter. In others it takes 'unsigned long long'. Use a define to * select between the two. @@ -682,12 +82,16 @@ main(main_param_t speid, main_param_t argp) ASSERT(sizeof(tile_t) == TILE_SIZE * TILE_SIZE * 4); ASSERT(sizeof(struct cell_command_render) % 8 == 0); + ASSERT(sizeof(struct cell_command_fragment_ops) % 8 == 0); + ASSERT(((unsigned long) &spu.fragment_program_code) % 8 == 0); one_time_init(); + spu_command_init(); - if (Debug) - printf("SPU: main() speid=%lu\n", (unsigned long) speid); + D_PRINTF(CELL_DEBUG_CMD, "main() speid=%lu\n", (unsigned long) speid); + D_PRINTF(CELL_DEBUG_FRAGMENT_OP_FALLBACK, "using fragment op fallback\n"); + /* get initialization data */ mfc_get(&spu.init, /* dest */ (unsigned int) argp, /* src */ sizeof(struct cell_init_info), /* bytes */ @@ -696,12 +100,18 @@ main(main_param_t speid, main_param_t argp) 0 /* rid */); wait_on_mask( 1 << tag ); + if (spu.init.id == 0) { + return_function_info(); + } + #if 0 if (spu.init.id==0) - spu_test_misc(); + spu_test_misc(spu.init.id); #endif - main_loop(); + command_loop(); + + spu_command_close(); return 0; } diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h index 2c7b625840..33767e7c51 100644 --- a/src/gallium/drivers/cell/spu/spu_main.h +++ b/src/gallium/drivers/cell/spu/spu_main.h @@ -36,9 +36,18 @@ #include "pipe/p_state.h" - -#define MAX_WIDTH 1024 -#define MAX_HEIGHT 1024 +#if DEBUG +/* These debug macros use the unusual construction ", ##__VA_ARGS__" + * which expands to the expected comma + args if variadic arguments + * are supplied, but swallows the comma if there are no variadic + * arguments (which avoids syntax errors that would otherwise occur). + */ +#define D_PRINTF(flag, format,...) \ + if (spu.init.debug_flags & (flag)) \ + printf("SPU %u: " format, spu.init.id, ##__VA_ARGS__) +#else +#define D_PRINTF(...) +#endif /** @@ -61,8 +70,11 @@ typedef union { /** Function for sampling textures */ -typedef vector float (*spu_sample_texture_func)(uint unit, - vector float texcoord); +typedef void (*spu_sample_texture_2d_func)(vector float s, + vector float t, + uint unit, uint level, uint face, + vector float colors[4]); + /** Function for performing per-fragment ops */ typedef void (*spu_fragment_ops_func)(uint x, uint y, @@ -76,9 +88,9 @@ typedef void (*spu_fragment_ops_func)(uint x, uint y, vector unsigned int mask); /** Function for running fragment program */ -typedef void (*spu_fragment_program_func)(vector float *inputs, - vector float *outputs, - vector float *constants); +typedef vector unsigned int (*spu_fragment_program_func)(vector float *inputs, + vector float *outputs, + vector float *constants); struct spu_framebuffer @@ -98,15 +110,27 @@ struct spu_framebuffer } ALIGN16_ATTRIB; -struct spu_texture +/** per-texture level info */ +struct spu_texture_level { void *start; - ushort width, height; + ushort width, height, depth; ushort tiles_per_row; - vector float tex_size; - vector unsigned int tex_size_mask; /**< == int(size - 1) */ - vector unsigned int tex_size_x_mask; /**< == int(size - 1) */ - vector unsigned int tex_size_y_mask; /**< == int(size - 1) */ + uint bytes_per_image; + /** texcoord scale factors */ + vector float scale_s, scale_t, scale_r; + /** texcoord masks (if REPEAT then size-1, else ~0) */ + vector signed int mask_s, mask_t, mask_r; + /** texcoord clamp limits */ + vector signed int max_s, max_t, max_r; +} ALIGN16_ATTRIB; + + +struct spu_texture +{ + struct spu_texture_level level[CELL_MAX_TEXTURE_LEVELS]; + uint max_level; + uint target; /**< PIPE_TEXTURE_x */ } ALIGN16_ATTRIB; @@ -124,7 +148,9 @@ struct spu_global struct spu_framebuffer fb; struct pipe_depth_stencil_alpha_state depth_stencil_alpha; struct pipe_blend_state blend; + struct pipe_blend_color blend_color; struct pipe_sampler_state sampler[PIPE_MAX_SAMPLERS]; + struct pipe_rasterizer_state rasterizer; struct spu_texture texture[PIPE_MAX_SAMPLERS]; struct vertex_info vertex_info; @@ -133,39 +159,38 @@ struct spu_global tile_t ztile ALIGN16_ATTRIB; /** Read depth/stencil tiles? */ - boolean read_depth; - boolean read_stencil; + boolean read_depth_stencil; /** Current tiles' status */ ubyte cur_ctile_status, cur_ztile_status; /** Status of all tiles in framebuffer */ - ubyte ctile_status[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB; - ubyte ztile_status[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB; + ubyte ctile_status[CELL_MAX_HEIGHT/TILE_SIZE][CELL_MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB; + ubyte ztile_status[CELL_MAX_HEIGHT/TILE_SIZE][CELL_MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB; - /** Current fragment ops machine code */ - uint fragment_ops_code[SPU_MAX_FRAGMENT_OPS_INSTS]; - /** Current fragment ops function */ - spu_fragment_ops_func fragment_ops; + /** Current fragment ops machine code, at 8-byte boundary */ + uint *fragment_ops_code; + uint fragment_ops_code_size; + /** Current fragment ops functions, 0 = frontfacing, 1 = backfacing */ + spu_fragment_ops_func fragment_ops[2]; - /** Current fragment program machine code */ - uint fragment_program_code[SPU_MAX_FRAGMENT_PROGRAM_INSTS]; + /** Current fragment program machine code, at 8-byte boundary */ + uint fragment_program_code[SPU_MAX_FRAGMENT_PROGRAM_INSTS] ALIGN8_ATTRIB; /** Current fragment ops function */ spu_fragment_program_func fragment_program; /** Current texture sampler function */ - spu_sample_texture_func sample_texture[CELL_MAX_SAMPLERS]; + spu_sample_texture_2d_func sample_texture_2d[CELL_MAX_SAMPLERS]; + spu_sample_texture_2d_func min_sample_texture_2d[CELL_MAX_SAMPLERS]; + spu_sample_texture_2d_func mag_sample_texture_2d[CELL_MAX_SAMPLERS]; - /** Fragment program constants (XXX preliminary/used) */ -#define MAX_CONSTANTS 32 - vector float constants[MAX_CONSTANTS]; + /** Fragment program constants */ + vector float constants[4 * CELL_MAX_CONSTANTS]; } ALIGN16_ATTRIB; extern struct spu_global spu; -extern boolean Debug; - @@ -184,7 +209,7 @@ extern boolean Debug; #define TAG_DCACHE1 21 #define TAG_DCACHE2 22 #define TAG_DCACHE3 23 - +#define TAG_FENCE 24 static INLINE void diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c index 03dd547845..eba9f95cf1 100644 --- a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c @@ -40,6 +40,24 @@ #define LINEAR_QUAD_LAYOUT 1 +static INLINE vector float +spu_min(vector float a, vector float b) +{ + vector unsigned int m; + m = spu_cmpgt(a, b); /* m = a > b ? ~0 : 0 */ + return spu_sel(a, b, m); +} + + +static INLINE vector float +spu_max(vector float a, vector float b) +{ + vector unsigned int m; + m = spu_cmpgt(a, b); /* m = a > b ? ~0 : 0 */ + return spu_sel(b, a, m); +} + + /** * Called by rasterizer for each quad after the shader has run. Do * all the per-fragment operations including alpha test, z test, @@ -60,11 +78,14 @@ spu_fallback_fragment_ops(uint x, uint y, vector unsigned int mask) { vector float frag_aos[4]; - unsigned int c0, c1, c2, c3; + unsigned int fbc0, fbc1, fbc2, fbc3 ; /* framebuffer/tile colors */ + unsigned int fragc0, fragc1, fragc2, fragc3; /* fragment colors */ - /* do alpha test */ + /* + * Do alpha test + */ if (spu.depth_stencil_alpha.alpha.enabled) { - vector float ref = spu_splats(spu.depth_stencil_alpha.alpha.ref); + vector float ref = spu_splats(spu.depth_stencil_alpha.alpha.ref_value); vector unsigned int amask; switch (spu.depth_stencil_alpha.alpha.func) { @@ -102,7 +123,10 @@ spu_fallback_fragment_ops(uint x, uint y, mask = spu_and(mask, amask); } - /* Z and/or stencil testing... */ + + /* + * Z and/or stencil testing... + */ if (spu.depth_stencil_alpha.depth.enabled || spu.depth_stencil_alpha.stencil[0].enabled) { @@ -178,6 +202,32 @@ spu_fallback_fragment_ops(uint x, uint y, } } + + /* + * If we'll need the current framebuffer/tile colors for blending + * or logicop or colormask, fetch them now. + */ + if (spu.blend.blend_enable || + spu.blend.logicop_enable || + spu.blend.colormask != 0xf) { + +#if LINEAR_QUAD_LAYOUT /* See comments/diagram below */ + fbc0 = colorTile->ui[y][x*2+0]; + fbc1 = colorTile->ui[y][x*2+1]; + fbc2 = colorTile->ui[y][x*2+2]; + fbc3 = colorTile->ui[y][x*2+3]; +#else + fbc0 = colorTile->ui[y+0][x+0]; + fbc1 = colorTile->ui[y+0][x+1]; + fbc2 = colorTile->ui[y+1][x+0]; + fbc3 = colorTile->ui[y+1][x+1]; +#endif + } + + + /* + * Do blending + */ if (spu.blend.blend_enable) { /* blending terms, misc regs */ vector float term1r, term1g, term1b, term1a; @@ -186,43 +236,30 @@ spu_fallback_fragment_ops(uint x, uint y, vector float fbRGBA[4]; /* current framebuffer colors */ - /* get colors from framebuffer/tile */ + /* convert framebuffer colors from packed int to vector float */ { - vector float fc[4]; - uint c0, c1, c2, c3; - -#if LINEAR_QUAD_LAYOUT /* See comments/diagram below */ - c0 = colorTile->ui[y][x*2+0]; - c1 = colorTile->ui[y][x*2+1]; - c2 = colorTile->ui[y][x*2+2]; - c3 = colorTile->ui[y][x*2+3]; -#else - c0 = colorTile->ui[y+0][x+0]; - c1 = colorTile->ui[y+0][x+1]; - c2 = colorTile->ui[y+1][x+0]; - c3 = colorTile->ui[y+1][x+1]; -#endif + vector float temp[4]; /* float colors in AOS form */ switch (spu.fb.color_format) { case PIPE_FORMAT_B8G8R8A8_UNORM: - fc[0] = spu_unpack_B8G8R8A8(c0); - fc[1] = spu_unpack_B8G8R8A8(c1); - fc[2] = spu_unpack_B8G8R8A8(c2); - fc[3] = spu_unpack_B8G8R8A8(c3); + temp[0] = spu_unpack_B8G8R8A8(fbc0); + temp[1] = spu_unpack_B8G8R8A8(fbc1); + temp[2] = spu_unpack_B8G8R8A8(fbc2); + temp[3] = spu_unpack_B8G8R8A8(fbc3); break; case PIPE_FORMAT_A8R8G8B8_UNORM: - fc[0] = spu_unpack_A8R8G8B8(c0); - fc[1] = spu_unpack_A8R8G8B8(c1); - fc[2] = spu_unpack_A8R8G8B8(c2); - fc[3] = spu_unpack_A8R8G8B8(c3); + temp[0] = spu_unpack_A8R8G8B8(fbc0); + temp[1] = spu_unpack_A8R8G8B8(fbc1); + temp[2] = spu_unpack_A8R8G8B8(fbc2); + temp[3] = spu_unpack_A8R8G8B8(fbc3); break; default: ASSERT(0); } - _transpose_matrix4x4(fbRGBA, fc); + _transpose_matrix4x4(fbRGBA, temp); /* fbRGBA = transpose(temp) */ } /* - * Compute Src RGB terms + * Compute Src RGB terms (fragment color * factor) */ switch (spu.blend.rgb_src_factor) { case PIPE_BLENDFACTOR_ONE: @@ -245,13 +282,33 @@ spu_fallback_fragment_ops(uint x, uint y, term1g = spu_mul(fragG, fragA); term1b = spu_mul(fragB, fragA); break; + case PIPE_BLENDFACTOR_DST_COLOR: + term1r = spu_mul(fragR, fbRGBA[0]); + term1g = spu_mul(fragG, fbRGBA[1]); + term1b = spu_mul(fragB, fbRGBA[1]); + break; + case PIPE_BLENDFACTOR_DST_ALPHA: + term1r = spu_mul(fragR, fbRGBA[3]); + term1g = spu_mul(fragG, fbRGBA[3]); + term1b = spu_mul(fragB, fbRGBA[3]); + break; + case PIPE_BLENDFACTOR_CONST_COLOR: + term1r = spu_mul(fragR, spu_splats(spu.blend_color.color[0])); + term1g = spu_mul(fragG, spu_splats(spu.blend_color.color[1])); + term1b = spu_mul(fragB, spu_splats(spu.blend_color.color[2])); + break; + case PIPE_BLENDFACTOR_CONST_ALPHA: + term1r = spu_mul(fragR, spu_splats(spu.blend_color.color[3])); + term1g = spu_mul(fragG, spu_splats(spu.blend_color.color[3])); + term1b = spu_mul(fragB, spu_splats(spu.blend_color.color[3])); + break; /* XXX more cases */ default: ASSERT(0); } /* - * Compute Src Alpha term + * Compute Src Alpha term (fragment alpha * factor) */ switch (spu.blend.alpha_src_factor) { case PIPE_BLENDFACTOR_ONE: @@ -263,19 +320,29 @@ spu_fallback_fragment_ops(uint x, uint y, case PIPE_BLENDFACTOR_SRC_ALPHA: term1a = spu_mul(fragA, fragA); break; + case PIPE_BLENDFACTOR_DST_COLOR: + /* fall-through */ + case PIPE_BLENDFACTOR_DST_ALPHA: + term1a = spu_mul(fragA, fbRGBA[3]); + break; + case PIPE_BLENDFACTOR_CONST_COLOR: + /* fall-through */ + case PIPE_BLENDFACTOR_CONST_ALPHA: + term1a = spu_mul(fragR, spu_splats(spu.blend_color.color[3])); + break; /* XXX more cases */ default: ASSERT(0); } /* - * Compute Dest RGB terms + * Compute Dest RGB terms (framebuffer color * factor) */ switch (spu.blend.rgb_dst_factor) { case PIPE_BLENDFACTOR_ONE: - term2r = fragR; - term2g = fragG; - term2b = fragB; + term2r = fbRGBA[0]; + term2g = fbRGBA[1]; + term2b = fbRGBA[2]; break; case PIPE_BLENDFACTOR_ZERO: term2r = @@ -299,17 +366,37 @@ spu_fallback_fragment_ops(uint x, uint y, term2g = spu_mul(fbRGBA[1], tmp); term2b = spu_mul(fbRGBA[2], tmp); break; - /* XXX more cases */ + case PIPE_BLENDFACTOR_DST_COLOR: + term2r = spu_mul(fbRGBA[0], fbRGBA[0]); + term2g = spu_mul(fbRGBA[1], fbRGBA[1]); + term2b = spu_mul(fbRGBA[2], fbRGBA[2]); + break; + case PIPE_BLENDFACTOR_DST_ALPHA: + term2r = spu_mul(fbRGBA[0], fbRGBA[3]); + term2g = spu_mul(fbRGBA[1], fbRGBA[3]); + term2b = spu_mul(fbRGBA[2], fbRGBA[3]); + break; + case PIPE_BLENDFACTOR_CONST_COLOR: + term2r = spu_mul(fbRGBA[0], spu_splats(spu.blend_color.color[0])); + term2g = spu_mul(fbRGBA[1], spu_splats(spu.blend_color.color[1])); + term2b = spu_mul(fbRGBA[2], spu_splats(spu.blend_color.color[2])); + break; + case PIPE_BLENDFACTOR_CONST_ALPHA: + term2r = spu_mul(fbRGBA[0], spu_splats(spu.blend_color.color[3])); + term2g = spu_mul(fbRGBA[1], spu_splats(spu.blend_color.color[3])); + term2b = spu_mul(fbRGBA[2], spu_splats(spu.blend_color.color[3])); + break; + /* XXX more cases */ default: ASSERT(0); } /* - * Compute Dest Alpha term + * Compute Dest Alpha term (framebuffer alpha * factor) */ switch (spu.blend.alpha_dst_factor) { case PIPE_BLENDFACTOR_ONE: - term2a = fragA; + term2a = fbRGBA[3]; break; case PIPE_BLENDFACTOR_SRC_COLOR: term2a = spu_splats(0.0f); @@ -322,6 +409,16 @@ spu_fallback_fragment_ops(uint x, uint y, tmp = spu_sub(one, fragA); term2a = spu_mul(fbRGBA[3], tmp); break; + case PIPE_BLENDFACTOR_DST_COLOR: + /* fall-through */ + case PIPE_BLENDFACTOR_DST_ALPHA: + term2a = spu_mul(fbRGBA[3], fbRGBA[3]); + break; + case PIPE_BLENDFACTOR_CONST_COLOR: + /* fall-through */ + case PIPE_BLENDFACTOR_CONST_ALPHA: + term2a = spu_mul(fbRGBA[3], spu_splats(spu.blend_color.color[3])); + break; /* XXX more cases */ default: ASSERT(0); @@ -341,7 +438,21 @@ spu_fallback_fragment_ops(uint x, uint y, fragG = spu_sub(term1g, term2g); fragB = spu_sub(term1b, term2b); break; - /* XXX more cases */ + case PIPE_BLEND_REVERSE_SUBTRACT: + fragR = spu_sub(term2r, term1r); + fragG = spu_sub(term2g, term1g); + fragB = spu_sub(term2b, term1b); + break; + case PIPE_BLEND_MIN: + fragR = spu_min(term1r, term2r); + fragG = spu_min(term1g, term2g); + fragB = spu_min(term1b, term2b); + break; + case PIPE_BLEND_MAX: + fragR = spu_max(term1r, term2r); + fragG = spu_max(term1g, term2g); + fragB = spu_max(term1b, term2b); + break; default: ASSERT(0); } @@ -356,7 +467,15 @@ spu_fallback_fragment_ops(uint x, uint y, case PIPE_BLEND_SUBTRACT: fragA = spu_sub(term1a, term2a); break; - /* XXX more cases */ + case PIPE_BLEND_REVERSE_SUBTRACT: + fragA = spu_sub(term2a, term1a); + break; + case PIPE_BLEND_MIN: + fragA = spu_min(term1a, term2a); + break; + case PIPE_BLEND_MAX: + fragA = spu_max(term1a, term2a); + break; default: ASSERT(0); } @@ -384,21 +503,20 @@ spu_fallback_fragment_ops(uint x, uint y, #endif /* - * Pack float colors into 32-bit RGBA words. + * Pack fragment float colors into 32-bit RGBA words. */ switch (spu.fb.color_format) { case PIPE_FORMAT_A8R8G8B8_UNORM: - c0 = spu_pack_A8R8G8B8(frag_aos[0]); - c1 = spu_pack_A8R8G8B8(frag_aos[1]); - c2 = spu_pack_A8R8G8B8(frag_aos[2]); - c3 = spu_pack_A8R8G8B8(frag_aos[3]); + fragc0 = spu_pack_A8R8G8B8(frag_aos[0]); + fragc1 = spu_pack_A8R8G8B8(frag_aos[1]); + fragc2 = spu_pack_A8R8G8B8(frag_aos[2]); + fragc3 = spu_pack_A8R8G8B8(frag_aos[3]); break; - case PIPE_FORMAT_B8G8R8A8_UNORM: - c0 = spu_pack_B8G8R8A8(frag_aos[0]); - c1 = spu_pack_B8G8R8A8(frag_aos[1]); - c2 = spu_pack_B8G8R8A8(frag_aos[2]); - c3 = spu_pack_B8G8R8A8(frag_aos[3]); + fragc0 = spu_pack_B8G8R8A8(frag_aos[0]); + fragc1 = spu_pack_B8G8R8A8(frag_aos[1]); + fragc2 = spu_pack_B8G8R8A8(frag_aos[2]); + fragc3 = spu_pack_B8G8R8A8(frag_aos[3]); break; default: fprintf(stderr, "SPU: Bad pixel format in spu_default_fragment_ops\n"); @@ -407,20 +525,57 @@ spu_fallback_fragment_ops(uint x, uint y, /* - * Color masking + * Do color masking */ if (spu.blend.colormask != 0xf) { - /* XXX to do */ - /* apply color mask to 32-bit packed colors */ + uint cmask = 0x0; /* each byte corresponds to a color channel */ + + /* Form bitmask depending on color buffer format and colormask bits */ + switch (spu.fb.color_format) { + case PIPE_FORMAT_A8R8G8B8_UNORM: + if (spu.blend.colormask & PIPE_MASK_R) + cmask |= 0x00ff0000; /* red */ + if (spu.blend.colormask & PIPE_MASK_G) + cmask |= 0x0000ff00; /* green */ + if (spu.blend.colormask & PIPE_MASK_B) + cmask |= 0x000000ff; /* blue */ + if (spu.blend.colormask & PIPE_MASK_A) + cmask |= 0xff000000; /* alpha */ + break; + case PIPE_FORMAT_B8G8R8A8_UNORM: + if (spu.blend.colormask & PIPE_MASK_R) + cmask |= 0x0000ff00; /* red */ + if (spu.blend.colormask & PIPE_MASK_G) + cmask |= 0x00ff0000; /* green */ + if (spu.blend.colormask & PIPE_MASK_B) + cmask |= 0xff000000; /* blue */ + if (spu.blend.colormask & PIPE_MASK_A) + cmask |= 0x000000ff; /* alpha */ + break; + default: + ASSERT(0); + } + + /* + * Apply color mask to the 32-bit packed colors. + * if (cmask[i]) + * frag color[i] = frag color[i]; + * else + * frag color[i] = framebuffer color[i]; + */ + fragc0 = (fragc0 & cmask) | (fbc0 & ~cmask); + fragc1 = (fragc1 & cmask) | (fbc1 & ~cmask); + fragc2 = (fragc2 & cmask) | (fbc2 & ~cmask); + fragc3 = (fragc3 & cmask) | (fbc3 & ~cmask); } /* - * Logic Ops + * Do logic ops */ if (spu.blend.logicop_enable) { /* XXX to do */ - /* apply logicop to 32-bit packed colors */ + /* apply logicop to 32-bit packed colors (fragcx and fbcx) */ } @@ -431,45 +586,46 @@ spu_fallback_fragment_ops(uint x, uint y, spu.cur_ctile_status = TILE_STATUS_DIRTY; } else { + /* write no fragments */ return; } /* - * Write new quad colors to the framebuffer/tile. + * Write new fragment/quad colors to the framebuffer/tile. * Only write pixels where the corresponding mask word is set. */ #if LINEAR_QUAD_LAYOUT /* * Quad layout: * +--+--+--+--+ - * |p0|p1|p2|p3| + * |p0|p1|p2|p3|... * +--+--+--+--+ */ if (spu_extract(mask, 0)) - colorTile->ui[y][x*2] = c0; + colorTile->ui[y][x*2] = fragc0; if (spu_extract(mask, 1)) - colorTile->ui[y][x*2+1] = c1; + colorTile->ui[y][x*2+1] = fragc1; if (spu_extract(mask, 2)) - colorTile->ui[y][x*2+2] = c2; + colorTile->ui[y][x*2+2] = fragc2; if (spu_extract(mask, 3)) - colorTile->ui[y][x*2+3] = c3; + colorTile->ui[y][x*2+3] = fragc3; #else /* * Quad layout: * +--+--+ - * |p0|p1| + * |p0|p1|... * +--+--+ - * |p2|p3| + * |p2|p3|... * +--+--+ */ if (spu_extract(mask, 0)) - colorTile->ui[y+0][x+0] = c0; + colorTile->ui[y+0][x+0] = fragc0; if (spu_extract(mask, 1)) - colorTile->ui[y+0][x+1] = c1; + colorTile->ui[y+0][x+1] = fragc1; if (spu_extract(mask, 2)) - colorTile->ui[y+1][x+0] = c2; + colorTile->ui[y+1][x+0] = fragc2; if (spu_extract(mask, 3)) - colorTile->ui[y+1][x+1] = c3; + colorTile->ui[y+1][x+1] = fragc3; #endif } diff --git a/src/gallium/drivers/cell/spu/spu_render.c b/src/gallium/drivers/cell/spu/spu_render.c index 305dc98881..7c225e2f27 100644 --- a/src/gallium/drivers/cell/spu/spu_render.c +++ b/src/gallium/drivers/cell/spu/spu_render.c @@ -98,7 +98,7 @@ my_tile(uint tx, uint ty) static INLINE void get_cz_tiles(uint tx, uint ty) { - if (spu.read_depth) { + if (spu.read_depth_stencil) { if (spu.cur_ztile_status != TILE_STATUS_CLEAR) { //printf("SPU %u: getting Z tile %u, %u\n", spu.init.id, tx, ty); get_tile(tx, ty, &spu.ztile, TAG_READ_TILE_Z, 1); @@ -153,7 +153,7 @@ static INLINE void wait_put_cz_tiles(void) { wait_on_mask(1 << TAG_WRITE_TILE_COLOR); - if (spu.read_depth) { + if (spu.read_depth_stencil) { wait_on_mask(1 << TAG_WRITE_TILE_Z); } } @@ -175,22 +175,14 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr) const ubyte *vertices; const ushort *indexes; uint i, j; + uint num_tiles; - - if (Debug) { - printf("SPU %u: RENDER prim %u, num_vert=%u num_ind=%u " - "inline_vert=%u\n", - spu.init.id, - render->prim_type, - render->num_verts, - render->num_indexes, - render->inline_verts); - - /* - printf(" bound: %g, %g .. %g, %g\n", - render->xmin, render->ymin, render->xmax, render->ymax); - */ - } + D_PRINTF(CELL_DEBUG_CMD, + "RENDER prim=%u num_vert=%u num_ind=%u inline_vert=%u\n", + render->prim_type, + render->num_verts, + render->num_indexes, + render->inline_verts); ASSERT(sizeof(*render) % 4 == 0); ASSERT(total_vertex_bytes % 16 == 0); @@ -251,6 +243,8 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr) wait_on_mask(1 << TAG_SURFACE_CLEAR); /* XXX temporary */ + num_tiles = 0; + /** ** loop over tiles, rendering tris **/ @@ -264,6 +258,8 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr) if (!my_tile(tx, ty)) continue; + num_tiles++; + spu.cur_ctile_status = spu.ctile_status[ty][tx]; spu.cur_ztile_status = spu.ztile_status[ty][tx]; @@ -293,9 +289,7 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr) spu.ztile_status[ty][tx] = spu.cur_ztile_status; } - if (Debug) - printf("SPU %u: RENDER done\n", - spu.init.id); + D_PRINTF(CELL_DEBUG_CMD, + "RENDER done (%u tiles hit)\n", + num_tiles); } - - diff --git a/src/gallium/drivers/cell/spu/spu_shuffle.h b/src/gallium/drivers/cell/spu/spu_shuffle.h new file mode 100644 index 0000000000..74f2a0b6d2 --- /dev/null +++ b/src/gallium/drivers/cell/spu/spu_shuffle.h @@ -0,0 +1,186 @@ +#ifndef SPU_SHUFFLE_H +#define SPU_SHUFFLE_H + +/* + * Generate shuffle patterns with minimal fuss. + * + * Based on ideas from + * http://www.insomniacgames.com/tech/articles/0408/files/shuffles.pdf + * + * A-P indicates 0-15th position in first vector + * a-p indicates 0-15th position in second vector + * + * +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ + * |00|01|02|03|04|05|06|07|08|09|0a|0b|0c|0d|0e|0f| + * +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ + * | A| B| C| D| + * +-----+-----+-----+-----+-----+-----+-----+-----+ + * | A| B| C| D| E| F| G| H| + * +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ + * | A| B| C| D| E| F| G| H| I| J| K| L| M| N| O| P| + * +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ + * + * x or X indicates 0xff + * 8 indicates 0x80 + * 0 indicates 0x00 + * + * The macros SHUFFLE4() SHUFFLE8() and SHUFFLE16() provide a const vector + * unsigned char literal suitable for use with spu_shuffle(). + * + * The macros SHUFB4() SHUFB8() and SHUFB16() provide a const qword vector + * literal suitable for use with si_shufb(). + * + * + * For example : + * SHUFB4(A,A,A,A) + * expands to : + * ((const qword){0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3}) + * + * SHUFFLE8(A,B,a,b,C,c,8,8) + * expands to : + * ((const vector unsigned char){0x00,0x01,0x02,0x03,0x10,0x11,0x12,0x13, + * 0x04,0x05,0x14,0x15,0xe0,0xe0,0xe0,0xe0}) + * + */ + +#include <spu_intrinsics.h> + +#define SHUFFLE_PATTERN_4_A__ 0x00, 0x01, 0x02, 0x03 +#define SHUFFLE_PATTERN_4_B__ 0x04, 0x05, 0x06, 0x07 +#define SHUFFLE_PATTERN_4_C__ 0x08, 0x09, 0x0a, 0x0b +#define SHUFFLE_PATTERN_4_D__ 0x0c, 0x0d, 0x0e, 0x0f +#define SHUFFLE_PATTERN_4_a__ 0x10, 0x11, 0x12, 0x13 +#define SHUFFLE_PATTERN_4_b__ 0x14, 0x15, 0x16, 0x17 +#define SHUFFLE_PATTERN_4_c__ 0x18, 0x19, 0x1a, 0x1b +#define SHUFFLE_PATTERN_4_d__ 0x1c, 0x1d, 0x1e, 0x1f +#define SHUFFLE_PATTERN_4_X__ 0xc0, 0xc0, 0xc0, 0xc0 +#define SHUFFLE_PATTERN_4_x__ 0xc0, 0xc0, 0xc0, 0xc0 +#define SHUFFLE_PATTERN_4_0__ 0x80, 0x80, 0x80, 0x80 +#define SHUFFLE_PATTERN_4_8__ 0xe0, 0xe0, 0xe0, 0xe0 + +#define SHUFFLE_VECTOR_4__(A, B, C, D) \ + SHUFFLE_PATTERN_4_##A##__, \ + SHUFFLE_PATTERN_4_##B##__, \ + SHUFFLE_PATTERN_4_##C##__, \ + SHUFFLE_PATTERN_4_##D##__ + +#define SHUFFLE4(A, B, C, D) \ + ((const vector unsigned char){ \ + SHUFFLE_VECTOR_4__(A, B, C, D) \ + }) + +#define SHUFB4(A, B, C, D) \ + ((const qword){ \ + SHUFFLE_VECTOR_4__(A, B, C, D) \ + }) + + +#define SHUFFLE_PATTERN_8_A__ 0x00, 0x01 +#define SHUFFLE_PATTERN_8_B__ 0x02, 0x03 +#define SHUFFLE_PATTERN_8_C__ 0x04, 0x05 +#define SHUFFLE_PATTERN_8_D__ 0x06, 0x07 +#define SHUFFLE_PATTERN_8_E__ 0x08, 0x09 +#define SHUFFLE_PATTERN_8_F__ 0x0a, 0x0b +#define SHUFFLE_PATTERN_8_G__ 0x0c, 0x0d +#define SHUFFLE_PATTERN_8_H__ 0x0e, 0x0f +#define SHUFFLE_PATTERN_8_a__ 0x10, 0x11 +#define SHUFFLE_PATTERN_8_b__ 0x12, 0x13 +#define SHUFFLE_PATTERN_8_c__ 0x14, 0x15 +#define SHUFFLE_PATTERN_8_d__ 0x16, 0x17 +#define SHUFFLE_PATTERN_8_e__ 0x18, 0x19 +#define SHUFFLE_PATTERN_8_f__ 0x1a, 0x1b +#define SHUFFLE_PATTERN_8_g__ 0x1c, 0x1d +#define SHUFFLE_PATTERN_8_h__ 0x1e, 0x1f +#define SHUFFLE_PATTERN_8_X__ 0xc0, 0xc0 +#define SHUFFLE_PATTERN_8_x__ 0xc0, 0xc0 +#define SHUFFLE_PATTERN_8_0__ 0x80, 0x80 +#define SHUFFLE_PATTERN_8_8__ 0xe0, 0xe0 + + +#define SHUFFLE_VECTOR_8__(A, B, C, D, E, F, G, H) \ + SHUFFLE_PATTERN_8_##A##__, \ + SHUFFLE_PATTERN_8_##B##__, \ + SHUFFLE_PATTERN_8_##C##__, \ + SHUFFLE_PATTERN_8_##D##__, \ + SHUFFLE_PATTERN_8_##E##__, \ + SHUFFLE_PATTERN_8_##F##__, \ + SHUFFLE_PATTERN_8_##G##__, \ + SHUFFLE_PATTERN_8_##H##__ + +#define SHUFFLE8(A, B, C, D, E, F, G, H) \ + ((const vector unsigned char){ \ + SHUFFLE_VECTOR_8__(A, B, C, D, E, F, G, H) \ + }) + +#define SHUFB8(A, B, C, D, E, F, G, H) \ + ((const qword){ \ + SHUFFLE_VECTOR_8__(A, B, C, D, E, F, G, H) \ + }) + + +#define SHUFFLE_PATTERN_16_A__ 0x00 +#define SHUFFLE_PATTERN_16_B__ 0x01 +#define SHUFFLE_PATTERN_16_C__ 0x02 +#define SHUFFLE_PATTERN_16_D__ 0x03 +#define SHUFFLE_PATTERN_16_E__ 0x04 +#define SHUFFLE_PATTERN_16_F__ 0x05 +#define SHUFFLE_PATTERN_16_G__ 0x06 +#define SHUFFLE_PATTERN_16_H__ 0x07 +#define SHUFFLE_PATTERN_16_I__ 0x08 +#define SHUFFLE_PATTERN_16_J__ 0x09 +#define SHUFFLE_PATTERN_16_K__ 0x0a +#define SHUFFLE_PATTERN_16_L__ 0x0b +#define SHUFFLE_PATTERN_16_M__ 0x0c +#define SHUFFLE_PATTERN_16_N__ 0x0d +#define SHUFFLE_PATTERN_16_O__ 0x0e +#define SHUFFLE_PATTERN_16_P__ 0x0f +#define SHUFFLE_PATTERN_16_a__ 0x10 +#define SHUFFLE_PATTERN_16_b__ 0x11 +#define SHUFFLE_PATTERN_16_c__ 0x12 +#define SHUFFLE_PATTERN_16_d__ 0x13 +#define SHUFFLE_PATTERN_16_e__ 0x14 +#define SHUFFLE_PATTERN_16_f__ 0x15 +#define SHUFFLE_PATTERN_16_g__ 0x16 +#define SHUFFLE_PATTERN_16_h__ 0x17 +#define SHUFFLE_PATTERN_16_i__ 0x18 +#define SHUFFLE_PATTERN_16_j__ 0x19 +#define SHUFFLE_PATTERN_16_k__ 0x1a +#define SHUFFLE_PATTERN_16_l__ 0x1b +#define SHUFFLE_PATTERN_16_m__ 0x1c +#define SHUFFLE_PATTERN_16_n__ 0x1d +#define SHUFFLE_PATTERN_16_o__ 0x1e +#define SHUFFLE_PATTERN_16_p__ 0x1f +#define SHUFFLE_PATTERN_16_X__ 0xc0 +#define SHUFFLE_PATTERN_16_x__ 0xc0 +#define SHUFFLE_PATTERN_16_0__ 0x80 +#define SHUFFLE_PATTERN_16_8__ 0xe0 + +#define SHUFFLE_VECTOR_16__(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \ + SHUFFLE_PATTERN_16_##A##__, \ + SHUFFLE_PATTERN_16_##B##__, \ + SHUFFLE_PATTERN_16_##C##__, \ + SHUFFLE_PATTERN_16_##D##__, \ + SHUFFLE_PATTERN_16_##E##__, \ + SHUFFLE_PATTERN_16_##F##__, \ + SHUFFLE_PATTERN_16_##G##__, \ + SHUFFLE_PATTERN_16_##H##__, \ + SHUFFLE_PATTERN_16_##I##__, \ + SHUFFLE_PATTERN_16_##J##__, \ + SHUFFLE_PATTERN_16_##K##__, \ + SHUFFLE_PATTERN_16_##L##__, \ + SHUFFLE_PATTERN_16_##M##__, \ + SHUFFLE_PATTERN_16_##N##__, \ + SHUFFLE_PATTERN_16_##O##__, \ + SHUFFLE_PATTERN_16_##P##__ + +#define SHUFFLE16(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \ + ((const vector unsigned char){ \ + SHUFFLE_VECTOR_16__(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \ + }) + +#define SHUFB16(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \ + ((const qword){ \ + SHUFFLE_VECTOR_16__(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \ + }) + +#endif diff --git a/src/gallium/drivers/cell/spu/spu_texture.c b/src/gallium/drivers/cell/spu/spu_texture.c index 117b8a36f8..69784c8978 100644 --- a/src/gallium/drivers/cell/spu/spu_texture.c +++ b/src/gallium/drivers/cell/spu/spu_texture.c @@ -26,6 +26,8 @@ **************************************************************************/ +#include <math.h> + #include "pipe/p_compiler.h" #include "spu_main.h" #include "spu_texture.h" @@ -40,37 +42,19 @@ void invalidate_tex_cache(void) { - uint unit = 0; - uint bytes = 4 * spu.texture[unit].width - * spu.texture[unit].height; - - spu_dcache_mark_dirty((unsigned) spu.texture[unit].start, bytes); -} + uint lvl; + for (lvl = 0; lvl < CELL_MAX_TEXTURE_LEVELS; lvl++) { + uint unit = 0; + uint bytes = 4 * spu.texture[unit].level[lvl].width + * spu.texture[unit].level[lvl].height; + if (spu.texture[unit].target == PIPE_TEXTURE_CUBE) + bytes *= 6; + else if (spu.texture[unit].target == PIPE_TEXTURE_3D) + bytes *= spu.texture[unit].level[lvl].depth; -/** - * XXX look into getting texels for all four pixels in a quad at once. - */ -static uint -get_texel(uint unit, vec_uint4 coordinate) -{ - /* - * XXX we could do the "/ TILE_SIZE" and "% TILE_SIZE" operations as - * SIMD since X and Y are already in a SIMD register. - */ - const unsigned texture_ea = (uintptr_t) spu.texture[unit].start; - ushort x = spu_extract(coordinate, 0); - ushort y = spu_extract(coordinate, 1); - unsigned tile_offset = sizeof(tile_t) - * ((y / TILE_SIZE * spu.texture[unit].tiles_per_row) + (x / TILE_SIZE)); - ushort texel_offset = (ushort) 4 - * (ushort) (((ushort) (y % TILE_SIZE) * (ushort) TILE_SIZE) + (x % TILE_SIZE)); - vec_uint4 tmp; - - spu_dcache_fetch_unaligned((qword *) & tmp, - texture_ea + tile_offset + texel_offset, - 4); - return spu_extract(tmp, 0); + spu_dcache_mark_dirty((unsigned) spu.texture[unit].level[lvl].start, bytes); + } } @@ -88,15 +72,17 @@ get_texel(uint unit, vec_uint4 coordinate) * a time. */ static void -get_four_texels(uint unit, vec_uint4 x, vec_uint4 y, vec_uint4 *texels) +get_four_texels(const struct spu_texture_level *tlevel, uint face, + vec_int4 x, vec_int4 y, + vec_uint4 *texels) { - const unsigned texture_ea = (uintptr_t) spu.texture[unit].start; - vec_uint4 tile_x = spu_rlmask(x, -5); - vec_uint4 tile_y = spu_rlmask(y, -5); - const qword offset_x = si_andi((qword) x, 0x1f); - const qword offset_y = si_andi((qword) y, 0x1f); + unsigned texture_ea = (uintptr_t) tlevel->start; + const vec_int4 tile_x = spu_rlmask(x, -5); /* tile_x = x / 32 */ + const vec_int4 tile_y = spu_rlmask(y, -5); /* tile_y = y / 32 */ + const qword offset_x = si_andi((qword) x, 0x1f); /* offset_x = x & 0x1f */ + const qword offset_y = si_andi((qword) y, 0x1f); /* offset_y = y & 0x1f */ - const qword tiles_per_row = (qword) spu_splats(spu.texture[unit].tiles_per_row); + const qword tiles_per_row = (qword) spu_splats(tlevel->tiles_per_row); const qword tile_size = (qword) spu_splats((unsigned) sizeof(tile_t)); qword tile_offset = si_mpya((qword) tile_y, tiles_per_row, (qword) tile_x); @@ -107,6 +93,8 @@ get_four_texels(uint unit, vec_uint4 x, vec_uint4 y, vec_uint4 *texels) vec_uint4 offset = (vec_uint4) si_a(tile_offset, texel_offset); + texture_ea = texture_ea + face * tlevel->bytes_per_image; + spu_dcache_fetch_unaligned((qword *) & texels[0], texture_ea + spu_extract(offset, 0), 4); spu_dcache_fetch_unaligned((qword *) & texels[1], @@ -118,83 +106,536 @@ get_four_texels(uint unit, vec_uint4 x, vec_uint4 y, vec_uint4 *texels) } +/** clamp vec to [0, max] */ +static INLINE vector signed int +spu_clamp(vector signed int vec, vector signed int max) +{ + static const vector signed int zero = {0,0,0,0}; + vector unsigned int c; + c = spu_cmpgt(vec, zero); /* c = vec > zero ? ~0 : 0 */ + vec = spu_sel(zero, vec, c); + c = spu_cmpgt(vec, max); /* c = vec > max ? ~0 : 0 */ + vec = spu_sel(vec, max, c); + return vec; +} + + + /** - * Get texture sample at texcoord. + * Do nearest texture sampling for four pixels. + * \param colors returned colors in SOA format (rrrr, gggg, bbbb, aaaa). */ -vector float -sample_texture_nearest(uint unit, vector float texcoord) +void +sample_texture_2d_nearest(vector float s, vector float t, + uint unit, uint level, uint face, + vector float colors[4]) { - vector float tc = spu_mul(texcoord, spu.texture[unit].tex_size); - vector unsigned int itc = spu_convtu(tc, 0); /* convert to int */ - itc = spu_and(itc, spu.texture[unit].tex_size_mask); /* mask (GL_REPEAT) */ - uint texel = get_texel(unit, itc); - return spu_unpack_A8R8G8B8(texel); + const struct spu_texture_level *tlevel = &spu.texture[unit].level[level]; + vector float ss = spu_mul(s, tlevel->scale_s); + vector float tt = spu_mul(t, tlevel->scale_t); + vector signed int is = spu_convts(ss, 0); + vector signed int it = spu_convts(tt, 0); + vec_uint4 texels[4]; + + /* PIPE_TEX_WRAP_REPEAT */ + is = spu_and(is, tlevel->mask_s); + it = spu_and(it, tlevel->mask_t); + + /* PIPE_TEX_WRAP_CLAMP */ + is = spu_clamp(is, tlevel->max_s); + it = spu_clamp(it, tlevel->max_t); + + get_four_texels(tlevel, face, is, it, texels); + + /* convert four packed ARGBA pixels to float RRRR,GGGG,BBBB,AAAA */ + spu_unpack_A8R8G8B8_transpose4(texels, colors); } -vector float -sample_texture_bilinear(uint unit, vector float texcoord) +/** + * Do bilinear texture sampling for four pixels. + * \param colors returned colors in SOA format (rrrr, gggg, bbbb, aaaa). + */ +void +sample_texture_2d_bilinear(vector float s, vector float t, + uint unit, uint level, uint face, + vector float colors[4]) { - static const vec_uint4 offset_x = {0, 0, 1, 1}; - static const vec_uint4 offset_y = {0, 1, 0, 1}; + const struct spu_texture_level *tlevel = &spu.texture[unit].level[level]; + static const vector float half = {-0.5f, -0.5f, -0.5f, -0.5f}; - vector float tc = spu_mul(texcoord, spu.texture[unit].tex_size); - tc = spu_add(tc, spu_splats(-0.5f)); /* half texel bias */ + vector float ss = spu_madd(s, tlevel->scale_s, half); + vector float tt = spu_madd(t, tlevel->scale_t, half); - /* integer texcoords S,T: */ - vec_uint4 itc = spu_convtu(tc, 0); /* convert to int */ + vector signed int is0 = spu_convts(ss, 0); + vector signed int it0 = spu_convts(tt, 0); - vec_uint4 texels[4]; - - /* setup texcoords for quad: - * +-----+-----+ - * |x0,y0|x1,y1| - * +-----+-----+ - * |x2,y2|x3,y3| - * +-----+-----+ - */ - vec_uint4 x = spu_splats(spu_extract(itc, 0)); - vec_uint4 y = spu_splats(spu_extract(itc, 1)); - x = spu_add(x, offset_x); - y = spu_add(y, offset_y); + /* is + 1, it + 1 */ + vector signed int is1 = spu_add(is0, 1); + vector signed int it1 = spu_add(it0, 1); - /* GL_REPEAT wrap mode: */ - x = spu_and(x, spu.texture[unit].tex_size_x_mask); - y = spu_and(y, spu.texture[unit].tex_size_y_mask); + /* PIPE_TEX_WRAP_REPEAT */ + is0 = spu_and(is0, tlevel->mask_s); + it0 = spu_and(it0, tlevel->mask_t); + is1 = spu_and(is1, tlevel->mask_s); + it1 = spu_and(it1, tlevel->mask_t); - get_four_texels(unit, x, y, texels); + /* PIPE_TEX_WRAP_CLAMP */ + is0 = spu_clamp(is0, tlevel->max_s); + it0 = spu_clamp(it0, tlevel->max_t); + is1 = spu_clamp(is1, tlevel->max_s); + it1 = spu_clamp(it1, tlevel->max_t); - /* integer A8R8G8B8 to float texel conversion */ - vector float texel00 = spu_unpack_A8R8G8B8(spu_extract(texels[0], 0)); - vector float texel01 = spu_unpack_A8R8G8B8(spu_extract(texels[1], 0)); - vector float texel10 = spu_unpack_A8R8G8B8(spu_extract(texels[2], 0)); - vector float texel11 = spu_unpack_A8R8G8B8(spu_extract(texels[3], 0)); + /* get packed int texels */ + vector unsigned int texels[16]; + get_four_texels(tlevel, face, is0, it0, texels + 0); /* upper-left */ + get_four_texels(tlevel, face, is1, it0, texels + 4); /* upper-right */ + get_four_texels(tlevel, face, is0, it1, texels + 8); /* lower-left */ + get_four_texels(tlevel, face, is1, it1, texels + 12); /* lower-right */ + /* convert packed int texels to float colors */ + vector float ftexels[16]; + spu_unpack_A8R8G8B8_transpose4(texels + 0, ftexels + 0); + spu_unpack_A8R8G8B8_transpose4(texels + 4, ftexels + 4); + spu_unpack_A8R8G8B8_transpose4(texels + 8, ftexels + 8); + spu_unpack_A8R8G8B8_transpose4(texels + 12, ftexels + 12); /* Compute weighting factors in [0,1] * Multiply texcoord by 1024, AND with 1023, convert back to float. */ - vector float tc1024 = spu_mul(tc, spu_splats(1024.0f)); - vector signed int itc1024 = spu_convts(tc1024, 0); - itc1024 = spu_and(itc1024, spu_splats((1 << 10) - 1)); - vector float weight = spu_convtf(itc1024, 10); - - /* smeared frac and 1-frac */ - vector float sfrac = spu_splats(spu_extract(weight, 0)); - vector float tfrac = spu_splats(spu_extract(weight, 1)); - vector float sfrac1 = spu_sub(spu_splats(1.0f), sfrac); - vector float tfrac1 = spu_sub(spu_splats(1.0f), tfrac); - - /* multiply the samples (colors) by the S/T weights */ - texel00 = spu_mul(spu_mul(texel00, sfrac1), tfrac1); - texel10 = spu_mul(spu_mul(texel10, sfrac ), tfrac1); - texel01 = spu_mul(spu_mul(texel01, sfrac1), tfrac ); - texel11 = spu_mul(spu_mul(texel11, sfrac ), tfrac ); - - /* compute sum of weighted samples */ - vector float texel_sum = spu_add(texel00, texel01); - texel_sum = spu_add(texel_sum, texel10); - texel_sum = spu_add(texel_sum, texel11); - - return texel_sum; + vector float ss1024 = spu_mul(ss, spu_splats(1024.0f)); + vector signed int iss1024 = spu_convts(ss1024, 0); + iss1024 = spu_and(iss1024, 1023); + vector float sWeights0 = spu_convtf(iss1024, 10); + + vector float tt1024 = spu_mul(tt, spu_splats(1024.0f)); + vector signed int itt1024 = spu_convts(tt1024, 0); + itt1024 = spu_and(itt1024, 1023); + vector float tWeights0 = spu_convtf(itt1024, 10); + + /* 1 - sWeight and 1 - tWeight */ + vector float sWeights1 = spu_sub(spu_splats(1.0f), sWeights0); + vector float tWeights1 = spu_sub(spu_splats(1.0f), tWeights0); + + /* reds, for four pixels */ + ftexels[ 0] = spu_mul(ftexels[ 0], spu_mul(sWeights1, tWeights1)); /*ul*/ + ftexels[ 4] = spu_mul(ftexels[ 4], spu_mul(sWeights0, tWeights1)); /*ur*/ + ftexels[ 8] = spu_mul(ftexels[ 8], spu_mul(sWeights1, tWeights0)); /*ll*/ + ftexels[12] = spu_mul(ftexels[12], spu_mul(sWeights0, tWeights0)); /*lr*/ + colors[0] = spu_add(spu_add(ftexels[0], ftexels[4]), + spu_add(ftexels[8], ftexels[12])); + + /* greens, for four pixels */ + ftexels[ 1] = spu_mul(ftexels[ 1], spu_mul(sWeights1, tWeights1)); /*ul*/ + ftexels[ 5] = spu_mul(ftexels[ 5], spu_mul(sWeights0, tWeights1)); /*ur*/ + ftexels[ 9] = spu_mul(ftexels[ 9], spu_mul(sWeights1, tWeights0)); /*ll*/ + ftexels[13] = spu_mul(ftexels[13], spu_mul(sWeights0, tWeights0)); /*lr*/ + colors[1] = spu_add(spu_add(ftexels[1], ftexels[5]), + spu_add(ftexels[9], ftexels[13])); + + /* blues, for four pixels */ + ftexels[ 2] = spu_mul(ftexels[ 2], spu_mul(sWeights1, tWeights1)); /*ul*/ + ftexels[ 6] = spu_mul(ftexels[ 6], spu_mul(sWeights0, tWeights1)); /*ur*/ + ftexels[10] = spu_mul(ftexels[10], spu_mul(sWeights1, tWeights0)); /*ll*/ + ftexels[14] = spu_mul(ftexels[14], spu_mul(sWeights0, tWeights0)); /*lr*/ + colors[2] = spu_add(spu_add(ftexels[2], ftexels[6]), + spu_add(ftexels[10], ftexels[14])); + + /* alphas, for four pixels */ + ftexels[ 3] = spu_mul(ftexels[ 3], spu_mul(sWeights1, tWeights1)); /*ul*/ + ftexels[ 7] = spu_mul(ftexels[ 7], spu_mul(sWeights0, tWeights1)); /*ur*/ + ftexels[11] = spu_mul(ftexels[11], spu_mul(sWeights1, tWeights0)); /*ll*/ + ftexels[15] = spu_mul(ftexels[15], spu_mul(sWeights0, tWeights0)); /*lr*/ + colors[3] = spu_add(spu_add(ftexels[3], ftexels[7]), + spu_add(ftexels[11], ftexels[15])); +} + + + +/** + * Adapted from /opt/cell/sdk/usr/spu/include/transpose_matrix4x4.h + */ +static INLINE void +transpose(vector unsigned int *mOut0, + vector unsigned int *mOut1, + vector unsigned int *mOut2, + vector unsigned int *mOut3, + vector unsigned int *mIn) +{ + vector unsigned int abcd, efgh, ijkl, mnop; /* input vectors */ + vector unsigned int aeim, bfjn, cgko, dhlp; /* output vectors */ + vector unsigned int aibj, ckdl, emfn, gohp; /* intermediate vectors */ + + vector unsigned char shufflehi = ((vector unsigned char) { + 0x00, 0x01, 0x02, 0x03, + 0x10, 0x11, 0x12, 0x13, + 0x04, 0x05, 0x06, 0x07, + 0x14, 0x15, 0x16, 0x17}); + vector unsigned char shufflelo = ((vector unsigned char) { + 0x08, 0x09, 0x0A, 0x0B, + 0x18, 0x19, 0x1A, 0x1B, + 0x0C, 0x0D, 0x0E, 0x0F, + 0x1C, 0x1D, 0x1E, 0x1F}); + abcd = *(mIn+0); + efgh = *(mIn+1); + ijkl = *(mIn+2); + mnop = *(mIn+3); + + aibj = spu_shuffle(abcd, ijkl, shufflehi); + ckdl = spu_shuffle(abcd, ijkl, shufflelo); + emfn = spu_shuffle(efgh, mnop, shufflehi); + gohp = spu_shuffle(efgh, mnop, shufflelo); + + aeim = spu_shuffle(aibj, emfn, shufflehi); + bfjn = spu_shuffle(aibj, emfn, shufflelo); + cgko = spu_shuffle(ckdl, gohp, shufflehi); + dhlp = spu_shuffle(ckdl, gohp, shufflelo); + + *mOut0 = aeim; + *mOut1 = bfjn; + *mOut2 = cgko; + *mOut3 = dhlp; +} + + +/** + * Bilinear filtering, using int instead of float arithmetic for computing + * sample weights. + */ +void +sample_texture_2d_bilinear_int(vector float s, vector float t, + uint unit, uint level, uint face, + vector float colors[4]) +{ + const struct spu_texture_level *tlevel = &spu.texture[unit].level[level]; + static const vector float half = {-0.5f, -0.5f, -0.5f, -0.5f}; + + /* Scale texcoords by size of texture, and add half pixel bias */ + vector float ss = spu_madd(s, tlevel->scale_s, half); + vector float tt = spu_madd(t, tlevel->scale_t, half); + + /* convert float coords to fixed-pt coords with 7 fraction bits */ + vector signed int is = spu_convts(ss, 7); /* XXX really need floor() here */ + vector signed int it = spu_convts(tt, 7); /* XXX really need floor() here */ + + /* compute integer texel weights in [0, 127] */ + vector signed int sWeights0 = spu_and(is, 127); + vector signed int tWeights0 = spu_and(it, 127); + vector signed int sWeights1 = spu_sub(127, sWeights0); + vector signed int tWeights1 = spu_sub(127, tWeights0); + + /* texel coords: is0 = is / 128, it0 = is / 128 */ + vector signed int is0 = spu_rlmask(is, -7); + vector signed int it0 = spu_rlmask(it, -7); + + /* texel coords: i1 = is0 + 1, it1 = it0 + 1 */ + vector signed int is1 = spu_add(is0, 1); + vector signed int it1 = spu_add(it0, 1); + + /* PIPE_TEX_WRAP_REPEAT */ + is0 = spu_and(is0, tlevel->mask_s); + it0 = spu_and(it0, tlevel->mask_t); + is1 = spu_and(is1, tlevel->mask_s); + it1 = spu_and(it1, tlevel->mask_t); + + /* PIPE_TEX_WRAP_CLAMP */ + is0 = spu_clamp(is0, tlevel->max_s); + it0 = spu_clamp(it0, tlevel->max_t); + is1 = spu_clamp(is1, tlevel->max_s); + it1 = spu_clamp(it1, tlevel->max_t); + + /* get packed int texels */ + vector unsigned int texels[16]; + get_four_texels(tlevel, face, is0, it0, texels + 0); /* upper-left */ + get_four_texels(tlevel, face, is1, it0, texels + 4); /* upper-right */ + get_four_texels(tlevel, face, is0, it1, texels + 8); /* lower-left */ + get_four_texels(tlevel, face, is1, it1, texels + 12); /* lower-right */ + + /* twiddle packed 32-bit BGRA pixels into RGBA as four unsigned ints */ + { + static const unsigned char ZERO = 0x80; + int i; + for (i = 0; i < 16; i++) { + texels[i] = spu_shuffle(texels[i], texels[i], + ((vector unsigned char) { + ZERO, ZERO, ZERO, 1, + ZERO, ZERO, ZERO, 2, + ZERO, ZERO, ZERO, 3, + ZERO, ZERO, ZERO, 0})); + } + } + + /* convert RGBA,RGBA,RGBA,RGBA to RRRR,GGGG,BBBB,AAAA */ + vector unsigned int texel0, texel1, texel2, texel3, texel4, texel5, texel6, texel7, + texel8, texel9, texel10, texel11, texel12, texel13, texel14, texel15; + transpose(&texel0, &texel1, &texel2, &texel3, texels + 0); + transpose(&texel4, &texel5, &texel6, &texel7, texels + 4); + transpose(&texel8, &texel9, &texel10, &texel11, texels + 8); + transpose(&texel12, &texel13, &texel14, &texel15, texels + 12); + + /* computed weighted colors */ + vector unsigned int c0, c1, c2, c3, cSum; + + /* red */ + c0 = (vector unsigned int) si_mpy((qword) texel0, si_mpy((qword) sWeights1, (qword) tWeights1)); /*ul*/ + c1 = (vector unsigned int) si_mpy((qword) texel4, si_mpy((qword) sWeights0, (qword) tWeights1)); /*ur*/ + c2 = (vector unsigned int) si_mpy((qword) texel8, si_mpy((qword) sWeights1, (qword) tWeights0)); /*ll*/ + c3 = (vector unsigned int) si_mpy((qword) texel12, si_mpy((qword) sWeights0, (qword) tWeights0)); /*lr*/ + cSum = spu_add(spu_add(c0, c1), spu_add(c2, c3)); + colors[0] = spu_convtf(cSum, 22); + + /* green */ + c0 = (vector unsigned int) si_mpy((qword) texel1, si_mpy((qword) sWeights1, (qword) tWeights1)); /*ul*/ + c1 = (vector unsigned int) si_mpy((qword) texel5, si_mpy((qword) sWeights0, (qword) tWeights1)); /*ur*/ + c2 = (vector unsigned int) si_mpy((qword) texel9, si_mpy((qword) sWeights1, (qword) tWeights0)); /*ll*/ + c3 = (vector unsigned int) si_mpy((qword) texel13, si_mpy((qword) sWeights0, (qword) tWeights0)); /*lr*/ + cSum = spu_add(spu_add(c0, c1), spu_add(c2, c3)); + colors[1] = spu_convtf(cSum, 22); + + /* blue */ + c0 = (vector unsigned int) si_mpy((qword) texel2, si_mpy((qword) sWeights1, (qword) tWeights1)); /*ul*/ + c1 = (vector unsigned int) si_mpy((qword) texel6, si_mpy((qword) sWeights0, (qword) tWeights1)); /*ur*/ + c2 = (vector unsigned int) si_mpy((qword) texel10, si_mpy((qword) sWeights1, (qword) tWeights0)); /*ll*/ + c3 = (vector unsigned int) si_mpy((qword) texel14, si_mpy((qword) sWeights0, (qword) tWeights0)); /*lr*/ + cSum = spu_add(spu_add(c0, c1), spu_add(c2, c3)); + colors[2] = spu_convtf(cSum, 22); + + /* alpha */ + c0 = (vector unsigned int) si_mpy((qword) texel3, si_mpy((qword) sWeights1, (qword) tWeights1)); /*ul*/ + c1 = (vector unsigned int) si_mpy((qword) texel7, si_mpy((qword) sWeights0, (qword) tWeights1)); /*ur*/ + c2 = (vector unsigned int) si_mpy((qword) texel11, si_mpy((qword) sWeights1, (qword) tWeights0)); /*ll*/ + c3 = (vector unsigned int) si_mpy((qword) texel15, si_mpy((qword) sWeights0, (qword) tWeights0)); /*lr*/ + cSum = spu_add(spu_add(c0, c1), spu_add(c2, c3)); + colors[3] = spu_convtf(cSum, 22); +} + + + +/** + * Compute level of detail factor from texcoords. + */ +static INLINE float +compute_lambda_2d(uint unit, vector float s, vector float t) +{ + uint baseLevel = 0; + float width = spu.texture[unit].level[baseLevel].width; + float height = spu.texture[unit].level[baseLevel].width; + float dsdx = width * (spu_extract(s, 1) - spu_extract(s, 0)); + float dsdy = width * (spu_extract(s, 2) - spu_extract(s, 0)); + float dtdx = height * (spu_extract(t, 1) - spu_extract(t, 0)); + float dtdy = height * (spu_extract(t, 2) - spu_extract(t, 0)); +#if 0 + /* ideal value */ + float x = dsdx * dsdx + dtdx * dtdx; + float y = dsdy * dsdy + dtdy * dtdy; + float rho = x > y ? x : y; + rho = sqrtf(rho); +#else + /* approximation */ + dsdx = fabsf(dsdx); + dsdy = fabsf(dsdy); + dtdx = fabsf(dtdx); + dtdy = fabsf(dtdy); + float rho = (dsdx + dsdy + dtdx + dtdy) * 0.5; +#endif + float lambda = logf(rho) * 1.442695f; /* compute logbase2(rho) */ + return lambda; +} + + +/** + * Blend two sets of colors according to weight. + */ +static void +blend_colors(vector float c0[4], const vector float c1[4], float weight) +{ + vector float t = spu_splats(weight); + vector float dc0 = spu_sub(c1[0], c0[0]); + vector float dc1 = spu_sub(c1[1], c0[1]); + vector float dc2 = spu_sub(c1[2], c0[2]); + vector float dc3 = spu_sub(c1[3], c0[3]); + c0[0] = spu_madd(dc0, t, c0[0]); + c0[1] = spu_madd(dc1, t, c0[1]); + c0[2] = spu_madd(dc2, t, c0[2]); + c0[3] = spu_madd(dc3, t, c0[3]); +} + + +/** + * Texture sampling with level of detail selection and possibly mipmap + * interpolation. + */ +void +sample_texture_2d_lod(vector float s, vector float t, + uint unit, uint level_ignored, uint face, + vector float colors[4]) +{ + /* + * Note that we're computing a lambda/lod here that's used for all + * four pixels in the quad. + */ + float lambda = compute_lambda_2d(unit, s, t); + + (void) face; + (void) level_ignored; + + /* apply lod bias */ + lambda += spu.sampler[unit].lod_bias; + + /* clamp */ + if (lambda < spu.sampler[unit].min_lod) + lambda = spu.sampler[unit].min_lod; + else if (lambda > spu.sampler[unit].max_lod) + lambda = spu.sampler[unit].max_lod; + + if (lambda <= 0.0f) { + /* magnify */ + spu.mag_sample_texture_2d[unit](s, t, unit, 0, face, colors); + } + else { + /* minify */ + if (spu.sampler[unit].min_img_filter == PIPE_TEX_FILTER_LINEAR) { + /* sample two mipmap levels and interpolate */ + int level = (int) lambda; + if (level > (int) spu.texture[unit].max_level) + level = spu.texture[unit].max_level; + spu.min_sample_texture_2d[unit](s, t, unit, level, face, colors); + if (spu.sampler[unit].min_img_filter == PIPE_TEX_FILTER_LINEAR) { + /* sample second mipmap level */ + float weight = lambda - (float) level; + level++; + if (level <= (int) spu.texture[unit].max_level) { + vector float colors2[4]; + spu.min_sample_texture_2d[unit](s, t, unit, level, face, colors2); + blend_colors(colors, colors2, weight); + } + } + } + else { + /* sample one mipmap level */ + int level = (int) (lambda + 0.5f); + if (level > (int) spu.texture[unit].max_level) + level = spu.texture[unit].max_level; + spu.min_sample_texture_2d[unit](s, t, unit, level, face, colors); + } + } +} + + +/** XXX need a SIMD version of this */ +static unsigned +choose_cube_face(float rx, float ry, float rz, float *newS, float *newT) +{ + /* + major axis + direction target sc tc ma + ---------- ------------------------------- --- --- --- + +rx TEXTURE_CUBE_MAP_POSITIVE_X_EXT -rz -ry rx + -rx TEXTURE_CUBE_MAP_NEGATIVE_X_EXT +rz -ry rx + +ry TEXTURE_CUBE_MAP_POSITIVE_Y_EXT +rx +rz ry + -ry TEXTURE_CUBE_MAP_NEGATIVE_Y_EXT +rx -rz ry + +rz TEXTURE_CUBE_MAP_POSITIVE_Z_EXT +rx -ry rz + -rz TEXTURE_CUBE_MAP_NEGATIVE_Z_EXT -rx -ry rz + */ + const float arx = fabsf(rx); + const float ary = fabsf(ry); + const float arz = fabsf(rz); + unsigned face; + float sc, tc, ma; + + if (arx > ary && arx > arz) { + if (rx >= 0.0F) { + face = PIPE_TEX_FACE_POS_X; + sc = -rz; + tc = -ry; + ma = arx; + } + else { + face = PIPE_TEX_FACE_NEG_X; + sc = rz; + tc = -ry; + ma = arx; + } + } + else if (ary > arx && ary > arz) { + if (ry >= 0.0F) { + face = PIPE_TEX_FACE_POS_Y; + sc = rx; + tc = rz; + ma = ary; + } + else { + face = PIPE_TEX_FACE_NEG_Y; + sc = rx; + tc = -rz; + ma = ary; + } + } + else { + if (rz > 0.0F) { + face = PIPE_TEX_FACE_POS_Z; + sc = rx; + tc = -ry; + ma = arz; + } + else { + face = PIPE_TEX_FACE_NEG_Z; + sc = -rx; + tc = -ry; + ma = arz; + } + } + + *newS = (sc / ma + 1.0F) * 0.5F; + *newT = (tc / ma + 1.0F) * 0.5F; + + return face; +} + + + +void +sample_texture_cube(vector float s, vector float t, vector float r, + uint unit, vector float colors[4]) +{ + uint p, faces[4], level = 0; + float newS[4], newT[4]; + + /* Compute cube faces referenced by the four sets of texcoords. + * XXX we should SIMD-ize this. + */ + for (p = 0; p < 4; p++) { + float rx = spu_extract(s, p); + float ry = spu_extract(t, p); + float rz = spu_extract(r, p); + faces[p] = choose_cube_face(rx, ry, rz, &newS[p], &newT[p]); + } + + if (faces[0] == faces[1] && + faces[0] == faces[2] && + faces[0] == faces[3]) { + /* GOOD! All four texcoords refer to the same cube face */ + s = (vector float) {newS[0], newS[1], newS[2], newS[3]}; + t = (vector float) {newT[0], newT[1], newT[2], newT[3]}; + spu.sample_texture_2d[unit](s, t, unit, level, faces[0], colors); + } + else { + /* BAD! The four texcoords refer to different faces */ + for (p = 0; p < 4; p++) { + vector float c[4]; + + spu.sample_texture_2d[unit](spu_splats(newS[p]), spu_splats(newT[p]), + unit, level, faces[p], c); + + float red = spu_extract(c[0], p); + float green = spu_extract(c[1], p); + float blue = spu_extract(c[2], p); + float alpha = spu_extract(c[3], p); + + colors[0] = spu_insert(red, colors[0], p); + colors[1] = spu_insert(green, colors[1], p); + colors[2] = spu_insert(blue, colors[2], p); + colors[3] = spu_insert(alpha, colors[3], p); + } + } } diff --git a/src/gallium/drivers/cell/spu/spu_texture.h b/src/gallium/drivers/cell/spu/spu_texture.h index f7c9738be8..7b75b007b5 100644 --- a/src/gallium/drivers/cell/spu/spu_texture.h +++ b/src/gallium/drivers/cell/spu/spu_texture.h @@ -36,12 +36,32 @@ extern void invalidate_tex_cache(void); -extern vector float -sample_texture_nearest(uint unit, vector float texcoord); +extern void +sample_texture_2d_nearest(vector float s, vector float t, + uint unit, uint level, uint face, + vector float colors[4]); + + +extern void +sample_texture_2d_bilinear(vector float s, vector float t, + uint unit, uint level, uint face, + vector float colors[4]); + +extern void +sample_texture_2d_bilinear_int(vector float s, vector float t, + uint unit, uint level, uint face, + vector float colors[4]); + +extern void +sample_texture_2d_lod(vector float s, vector float t, + uint unit, uint level, uint face, + vector float colors[4]); -extern vector float -sample_texture_bilinear(uint unit, vector float texcoord); + +extern void +sample_texture_cube(vector float s, vector float t, vector float r, + uint unit, vector float colors[4]); #endif /* SPU_TEXTURE_H */ diff --git a/src/gallium/drivers/cell/spu/spu_tile.c b/src/gallium/drivers/cell/spu/spu_tile.c index 216a33126b..6905015a48 100644 --- a/src/gallium/drivers/cell/spu/spu_tile.c +++ b/src/gallium/drivers/cell/spu/spu_tile.c @@ -87,3 +87,40 @@ put_tile(uint tx, uint ty, const tile_t *tile, int tag, int zBuf) 0 /* rid */); } + +/** + * For tiles whose status is TILE_STATUS_CLEAR, write solid-filled + * tiles back to the main framebuffer. + */ +void +really_clear_tiles(uint surfaceIndex) +{ + const uint num_tiles = spu.fb.width_tiles * spu.fb.height_tiles; + uint i; + + if (surfaceIndex == 0) { + clear_c_tile(&spu.ctile); + + for (i = spu.init.id; i < num_tiles; i += spu.init.num_spus) { + uint tx = i % spu.fb.width_tiles; + uint ty = i / spu.fb.width_tiles; + if (spu.ctile_status[ty][tx] == TILE_STATUS_CLEAR) { + put_tile(tx, ty, &spu.ctile, TAG_SURFACE_CLEAR, 0); + } + } + } + else { + clear_z_tile(&spu.ztile); + + for (i = spu.init.id; i < num_tiles; i += spu.init.num_spus) { + uint tx = i % spu.fb.width_tiles; + uint ty = i / spu.fb.width_tiles; + if (spu.ztile_status[ty][tx] == TILE_STATUS_CLEAR) + put_tile(tx, ty, &spu.ctile, TAG_SURFACE_CLEAR, 1); + } + } + +#if 0 + wait_on_mask(1 << TAG_SURFACE_CLEAR); +#endif +} diff --git a/src/gallium/drivers/cell/spu/spu_tile.h b/src/gallium/drivers/cell/spu/spu_tile.h index 1b5491112d..7bfb52be8f 100644 --- a/src/gallium/drivers/cell/spu/spu_tile.h +++ b/src/gallium/drivers/cell/spu/spu_tile.h @@ -36,12 +36,14 @@ -void +extern void get_tile(uint tx, uint ty, tile_t *tile, int tag, int zBuf); -void +extern void put_tile(uint tx, uint ty, const tile_t *tile, int tag, int zBuf); +extern void +really_clear_tiles(uint surfaceIndex); static INLINE void diff --git a/src/gallium/drivers/cell/spu/spu_tri.c b/src/gallium/drivers/cell/spu/spu_tri.c index 8b93878192..d727268475 100644 --- a/src/gallium/drivers/cell/spu/spu_tri.c +++ b/src/gallium/drivers/cell/spu/spu_tri.c @@ -29,12 +29,12 @@ * Triangle rendering within a tile. */ -#include <transpose_matrix4x4.h> #include "pipe/p_compiler.h" #include "pipe/p_format.h" #include "util/u_math.h" #include "spu_colorpack.h" #include "spu_main.h" +#include "spu_shuffle.h" #include "spu_texture.h" #include "spu_tile.h" #include "spu_tri.h" @@ -43,11 +43,6 @@ /** Masks are uint[4] vectors with each element being 0 or 0xffffffff */ typedef vector unsigned int mask_t; -typedef union -{ - vector float v; - float f[4]; -} float4; /** @@ -61,7 +56,7 @@ struct vertex_header { /* XXX fix this */ #undef CEILF -#define CEILF(X) ((float) (int) ((X) + 0.99999)) +#define CEILF(X) ((float) (int) ((X) + 0.99999f)) #define QUAD_TOP_LEFT 0 @@ -75,14 +70,25 @@ struct vertex_header { #define MASK_ALL 0xf +#define CHAN0 0 +#define CHAN1 1 +#define CHAN2 2 +#define CHAN3 3 + + #define DEBUG_VERTS 0 /** * Triangle edge info */ struct edge { - float dx; /**< X(v1) - X(v0), used only during setup */ - float dy; /**< Y(v1) - Y(v0), used only during setup */ + union { + struct { + float dx; /**< X(v1) - X(v0), used only during setup */ + float dy; /**< Y(v1) - Y(v0), used only during setup */ + }; + vec_float4 ds; /**< vector accessor for dx and dy */ + }; float dxdy; /**< dx/dy */ float sx, sy; /**< first sample point coord */ int lines; /**< number of lines on this edge */ @@ -91,9 +97,9 @@ struct edge { struct interp_coef { - float4 a0; - float4 dadx; - float4 dady; + vector float a0; + vector float dadx; + vector float dady; }; @@ -107,34 +113,32 @@ struct setup_stage { * turn. Currently fixed at 4 floats, but should change in time. * Codegen will help cope with this. */ - const struct vertex_header *vmax; - const struct vertex_header *vmid; - const struct vertex_header *vmin; - const struct vertex_header *vprovoke; + union { + struct { + const struct vertex_header *vmin; + const struct vertex_header *vmid; + const struct vertex_header *vmax; + const struct vertex_header *vprovoke; + }; + qword vertex_headers; + }; struct edge ebot; struct edge etop; struct edge emaj; - float oneoverarea; + float oneOverArea; /* XXX maybe make into vector? */ - uint tx, ty; + uint facing; + + uint tx, ty; /**< position of current tile (x, y) */ int cliprect_minx, cliprect_maxx, cliprect_miny, cliprect_maxy; -#if 0 - struct tgsi_interp_coef coef[PIPE_MAX_SHADER_INPUTS]; -#else struct interp_coef coef[PIPE_MAX_SHADER_INPUTS]; -#endif - -#if 0 - struct quad_header quad; -#endif struct { - int left[2]; /**< [0] = row0, [1] = row1 */ - int right[2]; + vec_int4 quad; /**< [0] = row0, [1] = row1; {left[0],left[1],right[0],right[1]} */ int y; unsigned y_flags; unsigned mask; /**< mask of MASK_BOTTOM/TOP_LEFT/RIGHT bits */ @@ -142,118 +146,103 @@ struct setup_stage { }; - static struct setup_stage setup; +static INLINE vector float +splatx(vector float v) +{ + return spu_splats(spu_extract(v, CHAN0)); +} - -#if 0 -/** - * Basically a cast wrapper. - */ -static INLINE struct setup_stage *setup_stage( struct draw_stage *stage ) +static INLINE vector float +splaty(vector float v) { - return (struct setup_stage *)stage; + return spu_splats(spu_extract(v, CHAN1)); } -#endif -#if 0 -/** - * Clip setup.quad against the scissor/surface bounds. - */ -static INLINE void -quad_clip(struct setup_stage *setup) +static INLINE vector float +splatz(vector float v) { - const struct pipe_scissor_state *cliprect = &setup.softpipe->cliprect; - const int minx = (int) cliprect->minx; - const int maxx = (int) cliprect->maxx; - const int miny = (int) cliprect->miny; - const int maxy = (int) cliprect->maxy; - - if (setup.quad.x0 >= maxx || - setup.quad.y0 >= maxy || - setup.quad.x0 + 1 < minx || - setup.quad.y0 + 1 < miny) { - /* totally clipped */ - setup.quad.mask = 0x0; - return; - } - if (setup.quad.x0 < minx) - setup.quad.mask &= (MASK_BOTTOM_RIGHT | MASK_TOP_RIGHT); - if (setup.quad.y0 < miny) - setup.quad.mask &= (MASK_BOTTOM_LEFT | MASK_BOTTOM_RIGHT); - if (setup.quad.x0 == maxx - 1) - setup.quad.mask &= (MASK_BOTTOM_LEFT | MASK_TOP_LEFT); - if (setup.quad.y0 == maxy - 1) - setup.quad.mask &= (MASK_TOP_LEFT | MASK_TOP_RIGHT); + return spu_splats(spu_extract(v, CHAN2)); } -#endif -#if 0 -/** - * Emit a quad (pass to next stage) with clipping. - */ -static INLINE void -clip_emit_quad(struct setup_stage *setup) +static INLINE vector float +splatw(vector float v) { - quad_clip(setup); - if (setup.quad.mask) { - struct softpipe_context *sp = setup.softpipe; - sp->quad.first->run(sp->quad.first, &setup.quad); - } + return spu_splats(spu_extract(v, CHAN3)); } -#endif + /** - * Evaluate attribute coefficients (plane equations) to compute - * attribute values for the four fragments in a quad. - * Eg: four colors will be computed (in AoS format). + * Setup fragment shader inputs by evaluating triangle's vertex + * attribute coefficient info. + * \param x quad x pos + * \param y quad y pos + * \param fragZ returns quad Z values + * \param fragInputs returns fragment program inputs + * Note: this code could be incorporated into the fragment program + * itself to avoid the loop and switch. */ -static INLINE void -eval_coeff(uint slot, float x, float y, vector float result[4]) +static void +eval_inputs(float x, float y, vector float *fragZ, vector float fragInputs[]) { - switch (spu.vertex_info.interp_mode[slot]) { - case INTERP_CONSTANT: - result[QUAD_TOP_LEFT] = - result[QUAD_TOP_RIGHT] = - result[QUAD_BOTTOM_LEFT] = - result[QUAD_BOTTOM_RIGHT] = setup.coef[slot].a0.v; - break; + static const vector float deltaX = (const vector float) {0, 1, 0, 1}; + static const vector float deltaY = (const vector float) {0, 0, 1, 1}; + + const uint posSlot = 0; + const vector float pos = setup.coef[posSlot].a0; + const vector float dposdx = setup.coef[posSlot].dadx; + const vector float dposdy = setup.coef[posSlot].dady; + const vector float fragX = spu_splats(x) + deltaX; + const vector float fragY = spu_splats(y) + deltaY; + vector float fragW, wInv; + uint i; - case INTERP_LINEAR: - /* fall-through, for now */ - default: - { - register vector float dadx = setup.coef[slot].dadx.v; - register vector float dady = setup.coef[slot].dady.v; - register vector float topLeft - = spu_add(setup.coef[slot].a0.v, - spu_add(spu_mul(spu_splats(x), dadx), - spu_mul(spu_splats(y), dady))); - - result[QUAD_TOP_LEFT] = topLeft; - result[QUAD_TOP_RIGHT] = spu_add(topLeft, dadx); - result[QUAD_BOTTOM_LEFT] = spu_add(topLeft, dady); - result[QUAD_BOTTOM_RIGHT] = spu_add(spu_add(topLeft, dadx), dady); + *fragZ = splatz(pos) + fragX * splatz(dposdx) + fragY * splatz(dposdy); + fragW = splatw(pos) + fragX * splatw(dposdx) + fragY * splatw(dposdy); + wInv = spu_re(fragW); /* 1 / w */ + + /* loop over fragment program inputs */ + for (i = 0; i < spu.vertex_info.num_attribs; i++) { + uint attr = i + 1; + enum interp_mode interp = spu.vertex_info.attrib[attr].interp_mode; + + /* constant term */ + vector float a0 = setup.coef[attr].a0; + vector float r0 = splatx(a0); + vector float r1 = splaty(a0); + vector float r2 = splatz(a0); + vector float r3 = splatw(a0); + + if (interp == INTERP_LINEAR || interp == INTERP_PERSPECTIVE) { + /* linear term */ + vector float dadx = setup.coef[attr].dadx; + vector float dady = setup.coef[attr].dady; + /* Use SPU intrinsics here to get slightly better code. + * originally: r0 += fragX * splatx(dadx) + fragY * splatx(dady); + */ + r0 = spu_madd(fragX, splatx(dadx), spu_madd(fragY, splatx(dady), r0)); + r1 = spu_madd(fragX, splaty(dadx), spu_madd(fragY, splaty(dady), r1)); + r2 = spu_madd(fragX, splatz(dadx), spu_madd(fragY, splatz(dady), r2)); + r3 = spu_madd(fragX, splatw(dadx), spu_madd(fragY, splatw(dady), r3)); + if (interp == INTERP_PERSPECTIVE) { + /* perspective term */ + r0 *= wInv; + r1 *= wInv; + r2 *= wInv; + r3 *= wInv; + } } + fragInputs[CHAN0] = r0; + fragInputs[CHAN1] = r1; + fragInputs[CHAN2] = r2; + fragInputs[CHAN3] = r3; + fragInputs += 4; } } -static INLINE vector float -eval_z(float x, float y) -{ - const uint slot = 0; - const float dzdx = setup.coef[slot].dadx.f[2]; - const float dzdy = setup.coef[slot].dady.f[2]; - const float topLeft = setup.coef[slot].a0.f[2] + x * dzdx + y * dzdy; - const vector float topLeftv = spu_splats(topLeft); - const vector float derivs = (vector float) { 0.0, dzdx, dzdy, dzdx + dzdy }; - return spu_add(topLeftv, derivs); -} - - /** * Emit a quad (pass to next stage). No clipping is done. * Note: about 1/5 to 1/7 of the time, mask is zero and this function @@ -261,120 +250,51 @@ eval_z(float x, float y) * overall. */ static INLINE void -emit_quad( int x, int y, mask_t mask ) +emit_quad( int x, int y, mask_t mask) { /* If any bits in mask are set... */ if (spu_extract(spu_orx(mask), 0)) { const int ix = x - setup.cliprect_minx; const int iy = y - setup.cliprect_miny; - vector float colors[4]; spu.cur_ctile_status = TILE_STATUS_DIRTY; spu.cur_ztile_status = TILE_STATUS_DIRTY; - if (spu.texture[0].start) { - /* texture mapping */ - const uint unit = 0; - vector float texcoords[4]; - eval_coeff(2, (float) x, (float) y, texcoords); - - if (spu_extract(mask, 0)) - colors[0] = spu.sample_texture[unit](unit, texcoords[0]); - if (spu_extract(mask, 1)) - colors[1] = spu.sample_texture[unit](unit, texcoords[1]); - if (spu_extract(mask, 2)) - colors[2] = spu.sample_texture[unit](unit, texcoords[2]); - if (spu_extract(mask, 3)) - colors[3] = spu.sample_texture[unit](unit, texcoords[3]); - - - if (spu.texture[1].start) { - /* multi-texture mapping */ - const uint unit = 1; - vector float colors1[4]; - - eval_coeff(2, (float) x, (float) y, texcoords); - - if (spu_extract(mask, 0)) - colors1[0] = spu.sample_texture[unit](unit, texcoords[0]); - if (spu_extract(mask, 1)) - colors1[1] = spu.sample_texture[unit](unit, texcoords[1]); - if (spu_extract(mask, 2)) - colors1[2] = spu.sample_texture[unit](unit, texcoords[2]); - if (spu_extract(mask, 3)) - colors1[3] = spu.sample_texture[unit](unit, texcoords[3]); - - /* hack: modulate first texture by second */ - colors[0] = spu_mul(colors[0], colors1[0]); - colors[1] = spu_mul(colors[1], colors1[1]); - colors[2] = spu_mul(colors[2], colors1[2]); - colors[3] = spu_mul(colors[3], colors1[3]); - } - - } - else { - /* simple shading */ -#if 0 - eval_coeff(1, (float) x, (float) y, colors); - -#else - /* XXX new fragment program code */ - - if (spu.fragment_program) { - vector float inputs[4*4], outputs[2*4]; - - /* setup inputs */ - eval_coeff(1, (float) x, (float) y, inputs); - - /* Execute the current fragment program */ - spu.fragment_program(inputs, outputs, spu.constants); - - /* Copy outputs */ - colors[0] = outputs[0*4+0]; - colors[1] = outputs[0*4+1]; - colors[2] = outputs[0*4+2]; - colors[3] = outputs[0*4+3]; - - if (0 && spu.init.id==0 && y == 48) { - printf("colors[0] = %f %f %f %f\n", - spu_extract(colors[0], 0), - spu_extract(colors[0], 1), - spu_extract(colors[0], 2), - spu_extract(colors[0], 3)); - printf("colors[1] = %f %f %f %f\n", - spu_extract(colors[1], 0), - spu_extract(colors[1], 1), - spu_extract(colors[1], 2), - spu_extract(colors[1], 3)); - } - - } -#endif - } - - { - /* Convert fragment data from AoS to SoA format. - * I.e. (RGBA,RGBA,RGBA,RGBA) -> (RRRR,GGGG,BBBB,AAAA) - * This is temporary! + /* + * Run fragment shader, execute per-fragment ops, update fb/tile. */ - vector float soa_frag[4]; - _transpose_matrix4x4(soa_frag, colors); + vector float inputs[4*4], outputs[2*4]; + vector unsigned int kill_mask; + vector float fragZ; + + eval_inputs((float) x, (float) y, &fragZ, inputs); - float4 fragZ; + ASSERT(spu.fragment_program); + ASSERT(spu.fragment_ops); - fragZ.v = eval_z((float) x, (float) y); + /* Execute the current fragment program */ + kill_mask = spu.fragment_program(inputs, outputs, spu.constants); - /* Do all per-fragment/quad operations here, including: - * alpha test, z test, stencil test, blend and framebuffer writing. + mask = spu_andc(mask, kill_mask); + + /* Execute per-fragment/quad operations, including: + * alpha test, z test, stencil test, blend and framebuffer writing. + * Note that there are two different fragment operations functions + * that can be called, one for front-facing fragments, and one + * for back-facing fragments. (Often the two are the same; + * but in some cases, like two-sided stenciling, they can be + * very different.) So choose the correct function depending + * on the calculated facing. */ - spu.fragment_ops(ix, iy, &spu.ctile, &spu.ztile, - fragZ.v, - soa_frag[0], soa_frag[1], - soa_frag[2], soa_frag[3], + spu.fragment_ops[setup.facing](ix, iy, &spu.ctile, &spu.ztile, + fragZ, + outputs[0*4+0], + outputs[0*4+1], + outputs[0*4+2], + outputs[0*4+3], mask); } - } } @@ -383,64 +303,49 @@ emit_quad( int x, int y, mask_t mask ) * Given an X or Y coordinate, return the block/quad coordinate that it * belongs to. */ -static INLINE int block( int x ) +static INLINE int +block(int x) { return x & ~1; } /** - * Compute mask which indicates which pixels in the 2x2 quad are actually inside - * the triangle's bounds. - * The mask is a uint4 vector and each element will be 0 or 0xffffffff. - */ -static INLINE mask_t calculate_mask( int x ) -{ - /* This is a little tricky. - * Use & instead of && to avoid branches. - * Use negation to convert true/false to ~0/0 values. - */ - mask_t mask; - mask = spu_insert(-((x >= setup.span.left[0]) & (x < setup.span.right[0])), mask, 0); - mask = spu_insert(-((x+1 >= setup.span.left[0]) & (x+1 < setup.span.right[0])), mask, 1); - mask = spu_insert(-((x >= setup.span.left[1]) & (x < setup.span.right[1])), mask, 2); - mask = spu_insert(-((x+1 >= setup.span.left[1]) & (x+1 < setup.span.right[1])), mask, 3); - return mask; -} - - -/** * Render a horizontal span of quads */ -static void flush_spans( void ) +static void +flush_spans(void) { int minleft, maxright; - int x; + + const int l0 = spu_extract(setup.span.quad, 0); + const int l1 = spu_extract(setup.span.quad, 1); + const int r0 = spu_extract(setup.span.quad, 2); + const int r1 = spu_extract(setup.span.quad, 3); switch (setup.span.y_flags) { case 0x3: /* both odd and even lines written (both quad rows) */ - minleft = MIN2(setup.span.left[0], setup.span.left[1]); - maxright = MAX2(setup.span.right[0], setup.span.right[1]); + minleft = MIN2(l0, l1); + maxright = MAX2(r0, r1); break; case 0x1: /* only even line written (quad top row) */ - minleft = setup.span.left[0]; - maxright = setup.span.right[0]; + minleft = l0; + maxright = r0; break; case 0x2: /* only odd line written (quad bottom row) */ - minleft = setup.span.left[1]; - maxright = setup.span.right[1]; + minleft = l1; + maxright = r1; break; default: return; } - /* OK, we're very likely to need the tile data now. * clear or finish waiting if needed. */ @@ -457,7 +362,7 @@ static void flush_spans( void ) } ASSERT(spu.cur_ctile_status != TILE_STATUS_DEFINED); - if (spu.read_depth) { + if (spu.read_depth_stencil) { if (spu.cur_ztile_status == TILE_STATUS_GETTING) { /* wait for mfc_get() to complete */ //printf("SPU: %u: waiting for ztile\n", spu.init.id); @@ -472,93 +377,119 @@ static void flush_spans( void ) ASSERT(spu.cur_ztile_status != TILE_STATUS_DEFINED); } - /* XXX this loop could be moved into the above switch cases and - * calculate_mask() could be simplified a bit... - */ - for (x = block(minleft); x <= block(maxright); x += 2) { -#if 1 - emit_quad( x, setup.span.y, calculate_mask( x ) ); -#endif + /* XXX this loop could be moved into the above switch cases... */ + + /* Setup for mask calculation */ + const vec_int4 quad_LlRr = setup.span.quad; + const vec_int4 quad_RrLl = spu_rlqwbyte(quad_LlRr, 8); + const vec_int4 quad_LLll = spu_shuffle(quad_LlRr, quad_LlRr, SHUFFLE4(A,A,B,B)); + const vec_int4 quad_RRrr = spu_shuffle(quad_RrLl, quad_RrLl, SHUFFLE4(A,A,B,B)); + + const vec_int4 twos = spu_splats(2); + + const int x = block(minleft); + vec_int4 xs = {x, x+1, x, x+1}; + + for (; spu_extract(xs, 0) <= block(maxright); xs += twos) { + /** + * Computes mask to indicate which pixels in the 2x2 quad are actually + * inside the triangle's bounds. + */ + + /* Calculate ({x,x+1,x,x+1} >= {l[0],l[0],l[1],l[1]}) */ + const mask_t gt_LLll_xs = spu_cmpgt(quad_LLll, xs); + const mask_t gte_xs_LLll = spu_nand(gt_LLll_xs, gt_LLll_xs); + + /* Calculate ({r[0],r[0],r[1],r[1]} > {x,x+1,x,x+1}) */ + const mask_t gt_RRrr_xs = spu_cmpgt(quad_RRrr, xs); + + /* Combine results to create mask */ + const mask_t mask = spu_and(gte_xs_LLll, gt_RRrr_xs); + + emit_quad(spu_extract(xs, 0), setup.span.y, mask); } setup.span.y = 0; setup.span.y_flags = 0; - setup.span.right[0] = 0; - setup.span.right[1] = 0; + /* Zero right elements */ + setup.span.quad = spu_shuffle(setup.span.quad, setup.span.quad, SHUFFLE4(A,B,0,0)); } + #if DEBUG_VERTS -static void print_vertex(const struct vertex_header *v) +static void +print_vertex(const struct vertex_header *v) { - int i; - fprintf(stderr, "Vertex: (%p)\n", v); - for (i = 0; i < setup.quad.nr_attrs; i++) { - fprintf(stderr, " %d: %f %f %f %f\n", i, - v->data[i][0], v->data[i][1], v->data[i][2], v->data[i][3]); + uint i; + fprintf(stderr, " Vertex: (%p)\n", v); + for (i = 0; i < spu.vertex_info.num_attribs; i++) { + fprintf(stderr, " %d: %f %f %f %f\n", i, + spu_extract(v->data[i], 0), + spu_extract(v->data[i], 1), + spu_extract(v->data[i], 2), + spu_extract(v->data[i], 3)); } } #endif -static boolean setup_sort_vertices(const struct vertex_header *v0, - const struct vertex_header *v1, - const struct vertex_header *v2) +/** + * Sort vertices from top to bottom. + * Compute area and determine front vs. back facing. + * Do coarse clip test against tile bounds + * \return FALSE if tri is totally outside tile, TRUE otherwise + */ +static boolean +setup_sort_vertices(const struct vertex_header *v0, + const struct vertex_header *v1, + const struct vertex_header *v2) { + float area, sign; #if DEBUG_VERTS - fprintf(stderr, "Triangle:\n"); - print_vertex(v0); - print_vertex(v1); - print_vertex(v2); + if (spu.init.id==0) { + fprintf(stderr, "SPU %u: Triangle:\n", spu.init.id); + print_vertex(v0); + print_vertex(v1); + print_vertex(v2); + } #endif - setup.vprovoke = v2; - /* determine bottom to top order of vertices */ { - float y0 = spu_extract(v0->data[0], 1); - float y1 = spu_extract(v1->data[0], 1); - float y2 = spu_extract(v2->data[0], 1); - if (y0 <= y1) { - if (y1 <= y2) { - /* y0<=y1<=y2 */ - setup.vmin = v0; - setup.vmid = v1; - setup.vmax = v2; - } - else if (y2 <= y0) { - /* y2<=y0<=y1 */ - setup.vmin = v2; - setup.vmid = v0; - setup.vmax = v1; - } - else { - /* y0<=y2<=y1 */ - setup.vmin = v0; - setup.vmid = v2; - setup.vmax = v1; - } - } - else { - if (y0 <= y2) { - /* y1<=y0<=y2 */ - setup.vmin = v1; - setup.vmid = v0; - setup.vmax = v2; - } - else if (y2 <= y1) { - /* y2<=y1<=y0 */ - setup.vmin = v2; - setup.vmid = v1; - setup.vmax = v0; - } - else { - /* y1<=y2<=y0 */ - setup.vmin = v1; - setup.vmid = v2; - setup.vmax = v0; - } - } + /* A table of shuffle patterns for putting vertex_header pointers into + correct order. Quite magical. */ + const vec_uchar16 sort_order_patterns[] = { + SHUFFLE4(A,B,C,C), + SHUFFLE4(C,A,B,C), + SHUFFLE4(A,C,B,C), + SHUFFLE4(B,C,A,C), + SHUFFLE4(B,A,C,C), + SHUFFLE4(C,B,A,C) }; + + /* The vertex_header pointers, packed for easy shuffling later */ + const vec_uint4 vs = {(unsigned)v0, (unsigned)v1, (unsigned)v2}; + + /* Collate y values into two vectors for comparison. + Using only one shuffle constant! ;) */ + const vec_float4 y_02_ = spu_shuffle(v0->data[0], v2->data[0], SHUFFLE4(0,B,b,C)); + const vec_float4 y_10_ = spu_shuffle(v1->data[0], v0->data[0], SHUFFLE4(0,B,b,C)); + const vec_float4 y_012 = spu_shuffle(y_02_, v1->data[0], SHUFFLE4(0,B,b,C)); + const vec_float4 y_120 = spu_shuffle(y_10_, v2->data[0], SHUFFLE4(0,B,b,C)); + + /* Perform comparison: {y0,y1,y2} > {y1,y2,y0} */ + const vec_uint4 compare = spu_cmpgt(y_012, y_120); + /* Compress the result of the comparison into 4 bits */ + const vec_uint4 gather = spu_gather(compare); + /* Subtract one to attain the index into the LUT. Magical. */ + const unsigned int index = spu_extract(gather, 0) - 1; + + /* Load the appropriate pattern and construct the desired vector. */ + setup.vertex_headers = (qword)spu_shuffle(vs, vs, sort_order_patterns[index]); + + /* Using the result of the comparison, set sign. + Very magical. */ + sign = ((si_to_uint(si_cntb((qword)gather)) == 2) ? 1.0f : -1.0f); } /* Check if triangle is completely outside the tile bounds */ @@ -575,41 +506,28 @@ static boolean setup_sort_vertices(const struct vertex_header *v0, spu_extract(setup.vmax->data[0], 0) > setup.cliprect_maxx) return FALSE; - setup.ebot.dx = spu_extract(setup.vmid->data[0], 0) - spu_extract(setup.vmin->data[0], 0); - setup.ebot.dy = spu_extract(setup.vmid->data[0], 1) - spu_extract(setup.vmin->data[0], 1); - setup.emaj.dx = spu_extract(setup.vmax->data[0], 0) - spu_extract(setup.vmin->data[0], 0); - setup.emaj.dy = spu_extract(setup.vmax->data[0], 1) - spu_extract(setup.vmin->data[0], 1); - setup.etop.dx = spu_extract(setup.vmax->data[0], 0) - spu_extract(setup.vmid->data[0], 0); - setup.etop.dy = spu_extract(setup.vmax->data[0], 1) - spu_extract(setup.vmid->data[0], 1); + setup.ebot.ds = spu_sub(setup.vmid->data[0], setup.vmin->data[0]); + setup.emaj.ds = spu_sub(setup.vmax->data[0], setup.vmin->data[0]); + setup.etop.ds = spu_sub(setup.vmax->data[0], setup.vmid->data[0]); /* * Compute triangle's area. Use 1/area to compute partial * derivatives of attributes later. - * - * The area will be the same as prim->det, but the sign may be - * different depending on how the vertices get sorted above. - * - * To determine whether the primitive is front or back facing we - * use the prim->det value because its sign is correct. */ - { - const float area = (setup.emaj.dx * setup.ebot.dy - - setup.ebot.dx * setup.emaj.dy); - - setup.oneoverarea = 1.0f / area; - /* - _mesa_printf("%s one-over-area %f area %f det %f\n", - __FUNCTION__, setup.oneoverarea, area, prim->det ); - */ - } + area = setup.emaj.dx * setup.ebot.dy - setup.ebot.dx * setup.emaj.dy; + + setup.oneOverArea = 1.0f / area; -#if 0 - /* We need to know if this is a front or back-facing triangle for: - * - the GLSL gl_FrontFacing fragment attribute (bool) - * - two-sided stencil test + /* The product of area * sign indicates front/back orientation (0/1). + * Just in case someone gets the bright idea of switching the front + * and back constants without noticing that we're assuming their + * values in this operation, also assert that the values are + * what we think they are. */ - setup.quad.facing = (prim->det > 0.0) ^ (setup.softpipe->rasterizer->front_winding == PIPE_WINDING_CW); -#endif + ASSERT(CELL_FACING_FRONT == 0); + ASSERT(CELL_FACING_BACK == 1); + setup.facing = (area * sign > 0.0f) + ^ (spu.rasterizer.front_winding == PIPE_WINDING_CW); return TRUE; } @@ -622,63 +540,11 @@ static boolean setup_sort_vertices(const struct vertex_header *v0, * \param slot which attribute slot */ static INLINE void -const_coeff(uint slot) +const_coeff4(uint slot) { - setup.coef[slot].dadx.v = (vector float) {0.0, 0.0, 0.0, 0.0}; - setup.coef[slot].dady.v = (vector float) {0.0, 0.0, 0.0, 0.0}; - setup.coef[slot].a0.v = setup.vprovoke->data[slot]; -} - - -/** - * Compute a0, dadx and dady for a linearly interpolated coefficient, - * for a triangle. - */ -static INLINE void -tri_linear_coeff(uint slot, uint firstComp, uint lastComp) -{ - uint i; - const float *vmin_d = (float *) &setup.vmin->data[slot]; - const float *vmid_d = (float *) &setup.vmid->data[slot]; - const float *vmax_d = (float *) &setup.vmax->data[slot]; - const float x = spu_extract(setup.vmin->data[0], 0) - 0.5f; - const float y = spu_extract(setup.vmin->data[0], 1) - 0.5f; - - for (i = firstComp; i < lastComp; i++) { - float botda = vmid_d[i] - vmin_d[i]; - float majda = vmax_d[i] - vmin_d[i]; - float a = setup.ebot.dy * majda - botda * setup.emaj.dy; - float b = setup.emaj.dx * botda - majda * setup.ebot.dx; - - ASSERT(slot < PIPE_MAX_SHADER_INPUTS); - - setup.coef[slot].dadx.f[i] = a * setup.oneoverarea; - setup.coef[slot].dady.f[i] = b * setup.oneoverarea; - - /* calculate a0 as the value which would be sampled for the - * fragment at (0,0), taking into account that we want to sample at - * pixel centers, in other words (0.5, 0.5). - * - * this is neat but unfortunately not a good way to do things for - * triangles with very large values of dadx or dady as it will - * result in the subtraction and re-addition from a0 of a very - * large number, which means we'll end up loosing a lot of the - * fractional bits and precision from a0. the way to fix this is - * to define a0 as the sample at a pixel center somewhere near vmin - * instead - i'll switch to this later. - */ - setup.coef[slot].a0.f[i] = (vmin_d[i] - - (setup.coef[slot].dadx.f[i] * x + - setup.coef[slot].dady.f[i] * y)); - } - - /* - _mesa_printf("attr[%d].%c: %f dx:%f dy:%f\n", - slot, "xyzw"[i], - setup.coef[slot].a0[i], - setup.coef[slot].dadx.f[i], - setup.coef[slot].dady.f[i]); - */ + setup.coef[slot].dadx = (vector float) {0.0, 0.0, 0.0, 0.0}; + setup.coef[slot].dady = (vector float) {0.0, 0.0, 0.0, 0.0}; + setup.coef[slot].a0 = setup.vprovoke->data[slot]; } @@ -702,18 +568,16 @@ tri_linear_coeff4(uint slot) vector float b = spu_sub(spu_mul(spu_splats(setup.emaj.dx), botda), spu_mul(majda, spu_splats(setup.ebot.dx))); - setup.coef[slot].dadx.v = spu_mul(a, spu_splats(setup.oneoverarea)); - setup.coef[slot].dady.v = spu_mul(b, spu_splats(setup.oneoverarea)); + setup.coef[slot].dadx = spu_mul(a, spu_splats(setup.oneOverArea)); + setup.coef[slot].dady = spu_mul(b, spu_splats(setup.oneOverArea)); - vector float tempx = spu_mul(setup.coef[slot].dadx.v, xxxx); - vector float tempy = spu_mul(setup.coef[slot].dady.v, yyyy); + vector float tempx = spu_mul(setup.coef[slot].dadx, xxxx); + vector float tempy = spu_mul(setup.coef[slot].dady, yyyy); - setup.coef[slot].a0.v = spu_sub(vmin_d, spu_add(tempx, tempy)); + setup.coef[slot].a0 = spu_sub(vmin_d, spu_add(tempx, tempy)); } - -#if 0 /** * Compute a0, dadx and dady for a perspective-corrected interpolant, * for a triangle. @@ -722,82 +586,76 @@ tri_linear_coeff4(uint slot) * Later, when we compute the value at a particular fragment position we'll * divide the interpolated value by the interpolated W at that fragment. */ -static void tri_persp_coeff( unsigned slot, - unsigned i ) +static void +tri_persp_coeff4(uint slot) { - /* premultiply by 1/w: - */ - float mina = setup.vmin->data[slot][i] * setup.vmin->data[0][3]; - float mida = setup.vmid->data[slot][i] * setup.vmid->data[0][3]; - float maxa = setup.vmax->data[slot][i] * setup.vmax->data[0][3]; - - float botda = mida - mina; - float majda = maxa - mina; - float a = setup.ebot.dy * majda - botda * setup.emaj.dy; - float b = setup.emaj.dx * botda - majda * setup.ebot.dx; - - /* - printf("tri persp %d,%d: %f %f %f\n", slot, i, - setup.vmin->data[slot][i], - setup.vmid->data[slot][i], - setup.vmax->data[slot][i] - ); - */ + const vector float xxxx = spu_splats(spu_extract(setup.vmin->data[0], 0) - 0.5f); + const vector float yyyy = spu_splats(spu_extract(setup.vmin->data[0], 1) - 0.5f); + + const vector float vmin_w = spu_splats(spu_extract(setup.vmin->data[0], 3)); + const vector float vmid_w = spu_splats(spu_extract(setup.vmid->data[0], 3)); + const vector float vmax_w = spu_splats(spu_extract(setup.vmax->data[0], 3)); + + vector float vmin_d = setup.vmin->data[slot]; + vector float vmid_d = setup.vmid->data[slot]; + vector float vmax_d = setup.vmax->data[slot]; + + vmin_d = spu_mul(vmin_d, vmin_w); + vmid_d = spu_mul(vmid_d, vmid_w); + vmax_d = spu_mul(vmax_d, vmax_w); + + vector float botda = vmid_d - vmin_d; + vector float majda = vmax_d - vmin_d; - assert(slot < PIPE_MAX_SHADER_INPUTS); - assert(i <= 3); + vector float a = spu_sub(spu_mul(spu_splats(setup.ebot.dy), majda), + spu_mul(botda, spu_splats(setup.emaj.dy))); + vector float b = spu_sub(spu_mul(spu_splats(setup.emaj.dx), botda), + spu_mul(majda, spu_splats(setup.ebot.dx))); - setup.coef[slot].dadx.f[i] = a * setup.oneoverarea; - setup.coef[slot].dady.f[i] = b * setup.oneoverarea; - setup.coef[slot].a0.f[i] = (mina - - (setup.coef[slot].dadx.f[i] * (setup.vmin->data[0][0] - 0.5f) + - setup.coef[slot].dady.f[i] * (setup.vmin->data[0][1] - 0.5f))); + setup.coef[slot].dadx = spu_mul(a, spu_splats(setup.oneOverArea)); + setup.coef[slot].dady = spu_mul(b, spu_splats(setup.oneOverArea)); + + vector float tempx = spu_mul(setup.coef[slot].dadx, xxxx); + vector float tempy = spu_mul(setup.coef[slot].dady, yyyy); + + setup.coef[slot].a0 = spu_sub(vmin_d, spu_add(tempx, tempy)); } -#endif + /** * Compute the setup.coef[] array dadx, dady, a0 values. * Must be called after setup.vmin,vmid,vmax,vprovoke are initialized. */ -static void setup_tri_coefficients(void) +static void +setup_tri_coefficients(void) { -#if 1 uint i; for (i = 0; i < spu.vertex_info.num_attribs; i++) { - switch (spu.vertex_info.interp_mode[i]) { + switch (spu.vertex_info.attrib[i].interp_mode) { case INTERP_NONE: break; - case INTERP_POS: - /*tri_linear_coeff(i, 2, 3);*/ - /* XXX interp W if PERSPECTIVE... */ - tri_linear_coeff4(i); - break; case INTERP_CONSTANT: - const_coeff(i); + const_coeff4(i); break; + case INTERP_POS: + /* fall-through */ case INTERP_LINEAR: tri_linear_coeff4(i); break; case INTERP_PERSPECTIVE: - tri_linear_coeff4(i); /* temporary */ + tri_persp_coeff4(i); break; default: ASSERT(0); } } -#else - ASSERT(spu.vertex_info.interp_mode[0] == INTERP_POS); - ASSERT(spu.vertex_info.interp_mode[1] == INTERP_LINEAR || - spu.vertex_info.interp_mode[1] == INTERP_CONSTANT); - tri_linear_coeff(0, 2, 3); /* slot 0, z */ - tri_linear_coeff(1, 0, 4); /* slot 1, color */ -#endif } -static void setup_tri_edges(void) +static void +setup_tri_edges(void) { float vmin_x = spu_extract(setup.vmin->data[0], 0) + 0.5f; float vmid_x = spu_extract(setup.vmid->data[0], 0) + 0.5f; @@ -827,9 +685,8 @@ static void setup_tri_edges(void) * Render the upper or lower half of a triangle. * Scissoring/cliprect is applied here too. */ -static void subtriangle( struct edge *eleft, - struct edge *eright, - unsigned lines ) +static void +subtriangle(struct edge *eleft, struct edge *eright, unsigned lines) { const int minx = setup.cliprect_minx; const int maxx = setup.cliprect_maxx; @@ -881,9 +738,11 @@ static void subtriangle( struct edge *eleft, setup.span.y = block(_y); } - setup.span.left[_y&1] = left; - setup.span.right[_y&1] = right; - setup.span.y_flags |= 1<<(_y&1); + int offset = _y&1; + vec_int4 quad_LlRr = {left, left, right, right}; + /* Store left and right in 0 or 1 row of quad based on offset */ + setup.span.quad = spu_sel(quad_LlRr, setup.span.quad, spu_maskw(5<<offset)); + setup.span.y_flags |= 1<<offset; } } @@ -902,7 +761,8 @@ static void subtriangle( struct edge *eleft, * The tile data should have already been fetched. */ boolean -tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty) +tri_draw(const float *v0, const float *v1, const float *v2, + uint tx, uint ty) { setup.tx = tx; setup.ty = ty; @@ -924,21 +784,16 @@ tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty) setup.span.y = 0; setup.span.y_flags = 0; - setup.span.right[0] = 0; - setup.span.right[1] = 0; - /* setup.span.z_mode = tri_z_mode( setup.ctx ); */ + /* Zero right elements */ + setup.span.quad = spu_shuffle(setup.span.quad, setup.span.quad, SHUFFLE4(A,B,0,0)); - /* init_constant_attribs( setup ); */ - - if (setup.oneoverarea < 0.0) { - /* emaj on left: - */ + if (setup.oneOverArea < 0.0) { + /* emaj on left */ subtriangle( &setup.emaj, &setup.ebot, setup.ebot.lines ); subtriangle( &setup.emaj, &setup.etop, setup.etop.lines ); } else { - /* emaj on right: - */ + /* emaj on right */ subtriangle( &setup.ebot, &setup.emaj, setup.ebot.lines ); subtriangle( &setup.etop, &setup.emaj, setup.etop.lines ); } diff --git a/src/gallium/drivers/cell/spu/spu_util.c b/src/gallium/drivers/cell/spu/spu_util.c index b8a0d4a265..af25dd3718 100644 --- a/src/gallium/drivers/cell/spu/spu_util.c +++ b/src/gallium/drivers/cell/spu/spu_util.c @@ -1,7 +1,7 @@ #include "cell/common.h" #include "pipe/p_shader_tokens.h" -#include "pipe/p_debug.h" +#include "util/u_debug.h" #include "tgsi/tgsi_parse.h" //#include "tgsi_build.h" #include "tgsi/tgsi_util.h" |