diff options
Diffstat (limited to 'src/gallium/drivers/cell/ppu')
24 files changed, 4047 insertions, 610 deletions
diff --git a/src/gallium/drivers/cell/ppu/Makefile b/src/gallium/drivers/cell/ppu/Makefile index b28f4c5c31..9358a47284 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 \ diff --git a/src/gallium/drivers/cell/ppu/cell_batch.c b/src/gallium/drivers/cell/ppu/cell_batch.c index 16882c0129..962775cd33 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 +90,37 @@ 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; + } + + ASSERT(size + sizeof(struct cell_command_fence) <= CELL_BUFFER_SIZE); + + fence_cmd = (struct cell_command_fence *) (cell->buffer[batch] + size); + fence_cmd->opcode = CELL_CMD_FENCE; + fence_cmd->fence = fence; + + /* update batch buffer size */ + cell->buffer_size[batch] = size + sizeof(struct cell_command_fence); + assert(sizeof(struct cell_command_fence) % 8 == 0); +} + + +/** * 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 +130,7 @@ 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 +138,14 @@ 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); @@ -139,6 +186,7 @@ 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; } @@ -169,7 +217,7 @@ cell_batch_append(struct cell_context *cell, const void *data, uint bytes) size = cell->buffer_size[cell->cur_batch]; - if (size + bytes > CELL_BUFFER_SIZE) { + if (bytes > cell_batch_free_space(cell)) { cell_batch_flush(cell); size = 0; } @@ -223,7 +271,7 @@ cell_batch_alloc_aligned(struct cell_context *cell, uint bytes, padbytes = (alignment - (size % alignment)) % alignment; - if (padbytes + size + bytes > CELL_BUFFER_SIZE) { + if (padbytes + bytes > cell_batch_free_space(cell)) { cell_batch_flush(cell); size = 0; } diff --git a/src/gallium/drivers/cell/ppu/cell_clear.c b/src/gallium/drivers/cell/ppu/cell_clear.c index c9c0c721bb..037635e466 100644 --- a/src/gallium/drivers/cell/ppu/cell_clear.c +++ b/src/gallium/drivers/cell/ppu/cell_clear.c @@ -106,4 +106,17 @@ cell_clear_surface(struct pipe_context *pipe, struct pipe_surface *ps, 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..22d552d8e3 100644 --- a/src/gallium/drivers/cell/ppu/cell_context.c +++ b/src/gallium/drivers/cell/ppu/cell_context.c @@ -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, 0); + 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..eb1397bb3f 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; @@ -135,6 +134,11 @@ 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 +153,9 @@ 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 +167,14 @@ 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_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..a64967b4b9 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 ); diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fp.c b/src/gallium/drivers/cell/ppu/cell_gen_fp.c index 6ffe94eb14..96a1743fc1 100644 --- a/src/gallium/drivers/cell/ppu/cell_gen_fp.c +++ b/src/gallium/drivers/cell/ppu/cell_gen_fp.c @@ -37,7 +37,7 @@ * \author Brian Paul */ - +#include <math.h> #include "pipe/p_defines.h" #include "pipe/p_state.h" #include "pipe/p_shader_tokens.h" @@ -51,25 +51,43 @@ #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} */ /** Per-instruction temps / intermediate temps */ int num_itemps; - int itemps[3]; + int itemps[12]; + + /** Current IF/ELSE/ENDIF nesting level */ + int if_nesting; + /** Index of 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 +130,78 @@ 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 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 pixel execution mask. + * The register is allocated an initialized upon the first call. + * + * The pixel 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); + + spe_indent(gen->f, 4); + spe_comment(gen->f, -4, "INIT EXEC MASK = ~0:"); + + /* exec_mask = {~0, ~0, ~0, ~0} */ + spe_load_int(gen->f, gen->exec_mask_reg, ~0); + + spe_indent(gen->f, -4); + } + + return gen->exec_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 +213,93 @@ 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 { + assert(swizzle < 4); - /* XXX need to examine src swizzle info here. - * That will involve changing the channel var... + switch (src->SrcRegister.File) { + case TGSI_FILE_TEMPORARY: + reg = gen->temp_regs[src->SrcRegister.Index][swizzle]; + break; + case TGSI_FILE_INPUT: + { + /* offset is measured in quadwords, not bytes */ + int offset = src->SrcRegister.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[src->SrcRegister.Index][swizzle]; + break; + case TGSI_FILE_CONSTANT: + { + /* offset is measured in quadwords, not bytes */ + int offset = src->SrcRegister.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); + } + } + + /* + * 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 +318,14 @@ 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) + reg = get_itemp(gen); + else + reg = gen->temp_regs[dest->DstRegister.Index][channel]; break; case TGSI_FILE_OUTPUT: reg = get_itemp(gen); @@ -211,19 +349,59 @@ 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) { + 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) { + 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: @@ -232,27 +410,114 @@ 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, -4, "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, -4, "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); +} + + static boolean emit_MOV(struct codegen *gen, const struct tgsi_full_instruction *inst) { - int ch; + int ch, src_reg[4], dst_reg[4]; + + spe_comment(gen->f, -4, "MOV:"); 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]); + src_reg[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + dst_reg[ch] = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + } + } + + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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 addition instructions. Recall that a single TGSI_OPCODE_ADD * becomes (up to) four SPU "fa" instructions because we're doing SOA @@ -261,24 +526,25 @@ emit_MOV(struct codegen *gen, const struct tgsi_full_instruction *inst) static boolean emit_ADD(struct codegen *gen, const struct tgsi_full_instruction *inst) { - int ch; - /* Loop over Red/Green/Blue/Alpha channels */ + int ch, s1_reg[4], s2_reg[4], d_reg[4]; + + spe_comment(gen->f, -4, "ADD:"); + /* Loop over Red/Green/Blue/Alpha channels, fetch src operands */ 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]); - + 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 add, store results */ + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { /* 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); -#endif - + spe_fa(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); /* Store the result (a no-op for TGSI_FILE_TEMPORARY dests) */ - store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); /* Free any intermediate temps we allocated */ free_itemps(gen); } @@ -286,6 +552,99 @@ emit_ADD(struct codegen *gen, const struct tgsi_full_instruction *inst) return true; } +/** + * Emit subtract. See emit_ADD for comments. + */ +static boolean +emit_SUB(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s1_reg[4], s2_reg[4], d_reg[4]; + spe_comment(gen->f, -4, "SUB:"); + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + /* d = s1 - s2 */ + spe_fs(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + 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]; + spe_comment(gen->f, -4, "MAD:"); + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + /* d = s1 * s2 + s3 */ + spe_fma(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch], s3_reg[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]; + spe_comment(gen->f, -4, "LERP:"); + /* setup/get src/dst/temp regs */ + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + spe_fs(gen->f, tmp_reg[ch], s2_reg[ch], s3_reg[ch]); + } + } + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + spe_fma(gen->f, d_reg[ch], tmp_reg[ch], s1_reg[ch], s3_reg[ch]); + } + } + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + } + free_itemps(gen); + return true; +} /** * Emit multiply. See emit_ADD for comments. @@ -293,17 +652,63 @@ emit_ADD(struct codegen *gen, const struct tgsi_full_instruction *inst) static boolean emit_MUL(struct codegen *gen, const struct tgsi_full_instruction *inst) { + int ch, s1_reg[4], s2_reg[4], d_reg[4]; + spe_comment(gen->f, -4, "MUL:"); + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + /* d = s1 * s2 */ + spe_fm(gen->f, d_reg[ch], s1_reg[ch], s2_reg[ch]); + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + return true; +} + +/** + * Emit reciprocal. See emit_ADD for comments. + */ +static boolean +emit_RCP(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ int ch; + spe_comment(gen->f, -4, "RCP:"); 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 + /* d = 1/s1 */ + spe_frest(gen->f, d_reg, s1_reg); + spe_fi(gen->f, d_reg, s1_reg, d_reg); + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + return true; +} + +/** + * Emit reciprocal sqrt. See emit_ADD for comments. + */ +static boolean +emit_RSQ(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + spe_comment(gen->f, -4, "RSQ:"); + 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 d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + /* d = 1/s1 */ + spe_frsqest(gen->f, d_reg, s1_reg); + spe_fi(gen->f, d_reg, s1_reg, d_reg); store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); free_itemps(gen); } @@ -311,6 +716,270 @@ emit_MUL(struct codegen *gen, const struct tgsi_full_instruction *inst) 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; + spe_comment(gen->f, -4, "ABS:"); + 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 d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + const int bit31mask_reg = get_itemp(gen); + + /* mask with bit 31 set, the rest cleared */ + spe_load_uint(gen->f, bit31mask_reg, (1 << 31)); + + /* d = sign bit cleared in s1 */ + spe_andc(gen->f, d_reg, s1_reg, bit31mask_reg); + + store_dest_reg(gen, d_reg, 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; + 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); + + spe_comment(gen->f, -4, "DP3:"); + + 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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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); + + spe_comment(gen->f, -4, "DP4:"); + + 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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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; + spe_comment(gen->f, -4, "DPH:"); + + 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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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); + + spe_comment(gen->f, -4, "NRM3:"); + + 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 (ch = 0; ch < 3; ch++) { /* NOTE: omit W channel */ + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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) +{ + spe_comment(gen->f, -4, "XPD:"); + + 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 set-if-greater-than. @@ -323,6 +992,8 @@ emit_SGT(struct codegen *gen, const struct tgsi_full_instruction *inst) { int ch; + spe_comment(gen->f, -4, "SGT:"); + 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]); @@ -331,26 +1002,776 @@ emit_SGT(struct codegen *gen, const struct tgsi_full_instruction *inst) /* 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 /* 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)); + + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + + return true; +} + +/** + * Emit set-if_less-then. See emit_SGT for comments. + */ +static boolean +emit_SLT(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + + spe_comment(gen->f, -4, "SLT:"); + + 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, s2_reg, s1_reg); + + /* 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)); + + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + + return true; +} + +/** + * Emit set-if_greater-then-or-equal. See emit_SGT for comments. + */ +static boolean +emit_SGE(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + + spe_comment(gen->f, -4, "SGE:"); + + 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, s2_reg, s1_reg); + + /* convert d from 0x0/0xffffffff to 0.0/1.0 */ + /* d = ~d & one_reg */ + spe_andc(gen->f, d_reg, get_const_one_reg(gen), d_reg); + + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + + return true; +} + +/** + * Emit set-if_less-then-or-equal. See emit_SGT for comments. + */ +static boolean +emit_SLE(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + + spe_comment(gen->f, -4, "SLE:"); + + 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); + + /* convert d from 0x0/0xffffffff to 0.0/1.0 */ + /* d = ~d & one_reg */ + spe_andc(gen->f, d_reg, get_const_one_reg(gen), d_reg); + + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + + return true; +} + +/** + * Emit set-if_equal. See emit_SGT for comments. + */ +static boolean +emit_SEQ(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + + spe_comment(gen->f, -4, "SEQ:"); + + 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_fceq(gen->f, d_reg, s1_reg, s2_reg); + + /* 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)); + + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + + return true; +} + +/** + * Emit set-if_not_equal. See emit_SGT for comments. + */ +static boolean +emit_SNE(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + + spe_comment(gen->f, -4, "SNE:"); + + 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_fceq(gen->f, d_reg, s1_reg, s2_reg); + spe_nor(gen->f, d_reg, d_reg, d_reg); + + /* 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)); + + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + + return true; +} + +/** + * Emit compare. See emit_SGT for comments. + */ +static boolean +emit_CMP(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + + spe_comment(gen->f, -4, "CMP:"); + + 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 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_xor(gen->f, zero_reg, zero_reg, 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; + + spe_comment(gen->f, -4, "TRUNC:"); + + 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 d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + + /* Convert float to int */ + spe_cflts(gen->f, d_reg, s1_reg, 0); + + /* Convert int to float */ + spe_csflt(gen->f, d_reg, d_reg, 0); + + store_dest_reg(gen, d_reg, 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; + + spe_comment(gen->f, -4, "FLR:"); + + int zero_reg = get_itemp(gen); + spe_xor(gen->f, zero_reg, zero_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 d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + int tmp_reg = get_itemp(gen); + + /* If negative, subtract 1.0 */ + spe_fcgt(gen->f, tmp_reg, zero_reg, s1_reg); + spe_selb(gen->f, tmp_reg, zero_reg, get_const_one_reg(gen), tmp_reg); + spe_fs(gen->f, tmp_reg, s1_reg, tmp_reg); + + /* Convert float to int */ + spe_cflts(gen->f, tmp_reg, tmp_reg, 0); + + /* Convert int to float */ + spe_csflt(gen->f, d_reg, tmp_reg, 0); + + store_dest_reg(gen, d_reg, 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; + + spe_comment(gen->f, -4, "FRC:"); + + int zero_reg = get_itemp(gen); + spe_xor(gen->f, zero_reg, zero_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 d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + int tmp_reg = get_itemp(gen); + + /* If negative, subtract 1.0 */ + spe_fcgt(gen->f, tmp_reg, zero_reg, s1_reg); + spe_selb(gen->f, tmp_reg, zero_reg, get_const_one_reg(gen), tmp_reg); + spe_fs(gen->f, tmp_reg, s1_reg, tmp_reg); + + /* Convert float to int */ + spe_cflts(gen->f, tmp_reg, tmp_reg, 0); + + /* Convert int to float */ + spe_csflt(gen->f, tmp_reg, tmp_reg, 0); + + /* d = s1 - FLR(s1) */ + spe_fs(gen->f, d_reg, s1_reg, tmp_reg); + + store_dest_reg(gen, d_reg, 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 + +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]; + } + } + assert(addr && "spu function not found"); + return addr / 4; /* discard 2 least significant bits */ +} + + +/** + * 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_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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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++) { + 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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + store_dest_reg(gen, d_regs[ch], ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + + return TRUE; +} + + +/** + * KILL if any of src reg values are less than zero. + */ +static boolean +emit_KIL(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + int s_regs[4], kil_reg = -1, cmp_reg, zero_reg; + + spe_comment(gen->f, -4, "CALL kil:"); + + /* zero = {0,0,0,0} */ + zero_reg = get_itemp(gen); + spe_load_uint(gen->f, zero_reg, 0); + + cmp_reg = get_itemp(gen); + + /* get src regs */ + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + s_regs[ch] = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + } + } + + /* test if any src regs are < 0 */ + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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) { + /* may have been a conditional kil */ + spe_and(gen->f, kil_reg, kil_reg, gen->exec_mask_reg); + } + + /* 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 max. See emit_SGT for comments. + */ +static boolean +emit_MAX(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s0_reg[4], s1_reg[4], d_reg[4], tmp_reg[4]; + + spe_comment(gen->f, -4, "MAX:"); + + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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 (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + spe_fcgt(gen->f, tmp_reg[ch], s0_reg[ch], s1_reg[ch]); + } + } + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + spe_selb(gen->f, d_reg[ch], s1_reg[ch], s0_reg[ch], tmp_reg[ch]); + } + } + + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + } + + free_itemps(gen); + return true; +} + +/** + * Emit max. See emit_SGT for comments. + */ +static boolean +emit_MIN(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch, s0_reg[4], s1_reg[4], d_reg[4], tmp_reg[4]; + + spe_comment(gen->f, -4, "MIN:"); + + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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 = (s1 > s0) ? s0 : s1 */ + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + spe_fcgt(gen->f, tmp_reg[ch], s1_reg[ch], s0_reg[ch]); + } + } + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + spe_selb(gen->f, d_reg[ch], s1_reg[ch], s0_reg[ch], tmp_reg[ch]); + } + } + + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + store_dest_reg(gen, d_reg[ch], ch, &inst->FullDstRegisters[0]); + } + } + + free_itemps(gen); + return true; +} + +static boolean +emit_IF(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + const int channel = 0; + const int exec_reg = get_exec_mask_reg(gen); + + spe_comment(gen->f, -4, "IF:"); + + /* update 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); + /* exec_mask = exec_mask & tmp */ + spe_and(gen->f, exec_reg, exec_reg, tmp_reg); + + gen->if_nesting++; + + free_itemps(gen); + + return true; +} + + +static boolean +emit_ELSE(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + const int exec_reg = get_exec_mask_reg(gen); + + spe_comment(gen->f, -4, "ELSE:"); + + /* exec_mask = !exec_mask */ + spe_complement(gen->f, exec_reg, exec_reg); + + return true; +} + + +static boolean +emit_ENDIF(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + const int exec_reg = get_exec_mask_reg(gen); + + spe_comment(gen->f, -4, "ENDIF:"); + + /* XXX todo: pop execution mask */ + + spe_load_int(gen->f, exec_reg, ~0x0); + + gen->if_nesting--; + return true; +} + + +static boolean +emit_DDX_DDY(struct codegen *gen, const struct tgsi_full_instruction *inst, + boolean ddx) +{ + int ch; + + spe_comment(gen->f, -4, ddx ? "DDX:" : "DDY:"); + + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << 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; +} + + + + /** * Emit END instruction. * We just return from the shader function at this point. @@ -361,11 +1782,8 @@ 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 + spe_comment(gen->f, -4, "END:"); + emit_epilogue(gen); return true; } @@ -379,19 +1797,101 @@ emit_instruction(struct codegen *gen, { switch (inst->Instruction.Opcode) { 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: + return emit_SUB(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: + return emit_RCP(gen, inst); + case TGSI_OPCODE_RSQ: + return emit_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: + return emit_SLT(gen, inst); + case TGSI_OPCODE_SGE: + return emit_SGE(gen, inst); + case TGSI_OPCODE_SLE: + return emit_SLE(gen, inst); + case TGSI_OPCODE_SEQ: + return emit_SEQ(gen, inst); + case TGSI_OPCODE_SNE: + return emit_SNE(gen, inst); + case TGSI_OPCODE_CMP: + return emit_CMP(gen, inst); + case TGSI_OPCODE_MAX: + return emit_MAX(gen, inst); + case TGSI_OPCODE_MIN: + return emit_MIN(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_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: + fprintf(stderr, "Cell: unimplemented TGSI instruction %d!\n", + inst->Instruction.Opcode); return false; } @@ -401,48 +1901,93 @@ emit_instruction(struct codegen *gen, /** + * 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); + + spe_comment(gen->f, -4, "IMMEDIATE:"); + + 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 { + int reg = spe_allocate_available_register(gen->f); + + if (reg < 0) + return false; + + /* 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; +} + + + +/** * Emit "code" for a TGSI declaration. * 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, -4, 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()). @@ -460,6 +2005,7 @@ cell_gen_fragment_program(struct cell_context *cell, struct codegen gen; 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 +2018,50 @@ 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, 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 (!emit_immediate(&gen, &parse.FullToken.FullImmediate)) + gen.error = true; break; case TGSI_TOKEN_TYPE_DECLARATION: - emit_declaration(&gen, &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 )) { + 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..2c64eb1bcc 100644 --- a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c @@ -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; } @@ -229,7 +237,40 @@ gen_alpha_test(const struct pipe_depth_stencil_alpha_state *dsa, spe_release_register(f, amask_reg); } +/* 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, boolean *is_already_set, unsigned int *r) +{ + if (*is_already_set) return; + *r = spe_allocate_available_register(f); + *is_already_set = true; +} + +static inline void +release_optional_register(struct spe_function *f, boolean *is_already_set, unsigned int r) +{ + if (!*is_already_set) return; + spe_release_register(f, r); + *is_already_set = false; +} + +static inline void +setup_const_register(struct spe_function *f, boolean *is_already_set, unsigned int *r, float value) +{ + if (*is_already_set) return; + setup_optional_register(f, is_already_set, r); + spe_load_float(f, *r, value); +} +static inline void +release_const_register(struct spe_function *f, boolean *is_already_set, unsigned int r) +{ + release_optional_register(f, is_already_set, r); +} /** * Generate SPE code to implement the given blend mode for a quad of pixels. @@ -242,6 +283,7 @@ 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, @@ -262,10 +304,17 @@ 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 */ + /* 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. + */ + boolean one_reg_set = false; + unsigned int one_reg; + boolean constR_reg_set = false, constG_reg_set = false, + constB_reg_set = false, constA_reg_set = false; + unsigned int constR_reg, constG_reg, constB_reg, constA_reg; ASSERT(blend->blend_enable); @@ -346,127 +395,449 @@ gen_blend(const struct pipe_blend_state *blend, spe_release_register(f, mask_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_set, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg_set, &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_set, &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_set, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg_set, &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_set, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg_set, &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_set, &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_set, &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_set, &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_set, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg_set, &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_set, &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_set, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg_set, &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_set, &constR_reg, blend_color->color[0]); + setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]); + setup_const_register(f, &constB_reg_set, &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_set, &constA_reg, blend_color->color[3]); + /* factor = Ac, so term = Afb*Ac */ + spe_fm(f, term2A_reg, fbA_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_set, &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; - /* XXX more cases */ + + /* 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 +850,21 @@ 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 +879,15 @@ 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 +907,14 @@ 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_set, one_reg); + release_const_register(f, &constR_reg_set, constR_reg); + release_const_register(f, &constG_reg_set, constG_reg); + release_const_register(f, &constB_reg_set, constB_reg); + release_const_register(f, &constA_reg_set, constA_reg); } @@ -524,24 +923,74 @@ 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 +1006,16 @@ 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 +1047,713 @@ 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... + */ + unsigned int r_mask; + unsigned int g_mask; + unsigned int b_mask; + unsigned int 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, + unsigned int stencil_max_value, + unsigned int fragment_mask_reg, unsigned int fbS_reg, + unsigned 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->value_mask == 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)) */ + unsigned int tmp_masked_stencil = spe_allocate_available_register(f); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->value_mask); + spe_compare_equal_uint(f, stencil_pass_reg, tmp_masked_stencil, state->value_mask & 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->value_mask == 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)) */ + unsigned int tmp_masked_stencil = spe_allocate_available_register(f); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->value_mask); + spe_compare_equal_uint(f, stencil_pass_reg, tmp_masked_stencil, state->value_mask & 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->value_mask == 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)) */ + unsigned int tmp_masked_stencil = spe_allocate_available_register(f); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->value_mask); + spe_compare_greater_uint(f, stencil_pass_reg, tmp_masked_stencil, state->value_mask & 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->value_mask == 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. + */ + unsigned 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)) */ + unsigned int tmp_reg = spe_allocate_available_register(f); + unsigned int tmp_masked_stencil = spe_allocate_available_register(f); + spe_load_uint(f, tmp_reg, state->value_mask & state->ref_value); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->value_mask); + 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->value_mask == 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)) */ + unsigned int tmp_masked_stencil = spe_allocate_available_register(f); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->value_mask); + spe_compare_greater_uint(f, stencil_pass_reg, tmp_masked_stencil, state->value_mask & 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->value_mask == stencil_max_value) { + /* stencil_pass = fragment_mask & (reference <= s) ] + * = fragment_mask & ~(reference > s) */ + /* As above, we have to do this by loading a register */ + unsigned 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)) */ + unsigned int tmp_reg = spe_allocate_available_register(f); + unsigned int tmp_masked_stencil = spe_allocate_available_register(f); + spe_load_uint(f, tmp_reg, state->ref_value & state->value_mask); + spe_and_uint(f, tmp_masked_stencil, fbS_reg, state->value_mask); + 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, unsigned int stencil_op, + unsigned int stencil_ref_value, unsigned int stencil_max_value, + unsigned int fbS_reg, unsigned 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) */ + unsigned 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) */ + unsigned 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 unsigned int depth_enabled, + unsigned int fbS_reg, + unsigned int *fail_reg, unsigned int *zfail_reg, + unsigned int *zpass_reg) +{ + unsigned 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. + */ + unsigned int stencil_pass_reg, stencil_fail_reg; + unsigned int stencil_fail_values, stencil_pass_depth_fail_values, stencil_pass_depth_pass_values; + unsigned int stencil_writemask_reg; + unsigned int zmask_reg; + unsigned 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->write_mask == 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->write_mask == 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].write_mask); + } + + /* 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. + */ + unsigned int 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) { + unsigned int 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; +} /** @@ -621,14 +1774,19 @@ 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. */ @@ -643,6 +1801,8 @@ 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!!!! */ @@ -651,7 +1811,12 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) 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); @@ -674,8 +1839,9 @@ 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 +1850,196 @@ 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); } + /* If we need the stencil buffers (because one- or two-sided stencil is + * enabled) or the depth buffer (because the depth test is enabled), + * go grab them. Note that if either one- or two-sided stencil is + * enabled, dsa->stencil[0].enabled will be true. + */ 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 */ + /* We may or may not need to allocate a register for Z or stencil values */ + boolean fbS_reg_set = false, fbZ_reg_set = false; + unsigned int fbS_reg, fbZ_reg = 0; + + spe_comment(f, 0, "Fetching Z/stencil quad from tile"); /* fetch quad of depth/stencil values from tile at (x,y) */ /* Load: fbZS_reg = memory[depth_tile_reg + offset_reg] */ + /* XXX Not sure this is allowed if we've only got a 16-bit Z buffer... */ 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); - } + /* From the Z/stencil buffer format, pull out the bits we need for + * Z and/or stencil. We'll also convert the incoming fragment Z + * value in fragZ_reg from a floating point value in [0.0..1.0] to + * an unsigned integer value with the appropriate resolution. + * Note that even if depth or stencil is *not* enabled, if it's + * present in the buffer, we pull it out and put it back later; + * otherwise, we can inadvertently destroy the contents of + * buffers we're not supposed to touch (e.g., if the user is + * clearing the depth buffer but not the stencil buffer, a + * quad of constant depth is drawn over the surface; the stencil + * buffer must be maintained). + */ + switch(zs_format) { - /* 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); + case PIPE_FORMAT_S8Z24_UNORM: /* fall through */ + case PIPE_FORMAT_X8Z24_UNORM: + /* Pull out both Z and stencil */ + setup_optional_register(f, &fbZ_reg_set, &fbZ_reg); + setup_optional_register(f, &fbS_reg_set, &fbS_reg); - /* scale_reg[0,1,2,3] = float(2^24-1) */ - spe_load_float(f, scale_reg, (float) 0xffffff); + /* four 24-bit Z values in the low-order bits */ + spe_and_uint(f, fbZ_reg, fbZS_reg, 0x00ffffff); - /* 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 */ + /* Incoming fragZ_reg value is a float in 0.0...1.0; convert + * to a 24-bit unsigned integer + */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + spe_rotmi(f, fragZ_reg, fragZ_reg, -8); - spe_release_register(f, scale_reg); - } - else { - /* XXX handle 16-bit Z format */ - ASSERT(0); - } - } + /* four 8-bit stencil values in the high-order bits */ + spe_rotmi(f, fbS_reg, fbZS_reg, -24); + break; - 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); - } - } + case PIPE_FORMAT_Z24S8_UNORM: /* fall through */ + case PIPE_FORMAT_Z24X8_UNORM: + setup_optional_register(f, &fbZ_reg_set, &fbZ_reg); + setup_optional_register(f, &fbS_reg_set, &fbS_reg); + + /* shift by 8 to get the upper 24-bit values */ + spe_rotmi(f, fbS_reg, fbZS_reg, -8); + + /* Incoming fragZ_reg value is a float in 0.0...1.0; convert + * to a 24-bit unsigned integer + */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + spe_rotmi(f, fragZ_reg, fragZ_reg, -8); + + /* 8-bit stencil in the low-order bits - mask them out */ + spe_and_uint(f, fbS_reg, fbZS_reg, 0x000000ff); + break; + + case PIPE_FORMAT_Z32_UNORM: + setup_optional_register(f, &fbZ_reg_set, &fbZ_reg); + /* Copy over 4 32-bit values */ + spe_move(f, fbZ_reg, fbZS_reg); + + /* Incoming fragZ_reg value is a float in 0.0...1.0; convert + * to a 32-bit unsigned integer + */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + /* No stencil, so can't do anything there */ + break; + case PIPE_FORMAT_Z16_UNORM: + /* XXX Not sure this is correct, but it was here before, so we're + * going with it for now + */ + setup_optional_register(f, &fbZ_reg_set, &fbZ_reg); + /* Copy over 4 32-bit values */ + spe_move(f, fbZ_reg, fbZS_reg); + + /* Incoming fragZ_reg value is a float in 0.0...1.0; convert + * to a 16-bit unsigned integer + */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + spe_rotmi(f, fragZ_reg, fragZ_reg, -16); + /* No stencil */ + + 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) { - /* XXX this may involve depth testing too */ - // gen_stencil_test(dsa, f, ... ); - ASSERT(0); + /* 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_set); + 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); - gen_depth_test(dsa, f, mask_reg, fragZ_reg, fbZ_reg, zmask_reg); + ASSERT(fbZ_reg_set); + 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); } - - /* 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); + 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_S8Z24_UNORM || - zs_format == PIPE_FORMAT_X8Z24_UNORM) { - /* XXX to do */ - ASSERT(0); + 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) { - /* XXX to do */ - ASSERT(0); + spe_move(f, fbZS_reg, fbZ_reg); /* fbZS = fbZ */ } else if (zs_format == PIPE_FORMAT_S8_UNORM) { - /* XXX to do */ - ASSERT(0); + ASSERT(0); /* XXX to do */ } else { - /* bad zs_format */ - ASSERT(0); + 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); } - spe_release_register(f, fbZ_reg); - spe_release_register(f, fbS_reg); + /* Don't need these any more */ + release_optional_register(f, &fbZ_reg_set, fbZ_reg); + release_optional_register(f, &fbS_reg_set, fbS_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 +2052,21 @@ 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 +2078,7 @@ 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 +2088,14 @@ 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..81efd137c7 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; + } } @@ -263,27 +248,25 @@ 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; - - pipe_texture_reference((struct pipe_texture **) &cell->texture[i], tex); + struct pipe_texture *new_tex = i < num ? texture[i] : NULL; + if ((struct pipe_texture *) cell->texture[i] != new_tex) { + pipe_texture_reference((struct pipe_texture **) &cell->texture[i], + new_tex); + changed |= (1 << i); + } } - cell->num_textures = num; - cell_update_texture_mapping(cell); + cell->num_textures = num; - cell->dirty |= CELL_NEW_TEXTURE; + if (changed) { + cell->dirty |= CELL_NEW_TEXTURE; + cell->dirty_textures |= changed; + } } 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..d223557950 100644 --- a/src/gallium/drivers/cell/ppu/cell_screen.c +++ b/src/gallium/drivers/cell/ppu/cell_screen.c @@ -58,9 +58,9 @@ 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: @@ -68,21 +68,21 @@ 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; default: - return 0; + return 10; } } @@ -108,7 +108,7 @@ cell_get_paramf(struct pipe_screen *screen, int param) return 16.0; /* arbitrary */ default: - return 0; + return 10; } } 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..0a0af81f53 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_emit.c +++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c @@ -25,18 +25,155 @@ * **************************************************************************/ +#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 = 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) @@ -73,6 +210,13 @@ cell_emit_state(struct cell_context *cell) #endif } + if (cell->dirty & (CELL_NEW_RASTERIZER)) { + struct cell_command_rasterizer *rast = + cell_batch_alloc(cell, sizeof(*rast)); + rast->opcode = CELL_CMD_STATE_RASTERIZER; + rast->rasterizer = *cell->rasterizer; + } + if (cell->dirty & (CELL_NEW_FS)) { /* Send new fragment program to SPUs */ struct cell_command_fragment_program *fp @@ -90,59 +234,81 @@ 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].size / sizeof(float); + uint i, j; + float *buf = cell_batch_alloc(cell, 16 + num_const * sizeof(float)); + uint64_t *ibuf = (uint64_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[1] = num_const; + j = 4; + 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_alloc(cell, 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]) { + 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]; + } } } + 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)) { + struct cell_command_texture *texture + = cell_batch_alloc(cell, sizeof(*texture)); + texture->opcode = 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_shader.c b/src/gallium/drivers/cell/ppu/cell_state_shader.c index 3a0d066da2..cda39f8d59 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_shader.c +++ b/src/gallium/drivers/cell/ppu/cell_state_shader.c @@ -191,13 +191,18 @@ cell_set_constant_buffer(struct pipe_context *pipe, assert(shader < PIPE_SHADER_TYPES); assert(index == 0); + draw_flush(cell->draw); + /* note: reference counting */ winsys_buffer_reference(ws, &cell->constants[shader].buffer, buf->buffer); cell->constants[shader].size = buf->size; - cell->dirty |= CELL_NEW_CONSTANTS; + 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..9f83ab8fa4 100644 --- a/src/gallium/drivers/cell/ppu/cell_texture.c +++ b/src/gallium/drivers/cell/ppu/cell_texture.c @@ -28,6 +28,7 @@ * Authors: * Keith Whitwell <keith@tungstengraphics.com> * Michel Dänzer <michel@tungstengraphics.com> + * Brian Paul */ #include "pipe/p_context.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; } @@ -135,177 +136,310 @@ 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); - FREE(spt); + 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; + winsys_buffer_reference(screen->winsys, &ct->tiled_buffer[i], NULL); + } + } + + 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); + 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 = pipe_buffer_map(screen, surface->buffer, + 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 = 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); + pipe_buffer_unmap(screen, surface->buffer); +} + + +/** + * 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) +{ + 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 = pipe_buffer_map(screen, surface->buffer, + 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); + } + + dst = (uint *) ((ubyte *) ct->untiled_data[level] + offset); - pipe_surface_unmap(surf); + 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)); + } + } - pipe_surface_reference(&surf, NULL); + pipe_buffer_unmap(screen, surface->buffer); } -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; -#endif - uint i; - - for (i = 0; i < CELL_MAX_SAMPLERS; i++) { - if (cell->texture[i]) - cell_tile_texture(cell, cell->texture[i]); - } + struct pipe_winsys *ws = screen->winsys; + struct cell_texture *ct = cell_texture(pt); + struct pipe_surface *ps; -#if 0 - if (cell->tex_surf && cell->tex_map) { - pipe_surface_unmap(cell->tex_surf); - cell->tex_map = NULL; - } + ps = ws->surface_alloc(ws); + if (ps) { + assert(ps->refcount); + assert(ps->winsys); + winsys_buffer_reference(ws, &ps->buffer, ct->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 = ct->stride[level]; + ps->offset = ct->level_offset[level]; + ps->usage = usage; - /* XXX free old surface */ + /* XXX may need to override usage flags (see sp_texture.c) */ - cell->tex_surf = cell_get_tex_surface(&cell->pipe, - &cell->texture[0]->base, - face, level, zslice); + pipe_texture_reference(&ps->texture, pt); + ps->face = face; + ps->level = level; + ps->zslice = zslice; - cell->tex_map = pipe_surface_map(cell->tex_surf); -#endif + 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); + } + + if (ps->usage & PIPE_BUFFER_USAGE_CPU_READ) { + /* convert from tiled to linear layout */ + cell_untwiddle_texture(screen, ps); + } + } + return ps; } @@ -313,11 +447,17 @@ static void cell_tex_surface_release(struct pipe_screen *screen, struct pipe_surface **s) { - /* 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. - */ - assert ((*s)->texture); + struct cell_texture *ct = cell_texture((*s)->texture); + const uint level = (*s)->level; + + if (((*s)->usage & PIPE_BUFFER_USAGE_CPU_READ) && (ct->untiled_data[level])) + { + align_free(ct->untiled_data[level]); + ct->untiled_data[level] = NULL; + } + + /* XXX if done rendering to teximage, re-tile */ + pipe_texture_reference(&(*s)->texture, NULL); screen->winsys->surface_release(screen->winsys, s); @@ -325,11 +465,15 @@ cell_tex_surface_release(struct pipe_screen *screen, static void * -cell_surface_map( struct pipe_screen *screen, - struct pipe_surface *surface, - unsigned flags ) +cell_surface_map(struct pipe_screen *screen, + struct pipe_surface *surface, + unsigned flags) { ubyte *map; + struct cell_texture *ct = cell_texture(surface->texture); + const uint level = surface->level; + + assert(ct); if (flags & ~surface->usage) { assert(0); @@ -339,22 +483,15 @@ cell_surface_map( struct pipe_screen *screen, map = pipe_buffer_map( screen, surface->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)) + else { - /* Do something to notify sharing contexts of a texture change. - * In softpipe, that would mean flushing the texture cache. - */ -#if 0 - cell_screen(screen)->timestamp++; -#endif + 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); + } } - - return map + surface->offset; } @@ -362,17 +499,21 @@ static void cell_surface_unmap(struct pipe_screen *screen, struct pipe_surface *surface) { - pipe_buffer_unmap( screen, surface->buffer ); -} + struct cell_texture *ct = cell_texture(surface->texture); + assert(ct); -void -cell_init_texture_functions(struct cell_context *cell) -{ - /*cell->pipe.texture_update = cell_texture_update;*/ + if ((ct->base.tex_usage & PIPE_TEXTURE_USAGE_SAMPLER) && + (surface->usage & PIPE_BUFFER_USAGE_CPU_WRITE)) { + /* convert from linear to tiled layout */ + cell_twiddle_texture(screen, surface); + } + + pipe_buffer_unmap( screen, surface->buffer ); } + void cell_init_screen_texture_funcs(struct pipe_screen *screen) { diff --git a/src/gallium/drivers/cell/ppu/cell_texture.h b/src/gallium/drivers/cell/ppu/cell_texture.h index 6d37e95ebc..7018b0c9bf 100644 --- a/src/gallium/drivers/cell/ppu/cell_texture.h +++ b/src/gallium/drivers/cell/ppu/cell_texture.h @@ -40,15 +40,19 @@ 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]; + void *untiled_data[CELL_MAX_TEXTURE_LEVELS]; }; @@ -62,14 +66,6 @@ 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); - - -extern void cell_init_screen_texture_funcs(struct pipe_screen *screen); diff --git a/src/gallium/drivers/cell/ppu/cell_vbuf.c b/src/gallium/drivers/cell/ppu/cell_vbuf.c index aa63435b93..65ba51b6bb 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" @@ -108,6 +109,11 @@ 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) { struct cell_command_release_verts *release 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) |