diff options
author | Christoph Bumiller <e0425955@student.tuwien.ac.at> | 2010-09-02 18:31:49 +0200 |
---|---|---|
committer | Christoph Bumiller <e0425955@student.tuwien.ac.at> | 2010-09-02 18:31:49 +0200 |
commit | 222d2f2ac2c7d93cbc0643082c78278ad2c8cfce (patch) | |
tree | b79152c238022b2a901201c22e5809ac520732bf /src/mesa/drivers/dri/i965 | |
parent | 443abc80db9e1a288ce770e76cccd43664348098 (diff) | |
parent | e73c5501b2fe20290d1b691c85a5d82ac3a0431c (diff) |
Merge remote branch 'origin/master' into nv50-compiler
Conflicts:
src/gallium/drivers/nv50/nv50_program.c
Diffstat (limited to 'src/mesa/drivers/dri/i965')
34 files changed, 3307 insertions, 198 deletions
diff --git a/src/mesa/drivers/dri/i965/Makefile b/src/mesa/drivers/dri/i965/Makefile index e381a5c714..bea48e1313 100644 --- a/src/mesa/drivers/dri/i965/Makefile +++ b/src/mesa/drivers/dri/i965/Makefile @@ -104,6 +104,11 @@ C_SOURCES = \ $(COMMON_SOURCES) \ $(DRIVER_SOURCES) +CXX_SOURCES = \ + brw_fs.cpp \ + brw_fs_channel_expressions.cpp \ + brw_fs_vector_splitting.cpp + ASM_SOURCES = DRIVER_DEFINES = -I../intel diff --git a/src/mesa/drivers/dri/i965/brw_context.c b/src/mesa/drivers/dri/i965/brw_context.c index 6d064b822e..d2b20165f9 100644 --- a/src/mesa/drivers/dri/i965/brw_context.c +++ b/src/mesa/drivers/dri/i965/brw_context.c @@ -144,7 +144,8 @@ GLboolean brwCreateContext( int api, brw->CMD_VF_STATISTICS = CMD_VF_STATISTICS_GM45; brw->CMD_PIPELINE_SELECT = CMD_PIPELINE_SELECT_GM45; brw->has_surface_tile_offset = GL_TRUE; - brw->has_compr4 = GL_TRUE; + if (intel->gen < 6) + brw->has_compr4 = GL_TRUE; brw->has_aa_line_parameters = GL_TRUE; brw->has_pln = GL_TRUE; } else { @@ -153,7 +154,11 @@ GLboolean brwCreateContext( int api, } /* WM maximum threads is number of EUs times number of threads per EU. */ - if (intel->gen == 5) { + if (intel->gen >= 6) { + brw->urb.size = 1024; + brw->vs_max_threads = 60; + brw->wm_max_threads = 80; + } else if (intel->gen == 5) { brw->urb.size = 1024; brw->vs_max_threads = 72; brw->wm_max_threads = 12 * 6; diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h index cc4e6638e8..703a7de78d 100644 --- a/src/mesa/drivers/dri/i965/brw_context.h +++ b/src/mesa/drivers/dri/i965/brw_context.h @@ -179,6 +179,16 @@ struct brw_fragment_program { GLbitfield tex_units_used; }; +struct brw_shader { + struct gl_shader base; + + /** Shader IR transformed for native compile, at link time. */ + struct exec_list *ir; +}; + +struct brw_shader_program { + struct gl_shader_program base; +}; /* Data about a particular attempt to compile a program. Note that * there can be many of these, each in a different GL state @@ -654,7 +664,13 @@ struct brw_context drm_intel_bo *prog_bo; drm_intel_bo *state_bo; - drm_intel_bo *const_bo; + drm_intel_bo *const_bo; /* pull constant buffer. */ + /** + * This is the push constant BO on gen6. + * + * Pre-gen6, push constants live in the CURBE. + */ + drm_intel_bo *push_const_bo; } wm; @@ -686,7 +702,13 @@ struct brw_context #define BRW_PACKCOLOR8888(r,g,b,a) ((r<<24) | (g<<16) | (b<<8) | a) - +struct brw_instruction_info { + char *name; + int nsrc; + int ndst; + GLboolean is_arith; +}; +extern const struct brw_instruction_info brw_opcodes[128]; /*====================================================================== * brw_vtbl.c diff --git a/src/mesa/drivers/dri/i965/brw_defines.h b/src/mesa/drivers/dri/i965/brw_defines.h index f7a68cead7..6b8e9e05d0 100644 --- a/src/mesa/drivers/dri/i965/brw_defines.h +++ b/src/mesa/drivers/dri/i965/brw_defines.h @@ -686,6 +686,9 @@ #define BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_GEN5 1 #define BRW_SAMPLER_MESSAGE_SAMPLE_LOD_GEN5 2 #define BRW_SAMPLER_MESSAGE_SAMPLE_COMPARE_GEN5 3 +#define BRW_SAMPLER_MESSAGE_SAMPLE_DERIVS_GEN5 4 +#define BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE_GEN5 5 +#define BRW_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE_GEN5 6 /* for GEN5 only */ #define BRW_SAMPLER_SIMD_MODE_SIMD4X2 0 diff --git a/src/mesa/drivers/dri/i965/brw_disasm.c b/src/mesa/drivers/dri/i965/brw_disasm.c index d230714536..f74a236834 100644 --- a/src/mesa/drivers/dri/i965/brw_disasm.c +++ b/src/mesa/drivers/dri/i965/brw_disasm.c @@ -159,6 +159,11 @@ char *saturate[2] = { [1] = ".sat" }; +char *accwr[2] = { + [0] = "", + [1] = "AccWrEnable" +}; + char *exec_size[8] = { [0] = "1", [1] = "2", @@ -206,6 +211,7 @@ char *compr_ctrl[4] = { [0] = "", [1] = "sechalf", [2] = "compr", + [3] = "compr4", }; char *dep_ctrl[4] = { @@ -235,6 +241,16 @@ char *reg_encoding[8] = { [7] = "F" }; +int reg_type_size[8] = { + [0] = 4, + [1] = 4, + [2] = 2, + [3] = 2, + [4] = 1, + [5] = 1, + [7] = 4 +}; + char *imm_encoding[8] = { [0] = "UD", [1] = "D", @@ -423,6 +439,11 @@ static int print_opcode (FILE *file, int id) static int reg (FILE *file, GLuint _reg_file, GLuint _reg_nr) { int err = 0; + + /* Clear the Compr4 instruction compression bit. */ + if (_reg_file == BRW_MESSAGE_REGISTER_FILE) + _reg_nr &= ~(1 << 7); + if (_reg_file == BRW_ARCHITECTURE_REGISTER_FILE) { switch (_reg_nr & 0xf0) { case BRW_ARF_NULL: @@ -476,7 +497,8 @@ static int dest (FILE *file, struct brw_instruction *inst) if (err == -1) return 0; if (inst->bits1.da1.dest_subreg_nr) - format (file, ".%d", inst->bits1.da1.dest_subreg_nr); + format (file, ".%d", inst->bits1.da1.dest_subreg_nr / + reg_type_size[inst->bits1.da1.dest_reg_type]); format (file, "<%d>", inst->bits1.da1.dest_horiz_stride); err |= control (file, "dest reg encoding", reg_encoding, inst->bits1.da1.dest_reg_type, NULL); } @@ -484,7 +506,8 @@ static int dest (FILE *file, struct brw_instruction *inst) { string (file, "g[a0"); if (inst->bits1.ia1.dest_subreg_nr) - format (file, ".%d", inst->bits1.ia1.dest_subreg_nr); + format (file, ".%d", inst->bits1.ia1.dest_subreg_nr / + reg_type_size[inst->bits1.ia1.dest_reg_type]); if (inst->bits1.ia1.dest_indirect_offset) format (file, " %d", inst->bits1.ia1.dest_indirect_offset); string (file, "]"); @@ -500,7 +523,8 @@ static int dest (FILE *file, struct brw_instruction *inst) if (err == -1) return 0; if (inst->bits1.da16.dest_subreg_nr) - format (file, ".%d", inst->bits1.da16.dest_subreg_nr); + format (file, ".%d", inst->bits1.da16.dest_subreg_nr / + reg_type_size[inst->bits1.da16.dest_reg_type]); string (file, "<1>"); err |= control (file, "writemask", writemask, inst->bits1.da16.dest_writemask, NULL); err |= control (file, "dest reg encoding", reg_encoding, inst->bits1.da16.dest_reg_type, NULL); @@ -541,7 +565,7 @@ static int src_da1 (FILE *file, GLuint type, GLuint _reg_file, if (err == -1) return 0; if (sub_reg_num) - format (file, ".%d", sub_reg_num); + format (file, ".%d", sub_reg_num / reg_type_size[type]); /* use formal style like spec */ src_align1_region (file, _vert_stride, _width, _horiz_stride); err |= control (file, "src reg encoding", reg_encoding, type, NULL); return err; @@ -595,11 +619,12 @@ static int src_da16 (FILE *file, if (err == -1) return 0; if (_subreg_nr) - format (file, ".%d", _subreg_nr); + /* bit4 for subreg number byte addressing. Make this same meaning as + in da1 case, so output looks consistent. */ + format (file, ".%d", 16 / reg_type_size[_reg_type]); string (file, "<"); err |= control (file, "vert stride", vert_stride, _vert_stride, NULL); string (file, ",4,1>"); - err |= control (file, "src da16 reg type", reg_encoding, _reg_type, NULL); /* * Three kinds of swizzle display: * identity - nothing printed @@ -863,12 +888,25 @@ int brw_disasm (FILE *file, struct brw_instruction *inst, int gen) inst->bits3.math.precision, &space); break; case BRW_MESSAGE_TARGET_SAMPLER: - format (file, " (%d, %d, ", - inst->bits3.sampler.binding_table_index, - inst->bits3.sampler.sampler); - err |= control (file, "sampler target format", sampler_target_format, - inst->bits3.sampler.return_format, NULL); - string (file, ")"); + if (gen >= 5) { + format (file, " (%d, %d, %d, %d)", + inst->bits3.sampler_gen5.binding_table_index, + inst->bits3.sampler_gen5.sampler, + inst->bits3.sampler_gen5.msg_type, + inst->bits3.sampler_gen5.simd_mode); + } else if (0 /* FINISHME: is_g4x */) { + format (file, " (%d, %d)", + inst->bits3.sampler_g4x.binding_table_index, + inst->bits3.sampler_g4x.sampler); + } else { + format (file, " (%d, %d, ", + inst->bits3.sampler.binding_table_index, + inst->bits3.sampler.sampler); + err |= control (file, "sampler target format", + sampler_target_format, + inst->bits3.sampler.return_format, NULL); + string (file, ")"); + } break; case BRW_MESSAGE_TARGET_DATAPORT_READ: if (gen >= 6) { @@ -929,6 +967,11 @@ int brw_disasm (FILE *file, struct brw_instruction *inst, int gen) inst->bits3.urb.used, &space); err |= control (file, "urb complete", urb_complete, inst->bits3.urb.complete, &space); + if (gen >= 5) { + format (file, " mlen %d, rlen %d\n", + inst->bits3.urb_gen5.msg_length, + inst->bits3.urb_gen5.response_length); + } break; case BRW_MESSAGE_TARGET_THREAD_SPAWNER: break; @@ -957,8 +1000,19 @@ int brw_disasm (FILE *file, struct brw_instruction *inst, int gen) err |= control(file, "access mode", access_mode, inst->header.access_mode, &space); err |= control (file, "mask control", mask_ctrl, inst->header.mask_control, &space); err |= control (file, "dependency control", dep_ctrl, inst->header.dependency_control, &space); - err |= control (file, "compression control", compr_ctrl, inst->header.compression_control, &space); + + if (inst->header.compression_control == BRW_COMPRESSION_COMPRESSED && + opcode[inst->header.opcode].ndst > 0 && + inst->bits1.da1.dest_reg_file == BRW_MESSAGE_REGISTER_FILE && + inst->bits1.da1.dest_reg_nr & (1 << 7)) { + format (file, " compr4"); + } else { + err |= control (file, "compression control", compr_ctrl, + inst->header.compression_control, &space); + } err |= control (file, "thread control", thread_ctrl, inst->header.thread_control, &space); + if (gen >= 6) + err |= control (file, "acc write control", accwr, inst->header.acc_wr_control, &space); if (inst->header.opcode == BRW_OPCODE_SEND) err |= control (file, "end of thread", end_of_thread, inst->bits3.generic.end_of_thread, &space); diff --git a/src/mesa/drivers/dri/i965/brw_draw_upload.c b/src/mesa/drivers/dri/i965/brw_draw_upload.c index f07aab86e9..249e874ab1 100644 --- a/src/mesa/drivers/dri/i965/brw_draw_upload.c +++ b/src/mesa/drivers/dri/i965/brw_draw_upload.c @@ -476,7 +476,7 @@ static void brw_emit_vertices(struct brw_context *brw) if (brw->vb.nr_enabled == 0) { BEGIN_BATCH(3); OUT_BATCH((CMD_VERTEX_ELEMENT << 16) | 1); - if (IS_GEN6(intel->intelScreen->deviceID)) { + if (intel->gen >= 6) { OUT_BATCH((0 << GEN6_VE0_INDEX_SHIFT) | GEN6_VE0_VALID | (BRW_SURFACEFORMAT_R32G32B32A32_FLOAT << BRW_VE0_FORMAT_SHIFT) | @@ -553,7 +553,7 @@ static void brw_emit_vertices(struct brw_context *brw) break; } - if (IS_GEN6(intel->intelScreen->deviceID)) { + if (intel->gen >= 6) { OUT_BATCH((i << GEN6_VE0_INDEX_SHIFT) | GEN6_VE0_VALID | (format << BRW_VE0_FORMAT_SHIFT) | diff --git a/src/mesa/drivers/dri/i965/brw_eu.c b/src/mesa/drivers/dri/i965/brw_eu.c index 4e7c1226ad..2ff39e8e64 100644 --- a/src/mesa/drivers/dri/i965/brw_eu.c +++ b/src/mesa/drivers/dri/i965/brw_eu.c @@ -85,6 +85,12 @@ void brw_set_saturate( struct brw_compile *p, GLuint value ) p->current->header.saturate = value; } +void brw_set_acc_write_control(struct brw_compile *p, GLuint value) +{ + if (p->brw->intel.gen >= 6) + p->current->header.acc_wr_control = value; +} + void brw_push_insn_state( struct brw_compile *p ) { assert(p->current != &p->stack[BRW_EU_MAX_INSN_STACK-1]); diff --git a/src/mesa/drivers/dri/i965/brw_eu.h b/src/mesa/drivers/dri/i965/brw_eu.h index ffdddd0a38..c63db16460 100644 --- a/src/mesa/drivers/dri/i965/brw_eu.h +++ b/src/mesa/drivers/dri/i965/brw_eu.h @@ -633,6 +633,8 @@ static INLINE struct brw_reg brw_swizzle( struct brw_reg reg, GLuint z, GLuint w) { + assert(reg.file != BRW_IMMEDIATE_VALUE); + reg.dw1.bits.swizzle = BRW_SWIZZLE4(BRW_GET_SWZ(reg.dw1.bits.swizzle, x), BRW_GET_SWZ(reg.dw1.bits.swizzle, y), BRW_GET_SWZ(reg.dw1.bits.swizzle, z), @@ -650,6 +652,7 @@ static INLINE struct brw_reg brw_swizzle1( struct brw_reg reg, static INLINE struct brw_reg brw_writemask( struct brw_reg reg, GLuint mask ) { + assert(reg.file != BRW_IMMEDIATE_VALUE); reg.dw1.bits.writemask &= mask; return reg; } @@ -657,6 +660,7 @@ static INLINE struct brw_reg brw_writemask( struct brw_reg reg, static INLINE struct brw_reg brw_set_writemask( struct brw_reg reg, GLuint mask ) { + assert(reg.file != BRW_IMMEDIATE_VALUE); reg.dw1.bits.writemask = mask; return reg; } @@ -766,6 +770,7 @@ void brw_set_compression_control( struct brw_compile *p, GLboolean control ); void brw_set_predicate_control_flag_value( struct brw_compile *p, GLuint value ); void brw_set_predicate_control( struct brw_compile *p, GLuint pc ); void brw_set_conditionalmod( struct brw_compile *p, GLuint conditional ); +void brw_set_acc_write_control(struct brw_compile *p, GLuint value); void brw_init_compile( struct brw_context *, struct brw_compile *p ); const GLuint *brw_get_program( struct brw_compile *p, GLuint *sz ); @@ -840,6 +845,7 @@ void brw_ff_sync(struct brw_compile *p, GLboolean eot); void brw_fb_WRITE(struct brw_compile *p, + int dispatch_width, struct brw_reg dest, GLuint msg_reg_nr, struct brw_reg src0, @@ -925,8 +931,8 @@ struct brw_instruction *brw_DO(struct brw_compile *p, struct brw_instruction *brw_WHILE(struct brw_compile *p, struct brw_instruction *patch_insn); -struct brw_instruction *brw_BREAK(struct brw_compile *p); -struct brw_instruction *brw_CONT(struct brw_compile *p); +struct brw_instruction *brw_BREAK(struct brw_compile *p, int pop_count); +struct brw_instruction *brw_CONT(struct brw_compile *p, int pop_count); /* Forward jumps: */ void brw_land_fwd_jump(struct brw_compile *p, diff --git a/src/mesa/drivers/dri/i965/brw_eu_emit.c b/src/mesa/drivers/dri/i965/brw_eu_emit.c index 0d5d17f501..0906150613 100644 --- a/src/mesa/drivers/dri/i965/brw_eu_emit.c +++ b/src/mesa/drivers/dri/i965/brw_eu_emit.c @@ -75,6 +75,8 @@ static void brw_set_dest( struct brw_instruction *insn, else { insn->bits1.da16.dest_subreg_nr = dest.subnr / 16; insn->bits1.da16.dest_writemask = dest.dw1.bits.writemask; + /* even ignored in da16, still need to set as '01' */ + insn->bits1.da16.dest_horiz_stride = 1; } } else { @@ -90,6 +92,8 @@ static void brw_set_dest( struct brw_instruction *insn, } else { insn->bits1.ia16.dest_indirect_offset = dest.dw1.bits.indirect_offset; + /* even ignored in da16, still need to set as '01' */ + insn->bits1.ia16.dest_horiz_stride = 1; } } @@ -368,9 +372,23 @@ static void brw_set_dp_write_message( struct brw_context *brw, GLuint send_commit_msg) { struct intel_context *intel = &brw->intel; - brw_set_src1(insn, brw_imm_d(0)); + brw_set_src1(insn, brw_imm_ud(0)); - if (intel->gen == 5) { + if (intel->gen >= 6) { + insn->bits3.dp_render_cache.binding_table_index = binding_table_index; + insn->bits3.dp_render_cache.msg_control = msg_control; + insn->bits3.dp_render_cache.pixel_scoreboard_clear = pixel_scoreboard_clear; + insn->bits3.dp_render_cache.msg_type = msg_type; + insn->bits3.dp_render_cache.send_commit_msg = send_commit_msg; + insn->bits3.dp_render_cache.header_present = 0; /* XXX */ + insn->bits3.dp_render_cache.response_length = response_length; + insn->bits3.dp_render_cache.msg_length = msg_length; + insn->bits3.dp_render_cache.end_of_thread = end_of_thread; + insn->header.destreg__conditionalmod = BRW_MESSAGE_TARGET_DATAPORT_WRITE; + /* XXX really need below? */ + insn->bits2.send_gen5.sfid = BRW_MESSAGE_TARGET_DATAPORT_WRITE; + insn->bits2.send_gen5.end_of_thread = end_of_thread; + } else if (intel->gen == 5) { insn->bits3.dp_write_gen5.binding_table_index = binding_table_index; insn->bits3.dp_write_gen5.msg_control = msg_control; insn->bits3.dp_write_gen5.pixel_scoreboard_clear = pixel_scoreboard_clear; @@ -759,7 +777,7 @@ void brw_ENDIF(struct brw_compile *p, } } -struct brw_instruction *brw_BREAK(struct brw_compile *p) +struct brw_instruction *brw_BREAK(struct brw_compile *p, int pop_count) { struct brw_instruction *insn; insn = next_insn(p, BRW_OPCODE_BREAK); @@ -770,10 +788,11 @@ struct brw_instruction *brw_BREAK(struct brw_compile *p) insn->header.execution_size = BRW_EXECUTE_8; /* insn->header.mask_control = BRW_MASK_DISABLE; */ insn->bits3.if_else.pad0 = 0; + insn->bits3.if_else.pop_count = pop_count; return insn; } -struct brw_instruction *brw_CONT(struct brw_compile *p) +struct brw_instruction *brw_CONT(struct brw_compile *p, int pop_count) { struct brw_instruction *insn; insn = next_insn(p, BRW_OPCODE_CONTINUE); @@ -784,6 +803,7 @@ struct brw_instruction *brw_CONT(struct brw_compile *p) insn->header.execution_size = BRW_EXECUTE_8; /* insn->header.mask_control = BRW_MASK_DISABLE; */ insn->bits3.if_else.pad0 = 0; + insn->bits3.if_else.pop_count = pop_count; return insn; } @@ -1332,6 +1352,7 @@ void brw_dp_READ_4_vs_relative(struct brw_compile *p, void brw_fb_WRITE(struct brw_compile *p, + int dispatch_width, struct brw_reg dest, GLuint msg_reg_nr, struct brw_reg src0, @@ -1340,22 +1361,40 @@ void brw_fb_WRITE(struct brw_compile *p, GLuint response_length, GLboolean eot) { - struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND); - + struct intel_context *intel = &p->brw->intel; + struct brw_instruction *insn; + GLuint msg_control, msg_type; + + insn = next_insn(p, BRW_OPCODE_SEND); insn->header.predicate_control = 0; /* XXX */ - insn->header.compression_control = BRW_COMPRESSION_NONE; - insn->header.destreg__conditionalmod = msg_reg_nr; - + insn->header.compression_control = BRW_COMPRESSION_NONE; + + if (intel->gen >= 6) { + /* headerless version, just submit color payload */ + src0 = brw_message_reg(msg_reg_nr); + + msg_type = BRW_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE_GEN6; + } else { + insn->header.destreg__conditionalmod = msg_reg_nr; + + msg_type = BRW_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE; + } + + if (dispatch_width == 16) + msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE; + else + msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01; + brw_set_dest(insn, dest); brw_set_src0(insn, src0); brw_set_dp_write_message(p->brw, insn, binding_table_index, - BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE, /* msg_control */ - BRW_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE, /* msg_type */ + msg_control, + msg_type, msg_length, 1, /* pixel scoreboard */ - response_length, + response_length, eot, 0 /* send_commit_msg */); } diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp new file mode 100644 index 0000000000..34c5d5262f --- /dev/null +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp @@ -0,0 +1,1924 @@ +/* + * Copyright © 2010 Intel Corporation + * + * 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, sublicense, + * 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 NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS 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. + * + * Authors: + * Eric Anholt <eric@anholt.net> + * + */ + +extern "C" { + +#include <sys/types.h> + +#include "main/macros.h" +#include "main/shaderobj.h" +#include "program/prog_parameter.h" +#include "program/prog_print.h" +#include "program/prog_optimize.h" +#include "program/hash_table.h" +#include "brw_context.h" +#include "brw_eu.h" +#include "brw_wm.h" +#include "talloc.h" +} +#include "../glsl/glsl_types.h" +#include "../glsl/ir_optimization.h" +#include "../glsl/ir_print_visitor.h" + +enum register_file { + ARF = BRW_ARCHITECTURE_REGISTER_FILE, + GRF = BRW_GENERAL_REGISTER_FILE, + MRF = BRW_MESSAGE_REGISTER_FILE, + IMM = BRW_IMMEDIATE_VALUE, + FIXED_HW_REG, /* a struct brw_reg */ + UNIFORM, /* prog_data->params[hw_reg] */ + BAD_FILE +}; + +enum fs_opcodes { + FS_OPCODE_FB_WRITE = 256, + FS_OPCODE_RCP, + FS_OPCODE_RSQ, + FS_OPCODE_SQRT, + FS_OPCODE_EXP2, + FS_OPCODE_LOG2, + FS_OPCODE_POW, + FS_OPCODE_SIN, + FS_OPCODE_COS, + FS_OPCODE_DDX, + FS_OPCODE_DDY, + FS_OPCODE_LINTERP, + FS_OPCODE_TEX, + FS_OPCODE_TXB, + FS_OPCODE_TXL, + FS_OPCODE_DISCARD, +}; + +static int using_new_fs = -1; + +struct gl_shader * +brw_new_shader(GLcontext *ctx, GLuint name, GLuint type) +{ + struct brw_shader *shader; + + shader = talloc_zero(NULL, struct brw_shader); + if (shader) { + shader->base.Type = type; + shader->base.Name = name; + _mesa_init_shader(ctx, &shader->base); + } + + return &shader->base; +} + +struct gl_shader_program * +brw_new_shader_program(GLcontext *ctx, GLuint name) +{ + struct brw_shader_program *prog; + prog = talloc_zero(NULL, struct brw_shader_program); + if (prog) { + prog->base.Name = name; + _mesa_init_shader_program(ctx, &prog->base); + } + return &prog->base; +} + +GLboolean +brw_compile_shader(GLcontext *ctx, struct gl_shader *shader) +{ + if (!_mesa_ir_compile_shader(ctx, shader)) + return GL_FALSE; + + return GL_TRUE; +} + +GLboolean +brw_link_shader(GLcontext *ctx, struct gl_shader_program *prog) +{ + if (using_new_fs == -1) + using_new_fs = getenv("INTEL_NEW_FS") != NULL; + + for (unsigned i = 0; i < prog->_NumLinkedShaders; i++) { + struct brw_shader *shader = (struct brw_shader *)prog->_LinkedShaders[i]; + + if (using_new_fs && shader->base.Type == GL_FRAGMENT_SHADER) { + void *mem_ctx = talloc_new(NULL); + bool progress; + + if (shader->ir) + talloc_free(shader->ir); + shader->ir = new(shader) exec_list; + clone_ir_list(mem_ctx, shader->ir, shader->base.ir); + + do_mat_op_to_vec(shader->ir); + do_mod_to_fract(shader->ir); + do_div_to_mul_rcp(shader->ir); + do_sub_to_add_neg(shader->ir); + do_explog_to_explog2(shader->ir); + + brw_do_channel_expressions(shader->ir); + brw_do_vector_splitting(shader->ir); + + do { + progress = false; + + progress = do_common_optimization(shader->ir, true) || progress; + } while (progress); + + validate_ir_tree(shader->ir); + + reparent_ir(shader->ir, shader->ir); + talloc_free(mem_ctx); + } + } + + if (!_mesa_ir_link_shader(ctx, prog)) + return GL_FALSE; + + return GL_TRUE; +} + +static int +type_size(const struct glsl_type *type) +{ + unsigned int size, i; + + switch (type->base_type) { + case GLSL_TYPE_UINT: + case GLSL_TYPE_INT: + case GLSL_TYPE_FLOAT: + case GLSL_TYPE_BOOL: + return type->components(); + case GLSL_TYPE_ARRAY: + /* FINISHME: uniform/varying arrays. */ + return type_size(type->fields.array) * type->length; + case GLSL_TYPE_STRUCT: + size = 0; + for (i = 0; i < type->length; i++) { + size += type_size(type->fields.structure[i].type); + } + return size; + case GLSL_TYPE_SAMPLER: + /* Samplers take up no register space, since they're baked in at + * link time. + */ + return 0; + default: + assert(!"not reached"); + return 0; + } +} + +class fs_reg { +public: + /* Callers of this talloc-based new need not call delete. It's + * easier to just talloc_free 'ctx' (or any of its ancestors). */ + static void* operator new(size_t size, void *ctx) + { + void *node; + + node = talloc_size(ctx, size); + assert(node != NULL); + + return node; + } + + /** Generic unset register constructor. */ + fs_reg() + { + this->file = BAD_FILE; + this->reg = 0; + this->reg_offset = 0; + this->hw_reg = -1; + this->negate = 0; + this->abs = 0; + } + + /** Immediate value constructor. */ + fs_reg(float f) + { + this->file = IMM; + this->reg = 0; + this->hw_reg = 0; + this->type = BRW_REGISTER_TYPE_F; + this->imm.f = f; + this->negate = 0; + this->abs = 0; + } + + /** Immediate value constructor. */ + fs_reg(int32_t i) + { + this->file = IMM; + this->reg = 0; + this->hw_reg = 0; + this->type = BRW_REGISTER_TYPE_D; + this->imm.i = i; + this->negate = 0; + this->abs = 0; + } + + /** Immediate value constructor. */ + fs_reg(uint32_t u) + { + this->file = IMM; + this->reg = 0; + this->hw_reg = 0; + this->type = BRW_REGISTER_TYPE_UD; + this->imm.u = u; + this->negate = 0; + this->abs = 0; + } + + /** Fixed brw_reg Immediate value constructor. */ + fs_reg(struct brw_reg fixed_hw_reg) + { + this->file = FIXED_HW_REG; + this->fixed_hw_reg = fixed_hw_reg; + this->reg = 0; + this->hw_reg = 0; + this->type = fixed_hw_reg.type; + this->negate = 0; + this->abs = 0; + } + + fs_reg(enum register_file file, int hw_reg); + fs_reg(class fs_visitor *v, const struct glsl_type *type); + + /** Register file: ARF, GRF, MRF, IMM. */ + enum register_file file; + /** Abstract register number. 0 = fixed hw reg */ + int reg; + /** Offset within the abstract register. */ + int reg_offset; + /** HW register number. Generally unset until register allocation. */ + int hw_reg; + /** Register type. BRW_REGISTER_TYPE_* */ + int type; + bool negate; + bool abs; + struct brw_reg fixed_hw_reg; + + /** Value for file == BRW_IMMMEDIATE_FILE */ + union { + int32_t i; + uint32_t u; + float f; + } imm; +}; + +static const fs_reg reg_undef; +static const fs_reg reg_null(ARF, BRW_ARF_NULL); + +class fs_inst : public exec_node { +public: + /* Callers of this talloc-based new need not call delete. It's + * easier to just talloc_free 'ctx' (or any of its ancestors). */ + static void* operator new(size_t size, void *ctx) + { + void *node; + + node = talloc_zero_size(ctx, size); + assert(node != NULL); + + return node; + } + + void init() + { + this->opcode = BRW_OPCODE_NOP; + this->saturate = false; + this->conditional_mod = BRW_CONDITIONAL_NONE; + this->predicated = false; + this->sampler = 0; + this->shadow_compare = false; + } + + fs_inst() + { + init(); + } + + fs_inst(int opcode) + { + init(); + this->opcode = opcode; + } + + fs_inst(int opcode, fs_reg dst, fs_reg src0) + { + init(); + this->opcode = opcode; + this->dst = dst; + this->src[0] = src0; + } + + fs_inst(int opcode, fs_reg dst, fs_reg src0, fs_reg src1) + { + init(); + this->opcode = opcode; + this->dst = dst; + this->src[0] = src0; + this->src[1] = src1; + } + + fs_inst(int opcode, fs_reg dst, fs_reg src0, fs_reg src1, fs_reg src2) + { + init(); + this->opcode = opcode; + this->dst = dst; + this->src[0] = src0; + this->src[1] = src1; + this->src[2] = src2; + } + + int opcode; /* BRW_OPCODE_* or FS_OPCODE_* */ + fs_reg dst; + fs_reg src[3]; + bool saturate; + bool predicated; + int conditional_mod; /**< BRW_CONDITIONAL_* */ + + int mlen; /** SEND message length */ + int sampler; + bool shadow_compare; + + /** @{ + * Annotation for the generated IR. One of the two can be set. + */ + ir_instruction *ir; + const char *annotation; + /** @} */ +}; + +class fs_visitor : public ir_visitor +{ +public: + + fs_visitor(struct brw_wm_compile *c, struct brw_shader *shader) + { + this->c = c; + this->p = &c->func; + this->brw = p->brw; + this->intel = &brw->intel; + this->ctx = &intel->ctx; + this->mem_ctx = talloc_new(NULL); + this->shader = shader; + this->fail = false; + this->next_abstract_grf = 1; + this->variable_ht = hash_table_ctor(0, + hash_table_pointer_hash, + hash_table_pointer_compare); + + this->frag_color = NULL; + this->frag_data = NULL; + this->frag_depth = NULL; + this->first_non_payload_grf = 0; + + this->current_annotation = NULL; + this->annotation_string = NULL; + this->annotation_ir = NULL; + } + ~fs_visitor() + { + talloc_free(this->mem_ctx); + hash_table_dtor(this->variable_ht); + } + + fs_reg *variable_storage(ir_variable *var); + + void visit(ir_variable *ir); + void visit(ir_assignment *ir); + void visit(ir_dereference_variable *ir); + void visit(ir_dereference_record *ir); + void visit(ir_dereference_array *ir); + void visit(ir_expression *ir); + void visit(ir_texture *ir); + void visit(ir_if *ir); + void visit(ir_constant *ir); + void visit(ir_swizzle *ir); + void visit(ir_return *ir); + void visit(ir_loop *ir); + void visit(ir_loop_jump *ir); + void visit(ir_discard *ir); + void visit(ir_call *ir); + void visit(ir_function *ir); + void visit(ir_function_signature *ir); + + fs_inst *emit(fs_inst inst); + void assign_curb_setup(); + void assign_urb_setup(); + void assign_regs(); + void generate_code(); + void generate_fb_write(fs_inst *inst); + void generate_linterp(fs_inst *inst, struct brw_reg dst, + struct brw_reg *src); + void generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src); + void generate_math(fs_inst *inst, struct brw_reg dst, struct brw_reg *src); + void generate_discard(fs_inst *inst); + + void emit_dummy_fs(); + void emit_interpolation(); + void emit_pinterp(int location); + void emit_fb_writes(); + + struct brw_reg interp_reg(int location, int channel); + + struct brw_context *brw; + struct intel_context *intel; + GLcontext *ctx; + struct brw_wm_compile *c; + struct brw_compile *p; + struct brw_shader *shader; + void *mem_ctx; + exec_list instructions; + int next_abstract_grf; + struct hash_table *variable_ht; + ir_variable *frag_color, *frag_data, *frag_depth; + int first_non_payload_grf; + + /** @{ debug annotation info */ + const char *current_annotation; + ir_instruction *base_ir; + const char **annotation_string; + ir_instruction **annotation_ir; + /** @} */ + + bool fail; + + /* Result of last visit() method. */ + fs_reg result; + + fs_reg pixel_x; + fs_reg pixel_y; + fs_reg pixel_w; + fs_reg delta_x; + fs_reg delta_y; + fs_reg interp_attrs[64]; + + int grf_used; + +}; + +/** Fixed HW reg constructor. */ +fs_reg::fs_reg(enum register_file file, int hw_reg) +{ + this->file = file; + this->reg = 0; + this->reg_offset = 0; + this->hw_reg = hw_reg; + this->type = BRW_REGISTER_TYPE_F; + this->negate = 0; + this->abs = 0; +} + +/** Automatic reg constructor. */ +fs_reg::fs_reg(class fs_visitor *v, const struct glsl_type *type) +{ + this->file = GRF; + this->reg = v->next_abstract_grf; + this->reg_offset = 0; + v->next_abstract_grf += type_size(type); + this->hw_reg = -1; + this->negate = 0; + this->abs = 0; + + switch (type->base_type) { + case GLSL_TYPE_FLOAT: + this->type = BRW_REGISTER_TYPE_F; + break; + case GLSL_TYPE_INT: + case GLSL_TYPE_BOOL: + this->type = BRW_REGISTER_TYPE_D; + break; + case GLSL_TYPE_UINT: + this->type = BRW_REGISTER_TYPE_UD; + break; + default: + assert(!"not reached"); + this->type = BRW_REGISTER_TYPE_F; + break; + } +} + +fs_reg * +fs_visitor::variable_storage(ir_variable *var) +{ + return (fs_reg *)hash_table_find(this->variable_ht, var); +} + +void +fs_visitor::visit(ir_variable *ir) +{ + fs_reg *reg = NULL; + + if (strcmp(ir->name, "gl_FragColor") == 0) { + this->frag_color = ir; + } else if (strcmp(ir->name, "gl_FragData") == 0) { + this->frag_data = ir; + } else if (strcmp(ir->name, "gl_FragDepth") == 0) { + this->frag_depth = ir; + assert(!"FINISHME: this hangs currently."); + } + + if (ir->mode == ir_var_in) { + reg = &this->interp_attrs[ir->location]; + } + + if (ir->mode == ir_var_uniform) { + const float *vec_values; + int param_index = c->prog_data.nr_params; + + /* FINISHME: This is wildly incomplete. */ + assert(ir->type->is_scalar() || ir->type->is_vector() || + ir->type->is_sampler()); + + const struct gl_program *fp = &this->brw->fragment_program->Base; + /* Our support for uniforms is piggy-backed on the struct + * gl_fragment_program, because that's where the values actually + * get stored, rather than in some global gl_shader_program uniform + * store. + */ + vec_values = fp->Parameters->ParameterValues[ir->location]; + for (unsigned int i = 0; i < ir->type->vector_elements; i++) { + c->prog_data.param[c->prog_data.nr_params++] = &vec_values[i]; + } + + reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index); + } + + if (!reg) + reg = new(this->mem_ctx) fs_reg(this, ir->type); + + hash_table_insert(this->variable_ht, reg, ir); +} + +void +fs_visitor::visit(ir_dereference_variable *ir) +{ + fs_reg *reg = variable_storage(ir->var); + this->result = *reg; +} + +void +fs_visitor::visit(ir_dereference_record *ir) +{ + assert(!"FINISHME"); +} + +void +fs_visitor::visit(ir_dereference_array *ir) +{ + ir_constant *index; + int element_size; + + ir->array->accept(this); + index = ir->array_index->as_constant(); + + if (ir->type->is_matrix()) { + element_size = ir->type->vector_elements; + } else { + element_size = type_size(ir->type); + } + + if (index) { + assert(this->result.file == UNIFORM || + (this->result.file == GRF && + this->result.reg != 0)); + this->result.reg_offset += index->value.i[0] * element_size; + } else { + assert(!"FINISHME: non-constant matrix column"); + } +} + +void +fs_visitor::visit(ir_expression *ir) +{ + unsigned int operand; + fs_reg op[2], temp; + fs_reg result; + fs_inst *inst; + + for (operand = 0; operand < ir->get_num_operands(); operand++) { + ir->operands[operand]->accept(this); + if (this->result.file == BAD_FILE) { + ir_print_visitor v; + printf("Failed to get tree for expression operand:\n"); + ir->operands[operand]->accept(&v); + this->fail = true; + } + op[operand] = this->result; + + /* Matrix expression operands should have been broken down to vector + * operations already. + */ + assert(!ir->operands[operand]->type->is_matrix()); + /* And then those vector operands should have been broken down to scalar. + */ + assert(!ir->operands[operand]->type->is_vector()); + } + + /* Storage for our result. If our result goes into an assignment, it will + * just get copy-propagated out, so no worries. + */ + this->result = fs_reg(this, ir->type); + + switch (ir->operation) { + case ir_unop_logic_not: + emit(fs_inst(BRW_OPCODE_ADD, this->result, op[0], fs_reg(-1))); + break; + case ir_unop_neg: + op[0].negate = ~op[0].negate; + this->result = op[0]; + break; + case ir_unop_abs: + op[0].abs = true; + this->result = op[0]; + break; + case ir_unop_sign: + temp = fs_reg(this, ir->type); + + emit(fs_inst(BRW_OPCODE_MOV, this->result, fs_reg(0.0f))); + + inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null, op[0], fs_reg(0.0f))); + inst->conditional_mod = BRW_CONDITIONAL_G; + inst = emit(fs_inst(BRW_OPCODE_MOV, this->result, fs_reg(1.0f))); + inst->predicated = true; + + inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null, op[0], fs_reg(0.0f))); + inst->conditional_mod = BRW_CONDITIONAL_L; + inst = emit(fs_inst(BRW_OPCODE_MOV, this->result, fs_reg(-1.0f))); + inst->predicated = true; + + break; + case ir_unop_rcp: + emit(fs_inst(FS_OPCODE_RCP, this->result, op[0])); + break; + + case ir_unop_exp2: + emit(fs_inst(FS_OPCODE_EXP2, this->result, op[0])); + break; + case ir_unop_log2: + emit(fs_inst(FS_OPCODE_LOG2, this->result, op[0])); + break; + case ir_unop_exp: + case ir_unop_log: + assert(!"not reached: should be handled by ir_explog_to_explog2"); + break; + case ir_unop_sin: + emit(fs_inst(FS_OPCODE_SIN, this->result, op[0])); + break; + case ir_unop_cos: + emit(fs_inst(FS_OPCODE_COS, this->result, op[0])); + break; + + case ir_unop_dFdx: + emit(fs_inst(FS_OPCODE_DDX, this->result, op[0])); + break; + case ir_unop_dFdy: + emit(fs_inst(FS_OPCODE_DDY, this->result, op[0])); + break; + + case ir_binop_add: + emit(fs_inst(BRW_OPCODE_ADD, this->result, op[0], op[1])); + break; + case ir_binop_sub: + assert(!"not reached: should be handled by ir_sub_to_add_neg"); + break; + + case ir_binop_mul: + emit(fs_inst(BRW_OPCODE_MUL, this->result, op[0], op[1])); + break; + case ir_binop_div: + assert(!"not reached: should be handled by ir_div_to_mul_rcp"); + break; + case ir_binop_mod: + assert(!"ir_binop_mod should have been converted to b * fract(a/b)"); + break; + + case ir_binop_less: + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], op[1])); + inst->conditional_mod = BRW_CONDITIONAL_L; + emit(fs_inst(BRW_OPCODE_AND, this->result, this->result, fs_reg(0x1))); + break; + case ir_binop_greater: + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], op[1])); + inst->conditional_mod = BRW_CONDITIONAL_G; + emit(fs_inst(BRW_OPCODE_AND, this->result, this->result, fs_reg(0x1))); + break; + case ir_binop_lequal: + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], op[1])); + inst->conditional_mod = BRW_CONDITIONAL_LE; + emit(fs_inst(BRW_OPCODE_AND, this->result, this->result, fs_reg(0x1))); + break; + case ir_binop_gequal: + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], op[1])); + inst->conditional_mod = BRW_CONDITIONAL_GE; + emit(fs_inst(BRW_OPCODE_AND, this->result, this->result, fs_reg(0x1))); + break; + case ir_binop_equal: + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], op[1])); + inst->conditional_mod = BRW_CONDITIONAL_Z; + emit(fs_inst(BRW_OPCODE_AND, this->result, this->result, fs_reg(0x1))); + break; + case ir_binop_nequal: + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], op[1])); + inst->conditional_mod = BRW_CONDITIONAL_NZ; + emit(fs_inst(BRW_OPCODE_AND, this->result, this->result, fs_reg(0x1))); + break; + + case ir_binop_logic_xor: + emit(fs_inst(BRW_OPCODE_XOR, this->result, op[0], op[1])); + break; + + case ir_binop_logic_or: + emit(fs_inst(BRW_OPCODE_OR, this->result, op[0], op[1])); + break; + + case ir_binop_logic_and: + emit(fs_inst(BRW_OPCODE_AND, this->result, op[0], op[1])); + break; + + case ir_binop_dot: + case ir_binop_cross: + case ir_unop_any: + assert(!"not reached: should be handled by brw_channel_expressions"); + break; + + case ir_unop_sqrt: + emit(fs_inst(FS_OPCODE_SQRT, this->result, op[0])); + break; + + case ir_unop_rsq: + emit(fs_inst(FS_OPCODE_RSQ, this->result, op[0])); + break; + + case ir_unop_i2f: + case ir_unop_b2f: + case ir_unop_b2i: + emit(fs_inst(BRW_OPCODE_MOV, this->result, op[0])); + break; + case ir_unop_f2i: + emit(fs_inst(BRW_OPCODE_RNDZ, this->result, op[0])); + break; + case ir_unop_f2b: + case ir_unop_i2b: + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], fs_reg(0.0f))); + inst->conditional_mod = BRW_CONDITIONAL_NZ; + + case ir_unop_trunc: + emit(fs_inst(BRW_OPCODE_RNDD, this->result, op[0])); + break; + case ir_unop_ceil: + op[0].negate = ~op[0].negate; + inst = emit(fs_inst(BRW_OPCODE_RNDD, this->result, op[0])); + this->result.negate = true; + break; + case ir_unop_floor: + inst = emit(fs_inst(BRW_OPCODE_RNDD, this->result, op[0])); + break; + case ir_unop_fract: + inst = emit(fs_inst(BRW_OPCODE_FRC, this->result, op[0])); + break; + + case ir_binop_min: + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], op[1])); + inst->conditional_mod = BRW_CONDITIONAL_L; + + inst = emit(fs_inst(BRW_OPCODE_SEL, this->result, op[0], op[1])); + inst->predicated = true; + break; + case ir_binop_max: + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], op[1])); + inst->conditional_mod = BRW_CONDITIONAL_G; + + inst = emit(fs_inst(BRW_OPCODE_SEL, this->result, op[0], op[1])); + inst->predicated = true; + break; + + case ir_binop_pow: + inst = emit(fs_inst(FS_OPCODE_POW, this->result, op[0], op[1])); + break; + + case ir_unop_bit_not: + case ir_unop_u2f: + case ir_binop_lshift: + case ir_binop_rshift: + case ir_binop_bit_and: + case ir_binop_bit_xor: + case ir_binop_bit_or: + assert(!"GLSL 1.30 features unsupported"); + break; + } +} + +void +fs_visitor::visit(ir_assignment *ir) +{ + struct fs_reg l, r; + int i; + int write_mask; + fs_inst *inst; + + /* FINISHME: arrays on the lhs */ + ir->lhs->accept(this); + l = this->result; + + ir->rhs->accept(this); + r = this->result; + + /* FINISHME: This should really set to the correct maximal writemask for each + * FINISHME: component written (in the loops below). This case can only + * FINISHME: occur for matrices, arrays, and structures. + */ + if (ir->write_mask == 0) { + assert(!ir->lhs->type->is_scalar() && !ir->lhs->type->is_vector()); + write_mask = WRITEMASK_XYZW; + } else { + assert(ir->lhs->type->is_vector() || ir->lhs->type->is_scalar()); + write_mask = ir->write_mask; + } + + assert(l.file != BAD_FILE); + assert(r.file != BAD_FILE); + + if (ir->condition) { + /* Get the condition bool into the predicate. */ + ir->condition->accept(this); + inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, fs_reg(0))); + inst->conditional_mod = BRW_CONDITIONAL_NZ; + } + + for (i = 0; i < type_size(ir->lhs->type); i++) { + if (i >= 4 || (write_mask & (1 << i))) { + inst = emit(fs_inst(BRW_OPCODE_MOV, l, r)); + if (ir->condition) + inst->predicated = true; + } + l.reg_offset++; + r.reg_offset++; + } +} + +void +fs_visitor::visit(ir_texture *ir) +{ + int base_mrf = 2; + fs_inst *inst = NULL; + unsigned int mlen = 0; + + ir->coordinate->accept(this); + fs_reg coordinate = this->result; + + if (ir->projector) { + fs_reg inv_proj = fs_reg(this, glsl_type::float_type); + + ir->projector->accept(this); + emit(fs_inst(FS_OPCODE_RCP, inv_proj, this->result)); + + fs_reg proj_coordinate = fs_reg(this, ir->coordinate->type); + for (unsigned int i = 0; i < ir->coordinate->type->vector_elements; i++) { + emit(fs_inst(BRW_OPCODE_MUL, proj_coordinate, coordinate, inv_proj)); + coordinate.reg_offset++; + proj_coordinate.reg_offset++; + } + proj_coordinate.reg_offset = 0; + + coordinate = proj_coordinate; + } + + for (mlen = 0; mlen < ir->coordinate->type->vector_elements; mlen++) { + emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), coordinate)); + coordinate.reg_offset++; + } + + /* Pre-Ironlake, the 8-wide sampler always took u,v,r. */ + if (intel->gen < 5) + mlen = 3; + + if (ir->shadow_comparitor) { + /* For shadow comparisons, we have to supply u,v,r. */ + mlen = 3; + + ir->shadow_comparitor->accept(this); + emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result)); + mlen++; + } + + /* Do we ever want to handle writemasking on texture samples? Is it + * performance relevant? + */ + fs_reg dst = fs_reg(this, glsl_type::vec4_type); + + switch (ir->op) { + case ir_tex: + inst = emit(fs_inst(FS_OPCODE_TEX, dst, fs_reg(MRF, base_mrf))); + break; + case ir_txb: + ir->lod_info.bias->accept(this); + emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result)); + mlen++; + + inst = emit(fs_inst(FS_OPCODE_TXB, dst, fs_reg(MRF, base_mrf))); + break; + case ir_txl: + ir->lod_info.lod->accept(this); + emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result)); + mlen++; + + inst = emit(fs_inst(FS_OPCODE_TXL, dst, fs_reg(MRF, base_mrf))); + break; + case ir_txd: + case ir_txf: + assert(!"GLSL 1.30 features unsupported"); + break; + } + + this->result = dst; + + if (ir->shadow_comparitor) + inst->shadow_compare = true; + inst->mlen = mlen; +} + +void +fs_visitor::visit(ir_swizzle *ir) +{ + ir->val->accept(this); + fs_reg val = this->result; + + fs_reg result = fs_reg(this, ir->type); + this->result = result; + + for (unsigned int i = 0; i < ir->type->vector_elements; i++) { + fs_reg channel = val; + int swiz = 0; + + switch (i) { + case 0: + swiz = ir->mask.x; + break; + case 1: + swiz = ir->mask.y; + break; + case 2: + swiz = ir->mask.z; + break; + case 3: + swiz = ir->mask.w; + break; + } + + channel.reg_offset += swiz; + emit(fs_inst(BRW_OPCODE_MOV, result, channel)); + result.reg_offset++; + } +} + +void +fs_visitor::visit(ir_discard *ir) +{ + assert(ir->condition == NULL); /* FINISHME */ + + emit(fs_inst(FS_OPCODE_DISCARD)); +} + +void +fs_visitor::visit(ir_constant *ir) +{ + fs_reg reg(this, ir->type); + this->result = reg; + + for (unsigned int i = 0; i < ir->type->vector_elements; i++) { + switch (ir->type->base_type) { + case GLSL_TYPE_FLOAT: + emit(fs_inst(BRW_OPCODE_MOV, reg, fs_reg(ir->value.f[i]))); + break; + case GLSL_TYPE_UINT: + emit(fs_inst(BRW_OPCODE_MOV, reg, fs_reg(ir->value.u[i]))); + break; + case GLSL_TYPE_INT: + emit(fs_inst(BRW_OPCODE_MOV, reg, fs_reg(ir->value.i[i]))); + break; + case GLSL_TYPE_BOOL: + emit(fs_inst(BRW_OPCODE_MOV, reg, fs_reg((int)ir->value.b[i]))); + break; + default: + assert(!"Non-float/uint/int/bool constant"); + } + reg.reg_offset++; + } +} + +void +fs_visitor::visit(ir_if *ir) +{ + fs_inst *inst; + + /* Don't point the annotation at the if statement, because then it plus + * the then and else blocks get printed. + */ + this->base_ir = ir->condition; + + /* Generate the condition into the condition code. */ + ir->condition->accept(this); + inst = emit(fs_inst(BRW_OPCODE_MOV, fs_reg(brw_null_reg()), this->result)); + inst->conditional_mod = BRW_CONDITIONAL_NZ; + + inst = emit(fs_inst(BRW_OPCODE_IF)); + inst->predicated = true; + + foreach_iter(exec_list_iterator, iter, ir->then_instructions) { + ir_instruction *ir = (ir_instruction *)iter.get(); + this->base_ir = ir; + + ir->accept(this); + } + + if (!ir->else_instructions.is_empty()) { + emit(fs_inst(BRW_OPCODE_ELSE)); + + foreach_iter(exec_list_iterator, iter, ir->else_instructions) { + ir_instruction *ir = (ir_instruction *)iter.get(); + this->base_ir = ir; + + ir->accept(this); + } + } + + emit(fs_inst(BRW_OPCODE_ENDIF)); +} + +void +fs_visitor::visit(ir_loop *ir) +{ + assert(!ir->from); + assert(!ir->to); + assert(!ir->increment); + assert(!ir->counter); + + emit(fs_inst(BRW_OPCODE_DO)); + + /* Start a safety counter. If the user messed up their loop + * counting, we don't want to hang the GPU. + */ + fs_reg max_iter = fs_reg(this, glsl_type::int_type); + emit(fs_inst(BRW_OPCODE_MOV, max_iter, fs_reg(10000))); + + foreach_iter(exec_list_iterator, iter, ir->body_instructions) { + ir_instruction *ir = (ir_instruction *)iter.get(); + fs_inst *inst; + + this->base_ir = ir; + ir->accept(this); + + /* Check the maximum loop iters counter. */ + inst = emit(fs_inst(BRW_OPCODE_ADD, max_iter, max_iter, fs_reg(-1))); + inst->conditional_mod = BRW_CONDITIONAL_Z; + + inst = emit(fs_inst(BRW_OPCODE_BREAK)); + inst->predicated = true; + } + + emit(fs_inst(BRW_OPCODE_WHILE)); +} + +void +fs_visitor::visit(ir_loop_jump *ir) +{ + switch (ir->mode) { + case ir_loop_jump::jump_break: + emit(fs_inst(BRW_OPCODE_BREAK)); + break; + case ir_loop_jump::jump_continue: + emit(fs_inst(BRW_OPCODE_CONTINUE)); + break; + } +} + +void +fs_visitor::visit(ir_call *ir) +{ + assert(!"FINISHME"); +} + +void +fs_visitor::visit(ir_return *ir) +{ + assert(!"FINISHME"); +} + +void +fs_visitor::visit(ir_function *ir) +{ + /* Ignore function bodies other than main() -- we shouldn't see calls to + * them since they should all be inlined before we get to ir_to_mesa. + */ + if (strcmp(ir->name, "main") == 0) { + const ir_function_signature *sig; + exec_list empty; + + sig = ir->matching_signature(&empty); + + assert(sig); + + foreach_iter(exec_list_iterator, iter, sig->body) { + ir_instruction *ir = (ir_instruction *)iter.get(); + this->base_ir = ir; + + ir->accept(this); + } + } +} + +void +fs_visitor::visit(ir_function_signature *ir) +{ + assert(!"not reached"); + (void)ir; +} + +fs_inst * +fs_visitor::emit(fs_inst inst) +{ + fs_inst *list_inst = new(mem_ctx) fs_inst; + *list_inst = inst; + + list_inst->annotation = this->current_annotation; + list_inst->ir = this->base_ir; + + this->instructions.push_tail(list_inst); + + return list_inst; +} + +/** Emits a dummy fragment shader consisting of magenta for bringup purposes. */ +void +fs_visitor::emit_dummy_fs() +{ + /* Everyone's favorite color. */ + emit(fs_inst(BRW_OPCODE_MOV, + fs_reg(MRF, 2), + fs_reg(1.0f))); + emit(fs_inst(BRW_OPCODE_MOV, + fs_reg(MRF, 3), + fs_reg(0.0f))); + emit(fs_inst(BRW_OPCODE_MOV, + fs_reg(MRF, 4), + fs_reg(1.0f))); + emit(fs_inst(BRW_OPCODE_MOV, + fs_reg(MRF, 5), + fs_reg(0.0f))); + + fs_inst *write; + write = emit(fs_inst(FS_OPCODE_FB_WRITE, + fs_reg(0), + fs_reg(0))); +} + +/* The register location here is relative to the start of the URB + * data. It will get adjusted to be a real location before + * generate_code() time. + */ +struct brw_reg +fs_visitor::interp_reg(int location, int channel) +{ + int regnr = location * 2 + channel / 2; + int stride = (channel & 1) * 4; + + return brw_vec1_grf(regnr, stride); +} + +/** Emits the interpolation for the varying inputs. */ +void +fs_visitor::emit_interpolation() +{ + struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW); + /* For now, the source regs for the setup URB data will be unset, + * since we don't know until codegen how many push constants we'll + * use, and therefore what the setup URB offset is. + */ + fs_reg src_reg = reg_undef; + + this->current_annotation = "compute pixel centers"; + this->pixel_x = fs_reg(this, glsl_type::uint_type); + this->pixel_y = fs_reg(this, glsl_type::uint_type); + emit(fs_inst(BRW_OPCODE_ADD, + this->pixel_x, + fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)), + fs_reg(brw_imm_v(0x10101010)))); + emit(fs_inst(BRW_OPCODE_ADD, + this->pixel_y, + fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)), + fs_reg(brw_imm_v(0x11001100)))); + + this->current_annotation = "compute pixel deltas from v0"; + this->delta_x = fs_reg(this, glsl_type::float_type); + this->delta_y = fs_reg(this, glsl_type::float_type); + emit(fs_inst(BRW_OPCODE_ADD, + this->delta_x, + this->pixel_x, + fs_reg(negate(brw_vec1_grf(1, 0))))); + emit(fs_inst(BRW_OPCODE_ADD, + this->delta_y, + this->pixel_y, + fs_reg(brw_vec1_grf(1, 1)))); + + this->current_annotation = "compute pos.w and 1/pos.w"; + /* Compute wpos. Unlike many other varying inputs, we usually need it + * to produce 1/w, and the varying variable wouldn't show up. + */ + fs_reg wpos = fs_reg(this, glsl_type::vec4_type); + this->interp_attrs[FRAG_ATTRIB_WPOS] = wpos; + emit(fs_inst(BRW_OPCODE_MOV, wpos, this->pixel_x)); /* FINISHME: ARB_fcc */ + wpos.reg_offset++; + emit(fs_inst(BRW_OPCODE_MOV, wpos, this->pixel_y)); /* FINISHME: ARB_fcc */ + wpos.reg_offset++; + emit(fs_inst(FS_OPCODE_LINTERP, wpos, this->delta_x, this->delta_y, + interp_reg(FRAG_ATTRIB_WPOS, 2))); + wpos.reg_offset++; + emit(fs_inst(FS_OPCODE_LINTERP, wpos, this->delta_x, this->delta_y, + interp_reg(FRAG_ATTRIB_WPOS, 3))); + /* Compute the pixel W value from wpos.w. */ + this->pixel_w = fs_reg(this, glsl_type::float_type); + emit(fs_inst(FS_OPCODE_RCP, this->pixel_w, wpos)); + + /* FINISHME: gl_FrontFacing */ + + foreach_iter(exec_list_iterator, iter, *this->shader->ir) { + ir_instruction *ir = (ir_instruction *)iter.get(); + ir_variable *var = ir->as_variable(); + + if (!var) + continue; + + if (var->mode != ir_var_in) + continue; + + /* If it's already set up (WPOS), skip. */ + if (var->location == 0) + continue; + + this->current_annotation = talloc_asprintf(this->mem_ctx, + "interpolate %s " + "(FRAG_ATTRIB[%d])", + var->name, + var->location); + emit_pinterp(var->location); + } + this->current_annotation = NULL; +} + +void +fs_visitor::emit_pinterp(int location) +{ + fs_reg interp_attr = fs_reg(this, glsl_type::vec4_type); + this->interp_attrs[location] = interp_attr; + + for (unsigned int i = 0; i < 4; i++) { + struct brw_reg interp = interp_reg(location, i); + emit(fs_inst(FS_OPCODE_LINTERP, + interp_attr, + this->delta_x, + this->delta_y, + fs_reg(interp))); + interp_attr.reg_offset++; + } + interp_attr.reg_offset -= 4; + + for (unsigned int i = 0; i < 4; i++) { + emit(fs_inst(BRW_OPCODE_MUL, + interp_attr, + interp_attr, + this->pixel_w)); + interp_attr.reg_offset++; + } +} + +void +fs_visitor::emit_fb_writes() +{ + this->current_annotation = "FB write"; + + assert(this->frag_color || !"FINISHME: MRT"); + fs_reg color = *(variable_storage(this->frag_color)); + + for (int i = 0; i < 4; i++) { + emit(fs_inst(BRW_OPCODE_MOV, + fs_reg(MRF, 2 + i), + color)); + color.reg_offset++; + } + + emit(fs_inst(FS_OPCODE_FB_WRITE, + fs_reg(0), + fs_reg(0))); + + this->current_annotation = NULL; +} + +void +fs_visitor::generate_fb_write(fs_inst *inst) +{ + GLboolean eot = 1; /* FINISHME: MRT */ + /* FINISHME: AADS */ + + /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied + * move, here's g1. + */ + brw_push_insn_state(p); + brw_set_mask_control(p, BRW_MASK_DISABLE); + brw_set_compression_control(p, BRW_COMPRESSION_NONE); + brw_MOV(p, + brw_message_reg(1), + brw_vec8_grf(1, 0)); + brw_pop_insn_state(p); + + int nr = 2 + 4; + + brw_fb_WRITE(p, + 8, /* dispatch_width */ + retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW), + 0, /* base MRF */ + retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW), + 0, /* FINISHME: MRT target */ + nr, + 0, + eot); +} + +void +fs_visitor::generate_linterp(fs_inst *inst, + struct brw_reg dst, struct brw_reg *src) +{ + struct brw_reg delta_x = src[0]; + struct brw_reg delta_y = src[1]; + struct brw_reg interp = src[2]; + + if (brw->has_pln && + delta_y.nr == delta_x.nr + 1 && + (intel->gen >= 6 || (delta_x.nr & 1) == 0)) { + brw_PLN(p, dst, interp, delta_x); + } else { + brw_LINE(p, brw_null_reg(), interp, delta_x); + brw_MAC(p, dst, suboffset(interp, 1), delta_y); + } +} + +void +fs_visitor::generate_math(fs_inst *inst, + struct brw_reg dst, struct brw_reg *src) +{ + int op; + + switch (inst->opcode) { + case FS_OPCODE_RCP: + op = BRW_MATH_FUNCTION_INV; + break; + case FS_OPCODE_RSQ: + op = BRW_MATH_FUNCTION_RSQ; + break; + case FS_OPCODE_SQRT: + op = BRW_MATH_FUNCTION_SQRT; + break; + case FS_OPCODE_EXP2: + op = BRW_MATH_FUNCTION_EXP; + break; + case FS_OPCODE_LOG2: + op = BRW_MATH_FUNCTION_LOG; + break; + case FS_OPCODE_POW: + op = BRW_MATH_FUNCTION_POW; + break; + case FS_OPCODE_SIN: + op = BRW_MATH_FUNCTION_SIN; + break; + case FS_OPCODE_COS: + op = BRW_MATH_FUNCTION_COS; + break; + default: + assert(!"not reached: unknown math function"); + op = 0; + break; + } + + if (inst->opcode == FS_OPCODE_POW) { + brw_MOV(p, brw_message_reg(3), src[1]); + } + + brw_math(p, dst, + op, + inst->saturate ? BRW_MATH_SATURATE_SATURATE : + BRW_MATH_SATURATE_NONE, + 2, src[0], + BRW_MATH_DATA_VECTOR, + BRW_MATH_PRECISION_FULL); +} + +void +fs_visitor::generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src) +{ + int msg_type = -1; + int rlen = 4; + + if (intel->gen == 5) { + switch (inst->opcode) { + case FS_OPCODE_TEX: + if (inst->shadow_compare) { + msg_type = BRW_SAMPLER_MESSAGE_SAMPLE_COMPARE_GEN5; + } else { + msg_type = BRW_SAMPLER_MESSAGE_SAMPLE_GEN5; + } + break; + case FS_OPCODE_TXB: + if (inst->shadow_compare) { + msg_type = BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE_GEN5; + } else { + msg_type = BRW_SAMPLER_MESSAGE_SAMPLE_BIAS_GEN5; + } + break; + } + } else { + switch (inst->opcode) { + case FS_OPCODE_TEX: + /* Note that G45 and older determines shadow compare and dispatch width + * from message length for most messages. + */ + if (inst->shadow_compare) { + msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE; + } else { + msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE; + } + case FS_OPCODE_TXB: + if (inst->shadow_compare) { + assert(!"FINISHME: shadow compare with bias."); + msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS; + } else { + msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS; + rlen = 8; + } + break; + } + } + assert(msg_type != -1); + + /* g0 header. */ + src.nr--; + + brw_SAMPLE(p, + retype(dst, BRW_REGISTER_TYPE_UW), + src.nr, + retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW), + SURF_INDEX_TEXTURE(inst->sampler), + inst->sampler, + WRITEMASK_XYZW, + msg_type, + rlen, + inst->mlen + 1, + 0, + 1, + BRW_SAMPLER_SIMD_MODE_SIMD8); +} + +void +fs_visitor::generate_discard(fs_inst *inst) +{ + struct brw_reg g0 = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW); + brw_push_insn_state(p); + brw_set_mask_control(p, BRW_MASK_DISABLE); + brw_NOT(p, c->emit_mask_reg, brw_mask_reg(1)); /* IMASK */ + brw_AND(p, g0, c->emit_mask_reg, g0); + brw_pop_insn_state(p); +} + +static void +trivial_assign_reg(int header_size, fs_reg *reg) +{ + if (reg->file == GRF && reg->reg != 0) { + reg->hw_reg = header_size + reg->reg - 1 + reg->reg_offset; + reg->reg = 0; + } +} + +void +fs_visitor::assign_curb_setup() +{ + c->prog_data.first_curbe_grf = c->key.nr_payload_regs; + c->prog_data.curb_read_length = ALIGN(c->prog_data.nr_params, 8) / 8; + + /* Map the offsets in the UNIFORM file to fixed HW regs. */ + foreach_iter(exec_list_iterator, iter, this->instructions) { + fs_inst *inst = (fs_inst *)iter.get(); + + for (unsigned int i = 0; i < 3; i++) { + if (inst->src[i].file == UNIFORM) { + int constant_nr = inst->src[i].hw_reg + inst->src[i].reg_offset; + struct brw_reg brw_reg = brw_vec1_grf(c->prog_data.first_curbe_grf + + constant_nr / 8, + constant_nr % 8); + + inst->src[i].file = FIXED_HW_REG; + inst->src[i].fixed_hw_reg = brw_reg; + } + } + } +} + +void +fs_visitor::assign_urb_setup() +{ + int urb_start = c->prog_data.first_curbe_grf + c->prog_data.curb_read_length; + int interp_reg_nr[FRAG_ATTRIB_MAX]; + + c->prog_data.urb_read_length = 0; + + /* Figure out where each of the incoming setup attributes lands. */ + for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) { + interp_reg_nr[i] = -1; + + if (i != FRAG_ATTRIB_WPOS && + !(brw->fragment_program->Base.InputsRead & BITFIELD64_BIT(i))) + continue; + + /* Each attribute is 4 setup channels, each of which is half a reg. */ + interp_reg_nr[i] = urb_start + c->prog_data.urb_read_length; + c->prog_data.urb_read_length += 2; + } + + /* Map the register numbers for FS_OPCODE_LINTERP so that it uses + * the correct setup input. + */ + foreach_iter(exec_list_iterator, iter, this->instructions) { + fs_inst *inst = (fs_inst *)iter.get(); + + if (inst->opcode != FS_OPCODE_LINTERP) + continue; + + assert(inst->src[2].file == FIXED_HW_REG); + + int location = inst->src[2].fixed_hw_reg.nr / 2; + assert(interp_reg_nr[location] != -1); + inst->src[2].fixed_hw_reg.nr = (interp_reg_nr[location] + + (inst->src[2].fixed_hw_reg.nr & 1)); + } + + this->first_non_payload_grf = urb_start + c->prog_data.urb_read_length; +} + +void +fs_visitor::assign_regs() +{ + int header_size = this->first_non_payload_grf; + int last_grf = 0; + + /* FINISHME: trivial assignment of register numbers */ + foreach_iter(exec_list_iterator, iter, this->instructions) { + fs_inst *inst = (fs_inst *)iter.get(); + + trivial_assign_reg(header_size, &inst->dst); + trivial_assign_reg(header_size, &inst->src[0]); + trivial_assign_reg(header_size, &inst->src[1]); + + last_grf = MAX2(last_grf, inst->dst.hw_reg); + last_grf = MAX2(last_grf, inst->src[0].hw_reg); + last_grf = MAX2(last_grf, inst->src[1].hw_reg); + } + + this->grf_used = last_grf + 1; +} + +static struct brw_reg brw_reg_from_fs_reg(fs_reg *reg) +{ + struct brw_reg brw_reg; + + switch (reg->file) { + case GRF: + case ARF: + case MRF: + brw_reg = brw_vec8_reg(reg->file, + reg->hw_reg, 0); + brw_reg = retype(brw_reg, reg->type); + break; + case IMM: + switch (reg->type) { + case BRW_REGISTER_TYPE_F: + brw_reg = brw_imm_f(reg->imm.f); + break; + case BRW_REGISTER_TYPE_D: + brw_reg = brw_imm_d(reg->imm.i); + break; + case BRW_REGISTER_TYPE_UD: + brw_reg = brw_imm_ud(reg->imm.u); + break; + default: + assert(!"not reached"); + break; + } + break; + case FIXED_HW_REG: + brw_reg = reg->fixed_hw_reg; + break; + case BAD_FILE: + /* Probably unused. */ + brw_reg = brw_null_reg(); + break; + case UNIFORM: + assert(!"not reached"); + brw_reg = brw_null_reg(); + break; + } + if (reg->abs) + brw_reg = brw_abs(brw_reg); + if (reg->negate) + brw_reg = negate(brw_reg); + + return brw_reg; +} + +void +fs_visitor::generate_code() +{ + unsigned int annotation_len = 0; + int last_native_inst = 0; + struct brw_instruction *if_stack[16], *loop_stack[16]; + int if_stack_depth = 0, loop_stack_depth = 0; + int if_depth_in_loop[16]; + + if_depth_in_loop[loop_stack_depth] = 0; + + memset(&if_stack, 0, sizeof(if_stack)); + foreach_iter(exec_list_iterator, iter, this->instructions) { + fs_inst *inst = (fs_inst *)iter.get(); + struct brw_reg src[3], dst; + + for (unsigned int i = 0; i < 3; i++) { + src[i] = brw_reg_from_fs_reg(&inst->src[i]); + } + dst = brw_reg_from_fs_reg(&inst->dst); + + brw_set_conditionalmod(p, inst->conditional_mod); + brw_set_predicate_control(p, inst->predicated); + + switch (inst->opcode) { + case BRW_OPCODE_MOV: + brw_MOV(p, dst, src[0]); + break; + case BRW_OPCODE_ADD: + brw_ADD(p, dst, src[0], src[1]); + break; + case BRW_OPCODE_MUL: + brw_MUL(p, dst, src[0], src[1]); + break; + + case BRW_OPCODE_FRC: + brw_FRC(p, dst, src[0]); + break; + case BRW_OPCODE_RNDD: + brw_RNDD(p, dst, src[0]); + break; + case BRW_OPCODE_RNDZ: + brw_RNDZ(p, dst, src[0]); + break; + + case BRW_OPCODE_AND: + brw_AND(p, dst, src[0], src[1]); + break; + case BRW_OPCODE_OR: + brw_OR(p, dst, src[0], src[1]); + break; + case BRW_OPCODE_XOR: + brw_XOR(p, dst, src[0], src[1]); + break; + + case BRW_OPCODE_CMP: + brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]); + break; + case BRW_OPCODE_SEL: + brw_SEL(p, dst, src[0], src[1]); + break; + + case BRW_OPCODE_IF: + assert(if_stack_depth < 16); + if_stack[if_stack_depth] = brw_IF(p, BRW_EXECUTE_8); + if_stack_depth++; + break; + case BRW_OPCODE_ELSE: + if_stack[if_stack_depth - 1] = + brw_ELSE(p, if_stack[if_stack_depth - 1]); + break; + case BRW_OPCODE_ENDIF: + if_stack_depth--; + brw_ENDIF(p , if_stack[if_stack_depth]); + break; + + case BRW_OPCODE_DO: + loop_stack[loop_stack_depth++] = brw_DO(p, BRW_EXECUTE_8); + if_depth_in_loop[loop_stack_depth] = 0; + break; + + case BRW_OPCODE_BREAK: + brw_BREAK(p, if_depth_in_loop[loop_stack_depth]); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); + break; + case BRW_OPCODE_CONTINUE: + brw_CONT(p, if_depth_in_loop[loop_stack_depth]); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); + break; + + case BRW_OPCODE_WHILE: { + struct brw_instruction *inst0, *inst1; + GLuint br = 1; + + if (intel->gen == 5) + br = 2; + + assert(loop_stack_depth > 0); + loop_stack_depth--; + inst0 = inst1 = brw_WHILE(p, loop_stack[loop_stack_depth]); + /* patch all the BREAK/CONT instructions from last BGNLOOP */ + while (inst0 > loop_stack[loop_stack_depth]) { + inst0--; + if (inst0->header.opcode == BRW_OPCODE_BREAK && + inst0->bits3.if_else.jump_count == 0) { + inst0->bits3.if_else.jump_count = br * (inst1 - inst0 + 1); + } + else if (inst0->header.opcode == BRW_OPCODE_CONTINUE && + inst0->bits3.if_else.jump_count == 0) { + inst0->bits3.if_else.jump_count = br * (inst1 - inst0); + } + } + } + break; + + case FS_OPCODE_RCP: + case FS_OPCODE_RSQ: + case FS_OPCODE_SQRT: + case FS_OPCODE_EXP2: + case FS_OPCODE_LOG2: + case FS_OPCODE_POW: + case FS_OPCODE_SIN: + case FS_OPCODE_COS: + generate_math(inst, dst, src); + break; + case FS_OPCODE_LINTERP: + generate_linterp(inst, dst, src); + break; + case FS_OPCODE_TEX: + case FS_OPCODE_TXB: + case FS_OPCODE_TXL: + generate_tex(inst, dst, src[0]); + break; + case FS_OPCODE_DISCARD: + generate_discard(inst); + break; + case FS_OPCODE_FB_WRITE: + generate_fb_write(inst); + break; + default: + if (inst->opcode < (int)ARRAY_SIZE(brw_opcodes)) { + _mesa_problem(ctx, "Unsupported opcode `%s' in FS", + brw_opcodes[inst->opcode].name); + } else { + _mesa_problem(ctx, "Unsupported opcode %d in FS", inst->opcode); + } + this->fail = true; + } + + if (annotation_len < p->nr_insn) { + annotation_len *= 2; + if (annotation_len < 16) + annotation_len = 16; + + this->annotation_string = talloc_realloc(this->mem_ctx, + annotation_string, + const char *, + annotation_len); + this->annotation_ir = talloc_realloc(this->mem_ctx, + annotation_ir, + ir_instruction *, + annotation_len); + } + + for (unsigned int i = last_native_inst; i < p->nr_insn; i++) { + this->annotation_string[i] = inst->annotation; + this->annotation_ir[i] = inst->ir; + } + last_native_inst = p->nr_insn; + } +} + +GLboolean +brw_wm_fs_emit(struct brw_context *brw, struct brw_wm_compile *c) +{ + struct brw_compile *p = &c->func; + struct intel_context *intel = &brw->intel; + GLcontext *ctx = &intel->ctx; + struct brw_shader *shader = NULL; + struct gl_shader_program *prog = ctx->Shader.CurrentProgram; + + if (!prog) + return GL_FALSE; + + if (!using_new_fs) + return GL_FALSE; + + for (unsigned int i = 0; i < prog->_NumLinkedShaders; i++) { + if (prog->_LinkedShaders[i]->Type == GL_FRAGMENT_SHADER) { + shader = (struct brw_shader *)prog->_LinkedShaders[i]; + break; + } + } + if (!shader) + return GL_FALSE; + + /* We always use 8-wide mode, at least for now. For one, flow + * control only works in 8-wide. Also, when we're fragment shader + * bound, we're almost always under register pressure as well, so + * 8-wide would save us from the performance cliff of spilling + * regs. + */ + c->dispatch_width = 8; + + if (INTEL_DEBUG & DEBUG_WM) { + printf("GLSL IR for native fragment shader %d:\n", prog->Name); + _mesa_print_ir(shader->ir, NULL); + printf("\n"); + } + + /* Now the main event: Visit the shader IR and generate our FS IR for it. + */ + fs_visitor v(c, shader); + + if (0) { + v.emit_dummy_fs(); + } else { + v.emit_interpolation(); + + /* Generate FS IR for main(). (the visitor only descends into + * functions called "main"). + */ + foreach_iter(exec_list_iterator, iter, *shader->ir) { + ir_instruction *ir = (ir_instruction *)iter.get(); + v.base_ir = ir; + ir->accept(&v); + } + + if (v.fail) + return GL_FALSE; + + v.emit_fb_writes(); + v.assign_curb_setup(); + v.assign_urb_setup(); + v.assign_regs(); + } + + v.generate_code(); + + if (INTEL_DEBUG & DEBUG_WM) { + const char *last_annotation_string = NULL; + ir_instruction *last_annotation_ir = NULL; + + printf("Native code for fragment shader %d:\n", prog->Name); + for (unsigned int i = 0; i < p->nr_insn; i++) { + if (last_annotation_ir != v.annotation_ir[i]) { + last_annotation_ir = v.annotation_ir[i]; + if (last_annotation_ir) { + printf(" "); + last_annotation_ir->print(); + printf("\n"); + } + } + if (last_annotation_string != v.annotation_string[i]) { + last_annotation_string = v.annotation_string[i]; + if (last_annotation_string) + printf(" %s\n", last_annotation_string); + } + brw_disasm(stdout, &p->store[i], intel->gen); + } + printf("\n"); + } + + c->prog_data.total_grf = v.grf_used; + c->prog_data.total_scratch = 0; + + return GL_TRUE; +} diff --git a/src/mesa/drivers/dri/i965/brw_fs_channel_expressions.cpp b/src/mesa/drivers/dri/i965/brw_fs_channel_expressions.cpp new file mode 100644 index 0000000000..d8d58a9467 --- /dev/null +++ b/src/mesa/drivers/dri/i965/brw_fs_channel_expressions.cpp @@ -0,0 +1,365 @@ +/* + * Copyright © 2010 Intel Corporation + * + * 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, sublicense, + * 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 NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS 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. + */ + +/** + * \file brw_wm_channel_expressions.cpp + * + * Breaks vector operations down into operations on each component. + * + * The 965 fragment shader receives 8 or 16 pixels at a time, so each + * channel of a vector is laid out as 1 or 2 8-float registers. Each + * ALU operation operates on one of those channel registers. As a + * result, there is no value to the 965 fragment shader in tracking + * "vector" expressions in the sense of GLSL fragment shaders, when + * doing a channel at a time may help in constant folding, algebraic + * simplification, and reducing the liveness of channel registers. + * + * The exception to the desire to break everything down to floats is + * texturing. The texture sampler returns a writemasked masked + * 4/8-register sequence containing the texture values. We don't want + * to dispatch to the sampler separately for each channel we need, so + * we do retain the vector types in that case. + */ + +extern "C" { +#include "main/core.h" +#include "brw_wm.h" +} +#include "../glsl/ir.h" +#include "../glsl/ir_expression_flattening.h" +#include "../glsl/glsl_types.h" + +class ir_channel_expressions_visitor : public ir_hierarchical_visitor { +public: + ir_channel_expressions_visitor() + { + this->progress = false; + this->mem_ctx = NULL; + } + + ir_visitor_status visit_leave(ir_assignment *); + + ir_rvalue *get_element(ir_variable *var, unsigned int element); + void assign(ir_assignment *ir, int elem, ir_rvalue *val); + + bool progress; + void *mem_ctx; +}; + +static bool +channel_expressions_predicate(ir_instruction *ir) +{ + ir_expression *expr = ir->as_expression(); + unsigned int i; + + if (!expr) + return false; + + for (i = 0; i < expr->get_num_operands(); i++) { + if (expr->operands[i]->type->is_vector()) + return true; + } + + return false; +} + +extern "C" { +GLboolean +brw_do_channel_expressions(exec_list *instructions) +{ + ir_channel_expressions_visitor v; + + /* Pull out any matrix expression to a separate assignment to a + * temp. This will make our handling of the breakdown to + * operations on the matrix's vector components much easier. + */ + do_expression_flattening(instructions, channel_expressions_predicate); + + visit_list_elements(&v, instructions); + + return v.progress; +} +} + +ir_rvalue * +ir_channel_expressions_visitor::get_element(ir_variable *var, unsigned int elem) +{ + ir_dereference *deref; + + if (var->type->is_scalar()) + return new(mem_ctx) ir_dereference_variable(var); + + assert(elem < var->type->components()); + deref = new(mem_ctx) ir_dereference_variable(var); + return new(mem_ctx) ir_swizzle(deref, elem, 0, 0, 0, 1); +} + +void +ir_channel_expressions_visitor::assign(ir_assignment *ir, int elem, ir_rvalue *val) +{ + ir_dereference *lhs = ir->lhs->clone(mem_ctx, NULL); + ir_assignment *assign; + ir_swizzle *val_swiz; + + /* This assign-of-expression should have been generated by the + * expression flattening visitor (since we never short circit to + * not flatten, even for plain assignments of variables), so the + * writemask is always full. + */ + assert(ir->write_mask == (1 << ir->lhs->type->components()) - 1); + + /* Smear the float across all the channels for the masked write. */ + val_swiz = new(mem_ctx) ir_swizzle(val, 0, 0, 0, 0, + ir->lhs->type->components()); + assign = new(mem_ctx) ir_assignment(lhs, val_swiz, NULL, (1 << elem)); + ir->insert_before(assign); +} + +ir_visitor_status +ir_channel_expressions_visitor::visit_leave(ir_assignment *ir) +{ + ir_expression *expr = ir->rhs->as_expression(); + bool found_vector = false; + unsigned int i, vector_elements = 1; + ir_variable *op_var[2]; + + if (!expr) + return visit_continue; + + if (!this->mem_ctx) + this->mem_ctx = talloc_parent(ir); + + for (i = 0; i < expr->get_num_operands(); i++) { + if (expr->operands[i]->type->is_vector()) { + found_vector = true; + vector_elements = expr->operands[i]->type->vector_elements; + break; + } + } + if (!found_vector) + return visit_continue; + + /* Store the expression operands in temps so we can use them + * multiple times. + */ + for (i = 0; i < expr->get_num_operands(); i++) { + ir_assignment *assign; + ir_dereference *deref; + + assert(!expr->operands[i]->type->is_matrix()); + + op_var[i] = new(mem_ctx) ir_variable(expr->operands[i]->type, + "channel_expressions", + ir_var_temporary); + ir->insert_before(op_var[i]); + + deref = new(mem_ctx) ir_dereference_variable(op_var[i]); + assign = new(mem_ctx) ir_assignment(deref, + expr->operands[i], + NULL); + ir->insert_before(assign); + } + + const glsl_type *element_type = glsl_type::get_instance(ir->lhs->type->base_type, + 1, 1); + + /* OK, time to break down this vector operation. */ + switch (expr->operation) { + case ir_unop_bit_not: + case ir_unop_logic_not: + case ir_unop_neg: + case ir_unop_abs: + case ir_unop_sign: + case ir_unop_rcp: + case ir_unop_rsq: + case ir_unop_sqrt: + case ir_unop_exp: + case ir_unop_log: + case ir_unop_exp2: + case ir_unop_log2: + case ir_unop_f2i: + case ir_unop_i2f: + case ir_unop_f2b: + case ir_unop_b2f: + case ir_unop_i2b: + case ir_unop_b2i: + case ir_unop_u2f: + case ir_unop_trunc: + case ir_unop_ceil: + case ir_unop_floor: + case ir_unop_fract: + case ir_unop_sin: + case ir_unop_cos: + case ir_unop_dFdx: + case ir_unop_dFdy: + for (i = 0; i < vector_elements; i++) { + ir_rvalue *op0 = get_element(op_var[0], i); + + assign(ir, i, new(mem_ctx) ir_expression(expr->operation, + element_type, + op0, + NULL)); + } + break; + + case ir_binop_add: + case ir_binop_sub: + case ir_binop_mul: + case ir_binop_div: + case ir_binop_mod: + case ir_binop_min: + case ir_binop_max: + case ir_binop_pow: + case ir_binop_lshift: + case ir_binop_rshift: + case ir_binop_bit_and: + case ir_binop_bit_xor: + case ir_binop_bit_or: + for (i = 0; i < vector_elements; i++) { + ir_rvalue *op0 = get_element(op_var[0], i); + ir_rvalue *op1 = get_element(op_var[1], i); + + assign(ir, i, new(mem_ctx) ir_expression(expr->operation, + element_type, + op0, + op1)); + } + break; + + case ir_unop_any: { + ir_expression *temp; + temp = new(mem_ctx) ir_expression(ir_binop_logic_or, + element_type, + get_element(op_var[0], 0), + get_element(op_var[0], 1)); + + for (i = 2; i < vector_elements; i++) { + temp = new(mem_ctx) ir_expression(ir_binop_logic_or, + element_type, + get_element(op_var[0], i), + temp); + } + assign(ir, 0, temp); + break; + } + + case ir_binop_dot: { + ir_expression *last = NULL; + for (i = 0; i < vector_elements; i++) { + ir_rvalue *op0 = get_element(op_var[0], i); + ir_rvalue *op1 = get_element(op_var[1], i); + ir_expression *temp; + + temp = new(mem_ctx) ir_expression(ir_binop_mul, + element_type, + op0, + op1); + if (last) { + last = new(mem_ctx) ir_expression(ir_binop_add, + element_type, + temp, + last); + } else { + last = temp; + } + } + assign(ir, 0, last); + break; + } + + case ir_binop_cross: { + for (i = 0; i < vector_elements; i++) { + int swiz0 = (i + 1) % 3; + int swiz1 = (i + 2) % 3; + ir_expression *temp1, *temp2; + + temp1 = new(mem_ctx) ir_expression(ir_binop_mul, + element_type, + get_element(op_var[0], swiz0), + get_element(op_var[1], swiz1)); + + temp2 = new(mem_ctx) ir_expression(ir_binop_mul, + element_type, + get_element(op_var[1], swiz0), + get_element(op_var[0], swiz1)); + + temp2 = new(mem_ctx) ir_expression(ir_unop_neg, + element_type, + temp2, + NULL); + + assign(ir, i, new(mem_ctx) ir_expression(ir_binop_add, + element_type, + temp1, temp2)); + } + break; + } + + case ir_binop_less: + case ir_binop_greater: + case ir_binop_lequal: + case ir_binop_gequal: + case ir_binop_logic_and: + case ir_binop_logic_xor: + case ir_binop_logic_or: + ir->print(); + printf("\n"); + assert(!"not reached: expression operates on scalars only"); + break; + case ir_binop_equal: + case ir_binop_nequal: { + ir_expression *last = NULL; + for (i = 0; i < vector_elements; i++) { + ir_rvalue *op0 = get_element(op_var[0], i); + ir_rvalue *op1 = get_element(op_var[1], i); + ir_expression *temp; + ir_expression_operation join; + + if (expr->operation == ir_binop_equal) + join = ir_binop_logic_and; + else + join = ir_binop_logic_or; + + temp = new(mem_ctx) ir_expression(expr->operation, + element_type, + op0, + op1); + if (last) { + last = new(mem_ctx) ir_expression(join, + element_type, + temp, + last); + } else { + last = temp; + } + } + assign(ir, 0, last); + break; + } + } + + ir->remove(); + this->progress = true; + + return visit_continue; +} diff --git a/src/mesa/drivers/dri/i965/brw_fs_vector_splitting.cpp b/src/mesa/drivers/dri/i965/brw_fs_vector_splitting.cpp new file mode 100644 index 0000000000..00d5c20248 --- /dev/null +++ b/src/mesa/drivers/dri/i965/brw_fs_vector_splitting.cpp @@ -0,0 +1,391 @@ +/* + * Copyright © 2010 Intel Corporation + * + * 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, sublicense, + * 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 NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS 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. + */ + +/** + * \file brw_wm_vector_splitting.cpp + * + * If a vector is only ever referenced by its components, then + * split those components out to individual variables so they can be + * handled normally by other optimization passes. + * + * This skips vectors in uniforms and varyings, which need to be + * accessible as vectors for their access by the GL. Also, vector + * results of non-variable-derefs in assignments aren't handled + * because to do so we would have to store the vector result to a + * temporary in order to unload each channel, and to do so would just + * loop us back to where we started. For the 965, this is exactly the + * behavior we want for the results of texture lookups, but probably not for + */ + +extern "C" { +#include "main/core.h" +#include "intel_context.h" +} +#include "../glsl/ir.h" +#include "../glsl/ir_visitor.h" +#include "../glsl/ir_print_visitor.h" +#include "../glsl/ir_rvalue_visitor.h" +#include "../glsl/glsl_types.h" + +static bool debug = false; + +class variable_entry : public exec_node +{ +public: + variable_entry(ir_variable *var) + { + this->var = var; + this->whole_vector_access = 0; + this->declaration = false; + this->mem_ctx = NULL; + } + + ir_variable *var; /* The key: the variable's pointer. */ + + /** Number of times the variable is referenced, including assignments. */ + unsigned whole_vector_access; + + bool declaration; /* If the variable had a decl in the instruction stream */ + + ir_variable *components[4]; + + /** talloc_parent(this->var) -- the shader's talloc context. */ + void *mem_ctx; +}; + +class ir_vector_reference_visitor : public ir_hierarchical_visitor { +public: + ir_vector_reference_visitor(void) + { + this->mem_ctx = talloc_new(NULL); + this->variable_list.make_empty(); + } + + ~ir_vector_reference_visitor(void) + { + talloc_free(mem_ctx); + } + + virtual ir_visitor_status visit(ir_variable *); + virtual ir_visitor_status visit(ir_dereference_variable *); + virtual ir_visitor_status visit_enter(ir_swizzle *); + virtual ir_visitor_status visit_enter(ir_assignment *); + virtual ir_visitor_status visit_enter(ir_function_signature *); + + variable_entry *get_variable_entry(ir_variable *var); + + /* List of variable_entry */ + exec_list variable_list; + + void *mem_ctx; +}; + +variable_entry * +ir_vector_reference_visitor::get_variable_entry(ir_variable *var) +{ + assert(var); + + if (!var->type->is_vector()) + return NULL; + + switch (var->mode) { + case ir_var_uniform: + case ir_var_in: + case ir_var_out: + case ir_var_inout: + /* Can't split varyings or uniforms. Function in/outs won't get split + * either, so don't care about the ambiguity. + */ + return NULL; + case ir_var_auto: + case ir_var_temporary: + break; + } + + foreach_iter(exec_list_iterator, iter, this->variable_list) { + variable_entry *entry = (variable_entry *)iter.get(); + if (entry->var == var) + return entry; + } + + variable_entry *entry = new(mem_ctx) variable_entry(var); + this->variable_list.push_tail(entry); + return entry; +} + + +ir_visitor_status +ir_vector_reference_visitor::visit(ir_variable *ir) +{ + variable_entry *entry = this->get_variable_entry(ir); + + if (entry) + entry->declaration = true; + + return visit_continue; +} + +ir_visitor_status +ir_vector_reference_visitor::visit(ir_dereference_variable *ir) +{ + ir_variable *const var = ir->var; + variable_entry *entry = this->get_variable_entry(var); + + if (entry) + entry->whole_vector_access++; + + return visit_continue; +} + +ir_visitor_status +ir_vector_reference_visitor::visit_enter(ir_swizzle *ir) +{ + /* Don't descend into a vector ir_dereference_variable below. */ + if (ir->val->as_dereference_variable() && ir->type->is_scalar()) + return visit_continue_with_parent; + + return visit_continue; +} + +ir_visitor_status +ir_vector_reference_visitor::visit_enter(ir_assignment *ir) +{ + if (ir->lhs->as_dereference_variable() && + ir->rhs->as_dereference_variable() && + !ir->condition) { + /* We'll split copies of a vector to copies of channels, so don't + * descend to the ir_dereference_variables. + */ + return visit_continue_with_parent; + } + if (ir->lhs->as_dereference_variable() && + is_power_of_two(ir->write_mask) && + !ir->condition) { + /* If we're writing just a channel, then channel-splitting the LHS is OK. + */ + ir->rhs->accept(this); + return visit_continue_with_parent; + } + return visit_continue; +} + +ir_visitor_status +ir_vector_reference_visitor::visit_enter(ir_function_signature *ir) +{ + /* We don't want to descend into the function parameters and + * split them, so just accept the body here. + */ + visit_list_elements(this, &ir->body); + return visit_continue_with_parent; +} + +class ir_vector_splitting_visitor : public ir_rvalue_visitor { +public: + ir_vector_splitting_visitor(exec_list *vars) + { + this->variable_list = vars; + } + + virtual ir_visitor_status visit_leave(ir_assignment *); + + void handle_rvalue(ir_rvalue **rvalue); + struct variable_entry *get_splitting_entry(ir_variable *var); + + exec_list *variable_list; + void *mem_ctx; +}; + +struct variable_entry * +ir_vector_splitting_visitor::get_splitting_entry(ir_variable *var) +{ + assert(var); + + if (!var->type->is_vector()) + return NULL; + + foreach_iter(exec_list_iterator, iter, *this->variable_list) { + variable_entry *entry = (variable_entry *)iter.get(); + if (entry->var == var) { + return entry; + } + } + + return NULL; +} + +void +ir_vector_splitting_visitor::handle_rvalue(ir_rvalue **rvalue) +{ + if (!*rvalue) + return; + + ir_swizzle *swiz = (*rvalue)->as_swizzle(); + if (!swiz || !swiz->type->is_scalar()) + return; + + ir_dereference_variable *deref_var = swiz->val->as_dereference_variable(); + if (!deref_var) + return; + + variable_entry *entry = get_splitting_entry(deref_var->var); + if (!entry) + return; + + ir_variable *var = entry->components[swiz->mask.x]; + *rvalue = new(entry->mem_ctx) ir_dereference_variable(var); +} + +ir_visitor_status +ir_vector_splitting_visitor::visit_leave(ir_assignment *ir) +{ + ir_dereference_variable *lhs_deref = ir->lhs->as_dereference_variable(); + ir_dereference_variable *rhs_deref = ir->rhs->as_dereference_variable(); + variable_entry *lhs = lhs_deref ? get_splitting_entry(lhs_deref->var) : NULL; + variable_entry *rhs = rhs_deref ? get_splitting_entry(rhs_deref->var) : NULL; + + if (lhs_deref && rhs_deref && (lhs || rhs) && !ir->condition) { + /* Straight assignment of vector variables. */ + for (unsigned int i = 0; i < ir->rhs->type->vector_elements; i++) { + ir_dereference *new_lhs; + ir_rvalue *new_rhs; + void *mem_ctx = lhs ? lhs->mem_ctx : rhs->mem_ctx; + unsigned int writemask; + + if (lhs) { + new_lhs = new(mem_ctx) ir_dereference_variable(lhs->components[i]); + writemask = (ir->write_mask >> i) & 1; + } else { + new_lhs = ir->lhs->clone(mem_ctx, NULL); + writemask = ir->write_mask & (1 << i); + } + + if (rhs) { + new_rhs = new(mem_ctx) ir_dereference_variable(rhs->components[i]); + /* If we're writing into a writemask, smear it out to that channel. */ + if (!lhs) + new_rhs = new(mem_ctx) ir_swizzle(new_rhs, i, i, i, i, i + 1); + } else { + new_rhs = new(mem_ctx) ir_swizzle(ir->rhs->clone(mem_ctx, NULL), + i, i, i, i, 1); + } + + ir->insert_before(new(mem_ctx) ir_assignment(new_lhs, + new_rhs, + NULL, writemask)); + } + ir->remove(); + } else if (lhs) { + int elem = -1; + + switch (ir->write_mask) { + case (1 << 0): + elem = 0; + break; + case (1 << 1): + elem = 1; + break; + case (1 << 2): + elem = 2; + break; + case (1 << 3): + elem = 3; + break; + default: + ir->print(); + assert(!"not reached: non-channelwise dereference of LHS."); + } + + ir->lhs = new(mem_ctx) ir_dereference_variable(lhs->components[elem]); + ir->write_mask = (1 << 0); + + handle_rvalue(&ir->rhs); + ir->rhs = new(mem_ctx) ir_swizzle(ir->rhs, + elem, elem, elem, elem, 1); + } else { + handle_rvalue(&ir->rhs); + } + + handle_rvalue(&ir->condition); + + return visit_continue; +} + +extern "C" { +bool +brw_do_vector_splitting(exec_list *instructions) +{ + ir_vector_reference_visitor refs; + + visit_list_elements(&refs, instructions); + + /* Trim out variables we can't split. */ + foreach_iter(exec_list_iterator, iter, refs.variable_list) { + variable_entry *entry = (variable_entry *)iter.get(); + + if (debug) { + printf("vector %s@%p: decl %d, whole_access %d\n", + entry->var->name, (void *) entry->var, entry->declaration, + entry->whole_vector_access); + } + + if (!entry->declaration || entry->whole_vector_access) { + entry->remove(); + } + } + + if (refs.variable_list.is_empty()) + return false; + + void *mem_ctx = talloc_new(NULL); + + /* Replace the decls of the vectors to be split with their split + * components. + */ + foreach_iter(exec_list_iterator, iter, refs.variable_list) { + variable_entry *entry = (variable_entry *)iter.get(); + const struct glsl_type *type; + type = glsl_type::get_instance(entry->var->type->base_type, 1, 1); + + entry->mem_ctx = talloc_parent(entry->var); + + for (unsigned int i = 0; i < entry->var->type->vector_elements; i++) { + const char *name = talloc_asprintf(mem_ctx, "%s_%c", + entry->var->name, + "xyzw"[i]); + + entry->components[i] = new(entry->mem_ctx) ir_variable(type, name, + ir_var_temporary); + entry->var->insert_before(entry->components[i]); + } + + entry->var->remove(); + } + + ir_vector_splitting_visitor split(&refs.variable_list); + visit_list_elements(&split, instructions); + + talloc_free(mem_ctx); + + return true; +} +} diff --git a/src/mesa/drivers/dri/i965/brw_misc_state.c b/src/mesa/drivers/dri/i965/brw_misc_state.c index 572175f463..6eeaba7772 100644 --- a/src/mesa/drivers/dri/i965/brw_misc_state.c +++ b/src/mesa/drivers/dri/i965/brw_misc_state.c @@ -281,7 +281,7 @@ static void emit_depthbuffer(struct brw_context *brw) } assert(region->tiling != I915_TILING_X); - if (IS_GEN6(intel->intelScreen->deviceID)) + if (intel->gen >= 6) assert(region->tiling != I915_TILING_NONE); BEGIN_BATCH(len); @@ -295,7 +295,7 @@ static void emit_depthbuffer(struct brw_context *brw) I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER, 0); OUT_BATCH((BRW_SURFACE_MIPMAPLAYOUT_BELOW << 1) | - ((region->pitch - 1) << 6) | + ((region->width - 1) << 6) | ((region->height - 1) << 19)); OUT_BATCH(0); diff --git a/src/mesa/drivers/dri/i965/brw_optimize.c b/src/mesa/drivers/dri/i965/brw_optimize.c index 8aa6fb6cc6..cbed2bd5cb 100644 --- a/src/mesa/drivers/dri/i965/brw_optimize.c +++ b/src/mesa/drivers/dri/i965/brw_optimize.c @@ -32,12 +32,7 @@ #include "brw_defines.h" #include "brw_eu.h" -static const struct { - char *name; - int nsrc; - int ndst; - GLboolean is_arith; -} inst_opcode[128] = { +const struct brw_instruction_info brw_opcodes[128] = { [BRW_OPCODE_MOV] = { .name = "mov", .nsrc = 1, .ndst = 1, .is_arith = 1 }, [BRW_OPCODE_FRC] = { .name = "frc", .nsrc = 1, .ndst = 1, .is_arith = 1 }, [BRW_OPCODE_RNDU] = { .name = "rndu", .nsrc = 1, .ndst = 1, .is_arith = 1 }, @@ -94,7 +89,7 @@ static const struct { static INLINE GLboolean brw_is_arithmetic_inst(const struct brw_instruction *inst) { - return inst_opcode[inst->header.opcode].is_arith; + return brw_opcodes[inst->header.opcode].is_arith; } static const GLuint inst_stride[7] = { @@ -122,7 +117,7 @@ brw_is_grf_written(const struct brw_instruction *inst, int reg_index, int size, int gen) { - if (inst_opcode[inst->header.opcode].ndst == 0) + if (brw_opcodes[inst->header.opcode].ndst == 0) return GL_FALSE; if (inst->bits1.da1.dest_address_mode != BRW_ADDRESS_DIRECT) @@ -161,20 +156,19 @@ brw_is_grf_written(const struct brw_instruction *inst, return left < right; } -/* Specific path for message register since we need to handle the compr4 case */ -static INLINE GLboolean -brw_is_mrf_written(const struct brw_instruction *inst, int reg_index, int size) +static GLboolean +brw_is_mrf_written_alu(const struct brw_instruction *inst, + int reg_index, int size) { - if (inst_opcode[inst->header.opcode].ndst == 0) + if (brw_opcodes[inst->header.opcode].ndst == 0) return GL_FALSE; - if (inst->bits1.da1.dest_address_mode != BRW_ADDRESS_DIRECT) - if (inst->bits1.ia1.dest_reg_file == BRW_MESSAGE_REGISTER_FILE) - return GL_TRUE; - if (inst->bits1.da1.dest_reg_file != BRW_MESSAGE_REGISTER_FILE) return GL_FALSE; + if (inst->bits1.da1.dest_address_mode != BRW_ADDRESS_DIRECT) + return GL_TRUE; + const int reg_start = reg_index * REG_SIZE; const int reg_end = reg_start + size; @@ -188,8 +182,6 @@ brw_is_mrf_written(const struct brw_instruction *inst, int reg_index, int size) if (is_compr4 && inst->header.execution_size != BRW_EXECUTE_16) return GL_TRUE; - GLboolean is_written = GL_FALSE; - /* Here we write mrf_{i} and mrf_{i+4}. So we read two times 8 elements */ if (is_compr4) { const int length = 8 * type_size * inst->bits1.da1.dest_horiz_stride; @@ -210,7 +202,8 @@ brw_is_mrf_written(const struct brw_instruction *inst, int reg_index, int size) const int left1 = MAX2(write_start1, reg_start); const int right1 = MIN2(write_end1, reg_end); - is_written = left0 < right0 || left1 < right1; + if (left0 < right0 || left1 < right1) + return GL_TRUE; } else { int length; @@ -223,25 +216,41 @@ brw_is_mrf_written(const struct brw_instruction *inst, int reg_index, int size) + inst->bits1.da1.dest_subreg_nr; const int write_end = write_start + length; const int left = MAX2(write_start, reg_start); - const int right = MIN2(write_end, reg_end);; + const int right = MIN2(write_end, reg_end); - is_written = left < right; + if (left < right) + return GL_TRUE; } - /* SEND may perform an implicit mov to a mrf register */ - if (is_written == GL_FALSE && - inst->header.opcode == BRW_OPCODE_SEND && - inst->bits1.da1.src0_reg_file != 0) { + return GL_FALSE; +} - const int mrf_start = inst->header.destreg__conditionalmod; - const int write_start = mrf_start * REG_SIZE; - const int write_end = write_start + REG_SIZE; - const int left = MAX2(write_start, reg_start); - const int right = MIN2(write_end, reg_end);; - is_written = left < right; - } +/* SEND may perform an implicit mov to a mrf register */ +static GLboolean brw_is_mrf_written_send(const struct brw_instruction *inst, + int reg_index, int size) +{ + + const int reg_start = reg_index * REG_SIZE; + const int reg_end = reg_start + size; + const int mrf_start = inst->header.destreg__conditionalmod; + const int write_start = mrf_start * REG_SIZE; + const int write_end = write_start + REG_SIZE; + const int left = MAX2(write_start, reg_start); + const int right = MIN2(write_end, reg_end); + + if (inst->header.opcode != BRW_OPCODE_SEND || + inst->bits1.da1.src0_reg_file == 0) + return GL_FALSE; - return is_written; + return left < right; +} + +/* Specific path for message register since we need to handle the compr4 case */ +static INLINE GLboolean +brw_is_mrf_written(const struct brw_instruction *inst, int reg_index, int size) +{ + return (brw_is_mrf_written_alu(inst, reg_index, size) || + brw_is_mrf_written_send(inst, reg_index, size)); } static INLINE GLboolean @@ -284,7 +293,7 @@ static INLINE GLboolean brw_is_grf_read(const struct brw_instruction *inst, int reg_index, int size) { int i, j; - if (inst_opcode[inst->header.opcode].nsrc == 0) + if (brw_opcodes[inst->header.opcode].nsrc == 0) return GL_FALSE; /* Look at first source. We must take into account register regions to @@ -292,7 +301,7 @@ brw_is_grf_read(const struct brw_instruction *inst, int reg_index, int size) * since we do not take into account the fact that some complete registers * may be skipped */ - if (inst_opcode[inst->header.opcode].nsrc >= 1) { + if (brw_opcodes[inst->header.opcode].nsrc >= 1) { if (inst->bits2.da1.src0_address_mode != BRW_ADDRESS_DIRECT) if (inst->bits1.ia1.src0_reg_file == BRW_GENERAL_REGISTER_FILE) @@ -327,7 +336,7 @@ brw_is_grf_read(const struct brw_instruction *inst, int reg_index, int size) } /* Second src register */ - if (inst_opcode[inst->header.opcode].nsrc >= 2) { + if (brw_opcodes[inst->header.opcode].nsrc >= 2) { if (inst->bits3.da1.src1_address_mode != BRW_ADDRESS_DIRECT) if (inst->bits1.ia1.src1_reg_file == BRW_GENERAL_REGISTER_FILE) diff --git a/src/mesa/drivers/dri/i965/brw_program.c b/src/mesa/drivers/dri/i965/brw_program.c index 4b08d2599b..bc152204a4 100644 --- a/src/mesa/drivers/dri/i965/brw_program.c +++ b/src/mesa/drivers/dri/i965/brw_program.c @@ -36,6 +36,7 @@ #include "program/program.h" #include "program/programopt.h" #include "tnl/tnl.h" +#include "talloc.h" #include "brw_context.h" #include "brw_wm.h" @@ -114,10 +115,7 @@ shader_error(GLcontext *ctx, struct gl_program *prog, const char *msg) shader = _mesa_lookup_shader_program(ctx, prog->Id); if (shader) { - if (shader->InfoLog) { - free(shader->InfoLog); - } - shader->InfoLog = _mesa_strdup(msg); + shader->InfoLog = talloc_strdup_append(shader->InfoLog, msg); shader->LinkStatus = GL_FALSE; } } @@ -170,6 +168,9 @@ static GLboolean brwProgramStringNotify( GLcontext *ctx, * See piglit glsl-{vs,fs}-functions-[23] tests. */ for (i = 0; i < prog->NumInstructions; i++) { + struct prog_instruction *inst = prog->Instructions + i; + int r; + if (prog->Instructions[i].Opcode == OPCODE_CAL) { shader_error(ctx, prog, "i965 driver doesn't yet support uninlined function " @@ -177,16 +178,28 @@ static GLboolean brwProgramStringNotify( GLcontext *ctx, "the end of the function to work around it.\n"); return GL_FALSE; } - if (prog->Instructions[i].DstReg.RelAddr && - prog->Instructions[i].DstReg.File == PROGRAM_INPUT) { + + if (prog->Instructions[i].Opcode == OPCODE_RET) { shader_error(ctx, prog, - "Variable indexing of shader inputs unsupported\n"); + "i965 driver doesn't yet support \"return\" " + "from main().\n"); return GL_FALSE; } - if (prog->Instructions[i].DstReg.RelAddr && + + for (r = 0; r < _mesa_num_inst_src_regs(inst->Opcode); r++) { + if (prog->Instructions[i].SrcReg[r].RelAddr && + prog->Instructions[i].SrcReg[r].File == PROGRAM_INPUT) { + shader_error(ctx, prog, + "Variable indexing of shader inputs unsupported\n"); + return GL_FALSE; + } + } + + if (target == GL_FRAGMENT_PROGRAM_ARB && + prog->Instructions[i].DstReg.RelAddr && prog->Instructions[i].DstReg.File == PROGRAM_OUTPUT) { shader_error(ctx, prog, - "Variable indexing of shader outputs unsupported\n"); + "Variable indexing of FS outputs unsupported\n"); return GL_FALSE; } if (target == GL_FRAGMENT_PROGRAM_ARB) { @@ -218,5 +231,10 @@ void brwInitFragProgFuncs( struct dd_function_table *functions ) functions->DeleteProgram = brwDeleteProgram; functions->IsProgramNative = brwIsProgramNative; functions->ProgramStringNotify = brwProgramStringNotify; + + functions->NewShader = brw_new_shader; + functions->NewShaderProgram = brw_new_shader_program; + functions->CompileShader = brw_compile_shader; + functions->LinkShader = brw_link_shader; } diff --git a/src/mesa/drivers/dri/i965/brw_state.h b/src/mesa/drivers/dri/i965/brw_state.h index af08446f2d..c5d296b129 100644 --- a/src/mesa/drivers/dri/i965/brw_state.h +++ b/src/mesa/drivers/dri/i965/brw_state.h @@ -107,6 +107,7 @@ extern const struct brw_tracked_state gen6_sf_vp; extern const struct brw_tracked_state gen6_urb; extern const struct brw_tracked_state gen6_viewport_state; extern const struct brw_tracked_state gen6_vs_state; +extern const struct brw_tracked_state gen6_wm_constants; extern const struct brw_tracked_state gen6_wm_state; /*********************************************************************** diff --git a/src/mesa/drivers/dri/i965/brw_state_upload.c b/src/mesa/drivers/dri/i965/brw_state_upload.c index f92a19c2aa..a0c130557e 100644 --- a/src/mesa/drivers/dri/i965/brw_state_upload.c +++ b/src/mesa/drivers/dri/i965/brw_state_upload.c @@ -35,7 +35,6 @@ #include "brw_state.h" #include "intel_batchbuffer.h" #include "intel_buffers.h" -#include "intel_chipset.h" /* This is used to initialize brw->state.atoms[]. We could use this * list directly except for a single atom, brw_constant_buffer, which @@ -129,7 +128,7 @@ const struct brw_tracked_state *gen6_atoms[] = &gen6_cc_state_pointers, &brw_vs_constants, /* Before vs_surfaces and constant_buffer */ - &brw_wm_constants, /* Before wm_surfaces and constant_buffer */ + &gen6_wm_constants, /* Before wm_surfaces and constant_buffer */ &brw_vs_surfaces, /* must do before unit */ &brw_wm_constant_surface, /* must do before wm surfaces/bind bo */ @@ -351,7 +350,7 @@ void brw_validate_state( struct brw_context *brw ) brw_add_validated_bo(brw, intel->batch->buf); - if (IS_GEN6(intel->intelScreen->deviceID)) { + if (intel->gen >= 6) { atoms = gen6_atoms; num_atoms = ARRAY_SIZE(gen6_atoms); } else { @@ -425,7 +424,7 @@ void brw_upload_state(struct brw_context *brw) const struct brw_tracked_state **atoms; int num_atoms; - if (IS_GEN6(intel->intelScreen->deviceID)) { + if (intel->gen >= 6) { atoms = gen6_atoms; num_atoms = ARRAY_SIZE(gen6_atoms); } else { diff --git a/src/mesa/drivers/dri/i965/brw_structs.h b/src/mesa/drivers/dri/i965/brw_structs.h index 2fde42a706..2a118e01c5 100644 --- a/src/mesa/drivers/dri/i965/brw_structs.h +++ b/src/mesa/drivers/dri/i965/brw_structs.h @@ -750,7 +750,7 @@ struct gen6_depth_stencil_state } ds1; struct { - GLuint pad0:25; + GLuint pad0:26; GLuint depth_write_enable:1; GLuint depth_test_func:3; GLuint pad1:1; @@ -1305,13 +1305,14 @@ struct brw_instruction GLuint access_mode:1; GLuint mask_control:1; GLuint dependency_control:2; - GLuint compression_control:2; + GLuint compression_control:2; /* gen6: quater control */ GLuint thread_control:2; GLuint predicate_control:4; GLuint predicate_inverse:1; GLuint execution_size:3; GLuint destreg__conditionalmod:4; /* destreg - send, conditionalmod - others */ - GLuint pad0:2; + GLuint acc_wr_control:1; + GLuint cmpt_control:1; GLuint debug_control:1; GLuint saturate:1; } header; @@ -1359,7 +1360,7 @@ struct brw_instruction GLuint dest_writemask:4; GLuint dest_subreg_nr:1; GLuint dest_reg_nr:8; - GLuint pad1:2; + GLuint dest_horiz_stride:2; GLuint dest_address_mode:1; } da16; @@ -1373,7 +1374,7 @@ struct brw_instruction GLuint dest_writemask:4; GLint dest_indirect_offset:6; GLuint dest_subreg_nr:3; - GLuint pad1:2; + GLuint dest_horiz_stride:2; GLuint dest_address_mode:1; } ia16; } bits1; diff --git a/src/mesa/drivers/dri/i965/brw_vs.c b/src/mesa/drivers/dri/i965/brw_vs.c index 9a832af9a9..9f90e1e5e5 100644 --- a/src/mesa/drivers/dri/i965/brw_vs.c +++ b/src/mesa/drivers/dri/i965/brw_vs.c @@ -75,10 +75,10 @@ static void do_vs_prog( struct brw_context *brw, c.prog_data.outputs_written |= BITFIELD64_BIT(VERT_RESULT_TEX0 + i); } - if (0) - _mesa_print_program(&c.vp->program.Base); - - + if (0) { + _mesa_fprint_program_opt(stdout, &c.vp->program.Base, PROG_PRINT_DEBUG, + GL_TRUE); + } /* Emit GEN4 code. */ diff --git a/src/mesa/drivers/dri/i965/brw_vs_emit.c b/src/mesa/drivers/dri/i965/brw_vs_emit.c index b6b558e9a6..1d88c6b5a4 100644 --- a/src/mesa/drivers/dri/i965/brw_vs_emit.c +++ b/src/mesa/drivers/dri/i965/brw_vs_emit.c @@ -47,6 +47,7 @@ brw_vs_arg_can_be_immediate(enum prog_opcode opcode, int arg) [OPCODE_MOV] = 1, [OPCODE_ADD] = 2, [OPCODE_CMP] = 3, + [OPCODE_DP2] = 2, [OPCODE_DP3] = 2, [OPCODE_DP4] = 2, [OPCODE_DPH] = 2, @@ -97,6 +98,39 @@ static void release_tmps( struct brw_vs_compile *c ) c->last_tmp = c->first_tmp; } +static int +get_first_reladdr_output(struct gl_vertex_program *vp) +{ + int i; + int first_reladdr_output = VERT_RESULT_MAX; + + for (i = 0; i < vp->Base.NumInstructions; i++) { + struct prog_instruction *inst = vp->Base.Instructions + i; + + if (inst->DstReg.File == PROGRAM_OUTPUT && + inst->DstReg.RelAddr && + inst->DstReg.Index < first_reladdr_output) + first_reladdr_output = inst->DstReg.Index; + } + + return first_reladdr_output; +} + +/* Clears the record of which vp_const_buffer elements have been + * loaded into our constant buffer registers, for the starts of new + * blocks after control flow. + */ +static void +clear_current_const(struct brw_vs_compile *c) +{ + unsigned int i; + + if (c->vp->use_const_buffer) { + for (i = 0; i < 3; i++) { + c->current_const[i].index = -1; + } + } +} /** * Preallocate GRF register before code emit. @@ -108,6 +142,7 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) struct intel_context *intel = &c->func.brw->intel; GLuint i, reg = 0, mrf; int attributes_in_vue; + int first_reladdr_output; /* Determine whether to use a real constant buffer or use a block * of GRF registers for constants. The later is faster but only @@ -225,6 +260,7 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) else mrf = 4; + first_reladdr_output = get_first_reladdr_output(&c->vp->program); for (i = 0; i < VERT_RESULT_MAX; i++) { if (c->prog_data.outputs_written & BITFIELD64_BIT(i)) { c->nr_outputs++; @@ -253,15 +289,16 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) * For attributes beyond the compute-to-MRF, we compute to * GRFs and they will be written in the second URB_WRITE. */ - if (mrf < 15) { + if (first_reladdr_output > i && mrf < 15) { c->regs[PROGRAM_OUTPUT][i] = brw_message_reg(mrf); mrf++; } else { - if (!c->first_overflow_output) + if (mrf >= 15 && !c->first_overflow_output) c->first_overflow_output = i; c->regs[PROGRAM_OUTPUT][i] = brw_vec8_grf(reg, 0); reg++; + mrf++; } } } @@ -292,10 +329,10 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) if (c->vp->use_const_buffer) { for (i = 0; i < 3; i++) { - c->current_const[i].index = -1; c->current_const[i].reg = brw_vec8_grf(reg, 0); reg++; } + clear_current_const(c); } for (i = 0; i < 128; i++) { @@ -502,6 +539,23 @@ static void emit_cmp( struct brw_compile *p, brw_set_predicate_control(p, BRW_PREDICATE_NONE); } +static void emit_sign(struct brw_vs_compile *c, + struct brw_reg dst, + struct brw_reg arg0) +{ + struct brw_compile *p = &c->func; + + brw_MOV(p, dst, brw_imm_f(0)); + + brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_L, arg0, brw_imm_f(0)); + brw_MOV(p, dst, brw_imm_f(-1.0)); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); + + brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_G, arg0, brw_imm_f(0)); + brw_MOV(p, dst, brw_imm_f(1.0)); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); +} + static void emit_max( struct brw_compile *p, struct brw_reg dst, struct brw_reg arg0, @@ -1010,13 +1064,11 @@ move_to_reladdr_dst(struct brw_vs_compile *c, int reg_size = 32; struct brw_reg addr_reg = c->regs[PROGRAM_ADDRESS][0]; struct brw_reg vp_address = retype(vec1(addr_reg), BRW_REGISTER_TYPE_D); - struct brw_reg temp_base = c->regs[inst->DstReg.File][0]; - GLuint byte_offset = temp_base.nr * 32 + temp_base.subnr; + struct brw_reg base = c->regs[inst->DstReg.File][inst->DstReg.Index]; + GLuint byte_offset = base.nr * 32 + base.subnr; struct brw_reg indirect = brw_vec4_indirect(0,0); struct brw_reg acc = retype(vec1(get_tmp(c)), BRW_REGISTER_TYPE_UW); - byte_offset += inst->DstReg.Index * reg_size; - brw_push_insn_state(p); brw_set_access_mode(p, BRW_ALIGN_1); @@ -1162,10 +1214,12 @@ static struct brw_reg get_arg( struct brw_vs_compile *c, /* Convert 3-bit swizzle to 2-bit. */ - reg.dw1.bits.swizzle = BRW_SWIZZLE4(GET_SWZ(src->Swizzle, 0), - GET_SWZ(src->Swizzle, 1), - GET_SWZ(src->Swizzle, 2), - GET_SWZ(src->Swizzle, 3)); + if (reg.file != BRW_IMMEDIATE_VALUE) { + reg.dw1.bits.swizzle = BRW_SWIZZLE4(GET_SWZ(src->Swizzle, 0), + GET_SWZ(src->Swizzle, 1), + GET_SWZ(src->Swizzle, 2), + GET_SWZ(src->Swizzle, 3)); + } /* Note this is ok for non-swizzle instructions: */ @@ -1211,6 +1265,7 @@ static struct brw_reg get_dst( struct brw_vs_compile *c, reg = brw_null_reg(); } + assert(reg.type != BRW_IMMEDIATE_VALUE); reg.dw1.bits.writemask = dst.WriteMask; return reg; @@ -1299,6 +1354,7 @@ static void emit_vertex_write( struct brw_vs_compile *c) struct brw_reg ndc; int eot; GLuint len_vertex_header = 2; + int next_mrf, i; if (c->key.copy_edgeflag) { brw_MOV(p, @@ -1376,6 +1432,7 @@ static void emit_vertex_write( struct brw_vs_compile *c) * of zeros followed by two sets of NDC coordinates: */ brw_set_access_mode(p, BRW_ALIGN_1); + brw_set_acc_write_control(p, 0); /* The VUE layout is documented in Volume 2a. */ if (intel->gen >= 6) { @@ -1416,6 +1473,23 @@ static void emit_vertex_write( struct brw_vs_compile *c) len_vertex_header = 2; } + /* Move variable-addressed, non-overflow outputs to their MRFs. */ + next_mrf = 2 + len_vertex_header; + for (i = 0; i < VERT_RESULT_MAX; i++) { + if (c->first_overflow_output > 0 && i >= c->first_overflow_output) + break; + if (!(c->prog_data.outputs_written & BITFIELD64_BIT(i))) + continue; + + if (i >= VERT_RESULT_TEX0 && + c->regs[PROGRAM_OUTPUT][i].file == BRW_GENERAL_REGISTER_FILE) { + brw_MOV(p, brw_message_reg(next_mrf), c->regs[PROGRAM_OUTPUT][i]); + next_mrf++; + } else if (c->regs[PROGRAM_OUTPUT][i].file == BRW_MESSAGE_REGISTER_FILE) { + next_mrf = c->regs[PROGRAM_OUTPUT][i].nr + 1; + } + } + eot = (c->first_overflow_output == 0); brw_urb_WRITE(p, @@ -1541,18 +1615,23 @@ void brw_vs_emit(struct brw_vs_compile *c ) const GLuint nr_insns = c->vp->program.Base.NumInstructions; GLuint insn, if_depth = 0, loop_depth = 0; struct brw_instruction *if_inst[MAX_IF_DEPTH], *loop_inst[MAX_LOOP_DEPTH] = { 0 }; + int if_depth_in_loop[MAX_LOOP_DEPTH]; const struct brw_indirect stack_index = brw_indirect(0, 0); GLuint index; GLuint file; if (INTEL_DEBUG & DEBUG_VS) { printf("vs-mesa:\n"); - _mesa_print_program(&c->vp->program.Base); + _mesa_fprint_program_opt(stdout, &c->vp->program.Base, PROG_PRINT_DEBUG, + GL_TRUE); printf("\n"); } brw_set_compression_control(p, BRW_COMPRESSION_NONE); brw_set_access_mode(p, BRW_ALIGN_16); + if_depth_in_loop[loop_depth] = 0; + + brw_set_acc_write_control(p, 1); for (insn = 0; insn < nr_insns; insn++) { GLuint i; @@ -1591,7 +1670,7 @@ void brw_vs_emit(struct brw_vs_compile *c ) const struct prog_instruction *inst = &c->vp->program.Base.Instructions[insn]; struct brw_reg args[3], dst; GLuint i; - + #if 0 printf("%d: ", insn); _mesa_print_instruction(inst); @@ -1636,6 +1715,9 @@ void brw_vs_emit(struct brw_vs_compile *c ) case OPCODE_COS: emit_math1(c, BRW_MATH_FUNCTION_COS, dst, args[0], BRW_MATH_PRECISION_FULL); break; + case OPCODE_DP2: + brw_DP2(p, dst, args[0], args[1]); + break; case OPCODE_DP3: brw_DP3(p, dst, args[0], args[1]); break; @@ -1732,6 +1814,9 @@ void brw_vs_emit(struct brw_vs_compile *c ) case OPCODE_SLE: unalias2(c, dst, args[0], args[1], emit_sle); break; + case OPCODE_SSG: + unalias1(c, dst, args[0], emit_sign); + break; case OPCODE_SUB: brw_ADD(p, dst, args[0], negate(args[1])); break; @@ -1753,31 +1838,38 @@ void brw_vs_emit(struct brw_vs_compile *c ) if_inst[if_depth] = brw_IF(p, BRW_EXECUTE_8); /* Note that brw_IF smashes the predicate_control field. */ if_inst[if_depth]->header.predicate_control = get_predicate(inst); + if_depth_in_loop[loop_depth]++; if_depth++; break; case OPCODE_ELSE: + clear_current_const(c); assert(if_depth > 0); if_inst[if_depth-1] = brw_ELSE(p, if_inst[if_depth-1]); break; case OPCODE_ENDIF: + clear_current_const(c); assert(if_depth > 0); brw_ENDIF(p, if_inst[--if_depth]); + if_depth_in_loop[loop_depth]--; break; case OPCODE_BGNLOOP: + clear_current_const(c); loop_inst[loop_depth++] = brw_DO(p, BRW_EXECUTE_8); + if_depth_in_loop[loop_depth] = 0; break; case OPCODE_BRK: brw_set_predicate_control(p, get_predicate(inst)); - brw_BREAK(p); + brw_BREAK(p, if_depth_in_loop[loop_depth]); brw_set_predicate_control(p, BRW_PREDICATE_NONE); break; case OPCODE_CONT: brw_set_predicate_control(p, get_predicate(inst)); - brw_CONT(p); + brw_CONT(p, if_depth_in_loop[loop_depth]); brw_set_predicate_control(p, BRW_PREDICATE_NONE); break; case OPCODE_ENDLOOP: { + clear_current_const(c); struct brw_instruction *inst0, *inst1; GLuint br = 1; @@ -1793,12 +1885,10 @@ void brw_vs_emit(struct brw_vs_compile *c ) if (inst0->header.opcode == BRW_OPCODE_BREAK && inst0->bits3.if_else.jump_count == 0) { inst0->bits3.if_else.jump_count = br * (inst1 - inst0 + 1); - inst0->bits3.if_else.pop_count = 0; } else if (inst0->header.opcode == BRW_OPCODE_CONTINUE && inst0->bits3.if_else.jump_count == 0) { inst0->bits3.if_else.jump_count = br * (inst1 - inst0); - inst0->bits3.if_else.pop_count = 0; } } } @@ -1883,11 +1973,9 @@ void brw_vs_emit(struct brw_vs_compile *c ) } } - if (inst->DstReg.RelAddr && inst->DstReg.File == PROGRAM_TEMPORARY) { - /* We don't do RelAddr of PROGRAM_OUTPUT yet, because of the - * compute-to-mrf and the fact that we are allocating - * registers for only the used PROGRAM_OUTPUTs. - */ + if (inst->DstReg.RelAddr) { + assert(inst->DstReg.File == PROGRAM_TEMPORARY|| + inst->DstReg.File == PROGRAM_OUTPUT); move_to_reladdr_dst(c, inst, dst); } @@ -1903,7 +1991,7 @@ void brw_vs_emit(struct brw_vs_compile *c ) printf("vs-native:\n"); for (i = 0; i < p->nr_insn; i++) - brw_disasm(stderr, &p->store[i], intel->gen); + brw_disasm(stdout, &p->store[i], intel->gen); printf("\n"); } } diff --git a/src/mesa/drivers/dri/i965/brw_vtbl.c b/src/mesa/drivers/dri/i965/brw_vtbl.c index 14227a5133..8f1601d10f 100644 --- a/src/mesa/drivers/dri/i965/brw_vtbl.c +++ b/src/mesa/drivers/dri/i965/brw_vtbl.c @@ -101,6 +101,7 @@ static void brw_destroy_context( struct intel_context *intel ) dri_bo_release(&brw->wm.prog_bo); dri_bo_release(&brw->wm.state_bo); dri_bo_release(&brw->wm.const_bo); + dri_bo_release(&brw->wm.push_const_bo); dri_bo_release(&brw->cc.prog_bo); dri_bo_release(&brw->cc.state_bo); dri_bo_release(&brw->cc.vp_bo); diff --git a/src/mesa/drivers/dri/i965/brw_wm.c b/src/mesa/drivers/dri/i965/brw_wm.c index e182fc3202..d70be7bda2 100644 --- a/src/mesa/drivers/dri/i965/brw_wm.c +++ b/src/mesa/drivers/dri/i965/brw_wm.c @@ -32,7 +32,7 @@ #include "brw_context.h" #include "brw_wm.h" #include "brw_state.h" - +#include "main/formats.h" /** Return number of src args for given instruction */ GLuint brw_wm_nr_args( GLuint opcode ) @@ -68,6 +68,7 @@ GLuint brw_wm_is_scalar_result( GLuint opcode ) case OPCODE_RCP: case OPCODE_RSQ: case OPCODE_SIN: + case OPCODE_DP2: case OPCODE_DP3: case OPCODE_DP4: case OPCODE_DPH: @@ -177,17 +178,19 @@ static void do_wm_prog( struct brw_context *brw, /* temporary sanity check assertion */ ASSERT(fp->isGLSL == brw_wm_is_glsl(&c->fp->program)); - /* - * Shader which use GLSL features such as flow control are handled - * differently from "simple" shaders. - */ - if (fp->isGLSL) { - c->dispatch_width = 8; - brw_wm_glsl_emit(brw, c); - } - else { - c->dispatch_width = 16; - brw_wm_non_glsl_emit(brw, c); + if (!brw_wm_fs_emit(brw, c)) { + /* + * Shader which use GLSL features such as flow control are handled + * differently from "simple" shaders. + */ + if (fp->isGLSL) { + c->dispatch_width = 8; + brw_wm_glsl_emit(brw, c); + } + else { + c->dispatch_width = 16; + brw_wm_non_glsl_emit(brw, c); + } } if (INTEL_DEBUG & DEBUG_WM) diff --git a/src/mesa/drivers/dri/i965/brw_wm.h b/src/mesa/drivers/dri/i965/brw_wm.h index f40977fab8..2639d4f26b 100644 --- a/src/mesa/drivers/dri/i965/brw_wm.h +++ b/src/mesa/drivers/dri/i965/brw_wm.h @@ -61,7 +61,7 @@ struct brw_wm_prog_key { GLuint source_depth_reg:3; GLuint aa_dest_stencil_reg:3; GLuint dest_depth_reg:3; - GLuint nr_depth_regs:3; + GLuint nr_payload_regs:4; GLuint computes_depth:1; /* could be derived from program string */ GLuint source_depth_to_render_target:1; GLuint flat_shade:1; @@ -306,6 +306,7 @@ void brw_wm_lookup_iz( GLuint line_aa, GLboolean brw_wm_is_glsl(const struct gl_fragment_program *fp); void brw_wm_glsl_emit(struct brw_context *brw, struct brw_wm_compile *c); +GLboolean brw_wm_fs_emit(struct brw_context *brw, struct brw_wm_compile *c); /* brw_wm_emit.c */ void emit_alu1(struct brw_compile *p, @@ -343,6 +344,11 @@ void emit_delta_xy(struct brw_compile *p, const struct brw_reg *dst, GLuint mask, const struct brw_reg *arg0); +void emit_dp2(struct brw_compile *p, + const struct brw_reg *dst, + GLuint mask, + const struct brw_reg *arg0, + const struct brw_reg *arg1); void emit_dp3(struct brw_compile *p, const struct brw_reg *dst, GLuint mask, @@ -425,6 +431,10 @@ void emit_sop(struct brw_compile *p, GLuint cond, const struct brw_reg *arg0, const struct brw_reg *arg1); +void emit_sign(struct brw_compile *p, + const struct brw_reg *dst, + GLuint mask, + const struct brw_reg *arg0); void emit_tex(struct brw_wm_compile *c, struct brw_reg *dst, GLuint dst_flags, @@ -450,4 +460,13 @@ void emit_xpd(struct brw_compile *p, const struct brw_reg *arg0, const struct brw_reg *arg1); +GLboolean brw_compile_shader(GLcontext *ctx, + struct gl_shader *shader); +GLboolean brw_link_shader(GLcontext *ctx, struct gl_shader_program *prog); +struct gl_shader *brw_new_shader(GLcontext *ctx, GLuint name, GLuint type); +struct gl_shader_program *brw_new_shader_program(GLcontext *ctx, GLuint name); + +GLboolean brw_do_channel_expressions(struct exec_list *instructions); +GLboolean brw_do_vector_splitting(struct exec_list *instructions); + #endif diff --git a/src/mesa/drivers/dri/i965/brw_wm_debug.c b/src/mesa/drivers/dri/i965/brw_wm_debug.c index a78cc8b54e..6a91251a80 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_debug.c +++ b/src/mesa/drivers/dri/i965/brw_wm_debug.c @@ -44,16 +44,16 @@ void brw_wm_print_value( struct brw_wm_compile *c, printf("undef"); else if( value - c->vreg >= 0 && value - c->vreg < BRW_WM_MAX_VREG) - printf("r%d", value - c->vreg); + printf("r%ld", (long) (value - c->vreg)); else if (value - c->creg >= 0 && value - c->creg < BRW_WM_MAX_PARAM) - printf("c%d", value - c->creg); + printf("c%ld", (long) (value - c->creg)); else if (value - c->payload.input_interp >= 0 && value - c->payload.input_interp < FRAG_ATTRIB_MAX) - printf("i%d", value - c->payload.input_interp); + printf("i%ld", (long) (value - c->payload.input_interp)); else if (value - c->payload.depth >= 0 && value - c->payload.depth < FRAG_ATTRIB_MAX) - printf("d%d", value - c->payload.depth); + printf("d%ld", (long) (value - c->payload.depth)); else printf("?"); } diff --git a/src/mesa/drivers/dri/i965/brw_wm_emit.c b/src/mesa/drivers/dri/i965/brw_wm_emit.c index d9fa2e6335..f3ad01b3fe 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_emit.c +++ b/src/mesa/drivers/dri/i965/brw_wm_emit.c @@ -668,6 +668,28 @@ void emit_cmp(struct brw_compile *p, } } +void emit_sign(struct brw_compile *p, + const struct brw_reg *dst, + GLuint mask, + const struct brw_reg *arg0) +{ + GLuint i; + + for (i = 0; i < 4; i++) { + if (mask & (1<<i)) { + brw_MOV(p, dst[i], brw_imm_f(0.0)); + + brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_L, arg0[i], brw_imm_f(0)); + brw_MOV(p, dst[i], brw_imm_f(-1.0)); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); + + brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_G, arg0[i], brw_imm_f(0)); + brw_MOV(p, dst[i], brw_imm_f(1.0)); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); + } + } +} + void emit_max(struct brw_compile *p, const struct brw_reg *dst, GLuint mask, @@ -709,6 +731,27 @@ void emit_min(struct brw_compile *p, } +void emit_dp2(struct brw_compile *p, + const struct brw_reg *dst, + GLuint mask, + const struct brw_reg *arg0, + const struct brw_reg *arg1) +{ + int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1; + + if (!(mask & WRITEMASK_XYZW)) + return; /* Do not emit dead code */ + + assert(is_power_of_two(mask & WRITEMASK_XYZW)); + + brw_MUL(p, brw_null_reg(), arg0[0], arg1[0]); + + brw_set_saturate(p, (mask & SATURATE) ? 1 : 0); + brw_MAC(p, dst[dst_chan], arg0[1], arg1[1]); + brw_set_saturate(p, 0); +} + + void emit_dp3(struct brw_compile *p, const struct brw_reg *dst, GLuint mask, @@ -809,21 +852,28 @@ void emit_math1(struct brw_wm_compile *c, const struct brw_reg *arg0) { struct brw_compile *p = &c->func; + struct intel_context *intel = &p->brw->intel; int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1; GLuint saturate = ((mask & SATURATE) ? BRW_MATH_SATURATE_SATURATE : BRW_MATH_SATURATE_NONE); + struct brw_reg src; + + if (intel->gen >= 6 && arg0[0].hstride == BRW_HORIZONTAL_STRIDE_0) { + /* Gen6 math requires that source and dst horizontal stride be 1. + * + */ + src = *dst; + brw_MOV(p, src, arg0[0]); + } else { + src = arg0[0]; + } if (!(mask & WRITEMASK_XYZW)) return; /* Do not emit dead code */ assert(is_power_of_two(mask & WRITEMASK_XYZW)); - /* If compressed, this will write message reg 2,3 from arg0.x's 16 - * channels. - */ - brw_MOV(p, brw_message_reg(2), arg0[0]); - /* Send two messages to perform all 16 operations: */ brw_push_insn_state(p); @@ -833,7 +883,7 @@ void emit_math1(struct brw_wm_compile *c, function, saturate, 2, - brw_null_reg(), + src, BRW_MATH_DATA_VECTOR, BRW_MATH_PRECISION_FULL); @@ -844,7 +894,7 @@ void emit_math1(struct brw_wm_compile *c, function, saturate, 3, - brw_null_reg(), + sechalf(src), BRW_MATH_DATA_VECTOR, BRW_MATH_PRECISION_FULL); } @@ -873,13 +923,6 @@ void emit_math2(struct brw_wm_compile *c, brw_push_insn_state(p); brw_set_compression_control(p, BRW_COMPRESSION_NONE); - brw_MOV(p, brw_message_reg(2), arg0[0]); - if (c->dispatch_width == 16) { - brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF); - brw_MOV(p, brw_message_reg(4), sechalf(arg0[0])); - } - - brw_set_compression_control(p, BRW_COMPRESSION_NONE); brw_MOV(p, brw_message_reg(3), arg1[0]); if (c->dispatch_width == 16) { brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF); @@ -892,7 +935,7 @@ void emit_math2(struct brw_wm_compile *c, function, saturate, 2, - brw_null_reg(), + arg0[0], BRW_MATH_DATA_VECTOR, BRW_MATH_PRECISION_FULL); @@ -905,7 +948,7 @@ void emit_math2(struct brw_wm_compile *c, function, saturate, 4, - brw_null_reg(), + sechalf(arg0[0]), BRW_MATH_DATA_VECTOR, BRW_MATH_PRECISION_FULL); } @@ -1199,6 +1242,7 @@ static void fire_fb_write( struct brw_wm_compile *c, GLuint eot ) { struct brw_compile *p = &c->func; + struct intel_context *intel = &p->brw->intel; struct brw_reg dst; if (c->dispatch_width == 16) @@ -1209,6 +1253,7 @@ static void fire_fb_write( struct brw_wm_compile *c, /* Pass through control information: */ /* mov (8) m1.0<1>:ud r1.0<8;8,1>:ud { Align1 NoMask } */ + if (intel->gen < 6) /* gen6, use headerless for fb write */ { brw_push_insn_state(p); brw_set_mask_control(p, BRW_MASK_DISABLE); /* ? */ @@ -1222,6 +1267,7 @@ static void fire_fb_write( struct brw_wm_compile *c, /* Send framebuffer write message: */ /* send (16) null.0<1>:uw m0 r0.0<8;8,1>:uw 0x85a04000:ud { Align1 EOT } */ brw_fb_WRITE(p, + c->dispatch_width, dst, base_reg, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW), @@ -1263,8 +1309,10 @@ void emit_fb_write(struct brw_wm_compile *c, { struct brw_compile *p = &c->func; struct brw_context *brw = p->brw; + struct intel_context *intel = &brw->intel; GLuint nr = 2; GLuint channel; + int base_reg; /* For gen6 fb write with no header, starting from color payload directly!. */ /* Reserve a space for AA - may not be needed: */ @@ -1276,9 +1324,40 @@ void emit_fb_write(struct brw_wm_compile *c, */ brw_push_insn_state(p); + if (intel->gen >= 6) + base_reg = nr; + else + base_reg = 0; + for (channel = 0; channel < 4; channel++) { - if (c->dispatch_width == 16 && brw->has_compr4) { - /* By setting the high bit of the MRF register number, we indicate + if (intel->gen >= 6) { + /* gen6 SIMD16 single source DP write looks like: + * m + 0: r0 + * m + 1: r1 + * m + 2: g0 + * m + 3: g1 + * m + 4: b0 + * m + 5: b1 + * m + 6: a0 + * m + 7: a1 + */ + if (c->dispatch_width == 16) { + brw_MOV(p, brw_message_reg(nr + channel * 2), arg0[channel]); + } else { + brw_MOV(p, brw_message_reg(nr + channel), arg0[channel]); + } + } else if (c->dispatch_width == 16 && brw->has_compr4) { + /* pre-gen6 SIMD16 single source DP write looks like: + * m + 0: r0 + * m + 1: g0 + * m + 2: b0 + * m + 3: a0 + * m + 4: r1 + * m + 5: g1 + * m + 6: b1 + * m + 7: a1 + * + * By setting the high bit of the MRF register number, we indicate * that we want COMPR4 mode - instead of doing the usual destination * + 1 for the second half we get destination + 4. */ @@ -1303,7 +1382,11 @@ void emit_fb_write(struct brw_wm_compile *c, } /* skip over the regs populated above: */ - nr += 8; + if (c->dispatch_width == 16) + nr += 8; + else + nr += 4; + brw_pop_insn_state(p); if (c->key.source_depth_to_render_target) @@ -1336,11 +1419,16 @@ void emit_fb_write(struct brw_wm_compile *c, nr += 2; } + if (intel->gen >= 6) { + /* Subtract off the message header, since we send headerless. */ + nr -= 2; + } + if (!c->key.runtime_check_aads_emit) { if (c->key.aa_dest_stencil_reg) emit_aa(c, arg1, 2); - fire_fb_write(c, 0, nr, target, eot); + fire_fb_write(c, base_reg, nr, target, eot); } else { struct brw_reg v1_null_ud = vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD)); @@ -1562,6 +1650,10 @@ void brw_wm_emit( struct brw_wm_compile *c ) emit_ddxy(p, dst, dst_flags, GL_FALSE, args[0]); break; + case OPCODE_DP2: + emit_dp2(p, dst, dst_flags, args[0], args[1]); + break; + case OPCODE_DP3: emit_dp3(p, dst, dst_flags, args[0], args[1]); break; @@ -1673,6 +1765,10 @@ void brw_wm_emit( struct brw_wm_compile *c ) emit_sne(p, dst, dst_flags, args[0], args[1]); break; + case OPCODE_SSG: + emit_sign(p, dst, dst_flags, args[0]); + break; + case OPCODE_LIT: emit_lit(c, dst, dst_flags, args[0]); break; @@ -1724,7 +1820,7 @@ void brw_wm_emit( struct brw_wm_compile *c ) printf("wm-native:\n"); for (i = 0; i < p->nr_insn; i++) - brw_disasm(stderr, &p->store[i], p->brw->intel.gen); + brw_disasm(stdout, &p->store[i], p->brw->intel.gen); printf("\n"); } } diff --git a/src/mesa/drivers/dri/i965/brw_wm_fp.c b/src/mesa/drivers/dri/i965/brw_wm_fp.c index 0bef874b88..3870bf10fc 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_fp.c +++ b/src/mesa/drivers/dri/i965/brw_wm_fp.c @@ -88,6 +88,7 @@ static struct prog_src_register src_reg(GLuint file, GLuint idx) reg.RelAddr = 0; reg.Negate = NEGATE_NONE; reg.Abs = 0; + reg.HasIndex2 = 0; return reg; } @@ -1036,13 +1037,12 @@ static void print_insns( const struct prog_instruction *insn, for (i = 0; i < nr; i++, insn++) { printf("%3d: ", i); if (insn->Opcode < MAX_OPCODE) - _mesa_print_instruction(insn); + _mesa_fprint_instruction_opt(stdout, insn, 0, PROG_PRINT_DEBUG, NULL); else if (insn->Opcode < MAX_WM_OPCODE) { GLuint idx = insn->Opcode - MAX_OPCODE; - _mesa_print_alu_instruction(insn, - wm_opcode_strings[idx], - 3); + _mesa_fprint_alu_instruction(stdout, insn, wm_opcode_strings[idx], + 3, PROG_PRINT_DEBUG, NULL); } else printf("965 Opcode %d\n", insn->Opcode); @@ -1061,7 +1061,8 @@ void brw_wm_pass_fp( struct brw_wm_compile *c ) if (INTEL_DEBUG & DEBUG_WM) { printf("pre-fp:\n"); - _mesa_print_program(&fp->program.Base); + _mesa_fprint_program_opt(stdout, &fp->program.Base, PROG_PRINT_DEBUG, + GL_TRUE); printf("\n"); } diff --git a/src/mesa/drivers/dri/i965/brw_wm_glsl.c b/src/mesa/drivers/dri/i965/brw_wm_glsl.c index 575f89b17f..c1083c5942 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_glsl.c +++ b/src/mesa/drivers/dri/i965/brw_wm_glsl.c @@ -303,13 +303,13 @@ static void prealloc_reg(struct brw_wm_compile *c) c->first_free_grf = 0; for (i = 0; i < 4; i++) { - if (i < c->key.nr_depth_regs) + if (i < (c->key.nr_payload_regs + 1) / 2) reg = brw_vec8_grf(i * 2, 0); else reg = brw_vec8_grf(0, 0); set_reg(c, PROGRAM_PAYLOAD, PAYLOAD_DEPTH, i, reg); } - reg_index += 2 * c->key.nr_depth_regs; + reg_index += c->key.nr_payload_regs; /* constants */ { @@ -380,7 +380,7 @@ static void prealloc_reg(struct brw_wm_compile *c) } } - c->prog_data.first_curbe_grf = c->key.nr_depth_regs * 2; + c->prog_data.first_curbe_grf = c->key.nr_payload_regs; c->prog_data.urb_read_length = urb_read_length; c->prog_data.curb_read_length = c->nr_creg; c->emit_mask_reg = brw_uw1_reg(BRW_GENERAL_REGISTER_FILE, reg_index, 0); @@ -1803,12 +1803,15 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c) #define MAX_IF_DEPTH 32 #define MAX_LOOP_DEPTH 32 struct brw_instruction *if_inst[MAX_IF_DEPTH], *loop_inst[MAX_LOOP_DEPTH]; + int if_depth_in_loop[MAX_LOOP_DEPTH]; GLuint i, if_depth = 0, loop_depth = 0; struct brw_compile *p = &c->func; struct brw_indirect stack_index = brw_indirect(0, 0); c->out_of_regs = GL_FALSE; + if_depth_in_loop[loop_depth] = 0; + prealloc_reg(c); brw_set_compression_control(p, BRW_COMPRESSION_NONE); brw_MOV(p, get_addr_reg(stack_index), brw_address(c->stack)); @@ -1903,6 +1906,9 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c) case OPCODE_SWZ: emit_alu1(p, brw_MOV, dst, dst_flags, args[0]); break; + case OPCODE_DP2: + emit_dp2(p, dst, dst_flags, args[0], args[1]); + break; case OPCODE_DP3: emit_dp3(p, dst, dst_flags, args[0], args[1]); break; @@ -1971,6 +1977,9 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c) emit_sop(p, dst, dst_flags, BRW_CONDITIONAL_NEQ, args[0], args[1]); break; + case OPCODE_SSG: + emit_sign(p, dst, dst_flags, args[0]); + break; case OPCODE_MUL: emit_alu2(p, brw_MUL, dst, dst_flags, args[0], args[1]); break; @@ -2014,6 +2023,7 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c) case OPCODE_IF: assert(if_depth < MAX_IF_DEPTH); if_inst[if_depth++] = brw_IF(p, BRW_EXECUTE_8); + if_depth_in_loop[loop_depth]++; break; case OPCODE_ELSE: assert(if_depth > 0); @@ -2022,6 +2032,7 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c) case OPCODE_ENDIF: assert(if_depth > 0); brw_ENDIF(p, if_inst[--if_depth]); + if_depth_in_loop[loop_depth]--; break; case OPCODE_BGNSUB: brw_save_label(p, inst->Comment, p->nr_insn); @@ -2056,13 +2067,14 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c) case OPCODE_BGNLOOP: /* XXX may need to invalidate the current_constant regs */ loop_inst[loop_depth++] = brw_DO(p, BRW_EXECUTE_8); + if_depth_in_loop[loop_depth] = 0; break; case OPCODE_BRK: - brw_BREAK(p); + brw_BREAK(p, if_depth_in_loop[loop_depth]); brw_set_predicate_control(p, BRW_PREDICATE_NONE); break; case OPCODE_CONT: - brw_CONT(p); + brw_CONT(p, if_depth_in_loop[loop_depth]); brw_set_predicate_control(p, BRW_PREDICATE_NONE); break; case OPCODE_ENDLOOP: @@ -2082,12 +2094,10 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c) if (inst0->header.opcode == BRW_OPCODE_BREAK && inst0->bits3.if_else.jump_count == 0) { inst0->bits3.if_else.jump_count = br * (inst1 - inst0 + 1); - inst0->bits3.if_else.pop_count = 0; } else if (inst0->header.opcode == BRW_OPCODE_CONTINUE && inst0->bits3.if_else.jump_count == 0) { inst0->bits3.if_else.jump_count = br * (inst1 - inst0); - inst0->bits3.if_else.pop_count = 0; } } } @@ -2111,7 +2121,7 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c) if (INTEL_DEBUG & DEBUG_WM) { printf("wm-native:\n"); for (i = 0; i < p->nr_insn; i++) - brw_disasm(stderr, &p->store[i], intel->gen); + brw_disasm(stdout, &p->store[i], intel->gen); printf("\n"); } } diff --git a/src/mesa/drivers/dri/i965/brw_wm_iz.c b/src/mesa/drivers/dri/i965/brw_wm_iz.c index 5e399ac62a..8505ef1951 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_iz.c +++ b/src/mesa/drivers/dri/i965/brw_wm_iz.c @@ -152,6 +152,6 @@ void brw_wm_lookup_iz( GLuint line_aa, reg+=2; } - key->nr_depth_regs = (reg+1)/2; + key->nr_payload_regs = reg; } diff --git a/src/mesa/drivers/dri/i965/brw_wm_pass0.c b/src/mesa/drivers/dri/i965/brw_wm_pass0.c index 05de85a957..8fc960b445 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_pass0.c +++ b/src/mesa/drivers/dri/i965/brw_wm_pass0.c @@ -379,7 +379,7 @@ static void pass0_init_payload( struct brw_wm_compile *c ) GLuint i; for (i = 0; i < 4; i++) { - GLuint j = i >= c->key.nr_depth_regs ? 0 : i; + GLuint j = i >= (c->key.nr_payload_regs + 1) / 2 ? 0 : i; pass0_set_fpreg_value( c, PROGRAM_PAYLOAD, PAYLOAD_DEPTH, i, &c->payload.depth[j] ); } diff --git a/src/mesa/drivers/dri/i965/brw_wm_pass1.c b/src/mesa/drivers/dri/i965/brw_wm_pass1.c index b449394029..962515a99e 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_pass1.c +++ b/src/mesa/drivers/dri/i965/brw_wm_pass1.c @@ -158,6 +158,7 @@ void brw_wm_pass1( struct brw_wm_compile *c ) case OPCODE_FLR: case OPCODE_FRC: case OPCODE_MOV: + case OPCODE_SSG: case OPCODE_SWZ: case OPCODE_TRUNC: read0 = writemask; @@ -254,6 +255,11 @@ void brw_wm_pass1( struct brw_wm_compile *c ) read2 = WRITEMASK_W; /* pixel w */ break; + case OPCODE_DP2: + read0 = WRITEMASK_XY; + read1 = WRITEMASK_XY; + break; + case OPCODE_DP3: read0 = WRITEMASK_XYZ; read1 = WRITEMASK_XYZ; diff --git a/src/mesa/drivers/dri/i965/brw_wm_pass2.c b/src/mesa/drivers/dri/i965/brw_wm_pass2.c index 31303febf0..54acb3038b 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_pass2.c +++ b/src/mesa/drivers/dri/i965/brw_wm_pass2.c @@ -76,7 +76,7 @@ static void init_registers( struct brw_wm_compile *c ) for (j = 0; j < c->grf_limit; j++) c->pass2_grf[j].nextuse = BRW_WM_MAX_INSN; - for (j = 0; j < c->key.nr_depth_regs; j++) + for (j = 0; j < (c->key.nr_payload_regs + 1) / 2; j++) prealloc_reg(c, &c->payload.depth[j], i++); for (j = 0; j < c->nr_creg; j++) @@ -101,7 +101,7 @@ static void init_registers( struct brw_wm_compile *c ) assert(nr_interp_regs >= 1); - c->prog_data.first_curbe_grf = c->key.nr_depth_regs * 2; + c->prog_data.first_curbe_grf = ALIGN(c->key.nr_payload_regs, 2); c->prog_data.urb_read_length = nr_interp_regs * 2; c->prog_data.curb_read_length = c->nr_creg * 2; diff --git a/src/mesa/drivers/dri/i965/brw_wm_state.c b/src/mesa/drivers/dri/i965/brw_wm_state.c index c1cf4db1ca..6699d0a73e 100644 --- a/src/mesa/drivers/dri/i965/brw_wm_state.c +++ b/src/mesa/drivers/dri/i965/brw_wm_state.c @@ -104,8 +104,22 @@ wm_unit_populate_key(struct brw_context *brw, struct brw_wm_unit_key *key) key->uses_kill = fp->UsesKill || ctx->Color.AlphaEnabled; key->is_glsl = bfp->isGLSL; - /* temporary sanity check assertion */ - ASSERT(bfp->isGLSL == brw_wm_is_glsl(fp)); + /* If using the fragment shader backend, the program is always + * 8-wide. + */ + if (ctx->Shader.CurrentProgram) { + int i; + + for (i = 0; i < ctx->Shader.CurrentProgram->_NumLinkedShaders; i++) { + struct brw_shader *shader = + (struct brw_shader *)ctx->Shader.CurrentProgram->_LinkedShaders[i];; + + if (shader->base.Type == GL_FRAGMENT_SHADER && + shader->ir != NULL) { + key->is_glsl = GL_TRUE; + } + } + } /* _NEW_DEPTH */ key->stats_wm = intel->stats_wm; diff --git a/src/mesa/drivers/dri/i965/gen6_cc.c b/src/mesa/drivers/dri/i965/gen6_cc.c index f7acad6912..26f1070a16 100644 --- a/src/mesa/drivers/dri/i965/gen6_cc.c +++ b/src/mesa/drivers/dri/i965/gen6_cc.c @@ -267,9 +267,9 @@ static void upload_cc_state_pointers(struct brw_context *brw) BEGIN_BATCH(4); OUT_BATCH(CMD_3D_CC_STATE_POINTERS << 16 | (4 - 2)); - OUT_RELOC(brw->cc.state_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 1); OUT_RELOC(brw->cc.blend_state_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 1); OUT_RELOC(brw->cc.depth_stencil_state_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 1); + OUT_RELOC(brw->cc.state_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 1); ADVANCE_BATCH(); intel_batchbuffer_emit_mi_flush(intel->batch); diff --git a/src/mesa/drivers/dri/i965/gen6_wm_state.c b/src/mesa/drivers/dri/i965/gen6_wm_state.c index 863c85449d..2cd640de17 100644 --- a/src/mesa/drivers/dri/i965/gen6_wm_state.c +++ b/src/mesa/drivers/dri/i965/gen6_wm_state.c @@ -34,18 +34,59 @@ #include "intel_batchbuffer.h" static void +prepare_wm_constants(struct brw_context *brw) +{ + struct intel_context *intel = &brw->intel; + GLcontext *ctx = &intel->ctx; + const struct brw_fragment_program *fp = + brw_fragment_program_const(brw->fragment_program); + + drm_intel_bo_unreference(brw->wm.push_const_bo); + brw->wm.push_const_bo = NULL; + + /* Updates the ParamaterValues[i] pointers for all parameters of the + * basic type of PROGRAM_STATE_VAR. + */ + /* XXX: Should this happen somewhere before to get our state flag set? */ + _mesa_load_state_parameters(ctx, fp->program.Base.Parameters); + + if (brw->wm.prog_data->nr_params != 0) { + float *constants; + unsigned int i; + + brw->wm.push_const_bo = drm_intel_bo_alloc(intel->bufmgr, + "WM constant_bo", + brw->wm.prog_data->nr_params * + sizeof(float), + 4096); + drm_intel_gem_bo_map_gtt(brw->wm.push_const_bo); + constants = brw->wm.push_const_bo->virtual; + for (i = 0; i < brw->wm.prog_data->nr_params; i++) { + constants[i] = *brw->wm.prog_data->param[i]; + } + drm_intel_gem_bo_unmap_gtt(brw->wm.push_const_bo); + } +} + +const struct brw_tracked_state gen6_wm_constants = { + .dirty = { + .mesa = _NEW_PROGRAM_CONSTANTS, + .brw = 0, + .cache = 0, + }, + .prepare = prepare_wm_constants, +}; + +static void upload_wm_state(struct brw_context *brw) { struct intel_context *intel = &brw->intel; GLcontext *ctx = &intel->ctx; const struct brw_fragment_program *fp = brw_fragment_program_const(brw->fragment_program); - unsigned int nr_params = fp->program.Base.Parameters->NumParameters; - drm_intel_bo *constant_bo; - int i; uint32_t dw2, dw4, dw5, dw6; - if (fp->use_const_buffer || nr_params == 0) { + if (fp->use_const_buffer || brw->wm.prog_data->nr_params == 0) { /* Disable the push constant buffers. */ BEGIN_BATCH(5); OUT_BATCH(CMD_3D_CONSTANT_PS_STATE << 16 | (5 - 2)); @@ -55,35 +96,17 @@ upload_wm_state(struct brw_context *brw) OUT_BATCH(0); ADVANCE_BATCH(); } else { - /* Updates the ParamaterValues[i] pointers for all parameters of the - * basic type of PROGRAM_STATE_VAR. - */ - _mesa_load_state_parameters(ctx, fp->program.Base.Parameters); - - constant_bo = drm_intel_bo_alloc(intel->bufmgr, "WM constant_bo", - nr_params * 4 * sizeof(float), - 4096); - drm_intel_gem_bo_map_gtt(constant_bo); - for (i = 0; i < nr_params; i++) { - memcpy((char *)constant_bo->virtual + i * 4 * sizeof(float), - fp->program.Base.Parameters->ParameterValues[i], - 4 * sizeof(float)); - } - drm_intel_gem_bo_unmap_gtt(constant_bo); - BEGIN_BATCH(5); OUT_BATCH(CMD_3D_CONSTANT_PS_STATE << 16 | GEN6_CONSTANT_BUFFER_0_ENABLE | (5 - 2)); - OUT_RELOC(constant_bo, + OUT_RELOC(brw->wm.push_const_bo, I915_GEM_DOMAIN_RENDER, 0, /* XXX: bad domain */ - ALIGN(nr_params, 2) / 2 - 1); + ALIGN(brw->wm.prog_data->nr_params, 8) / 8 - 1); OUT_BATCH(0); OUT_BATCH(0); OUT_BATCH(0); ADVANCE_BATCH(); - - drm_intel_bo_unreference(constant_bo); } intel_batchbuffer_emit_mi_flush(intel->batch); |