Home
last modified time | relevance | path

Searched refs:num_waves (Results 1 - 13 of 13) sorted by relevance

/third_party/mesa3d/src/amd/compiler/
H A Daco_scheduler.cpp34 #define SMEM_WINDOW_SIZE (350 - ctx.num_waves * 35)
35 #define VMEM_WINDOW_SIZE (1024 - ctx.num_waves * 64)
37 #define SMEM_MAX_MOVES (64 - ctx.num_waves * 4)
38 #define VMEM_MAX_MOVES (256 - ctx.num_waves * 16)
39 /* creating clauses decreases def-use distances, so make it less aggressive the lower num_waves is */
40 #define VMEM_CLAUSE_MAX_GRAB_DIST (ctx.num_waves * 2)
125 int16_t num_waves; member
678 /* only move VMEM instructions below descriptor loads. be more aggressive at higher num_waves
681 (cursor.insert_idx - cursor.source_idx > (ctx.num_waves * 4) ||
783 ctx.last_SMEM_stall = 10 - ctx.num_waves
[all...]
H A Daco_live_var_analysis.cpp416 program->num_waves = 0; in update_vgpr_sgpr_demand()
419 program->num_waves = program->dev.physical_sgprs / get_sgpr_alloc(program, new_demand.sgpr); in update_vgpr_sgpr_demand()
422 program->num_waves = in update_vgpr_sgpr_demand()
423 std::min<uint16_t>(program->num_waves, program->dev.physical_vgprs / vgpr_demand); in update_vgpr_sgpr_demand()
425 program->num_waves = std::min(program->num_waves, max_waves); in update_vgpr_sgpr_demand()
428 program->num_waves = max_suitable_waves(program, program->num_waves); in update_vgpr_sgpr_demand()
429 program->max_reg_demand.vgpr = get_addr_vgpr_from_waves(program, program->num_waves); in update_vgpr_sgpr_demand()
430 program->max_reg_demand.sgpr = get_addr_sgpr_from_waves(program, program->num_waves); in update_vgpr_sgpr_demand()
[all...]
H A Daco_statistics.cpp534 double parallelism = program->num_waves; in collect_preasm_stats()
554 fprintf(stderr, "num_waves: %u\n", program->num_waves); in collect_preasm_stats()
H A Daco_spill.cpp1926 if (program->num_waves > 0)
1974 assert(program->num_waves > 0);
H A Daco_validate.cpp999 uint16_t sgpr_limit = get_addr_sgpr_from_waves(program, program->num_waves); in validate_ra()
H A Daco_ir.h2103 uint16_t num_waves = 0;
2286 /* adjust num_waves for workgroup size and LDS limits */
/third_party/mesa3d/src/gallium/drivers/radeonsi/
H A Dsi_debug.c887 unsigned num_waves, FILE *f) in si_print_annotated_shader()
899 for (i = 0; i < num_waves; i++) { in si_print_annotated_shader()
903 if (i == num_waves) in si_print_annotated_shader()
908 num_waves -= i; in si_print_annotated_shader()
945 while (num_waves && inst->addr == waves->pc) { in si_print_annotated_shader()
959 num_waves--; in si_print_annotated_shader()
972 unsigned num_waves = ac_get_wave_info(sctx->gfx_level, waves); in si_dump_annotated_shaders() local
974 fprintf(f, COLOR_CYAN "The number of active waves = %u" COLOR_RESET "\n\n", num_waves); in si_dump_annotated_shaders()
976 si_print_annotated_shader(sctx->shader.vs.current, waves, num_waves, f); in si_dump_annotated_shaders()
977 si_print_annotated_shader(sctx->shader.tcs.current, waves, num_waves, in si_dump_annotated_shaders()
886 si_print_annotated_shader(struct si_shader *shader, struct ac_wave_info *waves, unsigned num_waves, FILE *f) si_print_annotated_shader() argument
[all...]
H A Dgfx10_shader_ngg.c778 LLVMValueRef num_waves = get_tgsize(ctx); in load_vertex_counts() local
783 * 8 waves: valid_mask = ~0ull >> (64 - num_waves * 8) in load_vertex_counts()
784 * 4 waves: valid_mask = ~0 >> (32 - num_waves * 8) in load_vertex_counts()
786 LLVMValueRef num_waves8 = LLVMBuildShl(builder, num_waves, LLVMConstInt(ctx->ac.i32, 3, 0), ""); in load_vertex_counts()
/third_party/mesa3d/src/amd/vulkan/
H A Dradv_debug.c302 struct ac_wave_info *waves, unsigned num_waves, FILE *f) in radv_dump_annotated_shader()
314 for (i = 0; i < num_waves; i++) { in radv_dump_annotated_shader()
319 if (i == num_waves) in radv_dump_annotated_shader()
324 num_waves -= i; in radv_dump_annotated_shader()
345 while (num_waves && start_addr + inst->offset == waves->pc) { in radv_dump_annotated_shader()
359 num_waves--; in radv_dump_annotated_shader()
373 unsigned num_waves = ac_get_wave_info(gfx_level, waves); in radv_dump_annotated_shaders() local
375 fprintf(f, COLOR_CYAN "The number of active waves = %u" COLOR_RESET "\n\n", num_waves); in radv_dump_annotated_shaders()
382 radv_dump_annotated_shader(pipeline->shaders[stage], stage, waves, num_waves, f); in radv_dump_annotated_shaders()
388 for (i = 0; i < num_waves; in radv_dump_annotated_shaders()
301 radv_dump_annotated_shader(struct radv_shader *shader, gl_shader_stage stage, struct ac_wave_info *waves, unsigned num_waves, FILE *f) radv_dump_annotated_shader() argument
[all...]
/third_party/mesa3d/src/amd/common/
H A Dac_debug.c822 unsigned num_waves = 0; in ac_get_wave_info()
839 assert(num_waves < AC_MAX_WAVES_PER_CHIP); in ac_get_wave_info()
840 w = &waves[num_waves]; in ac_get_wave_info()
848 num_waves++; in ac_get_wave_info()
852 qsort(waves, num_waves, sizeof(struct ac_wave_info), compare_wave); in ac_get_wave_info()
855 return num_waves; in ac_get_wave_info()
H A Dac_nir_lower_ngg.c243 /* Broadcast the packed data we read from LDS (to the first 16 lanes, but we only care up to num_waves). */ in summarize_repack()
256 /* Broadcast the packed data we read from LDS (to the first 16 lanes, but we only care up to num_waves). */ in summarize_repack()
348 nir_ssa_def *num_waves = nir_load_num_subgroups(b); in repack_invocations_in_workgroup() local
352 nir_ssa_def *wg_num_repacked_invocations = nir_read_invocation(b, sum, num_waves); in repack_invocations_in_workgroup()
/third_party/mesa3d/src/gallium/drivers/r600/
H A Devergreen_compute.c610 unsigned num_waves; in evergreen_emit_dispatch() local
625 /* num_waves = ceil((tg_size.x * tg_size.y, tg_size.z) / (16 * num_pipes)) */ in evergreen_emit_dispatch()
626 num_waves = (info->block[0] * info->block[1] * info->block[2] + in evergreen_emit_dispatch()
632 num_pipes, num_waves, lds_size); in evergreen_emit_dispatch()
658 lds_size | (num_waves << 14)); in evergreen_emit_dispatch()
H A Devergreen_state.c4539 unsigned num_waves; in evergreen_setup_tess_constants() local
4600 num_waves = ceilf((float)(*num_patches * num_tcs_output_cp) / (float)wave_divisor); in evergreen_setup_tess_constants()
4602 rctx->lds_alloc = (lds_size | (num_waves << 14)); in evergreen_setup_tess_constants()

Completed in 32 milliseconds