Module: Mesa Branch: main Commit: a4b60295a797cda88be111029da146117040081c URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=a4b60295a797cda88be111029da146117040081c
Author: Qiang Yu <[email protected]> Date: Wed Mar 29 15:56:21 2023 +0800 aco,ac/llvm,radv,radeonsi: handle ps bc optimization in nir for radv The side effect is removing the aco/llvm backend bc optimization code and linear/persp_centroid variable. Reviewed-by: Timur Kristóf <[email protected]> Signed-off-by: Qiang Yu <[email protected]> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22199> --- src/amd/compiler/aco_instruction_selection.cpp | 59 +------------------------- src/amd/compiler/aco_instruction_selection.h | 3 -- src/amd/llvm/ac_nir_to_llvm.c | 4 +- src/amd/llvm/ac_shader_abi.h | 1 - src/amd/vulkan/radv_nir_to_llvm.c | 36 +--------------- src/amd/vulkan/radv_pipeline.c | 5 +++ src/gallium/drivers/radeonsi/si_shader_llvm.c | 3 -- 7 files changed, 9 insertions(+), 102 deletions(-) diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp index 3cabec8c13d..12c386c0639 100644 --- a/src/amd/compiler/aco_instruction_selection.cpp +++ b/src/amd/compiler/aco_instruction_selection.cpp @@ -8080,7 +8080,7 @@ get_interp_param(isel_context* ctx, nir_intrinsic_op intrin, intrin == nir_intrinsic_load_barycentric_at_offset) { return get_arg(ctx, linear ? ctx->args->linear_center : ctx->args->persp_center); } else if (intrin == nir_intrinsic_load_barycentric_centroid) { - return linear ? ctx->linear_centroid : ctx->persp_centroid; + return get_arg(ctx, linear ? ctx->args->linear_centroid : ctx->args->persp_centroid); } else { assert(intrin == nir_intrinsic_load_barycentric_sample); return get_arg(ctx, linear ? ctx->args->linear_sample : ctx->args->persp_sample); @@ -11200,60 +11200,6 @@ split_arguments(isel_context* ctx, Pseudo_instruction* startpgm) } } -void -handle_bc_optimize(isel_context* ctx) -{ - /* needed when SPI_PS_IN_CONTROL.BC_OPTIMIZE_DISABLE is set to 0 */ - Builder bld(ctx->program, ctx->block); - uint32_t spi_ps_input_ena = ctx->program->config->spi_ps_input_ena; - bool uses_center = - G_0286CC_PERSP_CENTER_ENA(spi_ps_input_ena) || G_0286CC_LINEAR_CENTER_ENA(spi_ps_input_ena); - bool uses_persp_centroid = G_0286CC_PERSP_CENTROID_ENA(spi_ps_input_ena); - bool uses_linear_centroid = G_0286CC_LINEAR_CENTROID_ENA(spi_ps_input_ena); - - if (uses_persp_centroid) - ctx->persp_centroid = get_arg(ctx, ctx->args->persp_centroid); - if (uses_linear_centroid) - ctx->linear_centroid = get_arg(ctx, ctx->args->linear_centroid); - - if (uses_center && (uses_persp_centroid || uses_linear_centroid)) { - Temp sel = bld.vopc_e64(aco_opcode::v_cmp_lt_i32, bld.def(bld.lm), - get_arg(ctx, ctx->args->prim_mask), Operand::zero()); - - if (uses_persp_centroid) { - Temp new_coord[2]; - for (unsigned i = 0; i < 2; i++) { - Temp persp_centroid = - emit_extract_vector(ctx, get_arg(ctx, ctx->args->persp_centroid), i, v1); - Temp persp_center = - emit_extract_vector(ctx, get_arg(ctx, ctx->args->persp_center), i, v1); - new_coord[i] = - bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), persp_centroid, persp_center, sel); - } - ctx->persp_centroid = bld.tmp(v2); - bld.pseudo(aco_opcode::p_create_vector, Definition(ctx->persp_centroid), - Operand(new_coord[0]), Operand(new_coord[1])); - emit_split_vector(ctx, ctx->persp_centroid, 2); - } - - if (uses_linear_centroid) { - Temp new_coord[2]; - for (unsigned i = 0; i < 2; i++) { - Temp linear_centroid = - emit_extract_vector(ctx, get_arg(ctx, ctx->args->linear_centroid), i, v1); - Temp linear_center = - emit_extract_vector(ctx, get_arg(ctx, ctx->args->linear_center), i, v1); - new_coord[i] = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), linear_centroid, - linear_center, sel); - } - ctx->linear_centroid = bld.tmp(v2); - bld.pseudo(aco_opcode::p_create_vector, Definition(ctx->linear_centroid), - Operand(new_coord[0]), Operand(new_coord[1])); - emit_split_vector(ctx, ctx->linear_centroid, 2); - } - } -} - void setup_fp_mode(isel_context* ctx, nir_shader* shader) { @@ -11444,9 +11390,6 @@ select_program(Program* program, unsigned shader_count, struct nir_shader* const } else if (ctx.stage == geometry_gs) ctx.gs_wave_id = get_arg(&ctx, args->gs_wave_id); - if (ctx.stage == fragment_fs) - handle_bc_optimize(&ctx); - visit_cf_list(&ctx, &func->body); if (nir->info.stage == MESA_SHADER_GEOMETRY && !ngg_gs) { diff --git a/src/amd/compiler/aco_instruction_selection.h b/src/amd/compiler/aco_instruction_selection.h index acb254111dd..65a32a9b561 100644 --- a/src/amd/compiler/aco_instruction_selection.h +++ b/src/amd/compiler/aco_instruction_selection.h @@ -91,9 +91,6 @@ struct isel_context { Temp arg_temps[AC_MAX_ARGS]; - /* FS inputs */ - Temp persp_centroid, linear_centroid; - /* GS inputs */ Temp gs_wave_id; diff --git a/src/amd/llvm/ac_nir_to_llvm.c b/src/amd/llvm/ac_nir_to_llvm.c index 5a36ee54fb3..3a3c7d5c950 100644 --- a/src/amd/llvm/ac_nir_to_llvm.c +++ b/src/amd/llvm/ac_nir_to_llvm.c @@ -3195,7 +3195,7 @@ static LLVMValueRef lookup_interp_param(struct ac_nir_context *ctx, enum glsl_in if (location == INTERP_CENTER) return ac_get_arg(&ctx->ac, ctx->args->persp_center); else if (location == INTERP_CENTROID) - return ctx->abi->persp_centroid; + return ac_get_arg(&ctx->ac, ctx->args->persp_centroid); else if (location == INTERP_SAMPLE) return ac_get_arg(&ctx->ac, ctx->args->persp_sample); break; @@ -3203,7 +3203,7 @@ static LLVMValueRef lookup_interp_param(struct ac_nir_context *ctx, enum glsl_in if (location == INTERP_CENTER) return ac_get_arg(&ctx->ac, ctx->args->linear_center); else if (location == INTERP_CENTROID) - return ctx->abi->linear_centroid; + return ac_get_arg(&ctx->ac, ctx->args->linear_centroid); else if (location == INTERP_SAMPLE) return ac_get_arg(&ctx->ac, ctx->args->linear_sample); break; diff --git a/src/amd/llvm/ac_shader_abi.h b/src/amd/llvm/ac_shader_abi.h index 0369d574a5f..ec37a008c52 100644 --- a/src/amd/llvm/ac_shader_abi.h +++ b/src/amd/llvm/ac_shader_abi.h @@ -46,7 +46,6 @@ struct ac_shader_abi { LLVMValueRef vertex_id; LLVMValueRef vs_rel_patch_id; LLVMValueRef instance_id; - LLVMValueRef persp_centroid, linear_centroid; LLVMValueRef user_data; /* replaced registers when culling enabled */ diff --git a/src/amd/vulkan/radv_nir_to_llvm.c b/src/amd/vulkan/radv_nir_to_llvm.c index 7ae308ea4f6..0bcbd29b53f 100644 --- a/src/amd/vulkan/radv_nir_to_llvm.c +++ b/src/amd/vulkan/radv_nir_to_llvm.c @@ -253,38 +253,6 @@ radv_get_sampler_desc(struct ac_shader_abi *abi, LLVMValueRef index, return radv_load_rsrc(ctx, index, v4 ? ctx->ac.v4i32 : ctx->ac.v8i32); } -static void -prepare_interp_optimize(struct radv_shader_context *ctx, struct nir_shader *nir) -{ - bool uses_center = false; - bool uses_centroid = false; - nir_foreach_shader_in_variable (variable, nir) { - if (glsl_get_base_type(glsl_without_array(variable->type)) != GLSL_TYPE_FLOAT || - variable->data.sample) - continue; - - if (variable->data.centroid) - uses_centroid = true; - else - uses_center = true; - } - - ctx->abi.persp_centroid = ac_get_arg(&ctx->ac, ctx->args->ac.persp_centroid); - ctx->abi.linear_centroid = ac_get_arg(&ctx->ac, ctx->args->ac.linear_centroid); - - if (uses_center && uses_centroid) { - LLVMValueRef sel = - LLVMBuildICmp(ctx->ac.builder, LLVMIntSLT, ac_get_arg(&ctx->ac, ctx->args->ac.prim_mask), - ctx->ac.i32_0, ""); - ctx->abi.persp_centroid = - LLVMBuildSelect(ctx->ac.builder, sel, ac_get_arg(&ctx->ac, ctx->args->ac.persp_center), - ctx->abi.persp_centroid, ""); - ctx->abi.linear_centroid = - LLVMBuildSelect(ctx->ac.builder, sel, ac_get_arg(&ctx->ac, ctx->args->ac.linear_center), - ctx->abi.linear_centroid, ""); - } -} - static void scan_shader_output_decl(struct radv_shader_context *ctx, struct nir_variable *variable, struct nir_shader *shader, gl_shader_stage stage) @@ -548,9 +516,7 @@ ac_translate_nir_to_llvm(struct ac_llvm_compiler *ac_llvm, LLVMPositionBuilderAtEnd(ctx.ac.builder, then_block); } - if (shaders[shader_idx]->info.stage == MESA_SHADER_FRAGMENT) - prepare_interp_optimize(&ctx, shaders[shader_idx]); - else if (shaders[shader_idx]->info.stage == MESA_SHADER_GEOMETRY && !info->is_ngg) + if (shaders[shader_idx]->info.stage == MESA_SHADER_GEOMETRY && !info->is_ngg) prepare_gs_input_vgprs(&ctx, shader_count >= 2); if (!ac_nir_translate(&ctx.ac, &ctx.abi, &args->ac, shaders[shader_idx])) { diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c index b5322b74b1d..a90a3627949 100644 --- a/src/amd/vulkan/radv_pipeline.c +++ b/src/amd/vulkan/radv_pipeline.c @@ -610,6 +610,11 @@ radv_postprocess_nir(struct radv_device *device, const struct radv_pipeline_layo .enable_mrt_output_nan_fixup = pipeline_key->ps.epilog.enable_mrt_output_nan_fixup, .no_color_export = stage->info.ps.has_epilog, + + .bc_optimize_for_persp = + stage->info.ps.reads_persp_center && stage->info.ps.reads_persp_centroid, + .bc_optimize_for_linear = + stage->info.ps.reads_linear_center && stage->info.ps.reads_linear_centroid, }; NIR_PASS_V(stage->nir, ac_nir_lower_ps, &options); diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm.c b/src/gallium/drivers/radeonsi/si_shader_llvm.c index 583abbd61e2..9625aea0249 100644 --- a/src/gallium/drivers/radeonsi/si_shader_llvm.c +++ b/src/gallium/drivers/radeonsi/si_shader_llvm.c @@ -250,9 +250,6 @@ void si_llvm_create_main_func(struct si_shader_context *ctx) */ if (shader->is_monolithic && shader->key.ge.part.vs.prolog.ls_vgpr_fix) ac_fixup_ls_hs_input_vgprs(&ctx->ac, &ctx->abi, &ctx->args->ac); - } else if (ctx->stage == MESA_SHADER_FRAGMENT) { - ctx->abi.persp_centroid = ac_get_arg(&ctx->ac, ctx->args->ac.persp_centroid); - ctx->abi.linear_centroid = ac_get_arg(&ctx->ac, ctx->args->ac.linear_centroid); } }
