radv: enable lowering of GS intrinsics for the LLVM backend
authorSamuel Pitoiset <samuel.pitoiset@gmail.com>
Fri, 13 Mar 2020 10:23:07 +0000 (11:23 +0100)
committerSamuel Pitoiset <samuel.pitoiset@gmail.com>
Wed, 8 Apr 2020 06:24:05 +0000 (08:24 +0200)
This replaces emit_vertex with:
   if (vertex_count < max_vertices) {
      emit_vertex_with_counter vertex_count ...
      vertex_count += 1
   }

Which is exactly what NIR->LLVM was doing but at NIR level. This
pass is already called by ACO.

pipeline-db changes on GFX10:
Totals from affected shaders:
SGPRS: 1952 -> 1912 (-2.05 %)
VGPRS: 2112 -> 2044 (-3.22 %)
Code Size: 189368 -> 185620 (-1.98 %) bytes
Max Waves: 494 -> 491 (-0.61 %)

No pipeline-db changes on other generations.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4182>

src/amd/llvm/ac_nir_to_llvm.c
src/amd/llvm/ac_shader_abi.h
src/amd/vulkan/radv_nir_to_llvm.c
src/amd/vulkan/radv_shader.c

index 00970638516ff908b5e56ae10da132fa12930ee6..627f5d2d931212ccff4dfc57db393c0348acc49d 100644 (file)
@@ -3938,7 +3938,16 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
        case nir_intrinsic_emit_vertex:
                ctx->abi->emit_vertex(ctx->abi, nir_intrinsic_stream_id(instr), ctx->abi->outputs);
                break;
+       case nir_intrinsic_emit_vertex_with_counter: {
+               unsigned stream = nir_intrinsic_stream_id(instr);
+               LLVMValueRef next_vertex = get_src(ctx, instr->src[0]);
+               ctx->abi->emit_vertex_with_counter(ctx->abi, stream,
+                                                  next_vertex,
+                                                  ctx->abi->outputs);
+               break;
+       }
        case nir_intrinsic_end_primitive:
+       case nir_intrinsic_end_primitive_with_counter:
                ctx->abi->emit_primitive(ctx->abi, nir_intrinsic_stream_id(instr));
                break;
        case nir_intrinsic_load_tess_coord:
index f73465b014f6a3f45e89382cd8d823589f01e0ed..18f85a7911cd216a1ec44115c13473f63ddcdf14 100644 (file)
@@ -80,6 +80,11 @@ struct ac_shader_abi {
        void (*emit_primitive)(struct ac_shader_abi *abi,
                               unsigned stream);
 
+       void (*emit_vertex_with_counter)(struct ac_shader_abi *abi,
+                                        unsigned stream,
+                                        LLVMValueRef vertexidx,
+                                        LLVMValueRef *addrs);
+
        LLVMValueRef (*load_inputs)(struct ac_shader_abi *abi,
                                    unsigned location,
                                    unsigned driver_location,
index 7cb8deddc105f5ca45437317d974b01341e03c61..900246d275ef57a13cf3240a1ae7360cfa45b573 100644 (file)
@@ -876,39 +876,21 @@ static LLVMValueRef load_sample_mask_in(struct ac_shader_abi *abi)
 
 static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context *ctx,
                                     unsigned stream,
+                                    LLVMValueRef vertexidx,
                                     LLVMValueRef *addrs);
 
 static void
-visit_emit_vertex(struct ac_shader_abi *abi, unsigned stream, LLVMValueRef *addrs)
+visit_emit_vertex_with_counter(struct ac_shader_abi *abi, unsigned stream,
+                              LLVMValueRef vertexidx, LLVMValueRef *addrs)
 {
-       LLVMValueRef gs_next_vertex;
-       LLVMValueRef can_emit;
        unsigned offset = 0;
        struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
 
        if (ctx->args->options->key.vs_common_out.as_ngg) {
-               gfx10_ngg_gs_emit_vertex(ctx, stream, addrs);
+               gfx10_ngg_gs_emit_vertex(ctx, stream, vertexidx, addrs);
                return;
        }
 
-       /* Write vertex attribute values to GSVS ring */
-       gs_next_vertex = LLVMBuildLoad(ctx->ac.builder,
-                                      ctx->gs_next_vertex[stream],
-                                      "");
-
-       /* If this thread has already emitted the declared maximum number of
-        * vertices, don't emit any more: excessive vertex emissions are not
-        * supposed to have any effect.
-        */
-       can_emit = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, gs_next_vertex,
-                                LLVMConstInt(ctx->ac.i32, ctx->shader->info.gs.vertices_out, false), "");
-
-       bool use_kill = !ctx->args->shader_info->gs.writes_memory;
-       if (use_kill)
-               ac_build_kill_if_false(&ctx->ac, can_emit);
-       else
-               ac_build_ifcc(&ctx->ac, can_emit, 6505);
-
        for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
                unsigned output_usage_mask =
                        ctx->args->shader_info->gs.output_usage_mask[i];
@@ -933,7 +915,7 @@ visit_emit_vertex(struct ac_shader_abi *abi, unsigned stream, LLVMValueRef *addr
 
                        offset++;
 
-                       voffset = LLVMBuildAdd(ctx->ac.builder, voffset, gs_next_vertex, "");
+                       voffset = LLVMBuildAdd(ctx->ac.builder, voffset, vertexidx, "");
                        voffset = LLVMBuildMul(ctx->ac.builder, voffset, LLVMConstInt(ctx->ac.i32, 4, false), "");
 
                        out_val = ac_to_integer(&ctx->ac, out_val);
@@ -949,16 +931,9 @@ visit_emit_vertex(struct ac_shader_abi *abi, unsigned stream, LLVMValueRef *addr
                }
        }
 
-       gs_next_vertex = LLVMBuildAdd(ctx->ac.builder, gs_next_vertex,
-                                     ctx->ac.i32_1, "");
-       LLVMBuildStore(ctx->ac.builder, gs_next_vertex, ctx->gs_next_vertex[stream]);
-
        ac_build_sendmsg(&ctx->ac,
                         AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
                         ctx->gs_wave_id);
-
-       if (!use_kill)
-               ac_build_endif(&ctx->ac, 6505);
 }
 
 static void
@@ -3309,25 +3284,11 @@ static void gfx10_ngg_gs_emit_epilogue_2(struct radv_shader_context *ctx)
 
 static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context *ctx,
                                     unsigned stream,
+                                    LLVMValueRef vertexidx,
                                     LLVMValueRef *addrs)
 {
        LLVMBuilderRef builder = ctx->ac.builder;
        LLVMValueRef tmp;
-       const LLVMValueRef vertexidx =
-               LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
-
-       /* If this thread has already emitted the declared maximum number of
-        * vertices, skip the write: excessive vertex emissions are not
-        * supposed to have any effect.
-        */
-       const LLVMValueRef can_emit =
-               LLVMBuildICmp(builder, LLVMIntULT, vertexidx,
-                             LLVMConstInt(ctx->ac.i32, ctx->shader->info.gs.vertices_out, false), "");
-       ac_build_ifcc(&ctx->ac, can_emit, 9001);
-
-       tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
-       tmp = LLVMBuildSelect(builder, can_emit, tmp, vertexidx, "");
-       LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
 
        const LLVMValueRef vertexptr =
                ngg_gs_emit_vertex_ptr(ctx, get_thread_id_in_tg(ctx), vertexidx);
@@ -3359,6 +3320,13 @@ static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context *ctx,
        }
        assert(out_idx * 4 <= ctx->args->shader_info->gs.gsvs_vertex_size);
 
+       /* Store the current number of emitted vertices to zero out remaining
+        * primitive flags in case the geometry shader doesn't emit the maximum
+        * number of vertices.
+        */
+       tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
+       LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
+
        /* Determine and store whether this vertex completed a primitive. */
        const LLVMValueRef curverts = LLVMBuildLoad(builder, ctx->gs_curprim_verts[stream], "");
 
@@ -3395,8 +3363,6 @@ static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context *ctx,
        tmp = LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
        tmp = LLVMBuildAdd(builder, tmp, LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i32, ""), "");
        LLVMBuildStore(builder, tmp, ctx->gs_generated_prims[stream]);
-
-       ac_build_endif(&ctx->ac, 9001);
 }
 
 static void
@@ -3948,7 +3914,7 @@ LLVMModuleRef ac_translate_nir_to_llvm(struct ac_llvm_compiler *ac_llvm,
 
        ctx.abi.inputs = &ctx.inputs[0];
        ctx.abi.emit_outputs = handle_shader_outputs_post;
-       ctx.abi.emit_vertex = visit_emit_vertex;
+       ctx.abi.emit_vertex_with_counter = visit_emit_vertex_with_counter;
        ctx.abi.load_ubo = radv_load_ubo;
        ctx.abi.load_ssbo = radv_load_ssbo;
        ctx.abi.load_sampler_desc = radv_get_sampler_desc;
index 537ab9f8b7f5e2a2537c4a463e635fba93a5a895..bb88b368d0531ef06a380c1d5010e59a9004628e 100644 (file)
@@ -453,8 +453,7 @@ radv_shader_compile_to_nir(struct radv_device *device,
 
        nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
 
-       if (nir->info.stage == MESA_SHADER_GEOMETRY &&
-           device->physical_device->use_aco)
+       if (nir->info.stage == MESA_SHADER_GEOMETRY)
                nir_lower_gs_intrinsics(nir, true);
 
        static const nir_lower_tex_options tex_options = {