radeonsi: pull iteration over vertex streams into GS copy shader logic
authorNicolai Hähnle <nicolai.haehnle@amd.com>
Tue, 29 Nov 2016 13:23:05 +0000 (14:23 +0100)
committerNicolai Hähnle <nicolai.haehnle@amd.com>
Mon, 12 Dec 2016 08:04:33 +0000 (09:04 +0100)
The iteration is not needed for normal vertex shaders.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
src/gallium/drivers/radeonsi/si_shader.c

index 091560d279e6afcbbd5f406c4c74e79b65b79d91..5a40c5ee91d13c87c8b1597a09c7a1ba982d762b 100644 (file)
@@ -2265,11 +2265,13 @@ static void emit_streamout_output(struct si_shader_context *ctx,
                                   stream_out->dst_offset * 4);
 }
 
-/* On SI, the vertex shader is responsible for writing streamout data
- * to buffers. */
+/**
+ * Write streamout data to buffers for vertex stream @p stream (different
+ * vertex streams can occur for GS copy shaders).
+ */
 static void si_llvm_emit_streamout(struct si_shader_context *ctx,
                                   struct si_shader_output_values *outputs,
-                                  unsigned noutput)
+                                  unsigned noutput, unsigned stream)
 {
        struct si_shader_selector *sel = ctx->shader->selector;
        struct pipe_stream_output_info *so = &sel->so;
@@ -2288,27 +2290,11 @@ static void si_llvm_emit_streamout(struct si_shader_context *ctx,
        LLVMValueRef can_emit =
                LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");
 
-       LLVMValueRef stream_id =
-               unpack_param(ctx, ctx->param_streamout_config, 24, 2);
-
        /* Emit the streamout code conditionally. This actually avoids
         * out-of-bounds buffer access. The hw tells us via the SGPR
         * (so_vtx_count) which threads are allowed to emit streamout data. */
        lp_build_if(&if_ctx, gallivm, can_emit);
-
-       for (int stream = 0; stream < 4; ++stream) {
-               struct lp_build_if_state if_ctx_stream;
-
-               if (!sel->info.num_stream_output_components[stream])
-                       continue;
-
-               LLVMValueRef is_stream =
-                       LLVMBuildICmp(builder, LLVMIntEQ,
-                                     stream_id,
-                                     lp_build_const_int32(gallivm, stream), "");
-
-               lp_build_if(&if_ctx_stream, gallivm, is_stream);
-
+       {
                /* The buffer offset is computed as follows:
                 *   ByteOffset = streamout_offset[buffer_id]*4 +
                 *                (streamout_write_index + thread_id)*stride[buffer_id] +
@@ -2360,8 +2346,6 @@ static void si_llvm_emit_streamout(struct si_shader_context *ctx,
                        emit_streamout_output(ctx, so_buffers, so_write_offset,
                                              &so->output[i], &outputs[reg]);
                }
-
-               lp_build_endif(&if_ctx_stream);
        }
        lp_build_endif(&if_ctx);
 }
@@ -2932,7 +2916,7 @@ static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context *bld_base)
                                     VS_EPILOG_PRIMID_LOC, "");
 
        if (ctx->shader->selector->so.num_outputs)
-               si_llvm_emit_streamout(ctx, outputs, i);
+               si_llvm_emit_streamout(ctx, outputs, i, 0);
        si_llvm_export_vs(bld_base, outputs, i);
        FREE(outputs);
 }
@@ -6354,6 +6338,7 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
        struct si_shader_context ctx;
        struct si_shader *shader;
        struct gallivm_state *gallivm = &ctx.gallivm;
+       LLVMBuilderRef builder;
        struct lp_build_tgsi_context *bld_base = &ctx.soa.bld_base;
        struct lp_build_context *uint = &bld_base->uint_bld;
        struct si_shader_output_values *outputs;
@@ -6379,6 +6364,8 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
        si_init_shader_ctx(&ctx, sscreen, shader, tm);
        ctx.type = PIPE_SHADER_VERTEX;
 
+       builder = gallivm->builder;
+
        create_meta_data(&ctx);
        create_function(&ctx);
        preload_ring_buffers(&ctx);
@@ -6395,6 +6382,14 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
        args[7] = uint->one;  /* SLC */
        args[8] = uint->zero; /* TFE */
 
+       /* Fetch the vertex stream ID.*/
+       LLVMValueRef stream_id;
+
+       if (gs_selector->so.num_outputs)
+               stream_id = unpack_param(&ctx, ctx.param_streamout_config, 24, 2);
+       else
+               stream_id = uint->zero;
+
        /* Fetch vertex data from GSVS ring */
        for (i = 0; i < gsinfo->num_outputs; ++i) {
                unsigned chan;
@@ -6420,8 +6415,25 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
                }
        }
 
-       if (gs_selector->so.num_outputs)
-               si_llvm_emit_streamout(&ctx, outputs, gsinfo->num_outputs);
+       if (gs_selector->so.num_outputs) {
+               for (int stream = 0; stream < 4; stream++) {
+                       struct lp_build_if_state if_ctx_stream;
+
+                       if (!gsinfo->num_stream_output_components[stream])
+                               continue;
+
+                       LLVMValueRef is_stream =
+                               LLVMBuildICmp(builder, LLVMIntEQ,
+                                             stream_id,
+                                             lp_build_const_int32(gallivm, stream), "");
+
+                       lp_build_if(&if_ctx_stream, gallivm, is_stream);
+                       si_llvm_emit_streamout(&ctx, outputs,
+                                              gsinfo->num_outputs,
+                                              stream);
+                       lp_build_endif(&if_ctx_stream);
+               }
+       }
        si_llvm_export_vs(bld_base, outputs, gsinfo->num_outputs);
 
        LLVMBuildRetVoid(gallivm->builder);