radeonsi: count and report temp arrays in scratch separately

v2: only do this if debug output of shader dumping is enabled

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com> (v1)

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 20f4a1d0776fd75260f0368ba5ad56c5ed85ed6a..e434f3af4a4e6e39157cdf428438fdb3ce4ad31c 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -5348,6 +5348,9 @@ static unsigned llvm_get_type_size(LLVMTypeRef type)
        case LLVMVectorTypeKind:
                return LLVMGetVectorSize(type) *
                       llvm_get_type_size(LLVMGetElementType(type));
+       case LLVMArrayTypeKind:
+               return LLVMGetArrayLength(type) *
+                      llvm_get_type_size(LLVMGetElementType(type));
        default:
                assert(0);
                return 0;
@@ -5996,13 +5999,15 @@ static void si_shader_dump_stats(struct si_screen *sscreen,
                        "VGPRS: %d\n"
                        "Spilled SGPRs: %d\n"
                        "Spilled VGPRs: %d\n"
+                       "Private memory VGPRs: %d\n"
                        "Code Size: %d bytes\n"
                        "LDS: %d blocks\n"
                        "Scratch: %d bytes per wave\n"
                        "Max Waves: %d\n"
                        "********************\n\n\n",
                        conf->num_sgprs, conf->num_vgprs,
-                       conf->spilled_sgprs, conf->spilled_vgprs, code_size,
+                       conf->spilled_sgprs, conf->spilled_vgprs,
+                       conf->private_mem_vgprs, code_size,
                        conf->lds_size, conf->scratch_bytes_per_wave,
                        max_simd_waves);
        }
@@ -6010,11 +6015,11 @@ static void si_shader_dump_stats(struct si_screen *sscreen,
        pipe_debug_message(debug, SHADER_INFO,
                           "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
                           "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
-                          "Spilled VGPRs: %d",
+                          "Spilled VGPRs: %d PrivMem VGPRs: %d",
                           conf->num_sgprs, conf->num_vgprs, code_size,
                           conf->lds_size, conf->scratch_bytes_per_wave,
                           max_simd_waves, conf->spilled_sgprs,
-                          conf->spilled_vgprs);
+                          conf->spilled_vgprs, conf->private_mem_vgprs);
 }
 
 static const char *si_get_shader_name(struct si_shader *shader,
@@ -6571,6 +6576,32 @@ static void si_eliminate_const_vs_outputs(struct si_shader_context *ctx)
        }
 }
 
+static void si_count_scratch_private_memory(struct si_shader_context *ctx)
+{
+       ctx->shader->config.private_mem_vgprs = 0;
+
+       /* Process all LLVM instructions. */
+       LLVMBasicBlockRef bb = LLVMGetFirstBasicBlock(ctx->main_fn);
+       while (bb) {
+               LLVMValueRef next = LLVMGetFirstInstruction(bb);
+
+               while (next) {
+                       LLVMValueRef inst = next;
+                       next = LLVMGetNextInstruction(next);
+
+                       if (LLVMGetInstructionOpcode(inst) != LLVMAlloca)
+                               continue;
+
+                       LLVMTypeRef type = LLVMGetElementType(LLVMTypeOf(inst));
+                       /* No idea why LLVM aligns allocas to 4 elements. */
+                       unsigned alignment = LLVMGetAlignment(inst);
+                       unsigned dw_size = align(llvm_get_type_size(type) / 4, alignment);
+                       ctx->shader->config.private_mem_vgprs += dw_size;
+               }
+               bb = LLVMGetNextBasicBlock(bb);
+       }
+}
+
 static bool si_compile_tgsi_main(struct si_shader_context *ctx,
                                 struct si_shader *shader)
 {
@@ -7227,9 +7258,13 @@ int si_compile_tgsi_shader(struct si_screen *sscreen,
        si_llvm_finalize_module(&ctx,
                                    r600_extra_shader_checks(&sscreen->b, ctx.type));
 
-       /* Post-optimization transformations. */
+       /* Post-optimization transformations and analysis. */
        si_eliminate_const_vs_outputs(&ctx);
 
+       if ((debug && debug->debug_message) ||
+           r600_can_dump_shader(&sscreen->b, ctx.type))
+               si_count_scratch_private_memory(&ctx);
+
        /* Compile to bytecode. */
        r = si_compile_llvm(sscreen, &shader->binary, &shader->config, tm,
                            mod, debug, ctx.type, "TGSI shader");

diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index aa37676f887334419af574173ef2606ddda015b8..d4bc47b3878733375fc3ff733f5bf2b24d5d646b 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -448,6 +448,7 @@ struct si_shader_config {
        unsigned                        num_vgprs;
        unsigned                        spilled_sgprs;
        unsigned                        spilled_vgprs;
+       unsigned                        private_mem_vgprs;
        unsigned                        lds_size;
        unsigned                        spi_ps_input_ena;
        unsigned                        spi_ps_input_addr;