This will use vector loads (64 loads per 64 threads) instead of scalar loads (1 load per 64 threads). Vector loads are much slower. allow_smem should be set to allow scalar loads for constbufs.
Also, can_speculate should be true for constbuf loads. I think you really need to check shader-db or shader asm dumps, because this clearly decreases performance of all shaders using UBOs. Marek On Thu, Aug 17, 2017 at 1:03 PM, Timothy Arceri <tarc...@itsqueeze.com> wrote: > --- > src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c | 23 > +++++++++++++++-------- > 1 file changed, 15 insertions(+), 8 deletions(-) > > diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c > b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c > index f8c99ff7e7..05ca710094 100644 > --- a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c > +++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c > @@ -66,32 +66,36 @@ static LLVMValueRef get_buffer_size( > LLVMConstInt(ctx->i32, 0x3FFF, 0), ""); > > size = LLVMBuildUDiv(builder, size, stride, ""); > } > > return size; > } > > static LLVMValueRef > shader_buffer_fetch_rsrc(struct si_shader_context *ctx, > - const struct tgsi_full_src_register *reg) > + const struct tgsi_full_src_register *reg, > + bool ubo) > { > LLVMValueRef index; > > if (!reg->Register.Indirect) { > index = LLVMConstInt(ctx->i32, reg->Register.Index, false); > } else { > index = si_get_indirect_index(ctx, ®->Indirect, > reg->Register.Index); > } > > - return ctx->abi.load_ssbo(&ctx->abi, index, false); > + if (ubo) > + return ctx->abi.load_ubo(&ctx->abi, index); > + else > + return ctx->abi.load_ssbo(&ctx->abi, index, false); > } > > static bool tgsi_is_array_sampler(unsigned target) > { > return target == TGSI_TEXTURE_1D_ARRAY || > target == TGSI_TEXTURE_SHADOW1D_ARRAY || > target == TGSI_TEXTURE_2D_ARRAY || > target == TGSI_TEXTURE_SHADOW2D_ARRAY || > target == TGSI_TEXTURE_CUBE_ARRAY || > target == TGSI_TEXTURE_SHADOWCUBE_ARRAY || > @@ -356,26 +360,28 @@ static void load_fetch_args( > struct lp_build_emit_data * emit_data) > { > struct si_shader_context *ctx = si_shader_context(bld_base); > struct gallivm_state *gallivm = &ctx->gallivm; > const struct tgsi_full_instruction * inst = emit_data->inst; > unsigned target = inst->Memory.Texture; > LLVMValueRef rsrc; > > emit_data->dst_type = ctx->v4f32; > > - if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) { > + if (inst->Src[0].Register.File == TGSI_FILE_BUFFER || > + inst->Src[0].Register.File == TGSI_FILE_UBO) { > LLVMBuilderRef builder = gallivm->builder; > LLVMValueRef offset; > LLVMValueRef tmp; > > - rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]); > + bool ubo = inst->Src[0].Register.File == TGSI_FILE_UBO; > + rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0], ubo); > > tmp = lp_build_emit_fetch(bld_base, inst, 1, 0); > offset = LLVMBuildBitCast(builder, tmp, ctx->i32, ""); > > buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0, > offset, false, false); > } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE || > tgsi_is_bindless_image_file(inst->Src[0].Register.File)) { > LLVMValueRef coords; > > @@ -574,21 +580,22 @@ static void load_emit( > if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) > si_emit_waitcnt(ctx, VM_CNT); > > can_speculate = !(inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) && > is_oneway_access_only(inst, info, > info->shader_buffers_store | > info->shader_buffers_atomic, > info->images_store | > info->images_atomic); > > - if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) { > + if (inst->Src[0].Register.File == TGSI_FILE_BUFFER || > + inst->Src[0].Register.File == TGSI_FILE_UBO) { > load_emit_buffer(ctx, emit_data, can_speculate); > return; > } > > if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) { > emit_data->output[emit_data->chan] = > lp_build_intrinsic( > builder, > "llvm.amdgcn.buffer.load.format.v4f32", emit_data->dst_type, > emit_data->args, emit_data->arg_count, > get_load_intr_attribs(can_speculate)); > @@ -629,21 +636,21 @@ static void store_fetch_args( > data = lp_build_gather_values(gallivm, chans, 4); > > emit_data->args[emit_data->arg_count++] = data; > > memory = tgsi_full_src_register_from_dst(&inst->Dst[0]); > > if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) { > LLVMValueRef offset; > LLVMValueRef tmp; > > - rsrc = shader_buffer_fetch_rsrc(ctx, &memory); > + rsrc = shader_buffer_fetch_rsrc(ctx, &memory, false); > > tmp = lp_build_emit_fetch(bld_base, inst, 0, 0); > offset = LLVMBuildBitCast(builder, tmp, ctx->i32, ""); > > buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0, > offset, false, false); > } else if (inst->Dst[0].Register.File == TGSI_FILE_IMAGE || > tgsi_is_bindless_image_file(inst->Dst[0].Register.File)) { > unsigned target = inst->Memory.Texture; > LLVMValueRef coords; > @@ -849,21 +856,21 @@ static void atomic_fetch_args( > /* llvm.amdgcn.image/buffer.atomic.cmpswap reflect the hardware order > * of arguments, which is reversed relative to TGSI (and GLSL) > */ > if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) > emit_data->args[emit_data->arg_count++] = data2; > emit_data->args[emit_data->arg_count++] = data1; > > if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) { > LLVMValueRef offset; > > - rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]); > + rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0], false); > > tmp = lp_build_emit_fetch(bld_base, inst, 1, 0); > offset = LLVMBuildBitCast(builder, tmp, ctx->i32, ""); > > buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0, > offset, true, false); > } else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE || > tgsi_is_bindless_image_file(inst->Src[0].Register.File)) { > unsigned target = inst->Memory.Texture; > LLVMValueRef coords; > @@ -1057,21 +1064,21 @@ static void resq_fetch_args( > struct lp_build_tgsi_context * bld_base, > struct lp_build_emit_data * emit_data) > { > struct si_shader_context *ctx = si_shader_context(bld_base); > const struct tgsi_full_instruction *inst = emit_data->inst; > const struct tgsi_full_src_register *reg = &inst->Src[0]; > > emit_data->dst_type = ctx->v4i32; > > if (reg->Register.File == TGSI_FILE_BUFFER) { > - emit_data->args[0] = shader_buffer_fetch_rsrc(ctx, reg); > + emit_data->args[0] = shader_buffer_fetch_rsrc(ctx, reg, > false); > emit_data->arg_count = 1; > } else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) { > image_fetch_rsrc(bld_base, reg, false, inst->Memory.Texture, > &emit_data->args[0]); > emit_data->arg_count = 1; > } else { > LLVMValueRef res_ptr; > unsigned image_target; > > if (inst->Memory.Texture == TGSI_TEXTURE_3D) > -- > 2.13.4 > > _______________________________________________ > mesa-dev mailing list > mesa-dev@lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/mesa-dev _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev