On 10.05.2016 11:36, Bas Nieuwenhuizen wrote:
On Tue, May 10, 2016 at 6:28 PM, Nicolai Hähnle <nhaeh...@gmail.com> wrote:On 10.05.2016 11:25, Bas Nieuwenhuizen wrote:On Tue, May 10, 2016 at 6:13 PM, Nicolai Hähnle <nhaeh...@gmail.com> wrote:On 10.05.2016 05:52, Bas Nieuwenhuizen wrote:Signed-off-by: Bas Nieuwenhuizen <b...@basnieuwenhuizen.nl> --- src/gallium/drivers/radeonsi/si_shader.c | 81 ++++++++++++++++++++++++++++++++ 1 file changed, 81 insertions(+) diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index 5897149..d3df4d6 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -733,6 +733,87 @@ static void build_tbuffer_store_dwords(struct si_shader_context *ctx, V_008F0C_BUF_NUM_FORMAT_UINT, 1, 0, 1, 1, 0); } +static LLVMValueRef build_buffer_load(struct si_shader_context *ctx, + LLVMValueRef rsrc, + int num_channels, + LLVMValueRef vindex, + LLVMValueRef voffset, + LLVMValueRef soffset, + unsigned inst_offset, + unsigned glc, + unsigned slc) +{ + struct gallivm_state *gallivm = &ctx->radeon_bld.gallivm; + LLVMValueRef args[] = { + LLVMBuildBitCast(gallivm->builder, rsrc, ctx->v16i8, ""), + voffset ? voffset : vindex, + soffset, + LLVMConstInt(ctx->i32, inst_offset, 0), + LLVMConstInt(ctx->i32, voffset ? 1 : 0, 0), // offen + LLVMConstInt(ctx->i32, vindex ? 1 : 0, 0), //idxen + LLVMConstInt(ctx->i32, glc, 0), + LLVMConstInt(ctx->i32, slc, 0), + LLVMConstInt(ctx->i32, 0, 0), // TFE + }; + + unsigned func = CLAMP(num_channels, 1, 3) - 1; + LLVMTypeRef types[] = {ctx->i32, LLVMVectorType(ctx->i32, 2), ctx->v4i32}; + const char *type_names[] = {"i32", "v2i32", "v4i32"}; + const char *arg_type = "i32"; + + if (voffset && vindex) { + LLVMValueRef vaddr[] = {vindex, voffset}; + + arg_type = "v2i32"; + args[1] = lp_build_gather_values(gallivm, vaddr, 2); + } + + char name[256]; + snprintf(name, sizeof(name), "llvm.SI.buffer.load.dword.%s.%s", + type_names[func], arg_type);We're generally trying to get away from the llvm.SI.* intrinsics and use the llvm.amdgcn.* intrinsic instead - in this case llvm.amdgcn.buffer.load.The llvm.amdgcn.buffer.load doesn't allow specifiying VGPR + SGPR + immediate offset separately though as far as I can see. Furthermore I was trying to avoid a LLVM 3.9 dependency, although I can solve that with if based on LLVM version.Fair enough on the LLVM version dependency. I also think you're right about llvm.amdgcn.buffer.load, but that's something that should be fixed on the LLVM side eventually without introducing a new intrinsic.I am not sure if LLVM can eventually be able to. I don't know if the vgpr + sgpr + immediate offset is wrapped around in 32-bit. Otherwise a v_add and the offsets have different behavior and LLVM should not sink the add into the load instruction.
That's a good point. I don't know the answer unfortunately... Nicolai
- BasI think we should go for the if (HAVE_LLVM) approach even if it produces slightly worse code for now (it really should be only one additional v_add at most). Nicolai- BasNicolai+ + return lp_build_intrinsic(gallivm->builder, name, types[func], args, + ARRAY_SIZE(args), LLVMReadOnlyAttribute | + LLVMNoUnwindAttribute); +} + +static LLVMValueRef buffer_load(struct lp_build_tgsi_context *bld_base, + enum tgsi_opcode_type type, unsigned swizzle, + LLVMValueRef buffer, LLVMValueRef offset, + LLVMValueRef base) +{ + struct si_shader_context *ctx = si_shader_context(bld_base); + struct gallivm_state *gallivm = bld_base->base.gallivm; + LLVMValueRef value, value2; + LLVMTypeRef llvm_type = tgsi2llvmtype(bld_base, type); + LLVMTypeRef vec_type = LLVMVectorType(llvm_type, 4); + + if (swizzle == ~0) { + + value = build_buffer_load(ctx, buffer, 4, NULL, base, offset, + 0, 1, 0); + + return LLVMBuildBitCast(gallivm->builder, value, vec_type, ""); + } + + if (type != TGSI_TYPE_DOUBLE) { + value = build_buffer_load(ctx, buffer, 4, NULL, base, offset, + 0, 1, 0); + + value = LLVMBuildBitCast(gallivm->builder, value, vec_type, ""); + return LLVMBuildExtractElement(gallivm->builder, value, + lp_build_const_int32(gallivm, swizzle), ""); + } + + value = build_buffer_load(ctx, buffer, 1, NULL, base, offset, + swizzle * 4, 1, 0); + + value2 = build_buffer_load(ctx, buffer, 1, NULL, base, offset, + swizzle * 4 + 4, 1, 0); + + return radeon_llvm_emit_fetch_double(bld_base, value, value2); +} + /** * Load from LDS. *
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
