diff options
Diffstat (limited to 'src/shader_recompiler/backend/spirv/emit_spirv_image.cpp')
| -rw-r--r-- | src/shader_recompiler/backend/spirv/emit_spirv_image.cpp | 65 |
1 files changed, 55 insertions, 10 deletions
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp index fb5799c42..7d901c04b 100644 --- a/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp +++ b/src/shader_recompiler/backend/spirv/emit_spirv_image.cpp @@ -201,6 +201,13 @@ Id Image(EmitContext& ctx, const IR::Value& index, IR::TextureInstInfo info) { } } +bool IsTextureMsaa(EmitContext& ctx, const IR::TextureInstInfo& info) { + if (info.type == TextureType::Buffer) { + return false; + } + return ctx.textures.at(info.descriptor_index).is_multisample; +} + Id Decorate(EmitContext& ctx, IR::Inst* inst, Id sample) { const auto info{inst->Flags<IR::TextureInstInfo>()}; if (info.relaxed_precision != 0) { @@ -254,6 +261,30 @@ Id BitTest(EmitContext& ctx, Id mask, Id bit) { const Id bit_value{ctx.OpBitwiseAnd(ctx.U32[1], shifted, ctx.Const(1u))}; return ctx.OpINotEqual(ctx.U1, bit_value, ctx.u32_zero_value); } + +Id ImageGatherSubpixelOffset(EmitContext& ctx, const IR::TextureInstInfo& info, Id texture, + Id coords) { + // Apply a subpixel offset of 1/512 the texel size of the texture to ensure same rounding on + // AMD hardware as on Maxwell or other Nvidia architectures. + const auto calculate_coords{[&](size_t dim) { + const Id nudge{ctx.Const(0x1p-9f)}; + const Id image_size{ctx.OpImageQuerySizeLod(ctx.U32[dim], texture, ctx.u32_zero_value)}; + Id offset{dim == 2 ? ctx.ConstantComposite(ctx.F32[dim], nudge, nudge) + : ctx.ConstantComposite(ctx.F32[dim], nudge, nudge, ctx.f32_zero_value)}; + offset = ctx.OpFDiv(ctx.F32[dim], offset, ctx.OpConvertUToF(ctx.F32[dim], image_size)); + return ctx.OpFAdd(ctx.F32[dim], coords, offset); + }}; + switch (info.type) { + case TextureType::Color2D: + case TextureType::Color2DRect: + return calculate_coords(2); + case TextureType::ColorArray2D: + case TextureType::ColorCube: + return calculate_coords(3); + default: + return coords; + } +} } // Anonymous namespace Id EmitBindlessImageSampleImplicitLod(EmitContext&) { @@ -416,6 +447,9 @@ Id EmitImageGather(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id const IR::Value& offset, const IR::Value& offset2) { const auto info{inst->Flags<IR::TextureInstInfo>()}; const ImageOperands operands(ctx, offset, offset2); + if (ctx.profile.need_gather_subpixel_offset) { + coords = ImageGatherSubpixelOffset(ctx, info, TextureImage(ctx, info, index), coords); + } return Emit(&EmitContext::OpImageSparseGather, &EmitContext::OpImageGather, ctx, inst, ctx.F32[4], Texture(ctx, info, index), coords, ctx.Const(info.gather_component), operands.MaskOptional(), operands.Span()); @@ -425,6 +459,9 @@ Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, const IR::Value& offset, const IR::Value& offset2, Id dref) { const auto info{inst->Flags<IR::TextureInstInfo>()}; const ImageOperands operands(ctx, offset, offset2); + if (ctx.profile.need_gather_subpixel_offset) { + coords = ImageGatherSubpixelOffset(ctx, info, TextureImage(ctx, info, index), coords); + } return Emit(&EmitContext::OpImageSparseDrefGather, &EmitContext::OpImageDrefGather, ctx, inst, ctx.F32[4], Texture(ctx, info, index), coords, dref, operands.MaskOptional(), operands.Span()); @@ -436,34 +473,42 @@ Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id c if (info.type == TextureType::Buffer) { lod = Id{}; } + if (Sirit::ValidId(ms)) { + // This image is multisampled, lod must be implicit + lod = Id{}; + } const ImageOperands operands(offset, lod, ms); return Emit(&EmitContext::OpImageSparseFetch, &EmitContext::OpImageFetch, ctx, inst, ctx.F32[4], TextureImage(ctx, info, index), coords, operands.MaskOptional(), operands.Span()); } -Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod) { +Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod, + const IR::Value& skip_mips_val) { const auto info{inst->Flags<IR::TextureInstInfo>()}; const Id image{TextureImage(ctx, info, index)}; const Id zero{ctx.u32_zero_value}; - const auto mips{[&] { return ctx.OpImageQueryLevels(ctx.U32[1], image); }}; + const bool skip_mips{skip_mips_val.U1()}; + const auto mips{[&] { return skip_mips ? zero : ctx.OpImageQueryLevels(ctx.U32[1], image); }}; + const bool is_msaa{IsTextureMsaa(ctx, info)}; + const bool uses_lod{!is_msaa && info.type != TextureType::Buffer}; + const auto query{[&](Id type) { + return uses_lod ? ctx.OpImageQuerySizeLod(type, image, lod) + : ctx.OpImageQuerySize(type, image); + }}; switch (info.type) { case TextureType::Color1D: - return ctx.OpCompositeConstruct(ctx.U32[4], ctx.OpImageQuerySizeLod(ctx.U32[1], image, lod), - zero, zero, mips()); + return ctx.OpCompositeConstruct(ctx.U32[4], query(ctx.U32[1]), zero, zero, mips()); case TextureType::ColorArray1D: case TextureType::Color2D: case TextureType::ColorCube: case TextureType::Color2DRect: - return ctx.OpCompositeConstruct(ctx.U32[4], ctx.OpImageQuerySizeLod(ctx.U32[2], image, lod), - zero, mips()); + return ctx.OpCompositeConstruct(ctx.U32[4], query(ctx.U32[2]), zero, mips()); case TextureType::ColorArray2D: case TextureType::Color3D: case TextureType::ColorArrayCube: - return ctx.OpCompositeConstruct(ctx.U32[4], ctx.OpImageQuerySizeLod(ctx.U32[3], image, lod), - mips()); + return ctx.OpCompositeConstruct(ctx.U32[4], query(ctx.U32[3]), mips()); case TextureType::Buffer: - return ctx.OpCompositeConstruct(ctx.U32[4], ctx.OpImageQuerySize(ctx.U32[1], image), zero, - zero, mips()); + return ctx.OpCompositeConstruct(ctx.U32[4], query(ctx.U32[1]), zero, zero, mips()); } throw LogicError("Unspecified image type {}", info.type.Value()); } |