shader_recompiler: Define fragment output type based on number format. (#1097)

* shader_recompiler: Define fragment output type based on number format.

* shader_recompiler: Fix GetAttribute SPIR-V output type.

* shader_recompiler: Don't bitcast on SetAttribute unless integer target.

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -49,12 +49,13 @@ Id OutputAttrPointer(EmitContext& ctx, IR::Attribute attr, u32 element) {
         if (info.num_components == 1) {
             return info.id;
         } else {
-            return ctx.OpAccessChain(ctx.output_f32, info.id, ctx.ConstU32(element));
+            return ctx.OpAccessChain(info.pointer_type, info.id, ctx.ConstU32(element));
         }
     }
     switch (attr) {
     case IR::Attribute::Position0: {
         return ctx.OpAccessChain(ctx.output_f32, ctx.output_position, ctx.ConstU32(element));
+    }
     case IR::Attribute::Position1:
     case IR::Attribute::Position2:
     case IR::Attribute::Position3: {
@@ -70,18 +71,48 @@ Id OutputAttrPointer(EmitContext& ctx, IR::Attribute attr, u32 element) {
     case IR::Attribute::RenderTarget6:
     case IR::Attribute::RenderTarget7: {
         const u32 index = u32(attr) - u32(IR::Attribute::RenderTarget0);
-        if (ctx.frag_num_comp[index] > 1) {
-            return ctx.OpAccessChain(ctx.output_f32, ctx.frag_color[index], ctx.ConstU32(element));
+        const auto& info{ctx.frag_outputs.at(index)};
+        if (info.num_components > 1) {
+            return ctx.OpAccessChain(info.pointer_type, info.id, ctx.ConstU32(element));
         } else {
-            return ctx.frag_color[index];
+            return info.id;
         }
     }
     case IR::Attribute::Depth:
         return ctx.frag_depth;
     default:
-        throw NotImplementedException("Read attribute {}", attr);
+        throw NotImplementedException("Write attribute {}", attr);
     }
 }
+
+std::pair<Id, bool> OutputAttrComponentType(EmitContext& ctx, IR::Attribute attr) {
+    if (IR::IsParam(attr)) {
+        const u32 index{u32(attr) - u32(IR::Attribute::Param0)};
+        const auto& info{ctx.output_params.at(index)};
+        return {info.component_type, info.is_integer};
+    }
+    switch (attr) {
+    case IR::Attribute::Position0:
+    case IR::Attribute::Position1:
+    case IR::Attribute::Position2:
+    case IR::Attribute::Position3:
+    case IR::Attribute::Depth:
+        return {ctx.F32[1], false};
+    case IR::Attribute::RenderTarget0:
+    case IR::Attribute::RenderTarget1:
+    case IR::Attribute::RenderTarget2:
+    case IR::Attribute::RenderTarget3:
+    case IR::Attribute::RenderTarget4:
+    case IR::Attribute::RenderTarget5:
+    case IR::Attribute::RenderTarget6:
+    case IR::Attribute::RenderTarget7: {
+        const u32 index = u32(attr) - u32(IR::Attribute::RenderTarget0);
+        const auto& info{ctx.frag_outputs.at(index)};
+        return {info.component_type, info.is_integer};
+    }
+    default:
+        throw NotImplementedException("Write attribute {}", attr);
+    }
 }
 } // Anonymous namespace
 
@@ -156,17 +187,21 @@ Id EmitGetAttribute(EmitContext& ctx, IR::Attribute attr, u32 comp) {
                 // Attribute is disabled or varying component is not written
                 return ctx.ConstF32(comp == 3 ? 1.0f : 0.0f);
             }
-            if (param.is_default) {
-                return ctx.OpCompositeExtract(param.component_type, param.id, comp);
-            }
 
-            if (param.num_components > 1) {
+            Id result;
+            if (param.is_default) {
+                result = ctx.OpCompositeExtract(param.component_type, param.id, comp);
+            } else if (param.num_components > 1) {
                 const Id pointer{
                     ctx.OpAccessChain(param.pointer_type, param.id, ctx.ConstU32(comp))};
-                return ctx.OpLoad(param.component_type, pointer);
+                result = ctx.OpLoad(param.component_type, pointer);
             } else {
-                return ctx.OpLoad(param.component_type, param.id);
+                result = ctx.OpLoad(param.component_type, param.id);
             }
+            if (param.is_integer) {
+                result = ctx.OpBitcast(ctx.F32[1], result);
+            }
+            return result;
         } else {
             const auto step_rate = EmitReadStepRate(ctx, param.id.value);
             const auto offset = ctx.OpIAdd(
@@ -222,7 +257,12 @@ void EmitSetAttribute(EmitContext& ctx, IR::Attribute attr, Id value, u32 elemen
         return;
     }
     const Id pointer{OutputAttrPointer(ctx, attr, element)};
-    ctx.OpStore(pointer, ctx.OpBitcast(ctx.F32[1], value));
+    const auto component_type{OutputAttrComponentType(ctx, attr)};
+    if (component_type.second) {
+        ctx.OpStore(pointer, ctx.OpBitcast(component_type.first, value));
+    } else {
+        ctx.OpStore(pointer, value);
+    }
 }
 
 template <u32 N>

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -120,6 +120,7 @@ void EmitContext::DefineArithmeticTypes() {
 
     output_f32 = Name(TypePointer(spv::StorageClass::Output, F32[1]), "output_f32");
     output_u32 = Name(TypePointer(spv::StorageClass::Output, U32[1]), "output_u32");
+    output_s32 = Name(TypePointer(spv::StorageClass::Output, S32[1]), "output_s32");
 
     full_result_i32x2 = Name(TypeStruct(S32[1], S32[1]), "full_result_i32x2");
     full_result_u32x2 = Name(TypeStruct(U32[1], U32[1]), "full_result_u32x2");
@@ -151,21 +152,21 @@ const VectorIds& GetAttributeType(EmitContext& ctx, AmdGpu::NumberFormat fmt) {
     UNREACHABLE_MSG("Invalid attribute type {}", fmt);
 }
 
-EmitContext::SpirvAttribute EmitContext::GetAttributeInfo(AmdGpu::NumberFormat fmt, Id id) {
+EmitContext::SpirvAttribute EmitContext::GetAttributeInfo(AmdGpu::NumberFormat fmt, Id id,
+                                                          bool output) {
     switch (fmt) {
     case AmdGpu::NumberFormat::Float:
     case AmdGpu::NumberFormat::Unorm:
     case AmdGpu::NumberFormat::Snorm:
     case AmdGpu::NumberFormat::SnormNz:
-        return {id, input_f32, F32[1], 4};
-    case AmdGpu::NumberFormat::Uint:
-        return {id, input_u32, U32[1], 4};
-    case AmdGpu::NumberFormat::Sint:
-        return {id, input_s32, S32[1], 4};
     case AmdGpu::NumberFormat::Sscaled:
-        return {id, input_f32, F32[1], 4};
     case AmdGpu::NumberFormat::Uscaled:
-        return {id, input_f32, F32[1], 4};
+    case AmdGpu::NumberFormat::Srgb:
+        return {id, output ? output_f32 : input_f32, F32[1], 4, false};
+    case AmdGpu::NumberFormat::Uint:
+        return {id, output ? output_u32 : input_u32, U32[1], 4, true};
+    case AmdGpu::NumberFormat::Sint:
+        return {id, output ? output_s32 : input_s32, S32[1], 4, true};
     default:
         break;
     }
@@ -236,9 +237,13 @@ void EmitContext::DefineInputs() {
                                                                                              : 1;
                 // Note that we pass index rather than Id
                 input_params[input.binding] = {
-                    rate_idx, input_u32,
-                    U32[1],   input.num_components,
-                    false,    input.instance_data_buf,
+                    rate_idx,
+                    input_u32,
+                    U32[1],
+                    input.num_components,
+                    true,
+                    false,
+                    input.instance_data_buf,
                 };
             } else {
                 Id id{DefineInput(type, input.binding)};
@@ -247,7 +252,7 @@ void EmitContext::DefineInputs() {
                 } else {
                     Name(id, fmt::format("vs_in_attr{}", input.binding));
                 }
-                input_params[input.binding] = GetAttributeInfo(input.fmt, id);
+                input_params[input.binding] = GetAttributeInfo(input.fmt, id, false);
                 interfaces.push_back(id);
             }
         }
@@ -320,10 +325,12 @@ void EmitContext::DefineOutputs() {
                 continue;
             }
             const u32 num_components = info.stores.NumComponents(mrt);
-            frag_color[i] = DefineOutput(F32[num_components], i);
-            frag_num_comp[i] = num_components;
-            Name(frag_color[i], fmt::format("frag_color{}", i));
-            interfaces.push_back(frag_color[i]);
+            const AmdGpu::NumberFormat num_format{runtime_info.fs_info.color_buffers[i].num_format};
+            const Id type{GetAttributeType(*this, num_format)[num_components]};
+            const Id id = DefineOutput(type, i);
+            Name(id, fmt::format("frag_color{}", i));
+            frag_outputs[i] = GetAttributeInfo(num_format, id, true);
+            interfaces.push_back(id);
         }
         break;
     default:

src/shader_recompiler/backend/spirv/spirv_emit_context.h

@@ -166,6 +166,7 @@ public:
     Id input_s32{};
     Id output_u32{};
     Id output_f32{};
+    Id output_s32{};
 
     boost::container::small_vector<Id, 16> interfaces;
 
@@ -177,8 +178,6 @@ public:
     Id frag_coord{};
     Id front_facing{};
     Id frag_depth{};
-    std::array<Id, 8> frag_color{};
-    std::array<u32, 8> frag_num_comp{};
     Id clip_distances{};
     Id cull_distances{};
 
@@ -237,11 +236,13 @@ public:
         Id pointer_type;
         Id component_type;
         u32 num_components;
+        bool is_integer{};
         bool is_default{};
         s32 buffer_handle{-1};
     };
     std::array<SpirvAttribute, 32> input_params{};
     std::array<SpirvAttribute, 32> output_params{};
+    std::array<SpirvAttribute, 8> frag_outputs{};
 
 private:
     void DefineArithmeticTypes();
@@ -254,7 +255,7 @@ private:
     void DefineImagesAndSamplers();
     void DefineSharedMemory();
 
-    SpirvAttribute GetAttributeInfo(AmdGpu::NumberFormat fmt, Id id);
+    SpirvAttribute GetAttributeInfo(AmdGpu::NumberFormat fmt, Id id, bool output);
 };
 
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/frontend/translate/export.cpp

@@ -25,7 +25,7 @@ void Translator::EmitExport(const GcnInst& inst) {
             return comp;
         }
         const u32 index = u32(attrib) - u32(IR::Attribute::RenderTarget0);
-        switch (runtime_info.fs_info.mrt_swizzles[index]) {
+        switch (runtime_info.fs_info.color_buffers[index].mrt_swizzle) {
         case MrtSwizzle::Identity:
             return comp;
         case MrtSwizzle::Alt:

src/shader_recompiler/runtime_info.h

@@ -80,10 +80,16 @@ struct FragmentRuntimeInfo {
         auto operator<=>(const PsInput&) const noexcept = default;
     };
     boost::container::static_vector<PsInput, 32> inputs;
-    std::array<MrtSwizzle, MaxColorBuffers> mrt_swizzles;
+    struct PsColorBuffer {
+        AmdGpu::NumberFormat num_format;
+        MrtSwizzle mrt_swizzle;
+
+        auto operator<=>(const PsColorBuffer&) const noexcept = default;
+    };
+    std::array<PsColorBuffer, MaxColorBuffers> color_buffers;
 
     bool operator==(const FragmentRuntimeInfo& other) const noexcept {
-        return std::ranges::equal(mrt_swizzles, other.mrt_swizzles) &&
+        return std::ranges::equal(color_buffers, other.color_buffers) &&
                std::ranges::equal(inputs, other.inputs);
     }
 };

src/video_core/renderer_vulkan/vk_graphics_pipeline.h

@@ -26,6 +26,7 @@ using Liverpool = AmdGpu::Liverpool;
 struct GraphicsPipelineKey {
     std::array<size_t, MaxShaderStages> stage_hashes;
     std::array<vk::Format, Liverpool::NumColorBuffers> color_formats;
+    std::array<AmdGpu::NumberFormat, Liverpool::NumColorBuffers> color_num_formats;
     std::array<Liverpool::ColorBuffer::SwapMode, Liverpool::NumColorBuffers> mrt_swizzles;
     vk::Format depth_format;
     vk::Format stencil_format;

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -95,10 +95,6 @@ Shader::RuntimeInfo PipelineCache::BuildRuntimeInfo(Shader::Stage stage) {
     case Shader::Stage::Fragment: {
         info.num_user_data = regs.ps_program.settings.num_user_regs;
         info.num_allocated_vgprs = regs.ps_program.settings.num_vgprs * 4;
-        std::ranges::transform(graphics_key.mrt_swizzles, info.fs_info.mrt_swizzles.begin(),
-                               [](Liverpool::ColorBuffer::SwapMode mode) {
-                                   return static_cast<Shader::MrtSwizzle>(mode);
-                               });
         const auto& ps_inputs = regs.ps_inputs;
         for (u32 i = 0; i < regs.num_interp; i++) {
             info.fs_info.inputs.push_back({
@@ -108,6 +104,12 @@ Shader::RuntimeInfo PipelineCache::BuildRuntimeInfo(Shader::Stage stage) {
                 .default_value = u8(ps_inputs[i].default_value),
             });
         }
+        for (u32 i = 0; i < Shader::MaxColorBuffers; i++) {
+            info.fs_info.color_buffers[i] = {
+                .num_format = graphics_key.color_num_formats[i],
+                .mrt_swizzle = static_cast<Shader::MrtSwizzle>(graphics_key.mrt_swizzles[i]),
+            };
+        }
         break;
     }
     case Shader::Stage::Compute: {
@@ -244,6 +246,7 @@ bool PipelineCache::RefreshGraphicsKey() {
     // attachments. This might be not a case as HW color buffers can be bound in an arbitrary
     // order. We need to do some arrays compaction at this stage
     key.color_formats.fill(vk::Format::eUndefined);
+    key.color_num_formats.fill(AmdGpu::NumberFormat::Unorm);
     key.blend_controls.fill({});
     key.write_masks.fill({});
     key.mrt_swizzles.fill(Liverpool::ColorBuffer::SwapMode::Standard);
@@ -261,6 +264,7 @@ bool PipelineCache::RefreshGraphicsKey() {
         const bool is_vo_surface = renderer->IsVideoOutSurface(col_buf);
         key.color_formats[remapped_cb] = LiverpoolToVK::AdjustColorBufferFormat(
             base_format, col_buf.info.comp_swap.Value(), false /*is_vo_surface*/);
+        key.color_num_formats[remapped_cb] = col_buf.NumFormat();
         if (base_format == key.color_formats[remapped_cb]) {
             key.mrt_swizzles[remapped_cb] = col_buf.info.comp_swap.Value();
         }