shader: Implement transform feedbacks and define file format

5 files changed, 97 insertions, 16 deletions

diff --git a/src/shader_recompiler/backend/spirv/emit_context.cpp b/src/shader_recompiler/backend/spirv/emit_context.cpp
index df53e58a8..74c42233d 100644
--- a/src/shader_recompiler/backend/spirv/emit_context.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_context.cpp
@@ -135,6 +135,45 @@ Id DefineOutput(EmitContext& ctx, Id type, std::optional<spv::BuiltIn> builtin =
     return DefineVariable(ctx, type, builtin, spv::StorageClass::Output);
 }
 
+void DefineGenericOutput(EmitContext& ctx, size_t index) {
+    static constexpr std::string_view swizzle{"xyzw"};
+    const size_t base_attr_index{static_cast<size_t>(IR::Attribute::Generic0X) + index * 4};
+    u32 element{0};
+    while (element < 4) {
+        const u32 remainder{4 - element};
+        const TransformFeedbackVarying* xfb_varying{};
+        if (!ctx.profile.xfb_varyings.empty()) {
+            xfb_varying = &ctx.profile.xfb_varyings[base_attr_index + element];
+            xfb_varying = xfb_varying && xfb_varying->components > 0 ? xfb_varying : nullptr;
+        }
+        const u32 num_components{xfb_varying ? xfb_varying->components : remainder};
+
+        const Id id{DefineOutput(ctx, ctx.F32[num_components])};
+        ctx.Decorate(id, spv::Decoration::Location, static_cast<u32>(index));
+        if (element > 0) {
+            ctx.Decorate(id, spv::Decoration::Component, element);
+        }
+        if (xfb_varying) {
+            ctx.Decorate(id, spv::Decoration::XfbBuffer, xfb_varying->buffer);
+            ctx.Decorate(id, spv::Decoration::XfbStride, xfb_varying->stride);
+            ctx.Decorate(id, spv::Decoration::Offset, xfb_varying->offset);
+        }
+        if (num_components < 4 || element > 0) {
+            ctx.Name(id, fmt::format("out_attr{}", index));
+        } else {
+            const std::string_view subswizzle{swizzle.substr(element, num_components)};
+            ctx.Name(id, fmt::format("out_attr{}_{}", index, subswizzle));
+        }
+        const GenericElementInfo info{
+            .id = id,
+            .first_element = element,
+            .num_components = num_components,
+        };
+        std::fill_n(ctx.output_generics[index].begin(), num_components, info);
+        element += num_components;
+    }
+}
+
 Id GetAttributeType(EmitContext& ctx, AttributeType type) {
     switch (type) {
     case AttributeType::Float:
@@ -663,12 +702,15 @@ void EmitContext::DefineAttributeMemAccess(const Info& info) {
             OpReturn();
             ++label_index;
         }
-        for (size_t i = 0; i < info.stores_generics.size(); i++) {
+        for (size_t i = 0; i < info.stores_generics.size(); ++i) {
             if (!info.stores_generics[i]) {
                 continue;
             }
+            if (output_generics[i][0].num_components != 4) {
+                throw NotImplementedException("Physical stores and transform feedbacks");
+            }
             AddLabel(labels[label_index]);
-            const Id generic_id{output_generics.at(i)};
+            const Id generic_id{output_generics[i][0].id};
             const Id pointer{OpAccessChain(output_f32, generic_id, masked_index)};
             OpStore(pointer, store_value);
             OpReturn();
@@ -1015,11 +1057,9 @@ void EmitContext::DefineOutputs(const Info& info) {
         }
         viewport_index = DefineOutput(*this, U32[1], spv::BuiltIn::ViewportIndex);
     }
-    for (size_t i = 0; i < info.stores_generics.size(); ++i) {
-        if (info.stores_generics[i]) {
-            output_generics[i] = DefineOutput(*this, F32[4]);
-            Decorate(output_generics[i], spv::Decoration::Location, static_cast<u32>(i));
-            Name(output_generics[i], fmt::format("out_attr{}", i));
+    for (size_t index = 0; index < info.stores_generics.size(); ++index) {
+        if (info.stores_generics[index]) {
+            DefineGenericOutput(*this, index);
         }
     }
     if (stage == Stage::Fragment) {
diff --git a/src/shader_recompiler/backend/spirv/emit_context.h b/src/shader_recompiler/backend/spirv/emit_context.h
index cade1fa0d..b27e5540c 100644
--- a/src/shader_recompiler/backend/spirv/emit_context.h
+++ b/src/shader_recompiler/backend/spirv/emit_context.h
@@ -79,6 +79,12 @@ struct StorageDefinitions {
     Id U32x4{};
 };
 
+struct GenericElementInfo {
+    Id id{};
+    u32 first_element{};
+    u32 num_components{};
+};
+
 class EmitContext final : public Sirit::Module {
 public:
     explicit EmitContext(const Profile& profile, IR::Program& program, u32& binding);
@@ -189,7 +195,7 @@ public:
 
     Id output_point_size{};
     Id output_position{};
-    std::array<Id, 32> output_generics{};
+    std::array<std::array<GenericElementInfo, 4>, 32> output_generics{};
 
     std::array<Id, 8> frag_color{};
     Id frag_depth{};
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv.cpp b/src/shader_recompiler/backend/spirv/emit_spirv.cpp
index 7ad00c434..444ba276f 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv.cpp
@@ -288,6 +288,9 @@ void SetupCapabilities(const Profile& profile, const Info& info, EmitContext& ct
     if (info.uses_typeless_image_writes) {
         ctx.AddCapability(spv::Capability::StorageImageWriteWithoutFormat);
     }
+    if (!ctx.profile.xfb_varyings.empty()) {
+        ctx.AddCapability(spv::Capability::TransformFeedback);
+    }
     // TODO: Track this usage
     ctx.AddCapability(spv::Capability::ImageGatherExtended);
     ctx.AddCapability(spv::Capability::ImageQuery);
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp
index a91b4c212..f9c151a5c 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp
@@ -40,11 +40,17 @@ Id AttrPointer(EmitContext& ctx, Id pointer_type, Id vertex, Id base, Args&&...
 }
 
 std::optional<Id> OutputAttrPointer(EmitContext& ctx, IR::Attribute attr) {
-    const u32 element{static_cast<u32>(attr) % 4};
-    const auto element_id{[&] { return ctx.Constant(ctx.U32[1], element); }};
     if (IR::IsGeneric(attr)) {
         const u32 index{IR::GenericAttributeIndex(attr)};
-        return ctx.OpAccessChain(ctx.output_f32, ctx.output_generics.at(index), element_id());
+        const u32 element{IR::GenericAttributeElement(attr)};
+        const GenericElementInfo& info{ctx.output_generics.at(index).at(element)};
+        if (info.num_components == 1) {
+            return info.id;
+        } else {
+            const u32 index_element{element - info.first_element};
+            const Id index_id{ctx.Constant(ctx.U32[1], index_element)};
+            return ctx.OpAccessChain(ctx.output_f32, info.id, index_id);
+        }
     }
     switch (attr) {
     case IR::Attribute::PointSize:
@@ -52,8 +58,11 @@ std::optional<Id> OutputAttrPointer(EmitContext& ctx, IR::Attribute attr) {
     case IR::Attribute::PositionX:
     case IR::Attribute::PositionY:
     case IR::Attribute::PositionZ:
-    case IR::Attribute::PositionW:
-        return ctx.OpAccessChain(ctx.output_f32, ctx.output_position, element_id());
+    case IR::Attribute::PositionW: {
+        const u32 element{static_cast<u32>(attr) % 4};
+        const Id element_id{ctx.Constant(ctx.U32[1], element)};
+        return ctx.OpAccessChain(ctx.output_f32, ctx.output_position, element_id);
+    }
     case IR::Attribute::ClipDistance0:
     case IR::Attribute::ClipDistance1:
     case IR::Attribute::ClipDistance2:
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_special.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_special.cpp
index fee740c08..7af29e4dd 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_special.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_special.cpp
@@ -22,6 +22,21 @@ void SetFixedPipelinePointSize(EmitContext& ctx) {
         ctx.OpStore(ctx.output_point_size, ctx.Constant(ctx.F32[1], point_size));
     }
 }
+
+Id DefaultVarying(EmitContext& ctx, u32 num_components, u32 element, Id zero, Id one,
+                  Id default_vector) {
+    switch (num_components) {
+    case 1:
+        return element == 3 ? one : zero;
+    case 2:
+        return ctx.ConstantComposite(ctx.F32[2], zero, element + 1 == 3 ? one : zero);
+    case 3:
+        return ctx.ConstantComposite(ctx.F32[3], zero, zero, element + 2 == 3 ? one : zero);
+    case 4:
+        return default_vector;
+    }
+    throw InvalidArgument("Bad element");
+}
 } // Anonymous namespace
 
 void EmitPrologue(EmitContext& ctx) {
@@ -30,9 +45,17 @@ void EmitPrologue(EmitContext& ctx) {
         const Id one{ctx.Constant(ctx.F32[1], 1.0f)};
         const Id default_vector{ctx.ConstantComposite(ctx.F32[4], zero, zero, zero, one)};
         ctx.OpStore(ctx.output_position, default_vector);
-        for (const Id generic_id : ctx.output_generics) {
-            if (Sirit::ValidId(generic_id)) {
-                ctx.OpStore(generic_id, default_vector);
+        for (const auto& info : ctx.output_generics) {
+            if (info[0].num_components == 0) {
+                continue;
+            }
+            u32 element{0};
+            while (element < 4) {
+                const auto& element_info{info[element]};
+                const u32 num{element_info.num_components};
+                const Id value{DefaultVarying(ctx, num, element, zero, one, default_vector)};
+                ctx.OpStore(element_info.id, value);
+                element += num;
             }
         }
     }