early-access version 2156

src/shader_recompiler/backend/spirv/emit_context.cpp

@@ -7,11 +7,14 @@
 #include <climits>
 #include <string_view>
 
+#include <boost/container/static_vector.hpp>
+
 #include <fmt/format.h>
 
 #include "common/common_types.h"
 #include "common/div_ceil.h"
 #include "shader_recompiler/backend/spirv/emit_context.h"
+#include "shader_recompiler/backend/spirv/emit_spirv.h"
 
 namespace Shader::Backend::SPIRV {
 namespace {
@@ -456,8 +459,9 @@ void VectorTypes::Define(Sirit::Module& sirit_ctx, Id base_type, std::string_vie
 
 EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_info_,
                          IR::Program& program, Bindings& bindings)
-    : Sirit::Module(profile_.supported_spirv), profile{profile_},
-      runtime_info{runtime_info_}, stage{program.stage} {
+    : Sirit::Module(profile_.supported_spirv), profile{profile_}, runtime_info{runtime_info_},
+      stage{program.stage}, texture_rescaling_index{bindings.texture_scaling_index},
+      image_rescaling_index{bindings.image_scaling_index} {
     const bool is_unified{profile.unified_descriptor_binding};
     u32& uniform_binding{is_unified ? bindings.unified : bindings.uniform_buffer};
     u32& storage_binding{is_unified ? bindings.unified : bindings.storage_buffer};
@@ -474,10 +478,11 @@ EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_inf
     DefineStorageBuffers(program.info, storage_binding);
     DefineTextureBuffers(program.info, texture_binding);
     DefineImageBuffers(program.info, image_binding);
-    DefineTextures(program.info, texture_binding);
-    DefineImages(program.info, image_binding);
+    DefineTextures(program.info, texture_binding, bindings.texture_scaling_index);
+    DefineImages(program.info, image_binding, bindings.image_scaling_index);
     DefineAttributeMemAccess(program.info);
     DefineGlobalMemoryFunctions(program.info);
+    DefineRescalingInput(program.info);
 }
 
 EmitContext::~EmitContext() = default;
@@ -920,6 +925,73 @@ void EmitContext::DefineGlobalMemoryFunctions(const Info& info) {
         define(&StorageDefinitions::U32x4, storage_types.U32x4, U32[4], sizeof(u32[4]));
 }
 
+void EmitContext::DefineRescalingInput(const Info& info) {
+    if (!info.uses_rescaling_uniform) {
+        return;
+    }
+    if (profile.unified_descriptor_binding) {
+        DefineRescalingInputPushConstant();
+    } else {
+        DefineRescalingInputUniformConstant();
+    }
+}
+
+void EmitContext::DefineRescalingInputPushConstant() {
+    boost::container::static_vector<Id, 3> members{};
+    u32 member_index{0};
+
+    rescaling_textures_type = TypeArray(U32[1], Const(4u));
+    Decorate(rescaling_textures_type, spv::Decoration::ArrayStride, 4u);
+    members.push_back(rescaling_textures_type);
+    rescaling_textures_member_index = member_index++;
+
+    rescaling_images_type = TypeArray(U32[1], Const(NUM_IMAGE_SCALING_WORDS));
+    Decorate(rescaling_images_type, spv::Decoration::ArrayStride, 4u);
+    members.push_back(rescaling_images_type);
+    rescaling_images_member_index = member_index++;
+
+    if (stage != Stage::Compute) {
+        members.push_back(F32[1]);
+        rescaling_downfactor_member_index = member_index++;
+    }
+    const Id push_constant_struct{TypeStruct(std::span(members.data(), members.size()))};
+    Decorate(push_constant_struct, spv::Decoration::Block);
+    Name(push_constant_struct, "ResolutionInfo");
+
+    MemberDecorate(push_constant_struct, rescaling_textures_member_index, spv::Decoration::Offset,
+                   static_cast<u32>(offsetof(RescalingLayout, rescaling_textures)));
+    MemberName(push_constant_struct, rescaling_textures_member_index, "rescaling_textures");
+
+    MemberDecorate(push_constant_struct, rescaling_images_member_index, spv::Decoration::Offset,
+                   static_cast<u32>(offsetof(RescalingLayout, rescaling_images)));
+    MemberName(push_constant_struct, rescaling_images_member_index, "rescaling_images");
+
+    if (stage != Stage::Compute) {
+        MemberDecorate(push_constant_struct, rescaling_downfactor_member_index,
+                       spv::Decoration::Offset,
+                       static_cast<u32>(offsetof(RescalingLayout, down_factor)));
+        MemberName(push_constant_struct, rescaling_downfactor_member_index, "down_factor");
+    }
+    const Id pointer_type{TypePointer(spv::StorageClass::PushConstant, push_constant_struct)};
+    rescaling_push_constants = AddGlobalVariable(pointer_type, spv::StorageClass::PushConstant);
+    Name(rescaling_push_constants, "rescaling_push_constants");
+
+    if (profile.supported_spirv >= 0x00010400) {
+        interfaces.push_back(rescaling_push_constants);
+    }
+}
+
+void EmitContext::DefineRescalingInputUniformConstant() {
+    const Id pointer_type{TypePointer(spv::StorageClass::UniformConstant, F32[4])};
+    rescaling_uniform_constant =
+        AddGlobalVariable(pointer_type, spv::StorageClass::UniformConstant);
+    Decorate(rescaling_uniform_constant, spv::Decoration::Location, 0u);
+
+    if (profile.supported_spirv >= 0x00010400) {
+        interfaces.push_back(rescaling_uniform_constant);
+    }
+}
+
 void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) {
     if (info.constant_buffer_descriptors.empty()) {
         return;
@@ -1108,7 +1180,7 @@ void EmitContext::DefineImageBuffers(const Info& info, u32& binding) {
     }
 }
 
-void EmitContext::DefineTextures(const Info& info, u32& binding) {
+void EmitContext::DefineTextures(const Info& info, u32& binding, u32& scaling_index) {
     textures.reserve(info.texture_descriptors.size());
     for (const TextureDescriptor& desc : info.texture_descriptors) {
         const Id image_type{ImageType(*this, desc)};
@@ -1130,13 +1202,14 @@ void EmitContext::DefineTextures(const Info& info, u32& binding) {
             interfaces.push_back(id);
         }
         ++binding;
+        ++scaling_index;
     }
     if (info.uses_atomic_image_u32) {
         image_u32 = TypePointer(spv::StorageClass::Image, U32[1]);
     }
 }
 
-void EmitContext::DefineImages(const Info& info, u32& binding) {
+void EmitContext::DefineImages(const Info& info, u32& binding, u32& scaling_index) {
     images.reserve(info.image_descriptors.size());
     for (const ImageDescriptor& desc : info.image_descriptors) {
         if (desc.count != 1) {
@@ -1157,6 +1230,7 @@ void EmitContext::DefineImages(const Info& info, u32& binding) {
             interfaces.push_back(id);
         }
         ++binding;
+        ++scaling_index;
     }
 }
 

src/shader_recompiler/backend/spirv/emit_context.h

@@ -235,6 +235,16 @@ public:
     Id indexed_load_func{};
     Id indexed_store_func{};
 
+    Id rescaling_uniform_constant{};
+    Id rescaling_push_constants{};
+    Id rescaling_textures_type{};
+    Id rescaling_images_type{};
+    u32 rescaling_textures_member_index{};
+    u32 rescaling_images_member_index{};
+    u32 rescaling_downfactor_member_index{};
+    u32 texture_rescaling_index{};
+    u32 image_rescaling_index{};
+
     Id local_memory{};
 
     Id shared_memory_u8{};
@@ -299,10 +309,13 @@ private:
     void DefineStorageBuffers(const Info& info, u32& binding);
     void DefineTextureBuffers(const Info& info, u32& binding);
     void DefineImageBuffers(const Info& info, u32& binding);
-    void DefineTextures(const Info& info, u32& binding);
-    void DefineImages(const Info& info, u32& binding);
+    void DefineTextures(const Info& info, u32& binding, u32& scaling_index);
+    void DefineImages(const Info& info, u32& binding, u32& scaling_index);
     void DefineAttributeMemAccess(const Info& info);
     void DefineGlobalMemoryFunctions(const Info& info);
+    void DefineRescalingInput(const Info& info);
+    void DefineRescalingInputPushConstant();
+    void DefineRescalingInputUniformConstant();
 
     void DefineInputs(const IR::Program& program);
     void DefineOutputs(const IR::Program& program);

src/shader_recompiler/backend/spirv/emit_spirv.h

@@ -16,6 +16,19 @@
 
 namespace Shader::Backend::SPIRV {
 
+constexpr u32 NUM_TEXTURE_SCALING_WORDS = 4;
+constexpr u32 NUM_IMAGE_SCALING_WORDS = 2;
+constexpr u32 NUM_TEXTURE_AND_IMAGE_SCALING_WORDS =
+    NUM_TEXTURE_SCALING_WORDS + NUM_IMAGE_SCALING_WORDS;
+
+struct RescalingLayout {
+    alignas(16) std::array<u32, NUM_TEXTURE_SCALING_WORDS> rescaling_textures;
+    alignas(16) std::array<u32, NUM_IMAGE_SCALING_WORDS> rescaling_images;
+    alignas(16) u32 down_factor;
+};
+constexpr u32 RESCALING_LAYOUT_WORDS_OFFSET = offsetof(RescalingLayout, rescaling_textures);
+constexpr u32 RESCALING_LAYOUT_DOWN_FACTOR_OFFSET = offsetof(RescalingLayout, down_factor);
+
 [[nodiscard]] std::vector<u32> EmitSPIRV(const Profile& profile, const RuntimeInfo& runtime_info,
                                          IR::Program& program, Bindings& bindings);
 

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -542,6 +542,18 @@ Id EmitYDirection(EmitContext& ctx) {
     return ctx.Const(ctx.runtime_info.y_negate ? -1.0f : 1.0f);
 }
 
+Id EmitResolutionDownFactor(EmitContext& ctx) {
+    if (ctx.profile.unified_descriptor_binding) {
+        const Id pointer_type{ctx.TypePointer(spv::StorageClass::PushConstant, ctx.F32[1])};
+        const Id index{ctx.Const(ctx.rescaling_downfactor_member_index)};
+        const Id pointer{ctx.OpAccessChain(pointer_type, ctx.rescaling_push_constants, index)};
+        return ctx.OpLoad(ctx.F32[1], pointer);
+    } else {
+        const Id composite{ctx.OpLoad(ctx.F32[4], ctx.rescaling_uniform_constant)};
+        return ctx.OpCompositeExtract(ctx.F32[1], composite, 2u);
+    }
+}
+
 Id EmitLoadLocal(EmitContext& ctx, Id word_offset) {
     const Id pointer{ctx.OpAccessChain(ctx.private_u32, ctx.local_memory, word_offset)};
     return ctx.OpLoad(ctx.U32[1], pointer);

src/shader_recompiler/backend/spirv/emit_spirv_image.cpp

@@ -224,6 +224,36 @@ Id Emit(MethodPtrType sparse_ptr, MethodPtrType non_sparse_ptr, EmitContext& ctx
     Decorate(ctx, inst, sample);
     return ctx.OpCompositeExtract(result_type, sample, 1U);
 }
+
+Id IsScaled(EmitContext& ctx, const IR::Value& index, Id member_index, u32 base_index) {
+    const Id push_constant_u32{ctx.TypePointer(spv::StorageClass::PushConstant, ctx.U32[1])};
+    Id bit{};
+    if (index.IsImmediate()) {
+        // Use BitwiseAnd instead of BitfieldExtract for better codegen on Nvidia OpenGL.
+        // LOP32I.NZ is used to set the predicate rather than BFE+ISETP.
+        const u32 index_value{index.U32() + base_index};
+        const Id word_index{ctx.Const(index_value / 32)};
+        const Id bit_index_mask{ctx.Const(1u << (index_value % 32))};
+        const Id pointer{ctx.OpAccessChain(push_constant_u32, ctx.rescaling_push_constants,
+                                           member_index, word_index)};
+        const Id word{ctx.OpLoad(ctx.U32[1], pointer)};
+        bit = ctx.OpBitwiseAnd(ctx.U32[1], word, bit_index_mask);
+    } else {
+        Id index_value{ctx.Def(index)};
+        if (base_index != 0) {
+            index_value = ctx.OpIAdd(ctx.U32[1], index_value, ctx.Const(base_index));
+        }
+        const Id bit_index{ctx.OpBitwiseAnd(ctx.U32[1], index_value, ctx.Const(31u))};
+        bit = ctx.OpBitFieldUExtract(ctx.U32[1], index_value, bit_index, ctx.Const(1u));
+    }
+    return ctx.OpINotEqual(ctx.U1, bit, ctx.u32_zero_value);
+}
+
+Id BitTest(EmitContext& ctx, Id mask, Id bit) {
+    const Id shifted{ctx.OpShiftRightLogical(ctx.U32[1], mask, 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);
+}
 } // Anonymous namespace
 
 Id EmitBindlessImageSampleImplicitLod(EmitContext&) {
@@ -470,4 +500,28 @@ void EmitImageWrite(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id
     ctx.OpImageWrite(Image(ctx, index, info), coords, color);
 }
 
+Id EmitIsTextureScaled(EmitContext& ctx, const IR::Value& index) {
+    if (ctx.profile.unified_descriptor_binding) {
+        const Id member_index{ctx.Const(ctx.rescaling_textures_member_index)};
+        return IsScaled(ctx, index, member_index, ctx.texture_rescaling_index);
+    } else {
+        const Id composite{ctx.OpLoad(ctx.F32[4], ctx.rescaling_uniform_constant)};
+        const Id mask_f32{ctx.OpCompositeExtract(ctx.F32[1], composite, 0u)};
+        const Id mask{ctx.OpBitcast(ctx.U32[1], mask_f32)};
+        return BitTest(ctx, mask, ctx.Def(index));
+    }
+}
+
+Id EmitIsImageScaled(EmitContext& ctx, const IR::Value& index) {
+    if (ctx.profile.unified_descriptor_binding) {
+        const Id member_index{ctx.Const(ctx.rescaling_images_member_index)};
+        return IsScaled(ctx, index, member_index, ctx.image_rescaling_index);
+    } else {
+        const Id composite{ctx.OpLoad(ctx.F32[4], ctx.rescaling_uniform_constant)};
+        const Id mask_f32{ctx.OpCompositeExtract(ctx.F32[1], composite, 1u)};
+        const Id mask{ctx.OpBitcast(ctx.U32[1], mask_f32)};
+        return BitTest(ctx, mask, ctx.Def(index));
+    }
+}
+
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -75,6 +75,7 @@ Id EmitInvocationId(EmitContext& ctx);
 Id EmitSampleId(EmitContext& ctx);
 Id EmitIsHelperInvocation(EmitContext& ctx);
 Id EmitYDirection(EmitContext& ctx);
+Id EmitResolutionDownFactor(EmitContext& ctx);
 Id EmitLoadLocal(EmitContext& ctx, Id word_offset);
 void EmitWriteLocal(EmitContext& ctx, Id word_offset, Id value);
 Id EmitUndefU1(EmitContext& ctx);
@@ -283,6 +284,8 @@ Id EmitIAdd64(EmitContext& ctx, Id a, Id b);
 Id EmitISub32(EmitContext& ctx, Id a, Id b);
 Id EmitISub64(EmitContext& ctx, Id a, Id b);
 Id EmitIMul32(EmitContext& ctx, Id a, Id b);
+Id EmitSDiv32(EmitContext& ctx, Id a, Id b);
+Id EmitUDiv32(EmitContext& ctx, Id a, Id b);
 Id EmitINeg32(EmitContext& ctx, Id value);
 Id EmitINeg64(EmitContext& ctx, Id value);
 Id EmitIAbs32(EmitContext& ctx, Id value);
@@ -510,6 +513,8 @@ Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, I
                      Id derivates, Id offset, Id lod_clamp);
 Id EmitImageRead(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords);
 void EmitImageWrite(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, Id color);
+Id EmitIsTextureScaled(EmitContext& ctx, const IR::Value& index);
+Id EmitIsImageScaled(EmitContext& ctx, const IR::Value& index);
 Id EmitBindlessImageAtomicIAdd32(EmitContext&);
 Id EmitBindlessImageAtomicSMin32(EmitContext&);
 Id EmitBindlessImageAtomicUMin32(EmitContext&);

src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp

@@ -72,6 +72,14 @@ Id EmitIMul32(EmitContext& ctx, Id a, Id b) {
     return ctx.OpIMul(ctx.U32[1], a, b);
 }
 
+Id EmitSDiv32(EmitContext& ctx, Id a, Id b) {
+    return ctx.OpSDiv(ctx.U32[1], a, b);
+}
+
+Id EmitUDiv32(EmitContext& ctx, Id a, Id b) {
+    return ctx.OpUDiv(ctx.U32[1], a, b);
+}
+
 Id EmitINeg32(EmitContext& ctx, Id value) {
     return ctx.OpSNegate(ctx.U32[1], value);
 }