Patch ImageQueryDimensions

src/shader_recompiler/info.h

@@ -87,6 +87,14 @@ struct SamplerResource {
 };
 using SamplerResourceList = boost::container::small_vector<SamplerResource, 16>;
 
+struct FMaskResource {
+    u32 sgpr_base;
+    u32 dword_offset;
+
+    constexpr AmdGpu::Image GetSharp(const Info& info) const noexcept;
+};
+using FMaskResourceList = boost::container::small_vector<FMaskResource, 16>;
+
 struct PushData {
     static constexpr u32 BufOffsetIndex = 2;
     static constexpr u32 UdRegsIndex = 4;
@@ -179,6 +187,7 @@ struct Info {
     TextureBufferResourceList texture_buffers;
     ImageResourceList images;
     SamplerResourceList samplers;
+    FMaskResourceList fmasks;
 
     std::span<const u32> user_data;
     Stage stage;
@@ -263,6 +272,10 @@ constexpr AmdGpu::Sampler SamplerResource::GetSharp(const Info& info) const noex
     return inline_sampler ? inline_sampler : info.ReadUd<AmdGpu::Sampler>(sgpr_base, dword_offset);
 }
 
+constexpr AmdGpu::Image FMaskResource::GetSharp(const Info& info) const noexcept {
+    return info.ReadUd<AmdGpu::Image>(sgpr_base, dword_offset);
+}
+
 } // namespace Shader
 
 template <>

src/shader_recompiler/ir/passes/resource_tracking_pass.cpp

@@ -147,7 +147,7 @@ public:
     explicit Descriptors(Info& info_)
         : info{info_}, buffer_resources{info_.buffers},
           texture_buffer_resources{info_.texture_buffers}, image_resources{info_.images},
-          sampler_resources{info_.samplers} {}
+          sampler_resources{info_.samplers}, fmask_resources(info_.fmasks) {}
 
     u32 Add(const BufferResource& desc) {
         const u32 index{Add(buffer_resources, desc, [&desc](const auto& existing) {
@@ -193,6 +193,14 @@ public:
         return index;
     }
 
+    u32 Add(const FMaskResource& desc) {
+        u32 index = Add(fmask_resources, desc, [&desc](const auto& existing) {
+            return desc.sgpr_base == existing.sgpr_base &&
+                   desc.dword_offset == existing.dword_offset;
+        });
+        return index;
+    }
+
 private:
     template <typename Descriptors, typename Descriptor, typename Func>
     static u32 Add(Descriptors& descriptors, const Descriptor& desc, Func&& pred) {
@@ -209,6 +217,7 @@ private:
     TextureBufferResourceList& texture_buffer_resources;
     ImageResourceList& image_resources;
     SamplerResourceList& sampler_resources;
+    FMaskResourceList& fmask_resources;
 };
 
 } // Anonymous namespace
@@ -673,6 +682,19 @@ void PatchImageInstruction(IR::Block& block, IR::Inst& inst, Info& info, Descrip
         case IR::Opcode::ImageQueryLod:
             inst.ReplaceUsesWith(ir.Imm32(1));
             return;
+        case IR::Opcode::ImageQueryDimensions: {
+            IR::Value dims = ir.CompositeConstruct(ir.Imm32(static_cast<u32>(image.width)), // x
+                                                   ir.Imm32(static_cast<u32>(image.width)), // y
+                                                   ir.Imm32(1), ir.Imm32(1)); // depth, mip
+            inst.ReplaceUsesWith(dims);
+
+            // Track FMask resource to do specialization.
+            descriptors.Add(FMaskResource{
+                .sgpr_base = tsharp.sgpr_base,
+                .dword_offset = tsharp.dword_offset,
+            });
+            return;
+        }
         default:
             UNREACHABLE_MSG("Can't patch fmask instruction {}", inst.GetOpcode());
         }

src/shader_recompiler/specialization.h

@@ -31,6 +31,13 @@ struct ImageSpecialization {
     auto operator<=>(const ImageSpecialization&) const = default;
 };
 
+struct FMaskSpecialization {
+    u32 width;
+    u32 height;
+
+    auto operator<=>(const FMaskSpecialization&) const = default;
+};
+
 /**
  * Alongside runtime information, this structure also checks bound resources
  * for compatibility. Can be used as a key for storing shader permutations.
@@ -46,6 +53,7 @@ struct StageSpecialization {
     boost::container::small_vector<BufferSpecialization, 16> buffers;
     boost::container::small_vector<TextureBufferSpecialization, 8> tex_buffers;
     boost::container::small_vector<ImageSpecialization, 16> images;
+    boost::container::small_vector<FMaskSpecialization, 8> fmasks;
     Backend::Bindings start{};
 
     explicit StageSpecialization(const Shader::Info& info_, RuntimeInfo runtime_info_,
@@ -67,6 +75,11 @@ struct StageSpecialization {
                                                               : sharp.GetType();
                          spec.is_integer = AmdGpu::IsInteger(sharp.GetNumberFmt());
                      });
+        ForEachSharp(binding, fmasks, info->fmasks,
+                     [](auto& spec, const auto& desc, AmdGpu::Image sharp) {
+                         spec.width = sharp.width;
+                         spec.height = sharp.height;
+                     });
     }
 
     void ForEachSharp(u32& binding, auto& spec_list, auto& desc_list, auto&& func) {
@@ -105,6 +118,11 @@ struct StageSpecialization {
                 return false;
             }
         }
+        for (u32 i = 0; i < fmasks.size(); i++) {
+            if (other.bitset[binding++] && fmasks[i] != other.fmasks[i]) {
+                return false;
+            }
+        }
         return true;
     }
 };