shader: Properly scale image reads and add GL SPIR-V support

Thanks for everything!

src/shader_recompiler/backend/bindings.h

@@ -14,6 +14,8 @@ struct Bindings {
     u32 storage_buffer{};
     u32 texture{};
     u32 image{};
+    u32 texture_scaling_index{};
+    u32 image_scaling_index{};
 };
 
 } // namespace Shader::Backend

src/shader_recompiler/backend/glasm/emit_context.cpp

@@ -6,6 +6,7 @@
 
 #include "shader_recompiler/backend/bindings.h"
 #include "shader_recompiler/backend/glasm/emit_context.h"
+#include "shader_recompiler/backend/glasm/emit_glasm.h"
 #include "shader_recompiler/frontend/ir/program.h"
 #include "shader_recompiler/profile.h"
 #include "shader_recompiler/runtime_info.h"
@@ -55,7 +56,8 @@ EmitContext::EmitContext(IR::Program& program, Bindings& bindings, const Profile
     }
     if (!runtime_info.glasm_use_storage_buffers) {
         if (const size_t num = info.storage_buffers_descriptors.size(); num > 0) {
-            Add("PARAM c[{}]={{program.local[0..{}]}};", num, num - 1);
+            const size_t index{num + PROGRAM_LOCAL_PARAMETER_STORAGE_BUFFER_BASE};
+            Add("PARAM c[{}]={{program.local[0..{}]}};", index, index - 1);
         }
     }
     stage = program.stage;

src/shader_recompiler/backend/glasm/emit_glasm.h

@@ -13,6 +13,8 @@
 
 namespace Shader::Backend::GLASM {
 
+constexpr u32 PROGRAM_LOCAL_PARAMETER_STORAGE_BUFFER_BASE = 1;
+
 [[nodiscard]] std::string EmitGLASM(const Profile& profile, const RuntimeInfo& runtime_info,
                                     IR::Program& program, Bindings& bindings);
 

src/shader_recompiler/backend/glasm/emit_glasm_image.cpp

@@ -617,6 +617,15 @@ void EmitIsTextureScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& inde
             1u << index.U32(), ctx.reg_alloc.Define(inst));
 }
 
+void EmitIsImageScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index) {
+    if (!index.IsImmediate()) {
+        throw NotImplementedException("Non-constant texture rescaling");
+    }
+    ctx.Add("AND.U RC.x,scaling[0].y,{};"
+            "SNE.S {},RC.x,0;",
+            1u << index.U32(), ctx.reg_alloc.Define(inst));
+}
+
 void EmitImageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Register coord,
                            ScalarU32 value) {
     ImageAtomic(ctx, inst, index, coord, value, "ADD.U32");

src/shader_recompiler/backend/glasm/emit_glasm_instructions.h

@@ -557,6 +557,7 @@ void EmitImageRead(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Reg
 void EmitImageWrite(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Register coord,
                     Register color);
 void EmitIsTextureScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index);
+void EmitIsImageScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index);
 void EmitBindlessImageAtomicIAdd32(EmitContext&);
 void EmitBindlessImageAtomicSMin32(EmitContext&);
 void EmitBindlessImageAtomicUMin32(EmitContext&);

src/shader_recompiler/backend/glasm/emit_glasm_not_implemented.cpp

@@ -211,7 +211,7 @@ void EmitYDirection(EmitContext& ctx, IR::Inst& inst) {
 }
 
 void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst) {
-    ctx.Add("MOV.F {}.x,scaling[0].y;", inst);
+    ctx.Add("MOV.F {}.x,scaling[0].z;", inst);
 }
 
 void EmitUndefU1(EmitContext& ctx, IR::Inst& inst) {

src/shader_recompiler/backend/glsl/emit_glsl_context_get_set.cpp

@@ -446,7 +446,7 @@ void EmitYDirection(EmitContext& ctx, IR::Inst& inst) {
 }
 
 void EmitResolutionDownFactor(EmitContext& ctx, IR::Inst& inst) {
-    ctx.AddF32("{}=scaling.y;", inst);
+    ctx.AddF32("{}=scaling.z;", inst);
 }
 
 void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, std::string_view word_offset) {

src/shader_recompiler/backend/glsl/emit_glsl_image.cpp

@@ -620,6 +620,14 @@ void EmitIsTextureScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& inde
     ctx.AddU1("{}=(ftou(scaling.x)&{})!=0;", inst, 1u << image_index);
 }
 
+void EmitIsImageScaled(EmitContext& ctx, IR::Inst& inst, const IR::Value& index) {
+    if (!index.IsImmediate()) {
+        throw NotImplementedException("Non-constant texture rescaling");
+    }
+    const u32 image_index{index.U32()};
+    ctx.AddU1("{}=(ftou(scaling.y)&{})!=0;", inst, 1u << image_index);
+}
+
 void EmitBindlessImageSampleImplicitLod(EmitContext&) {
     NotImplemented();
 }

src/shader_recompiler/backend/spirv/emit_context.cpp

@@ -14,6 +14,7 @@
 #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 {
@@ -476,8 +477,9 @@ 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_},
+    : Sirit::Module(profile_.supported_spirv), profile{profile_}, runtime_info{runtime_info_},
-      runtime_info{runtime_info_}, stage{program.stage} {
+      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};
@@ -494,8 +496,8 @@ 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);
+    DefineTextures(program.info, texture_binding, bindings.texture_scaling_index);
-    DefineImages(program.info, image_binding);
+    DefineImages(program.info, image_binding, bindings.image_scaling_index);
     DefineAttributeMemAccess(program.info);
     DefineGlobalMemoryFunctions(program.info);
     DefineRescalingInput(program.info);
@@ -1003,25 +1005,49 @@ void EmitContext::DefineRescalingInput(const Info& info) {
     if (!info.uses_rescaling_uniform) {
         return;
     }
-    boost::container::static_vector<Id, 2> members{F32[1]};
+    if (profile.unified_descriptor_binding) {
+        DefineRescalingInputPushConstant(info);
+    } else {
+        DefineRescalingInputUniformConstant();
+    }
+}
+
+void EmitContext::DefineRescalingInputPushConstant(const Info& info) {
+    boost::container::static_vector<Id, 3> members{F32[1]};
     u32 member_index{0};
-    const u32 num_texture_words{Common::DivCeil(runtime_info.num_textures, 32u)};
+    if (!info.texture_descriptors.empty()) {
-    if (runtime_info.num_textures > 0) {
+        rescaling_textures_type = TypeArray(U32[1], Const(4u));
-        rescaling_textures_type = TypeArray(U32[1], Const(num_texture_words));
         Decorate(rescaling_textures_type, spv::Decoration::ArrayStride, 4u);
         members.push_back(rescaling_textures_type);
         rescaling_textures_member_index = ++member_index;
     }
+    if (!info.image_descriptors.empty()) {
+        rescaling_images_type = TypeArray(U32[1], Const(NUM_IMAGE_SCALING_WORDS));
+        if (rescaling_textures_type.value != rescaling_images_type.value) {
+            Decorate(rescaling_images_type, spv::Decoration::ArrayStride, 4u);
+        }
+        members.push_back(rescaling_images_type);
+        rescaling_images_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, 0u, spv::Decoration::Offset, 0u);
     MemberName(push_constant_struct, 0u, "down_factor");
-    if (runtime_info.num_textures > 0) {
+
-        MemberDecorate(push_constant_struct, rescaling_textures_member_index,
+    const u32 offset_bias = stage == Stage::Compute ? sizeof(u32) : 0;
-                       spv::Decoration::Offset, 4u);
+    if (!info.texture_descriptors.empty()) {
+        MemberDecorate(
+            push_constant_struct, rescaling_textures_member_index, spv::Decoration::Offset,
+            static_cast<u32>(offsetof(RescalingLayout, rescaling_textures) - offset_bias));
         MemberName(push_constant_struct, rescaling_textures_member_index, "rescaling_textures");
     }
+    if (!info.image_descriptors.empty()) {
+        MemberDecorate(push_constant_struct, rescaling_images_member_index, spv::Decoration::Offset,
+                       static_cast<u32>(offsetof(RescalingLayout, rescaling_images) - offset_bias));
+        MemberName(push_constant_struct, rescaling_images_member_index, "rescaling_images");
+    }
     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");
@@ -1031,6 +1057,17 @@ void EmitContext::DefineRescalingInput(const Info& info) {
     }
 }
 
+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;
@@ -1219,7 +1256,7 @@ 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)};
@@ -1241,13 +1278,14 @@ 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) {
@@ -1268,6 +1306,7 @@ void EmitContext::DefineImages(const Info& info, u32& binding) {
             interfaces.push_back(id);
         }
         ++binding;
+        ++scaling_index;
     }
 }
 

src/shader_recompiler/backend/spirv/emit_context.h

@@ -238,9 +238,14 @@ 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 texture_rescaling_index{};
+    u32 image_rescaling_index{};
 
     Id local_memory{};
 
@@ -314,11 +319,13 @@ 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 DefineTextures(const Info& info, u32& binding, u32& scaling_index);
-    void DefineImages(const Info& info, u32& binding);
+    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(const Info& info);
+    void DefineRescalingInputUniformConstant();
 
     void DefineInputs(const IR::Program& program);
     void DefineOutputs(const IR::Program& program);

src/shader_recompiler/backend/spirv/emit_spirv.h

@@ -16,15 +16,23 @@
 
 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 {
+    u32 down_factor;
+    std::array<u32, NUM_TEXTURE_SCALING_WORDS> rescaling_textures;
+    std::array<u32, NUM_IMAGE_SCALING_WORDS> rescaling_images;
+};
+
 [[nodiscard]] std::vector<u32> EmitSPIRV(const Profile& profile, const RuntimeInfo& runtime_info,
                                          IR::Program& program, Bindings& bindings);
 
 [[nodiscard]] inline std::vector<u32> EmitSPIRV(const Profile& profile, IR::Program& program) {
-    RuntimeInfo runtime_info{};
-    runtime_info.num_textures = Shader::NumDescriptors(program.info.texture_descriptors);
-
     Bindings binding;
-    return EmitSPIRV(profile, runtime_info, program, binding);
+    return EmitSPIRV(profile, {}, program, binding);
 }
 
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -527,10 +527,15 @@ Id EmitYDirection(EmitContext& ctx) {
 }
 
 Id EmitResolutionDownFactor(EmitContext& ctx) {
+    if (ctx.profile.unified_descriptor_binding) {
         const Id pointer_type{ctx.TypePointer(spv::StorageClass::PushConstant, ctx.F32[1])};
         const Id pointer{
             ctx.OpAccessChain(pointer_type, ctx.rescaling_push_constants, ctx.u32_zero_value)};
         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) {

src/shader_recompiler/backend/spirv/emit_spirv_image.cpp

@@ -224,6 +224,40 @@ 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 word_index{ctx.OpShiftRightArithmetic(ctx.U32[1], index_value, ctx.Const(5u))};
+        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)};
+        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&) {
@@ -471,29 +505,27 @@ void EmitImageWrite(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id
 }
 
 Id EmitIsTextureScaled(EmitContext& ctx, const IR::Value& index) {
-    const Id push_constant_u32{ctx.TypePointer(spv::StorageClass::PushConstant, ctx.U32[1])};
+    if (ctx.profile.unified_descriptor_binding) {
         const Id member_index{ctx.Const(ctx.rescaling_textures_member_index)};
-    Id bit{};
+        return IsScaled(ctx, index, member_index, ctx.texture_rescaling_index);
-    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()};
-        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 {
-        const Id index_value{ctx.Def(index)};
+        const Id composite{ctx.OpLoad(ctx.F32[4], ctx.rescaling_uniform_constant)};
-        const Id word_index{ctx.OpShiftRightArithmetic(ctx.U32[1], index_value, ctx.Const(5u))};
+        const Id mask_f32{ctx.OpCompositeExtract(ctx.F32[1], composite, 0u)};
-        const Id pointer{ctx.OpAccessChain(push_constant_u32, ctx.rescaling_push_constants,
+        const Id mask{ctx.OpBitcast(ctx.U32[1], mask_f32)};
-                                           member_index, word_index)};
+        return BitTest(ctx, mask, ctx.Def(index));
-        const Id word{ctx.OpLoad(ctx.U32[1], pointer)};
+    }
-        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));
+
+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));
     }
-    return ctx.OpINotEqual(ctx.U1, bit, ctx.u32_zero_value);
 }
 
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -514,6 +514,7 @@ Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, I
 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/frontend/ir/ir_emitter.cpp

@@ -1950,6 +1950,10 @@ U1 IREmitter::IsTextureScaled(const U32& index) {
     return Inst<U1>(Opcode::IsTextureScaled, index);
 }
 
+U1 IREmitter::IsImageScaled(const U32& index) {
+    return Inst<U1>(Opcode::IsImageScaled, index);
+}
+
 U1 IREmitter::VoteAll(const U1& value) {
     return Inst<U1>(Opcode::VoteAll, value);
 }

src/shader_recompiler/frontend/ir/ir_emitter.h

@@ -361,6 +361,7 @@ public:
                                             const Value& value, TextureInstInfo info);
 
     [[nodiscard]] U1 IsTextureScaled(const U32& index);
+    [[nodiscard]] U1 IsImageScaled(const U32& index);
 
     [[nodiscard]] U1 VoteAll(const U1& value);
     [[nodiscard]] U1 VoteAny(const U1& value);

src/shader_recompiler/frontend/ir/opcodes.inc

@@ -494,6 +494,7 @@ OPCODE(ImageRead,                                           U32x4,          Opaq
 OPCODE(ImageWrite,                                          Void,           Opaque,         Opaque,         U32x4,                                          )
 
 OPCODE(IsTextureScaled,                                     U1,             U32,                                                                            )
+OPCODE(IsImageScaled,                                       U1,             U32,                                                                            )
 
 // Atomic Image operations
 

src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp

@@ -432,6 +432,7 @@ void VisitUsages(Info& info, IR::Inst& inst) {
         break;
     case IR::Opcode::ResolutionDownFactor:
     case IR::Opcode::IsTextureScaled:
+    case IR::Opcode::IsImageScaled:
         info.uses_rescaling_uniform = true;
         break;
     case IR::Opcode::LaneId:

src/shader_recompiler/ir_opt/rescaling_pass.cpp

@@ -129,8 +129,7 @@ void PatchImageFetch(IR::Block& block, IR::Inst& inst) {
 void PatchImageRead(IR::Block& block, IR::Inst& inst) {
     IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
     const auto info{inst.Flags<IR::TextureInstInfo>()};
-    // TODO: Scale conditionally
+    const IR::U1 is_scaled{ir.IsImageScaled(ir.Imm32(info.descriptor_index))};
-    const IR::U1 is_scaled{IR::Value{true}};
     ScaleIntegerCoord(ir, inst, is_scaled);
 }
 

src/shader_recompiler/runtime_info.h

@@ -63,8 +63,6 @@ struct RuntimeInfo {
     std::array<AttributeType, 32> generic_input_types{};
     VaryingState previous_stage_stores;
 
-    u32 num_textures{};
-
     bool convert_depth_mode{};
     bool force_early_z{};
 

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -5,6 +5,7 @@
 #include <algorithm>
 #include <span>
 
+#include "shader_recompiler/backend/glasm/emit_glasm.h"
 #include "video_core/buffer_cache/buffer_cache.h"
 #include "video_core/renderer_opengl/gl_buffer_cache.h"
 #include "video_core/renderer_opengl/gl_device.h"
@@ -229,7 +230,9 @@ void BufferCacheRuntime::BindStorageBuffer(size_t stage, u32 binding_index, Buff
             .padding = 0,
         };
         buffer.MakeResident(is_written ? GL_READ_WRITE : GL_READ_ONLY);
-        glProgramLocalParametersI4uivNV(PROGRAM_LUT[stage], binding_index, 1,
+        glProgramLocalParametersI4uivNV(
+            PROGRAM_LUT[stage],
+            Shader::Backend::GLASM::PROGRAM_LOCAL_PARAMETER_STORAGE_BUFFER_BASE + binding_index, 1,
             reinterpret_cast<const GLuint*>(&ssbo));
     }
 }
@@ -250,7 +253,9 @@ void BufferCacheRuntime::BindComputeStorageBuffer(u32 binding_index, Buffer& buf
             .padding = 0,
         };
         buffer.MakeResident(is_written ? GL_READ_WRITE : GL_READ_ONLY);
-        glProgramLocalParametersI4uivNV(GL_COMPUTE_PROGRAM_NV, binding_index, 1,
+        glProgramLocalParametersI4uivNV(
+            GL_COMPUTE_PROGRAM_NV,
+            Shader::Backend::GLASM::PROGRAM_LOCAL_PARAMETER_STORAGE_BUFFER_BASE + binding_index, 1,
             reinterpret_cast<const GLuint*>(&ssbo));
     }
 }

src/video_core/renderer_opengl/gl_compute_pipeline.cpp

@@ -181,33 +181,40 @@ void ComputePipeline::Configure() {
     texture_binding += num_texture_buffers;
     image_binding += num_image_buffers;
 
-    u32 scaling_mask{};
+    u32 texture_scaling_mask{};
     for (const auto& desc : info.texture_descriptors) {
         for (u32 index = 0; index < desc.count; ++index) {
             ImageView& image_view{texture_cache.GetImageView((views_it++)->id)};
             textures[texture_binding] = image_view.Handle(desc.type);
             if (texture_cache.IsRescaling(image_view)) {
-                scaling_mask |= 1u << texture_binding;
+                texture_scaling_mask |= 1u << texture_binding;
             }
             ++texture_binding;
         }
     }
+    u32 image_scaling_mask{};
     for (const auto& desc : info.image_descriptors) {
         for (u32 index = 0; index < desc.count; ++index) {
             ImageView& image_view{texture_cache.GetImageView((views_it++)->id)};
             if (desc.is_written) {
                 texture_cache.MarkModification(image_view.image_id);
             }
-            images[image_binding++] = image_view.StorageView(desc.type, desc.format);
+            images[image_binding] = image_view.StorageView(desc.type, desc.format);
+            if (texture_cache.IsRescaling(image_view)) {
+                image_scaling_mask |= 1u << image_binding;
+            }
+            ++image_binding;
         }
     }
     if (info.uses_rescaling_uniform) {
-        const f32 float_scaling_mask{Common::BitCast<f32>(scaling_mask)};
+        const f32 float_texture_scaling_mask{Common::BitCast<f32>(texture_scaling_mask)};
+        const f32 float_image_scaling_mask{Common::BitCast<f32>(image_scaling_mask)};
         if (assembly_program.handle != 0) {
-            glProgramLocalParameter4fARB(GL_COMPUTE_PROGRAM_NV, 0, float_scaling_mask, 0.0f, 0.0f,
+            glProgramLocalParameter4fARB(GL_COMPUTE_PROGRAM_NV, 0, float_texture_scaling_mask,
-                                         0.0f);
+                                         float_image_scaling_mask, 0.0f, 0.0f);
         } else {
-            glProgramUniform4f(source_program.handle, 0, float_scaling_mask, 0.0f, 0.0f, 0.0f);
+            glProgramUniform4f(source_program.handle, 0, float_texture_scaling_mask,
+                               float_image_scaling_mask, 0.0f, 0.0f);
         }
     }
     if (texture_binding != 0) {

src/video_core/renderer_opengl/gl_graphics_pipeline.cpp

@@ -464,8 +464,10 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
         views_it += num_texture_buffers[stage];
         views_it += num_image_buffers[stage];
 
-        u32 scaling_mask{};
+        u32 texture_scaling_mask{};
+        u32 image_scaling_mask{};
         u32 stage_texture_binding{};
+        u32 stage_image_binding{};
 
         const auto& info{stage_infos[stage]};
         for (const auto& desc : info.texture_descriptors) {
@@ -473,7 +475,7 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
                 ImageView& image_view{texture_cache.GetImageView((views_it++)->id)};
                 textures[texture_binding] = image_view.Handle(desc.type);
                 if (texture_cache.IsRescaling(image_view)) {
-                    scaling_mask |= 1u << stage_texture_binding;
+                    texture_scaling_mask |= 1u << stage_texture_binding;
                 }
                 ++texture_binding;
                 ++stage_texture_binding;
@@ -485,20 +487,26 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
                 if (desc.is_written) {
                     texture_cache.MarkModification(image_view.image_id);
                 }
-                images[image_binding++] = image_view.StorageView(desc.type, desc.format);
+                images[image_binding] = image_view.StorageView(desc.type, desc.format);
+                if (texture_cache.IsRescaling(image_view)) {
+                    image_scaling_mask |= 1u << stage_image_binding;
+                }
+                ++image_binding;
+                ++stage_image_binding;
             }
         }
         if (info.uses_rescaling_uniform) {
-            const f32 float_scaling_mask{Common::BitCast<f32>(scaling_mask)};
+            const f32 float_texture_scaling_mask{Common::BitCast<f32>(texture_scaling_mask)};
+            const f32 float_image_scaling_mask{Common::BitCast<f32>(image_scaling_mask)};
             const bool is_rescaling{texture_cache.IsRescaling()};
             const f32 config_down_factor{Settings::values.resolution_info.down_factor};
             const f32 down_factor{is_rescaling ? config_down_factor : 1.0f};
             if (use_assembly) {
-                glProgramLocalParameter4fARB(AssemblyStage(stage), 0, float_scaling_mask,
+                glProgramLocalParameter4fARB(AssemblyStage(stage), 0, float_texture_scaling_mask,
-                                             down_factor, 0.0f, 0.0f);
+                                             float_image_scaling_mask, down_factor, 0.0f);
             } else {
-                glProgramUniform4f(source_programs[stage].handle, 0, float_scaling_mask,
+                glProgramUniform4f(source_programs[stage].handle, 0, float_texture_scaling_mask,
-                                   down_factor, 0.0f, 0.0f);
+                                   float_image_scaling_mask, down_factor, 0.0f);
             }
         }
     }};

src/video_core/renderer_vulkan/pipeline_helper.h

@@ -10,6 +10,7 @@
 
 #include "common/assert.h"
 #include "common/common_types.h"
+#include "shader_recompiler/backend/spirv/emit_spirv.h"
 #include "shader_recompiler/shader_info.h"
 #include "video_core/renderer_vulkan/vk_texture_cache.h"
 #include "video_core/renderer_vulkan/vk_update_descriptor.h"
@@ -20,7 +21,7 @@
 
 namespace Vulkan {
 
-constexpr size_t MAX_RESCALING_WORDS = 4;
+using Shader::Backend::SPIRV::NUM_TEXTURE_AND_IMAGE_SCALING_WORDS;
 
 class DescriptorLayoutBuilder {
 public:
@@ -74,7 +75,8 @@ public:
             .stageFlags = static_cast<VkShaderStageFlags>(
                 is_compute ? VK_SHADER_STAGE_COMPUTE_BIT : VK_SHADER_STAGE_ALL_GRAPHICS),
             .offset = 0,
-            .size = (is_compute ? 0 : sizeof(f32)) + sizeof(std::array<u32, MAX_RESCALING_WORDS>),
+            .size = (is_compute ? 0 : sizeof(f32)) +
+                    sizeof(std::array<u32, NUM_TEXTURE_AND_IMAGE_SCALING_WORDS>),
         };
         return device->GetLogical().CreatePipelineLayout({
             .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
@@ -146,14 +148,25 @@ public:
         }
     }
 
-    const std::array<u32, MAX_RESCALING_WORDS>& Data() const noexcept {
+    void PushImage(bool is_rescaled) noexcept {
+        *image_ptr |= is_rescaled ? image_bit : 0;
+        image_bit <<= 1;
+        if (image_bit == 0) {
+            image_bit = 1u;
+            ++image_ptr;
+        }
+    }
+
+    const std::array<u32, NUM_TEXTURE_AND_IMAGE_SCALING_WORDS>& Data() const noexcept {
         return words;
     }
 
 private:
-    std::array<u32, MAX_RESCALING_WORDS> words{};
+    std::array<u32, NUM_TEXTURE_AND_IMAGE_SCALING_WORDS> words{};
     u32* texture_ptr{words.data()};
+    u32* image_ptr{words.data() + Shader::Backend::SPIRV::NUM_TEXTURE_SCALING_WORDS};
     u32 texture_bit{1u};
+    u32 image_bit{1u};
 };
 
 inline void PushImageDescriptors(TextureCache& texture_cache,
@@ -181,6 +194,7 @@ inline void PushImageDescriptors(TextureCache& texture_cache,
             }
             const VkImageView vk_image_view{image_view.StorageView(desc.type, desc.format)};
             update_descriptor_queue.AddImage(vk_image_view);
+            rescaling.PushImage(texture_cache.IsRescaling(image_view));
         }
     }
 }

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -139,9 +139,6 @@ Shader::RuntimeInfo MakeRuntimeInfo(std::span<const Shader::IR::Program> program
     } else {
         info.previous_stage_stores.mask.set();
     }
-    for (const auto& stage : programs) {
-        info.num_textures += Shader::NumDescriptors(stage.info.texture_descriptors);
-    }
     const Shader::Stage stage{program.stage};
     const bool has_geometry{key.unique_hashes[4] != 0 && !programs[4].is_geometry_passthrough};
     const bool gl_ndc{key.state.ndc_minus_one_to_one != 0};