spirv: Simplify shared memory handling (#427)

* spirv: Simplify shared memory handling

* spirv: Ignore clip plane

* spirv: Fix image offsets

* ir_pass: Implement shared memory lowering pass

* NVIDIA doesn't like using shared mem in fragment shader and softlocks driver

* spirv: Add log for ignoring pos1

CMakeLists.txt

@@ -421,6 +421,7 @@ set(SHADER_RECOMPILER src/shader_recompiler/exception.h
                       src/shader_recompiler/ir/passes/dead_code_elimination_pass.cpp
                       src/shader_recompiler/ir/passes/identity_removal_pass.cpp
                       src/shader_recompiler/ir/passes/ir_passes.h
+                      src/shader_recompiler/ir/passes/lower_shared_mem_to_registers.cpp
                       src/shader_recompiler/ir/passes/resource_tracking_pass.cpp
                       src/shader_recompiler/ir/passes/shader_info_collection_pass.cpp
                       src/shader_recompiler/ir/passes/ssa_rewrite_pass.cpp

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -214,6 +214,10 @@ Id EmitGetAttributeU32(EmitContext& ctx, IR::Attribute attr, u32 comp) {
 }
 
 void EmitSetAttribute(EmitContext& ctx, IR::Attribute attr, Id value, u32 element) {
+    if (attr == IR::Attribute::Position1) {
+        LOG_WARNING(Render_Vulkan, "Ignoring pos1 export");
+        return;
+    }
     const Id pointer{OutputAttrPointer(ctx, attr, element)};
     ctx.OpStore(pointer, ctx.OpBitcast(ctx.F32[1], value));
 }

src/shader_recompiler/backend/spirv/emit_spirv_image.cpp

@@ -17,113 +17,133 @@ struct ImageOperands {
         operands.push_back(value);
     }
 
+    void AddOffset(EmitContext& ctx, const IR::Value& offset,
+                   bool can_use_runtime_offsets = false) {
+        if (offset.IsEmpty()) {
+            return;
+        }
+        if (offset.IsImmediate()) {
+            const s32 operand = offset.U32();
+            Add(spv::ImageOperandsMask::ConstOffset, ctx.ConstS32(operand));
+            return;
+        }
+        IR::Inst* const inst{offset.InstRecursive()};
+        if (inst->AreAllArgsImmediates()) {
+            switch (inst->GetOpcode()) {
+            case IR::Opcode::CompositeConstructU32x2:
+                Add(spv::ImageOperandsMask::ConstOffset,
+                    ctx.ConstS32(static_cast<s32>(inst->Arg(0).U32()),
+                                 static_cast<s32>(inst->Arg(1).U32())));
+                return;
+            case IR::Opcode::CompositeConstructU32x3:
+                Add(spv::ImageOperandsMask::ConstOffset,
+                    ctx.ConstS32(static_cast<s32>(inst->Arg(0).U32()),
+                                 static_cast<s32>(inst->Arg(1).U32()),
+                                 static_cast<s32>(inst->Arg(2).U32())));
+                return;
+            default:
+                break;
+            }
+        }
+        if (can_use_runtime_offsets) {
+            Add(spv::ImageOperandsMask::Offset, ctx.Def(offset));
+        } else {
+            LOG_WARNING(Render_Vulkan,
+                        "Runtime offset provided to unsupported image sample instruction");
+        }
+    }
+
     spv::ImageOperandsMask mask{};
     boost::container::static_vector<Id, 4> operands;
 };
 
 Id EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id bias,
-                              Id offset) {
+                              const IR::Value& offset) {
     const auto& texture = ctx.images[handle & 0xFFFF];
     const Id image = ctx.OpLoad(texture.image_type, texture.id);
     const Id sampler = ctx.OpLoad(ctx.sampler_type, ctx.samplers[handle >> 16]);
     const Id sampled_image = ctx.OpSampledImage(texture.sampled_type, image, sampler);
     ImageOperands operands;
-    if (Sirit::ValidId(bias)) {
+    operands.Add(spv::ImageOperandsMask::Bias, bias);
-        operands.Add(spv::ImageOperandsMask::Bias, bias);
+    operands.AddOffset(ctx, offset);
-    }
-    if (Sirit::ValidId(offset)) {
-        operands.Add(spv::ImageOperandsMask::Offset, offset);
-    }
     return ctx.OpImageSampleImplicitLod(ctx.F32[4], sampled_image, coords, operands.mask,
                                         operands.operands);
 }
 
 Id EmitImageSampleExplicitLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id lod,
-                              Id offset) {
+                              const IR::Value& offset) {
     const auto& texture = ctx.images[handle & 0xFFFF];
     const Id image = ctx.OpLoad(texture.image_type, texture.id);
     const Id sampler = ctx.OpLoad(ctx.sampler_type, ctx.samplers[handle >> 16]);
     const Id sampled_image = ctx.OpSampledImage(texture.sampled_type, image, sampler);
     ImageOperands operands;
-    if (Sirit::ValidId(lod)) {
+    operands.Add(spv::ImageOperandsMask::Lod, lod);
-        operands.Add(spv::ImageOperandsMask::Lod, lod);
+    operands.AddOffset(ctx, offset);
-    }
-    if (Sirit::ValidId(offset)) {
-        operands.Add(spv::ImageOperandsMask::Offset, offset);
-    }
     return ctx.OpImageSampleExplicitLod(ctx.F32[4], sampled_image, coords, operands.mask,
                                         operands.operands);
 }
 
 Id EmitImageSampleDrefImplicitLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id dref,
-                                  Id bias, Id offset) {
+                                  Id bias, const IR::Value& offset) {
     const auto& texture = ctx.images[handle & 0xFFFF];
     const Id image = ctx.OpLoad(texture.image_type, texture.id);
     const Id sampler = ctx.OpLoad(ctx.sampler_type, ctx.samplers[handle >> 16]);
     const Id sampled_image = ctx.OpSampledImage(texture.sampled_type, image, sampler);
     ImageOperands operands;
-    if (Sirit::ValidId(bias)) {
+    operands.Add(spv::ImageOperandsMask::Bias, bias);
-        operands.Add(spv::ImageOperandsMask::Bias, bias);
+    operands.AddOffset(ctx, offset);
-    }
-    if (Sirit::ValidId(offset)) {
-        operands.Add(spv::ImageOperandsMask::Offset, offset);
-    }
     return ctx.OpImageSampleDrefImplicitLod(ctx.F32[1], sampled_image, coords, dref, operands.mask,
                                             operands.operands);
 }
 
 Id EmitImageSampleDrefExplicitLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id dref,
-                                  Id lod, Id offset) {
+                                  Id lod, const IR::Value& offset) {
     const auto& texture = ctx.images[handle & 0xFFFF];
     const Id image = ctx.OpLoad(texture.image_type, texture.id);
     const Id sampler = ctx.OpLoad(ctx.sampler_type, ctx.samplers[handle >> 16]);
     const Id sampled_image = ctx.OpSampledImage(texture.sampled_type, image, sampler);
     ImageOperands operands;
-    if (Sirit::ValidId(lod)) {
+    operands.AddOffset(ctx, offset);
-        operands.Add(spv::ImageOperandsMask::Lod, lod);
+    operands.Add(spv::ImageOperandsMask::Lod, lod);
-    }
-    if (Sirit::ValidId(offset)) {
-        operands.Add(spv::ImageOperandsMask::Offset, offset);
-    }
     return ctx.OpImageSampleDrefExplicitLod(ctx.F32[1], sampled_image, coords, dref, operands.mask,
                                             operands.operands);
 }
 
-Id EmitImageGather(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id offset, Id offset2) {
+Id EmitImageGather(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords,
+                   const IR::Value& offset) {
     const auto& texture = ctx.images[handle & 0xFFFF];
     const Id image = ctx.OpLoad(texture.image_type, texture.id);
     const Id sampler = ctx.OpLoad(ctx.sampler_type, ctx.samplers[handle >> 16]);
     const Id sampled_image = ctx.OpSampledImage(texture.sampled_type, image, sampler);
     const u32 comp = inst->Flags<IR::TextureInstInfo>().gather_comp.Value();
     ImageOperands operands;
-    operands.Add(spv::ImageOperandsMask::Offset, offset);
+    operands.AddOffset(ctx, offset);
     return ctx.OpImageGather(ctx.F32[4], sampled_image, coords, ctx.ConstU32(comp), operands.mask,
                              operands.operands);
 }
 
-Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id offset,
+Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords,
-                       Id offset2, Id dref) {
+                       const IR::Value& offset, Id dref) {
     const auto& texture = ctx.images[handle & 0xFFFF];
     const Id image = ctx.OpLoad(texture.image_type, texture.id);
     const Id sampler = ctx.OpLoad(ctx.sampler_type, ctx.samplers[handle >> 16]);
     const Id sampled_image = ctx.OpSampledImage(texture.sampled_type, image, sampler);
     ImageOperands operands;
-    operands.Add(spv::ImageOperandsMask::Offset, offset);
+    operands.AddOffset(ctx, offset);
     return ctx.OpImageDrefGather(ctx.F32[4], sampled_image, coords, dref, operands.mask,
                                  operands.operands);
 }
 
-Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id offset, Id lod,
+Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, const IR::Value& offset,
-                  Id ms) {
+                  Id lod, Id ms) {
     const auto& texture = ctx.images[handle & 0xFFFF];
     const Id image = ctx.OpLoad(texture.image_type, texture.id);
     const Id result_type = texture.data_types->Get(4);
-    if (Sirit::ValidId(lod)) {
+    ImageOperands operands;
-        return ctx.OpBitcast(ctx.F32[4], ctx.OpImageFetch(result_type, image, coords,
+    operands.AddOffset(ctx, offset);
-                                                          spv::ImageOperandsMask::Lod, lod));
+    operands.Add(spv::ImageOperandsMask::Lod, lod);
-    } else {
+    return ctx.OpBitcast(
-        return ctx.OpBitcast(ctx.F32[4], ctx.OpImageFetch(result_type, image, coords));
+        ctx.F32[4], ctx.OpImageFetch(result_type, image, coords, operands.mask, operands.operands));
-    }
 }
 
 Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, u32 handle, Id lod, bool skip_mips) {

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -93,15 +93,9 @@ Id EmitUndefU8(EmitContext& ctx);
 Id EmitUndefU16(EmitContext& ctx);
 Id EmitUndefU32(EmitContext& ctx);
 Id EmitUndefU64(EmitContext& ctx);
-Id EmitLoadSharedU8(EmitContext& ctx, Id offset);
-Id EmitLoadSharedS8(EmitContext& ctx, Id offset);
-Id EmitLoadSharedU16(EmitContext& ctx, Id offset);
-Id EmitLoadSharedS16(EmitContext& ctx, Id offset);
 Id EmitLoadSharedU32(EmitContext& ctx, Id offset);
 Id EmitLoadSharedU64(EmitContext& ctx, Id offset);
 Id EmitLoadSharedU128(EmitContext& ctx, Id offset);
-void EmitWriteSharedU8(EmitContext& ctx, Id offset, Id value);
-void EmitWriteSharedU16(EmitContext& ctx, Id offset, Id value);
 void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value);
 void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value);
 void EmitWriteSharedU128(EmitContext& ctx, Id offset, Id value);
@@ -358,18 +352,19 @@ Id EmitConvertU16U32(EmitContext& ctx, Id value);
 Id EmitConvertU32U16(EmitContext& ctx, Id value);
 
 Id EmitImageSampleImplicitLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id bias,
-                              Id offset);
+                              const IR::Value& offset);
 Id EmitImageSampleExplicitLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id lod,
-                              Id offset);
+                              const IR::Value& offset);
 Id EmitImageSampleDrefImplicitLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id dref,
-                                  Id bias, Id offset);
+                                  Id bias, const IR::Value& offset);
 Id EmitImageSampleDrefExplicitLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id dref,
-                                  Id lod, Id offset);
+                                  Id lod, const IR::Value& offset);
-Id EmitImageGather(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id offset, Id offset2);
+Id EmitImageGather(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords,
-Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id offset,
+                   const IR::Value& offset);
-                       Id offset2, Id dref);
+Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords,
-Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id offset, Id lod,
+                       const IR::Value& offset, Id dref);
-                  Id ms);
+Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, const IR::Value& offset,
+                  Id lod, Id ms);
 Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, u32 handle, Id lod, bool skip_mips);
 Id EmitImageQueryLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords);
 Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,

src/shader_recompiler/backend/spirv/emit_spirv_shared_memory.cpp

@@ -5,99 +5,25 @@
 #include "shader_recompiler/backend/spirv/spirv_emit_context.h"
 
 namespace Shader::Backend::SPIRV {
-namespace {
-Id Pointer(EmitContext& ctx, Id pointer_type, Id array, Id offset, u32 shift) {
-    const Id shift_id{ctx.ConstU32(shift)};
-    const Id index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
-    return ctx.OpAccessChain(pointer_type, array, ctx.u32_zero_value, index);
-}
 
-Id Word(EmitContext& ctx, Id offset) {
+Id EmitLoadSharedU32(EmitContext& ctx, Id offset) {
     const Id shift_id{ctx.ConstU32(2U)};
     const Id index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
     const Id pointer{ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, index)};
     return ctx.OpLoad(ctx.U32[1], pointer);
 }
 
-std::pair<Id, Id> ExtractArgs(EmitContext& ctx, Id offset, u32 mask, u32 count) {
-    const Id shift{ctx.OpShiftLeftLogical(ctx.U32[1], offset, ctx.ConstU32(3U))};
-    const Id bit{ctx.OpBitwiseAnd(ctx.U32[1], shift, ctx.ConstU32(mask))};
-    const Id count_id{ctx.ConstU32(count)};
-    return {bit, count_id};
-}
-} // Anonymous namespace
-
-Id EmitLoadSharedU8(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
-        const Id pointer{
-            ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
-        return ctx.OpUConvert(ctx.U32[1], ctx.OpLoad(ctx.U8, pointer));
-    } else {
-        const auto [bit, count]{ExtractArgs(ctx, offset, 24, 8)};
-        return ctx.OpBitFieldUExtract(ctx.U32[1], Word(ctx, offset), bit, count);
-    }
-}
-
-Id EmitLoadSharedS8(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
-        const Id pointer{
-            ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
-        return ctx.OpSConvert(ctx.U32[1], ctx.OpLoad(ctx.U8, pointer));
-    } else {
-        const auto [bit, count]{ExtractArgs(ctx, offset, 24, 8)};
-        return ctx.OpBitFieldSExtract(ctx.U32[1], Word(ctx, offset), bit, count);
-    }
-}
-
-Id EmitLoadSharedU16(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
-        const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
-        return ctx.OpUConvert(ctx.U32[1], ctx.OpLoad(ctx.U16, pointer));
-    } else {
-        const auto [bit, count]{ExtractArgs(ctx, offset, 16, 16)};
-        return ctx.OpBitFieldUExtract(ctx.U32[1], Word(ctx, offset), bit, count);
-    }
-}
-
-Id EmitLoadSharedS16(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
-        const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
-        return ctx.OpSConvert(ctx.U32[1], ctx.OpLoad(ctx.U16, pointer));
-    } else {
-        const auto [bit, count]{ExtractArgs(ctx, offset, 16, 16)};
-        return ctx.OpBitFieldSExtract(ctx.U32[1], Word(ctx, offset), bit, count);
-    }
-}
-
-Id EmitLoadSharedU32(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
-        const Id pointer{Pointer(ctx, ctx.shared_u32, ctx.shared_memory_u32, offset, 2)};
-        return ctx.OpLoad(ctx.U32[1], pointer);
-    } else {
-        return Word(ctx, offset);
-    }
-}
-
 Id EmitLoadSharedU64(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    const Id shift_id{ctx.ConstU32(2U)};
-        const Id pointer{Pointer(ctx, ctx.shared_u32x2, ctx.shared_memory_u32x2, offset, 3)};
+    const Id base_index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
-        return ctx.OpLoad(ctx.U32[2], pointer);
+    const Id next_index{ctx.OpIAdd(ctx.U32[1], base_index, ctx.ConstU32(1U))};
-    } else {
+    const Id lhs_pointer{ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, base_index)};
-        const Id shift_id{ctx.ConstU32(2U)};
+    const Id rhs_pointer{ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, next_index)};
-        const Id base_index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
+    return ctx.OpCompositeConstruct(ctx.U32[2], ctx.OpLoad(ctx.U32[1], lhs_pointer),
-        const Id next_index{ctx.OpIAdd(ctx.U32[1], base_index, ctx.ConstU32(1U))};
+                                    ctx.OpLoad(ctx.U32[1], rhs_pointer));
-        const Id lhs_pointer{ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, base_index)};
-        const Id rhs_pointer{ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, next_index)};
-        return ctx.OpCompositeConstruct(ctx.U32[2], ctx.OpLoad(ctx.U32[1], lhs_pointer),
-                                        ctx.OpLoad(ctx.U32[1], rhs_pointer));
-    }
 }
 
 Id EmitLoadSharedU128(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
-        const Id pointer{Pointer(ctx, ctx.shared_u32x4, ctx.shared_memory_u32x4, offset, 4)};
-        return ctx.OpLoad(ctx.U32[4], pointer);
-    }
     const Id shift_id{ctx.ConstU32(2U)};
     const Id base_index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
     std::array<Id, 4> values{};
@@ -109,35 +35,14 @@ Id EmitLoadSharedU128(EmitContext& ctx, Id offset) {
     return ctx.OpCompositeConstruct(ctx.U32[4], values);
 }
 
-void EmitWriteSharedU8(EmitContext& ctx, Id offset, Id value) {
-    const Id pointer{
-        ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
-    ctx.OpStore(pointer, ctx.OpUConvert(ctx.U8, value));
-}
-
-void EmitWriteSharedU16(EmitContext& ctx, Id offset, Id value) {
-    const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
-    ctx.OpStore(pointer, ctx.OpUConvert(ctx.U16, value));
-}
-
 void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value) {
-    Id pointer{};
+    const Id shift{ctx.ConstU32(2U)};
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    const Id word_offset{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift)};
-        pointer = Pointer(ctx, ctx.shared_u32, ctx.shared_memory_u32, offset, 2);
+    const Id pointer = ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, word_offset);
-    } else {
-        const Id shift{ctx.ConstU32(2U)};
-        const Id word_offset{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift)};
-        pointer = ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, word_offset);
-    }
     ctx.OpStore(pointer, value);
 }
 
 void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
-        const Id pointer{Pointer(ctx, ctx.shared_u32x2, ctx.shared_memory_u32x2, offset, 3)};
-        ctx.OpStore(pointer, value);
-        return;
-    }
     const Id shift{ctx.ConstU32(2U)};
     const Id word_offset{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift)};
     const Id next_offset{ctx.OpIAdd(ctx.U32[1], word_offset, ctx.ConstU32(1U))};
@@ -148,11 +53,6 @@ void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value) {
 }
 
 void EmitWriteSharedU128(EmitContext& ctx, Id offset, Id value) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
-        const Id pointer{Pointer(ctx, ctx.shared_u32x4, ctx.shared_memory_u32x4, offset, 4)};
-        ctx.OpStore(pointer, value);
-        return;
-    }
     const Id shift{ctx.ConstU32(2U)};
     const Id base_index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift)};
     for (u32 i = 0; i < 4; ++i) {

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -513,43 +513,9 @@ void EmitContext::DefineSharedMemory() {
     if (info.shared_memory_size == 0) {
         info.shared_memory_size = DefaultSharedMemSize;
     }
-    const auto make{[&](Id element_type, u32 element_size) {
-        const u32 num_elements{Common::DivCeil(info.shared_memory_size, element_size)};
-        const Id array_type{TypeArray(element_type, ConstU32(num_elements))};
-        Decorate(array_type, spv::Decoration::ArrayStride, element_size);
-
-        const Id struct_type{TypeStruct(array_type)};
-        MemberDecorate(struct_type, 0U, spv::Decoration::Offset, 0U);
-        Decorate(struct_type, spv::Decoration::Block);
-
-        const Id pointer{TypePointer(spv::StorageClass::Workgroup, struct_type)};
-        const Id element_pointer{TypePointer(spv::StorageClass::Workgroup, element_type)};
-        const Id variable{AddGlobalVariable(pointer, spv::StorageClass::Workgroup)};
-        Decorate(variable, spv::Decoration::Aliased);
-        interfaces.push_back(variable);
-
-        return std::make_tuple(variable, element_pointer, pointer);
-    }};
-    if (profile.support_explicit_workgroup_layout) {
-        AddExtension("SPV_KHR_workgroup_memory_explicit_layout");
-        AddCapability(spv::Capability::WorkgroupMemoryExplicitLayoutKHR);
-        if (info.uses_shared_u8) {
-            AddCapability(spv::Capability::WorkgroupMemoryExplicitLayout8BitAccessKHR);
-            std::tie(shared_memory_u8, shared_u8, std::ignore) = make(U8, 1);
-        }
-        if (info.uses_shared_u16) {
-            AddCapability(spv::Capability::WorkgroupMemoryExplicitLayout16BitAccessKHR);
-            std::tie(shared_memory_u16, shared_u16, std::ignore) = make(U16, 2);
-        }
-        std::tie(shared_memory_u32, shared_u32, shared_memory_u32_type) = make(U32[1], 4);
-        std::tie(shared_memory_u32x2, shared_u32x2, std::ignore) = make(U32[2], 8);
-        std::tie(shared_memory_u32x4, shared_u32x4, std::ignore) = make(U32[4], 16);
-        return;
-    }
     const u32 num_elements{Common::DivCeil(info.shared_memory_size, 4U)};
     const Id type{TypeArray(U32[1], ConstU32(num_elements))};
     shared_memory_u32_type = TypePointer(spv::StorageClass::Workgroup, type);
-
     shared_u32 = TypePointer(spv::StorageClass::Workgroup, U32[1]);
     shared_memory_u32 = AddGlobalVariable(shared_memory_u32_type, spv::StorageClass::Workgroup);
     interfaces.push_back(shared_memory_u32);

src/shader_recompiler/frontend/translate/vector_memory.cpp

@@ -250,10 +250,10 @@ void Translator::IMAGE_GATHER(const GcnInst& inst) {
     const IR::Value texel = [&]() -> IR::Value {
         const IR::F32 lod = flags.test(MimgModifier::Level0) ? ir.Imm32(0.f) : IR::F32{};
         if (!flags.test(MimgModifier::Pcf)) {
-            return ir.ImageGather(handle, body, offset, {}, info);
+            return ir.ImageGather(handle, body, offset, info);
         }
         ASSERT(mimg.dmask & 1); // should be always 1st (R) component
-        return ir.ImageGatherDref(handle, body, offset, {}, dref, info);
+        return ir.ImageGatherDref(handle, body, offset, dref, info);
     }();
 
     // For gather4 instructions dmask selects which component to read and must have

src/shader_recompiler/ir/ir_emitter.cpp

@@ -259,10 +259,6 @@ void IREmitter::SetAttribute(IR::Attribute attribute, const F32& value, u32 comp
 
 Value IREmitter::LoadShared(int bit_size, bool is_signed, const U32& offset) {
     switch (bit_size) {
-    case 8:
-        return Inst<U32>(is_signed ? Opcode::LoadSharedS8 : Opcode::LoadSharedU8, offset);
-    case 16:
-        return Inst<U32>(is_signed ? Opcode::LoadSharedS16 : Opcode::LoadSharedU16, offset);
     case 32:
         return Inst<U32>(Opcode::LoadSharedU32, offset);
     case 64:
@@ -276,12 +272,6 @@ Value IREmitter::LoadShared(int bit_size, bool is_signed, const U32& offset) {
 
 void IREmitter::WriteShared(int bit_size, const Value& value, const U32& offset) {
     switch (bit_size) {
-    case 8:
-        Inst(Opcode::WriteSharedU8, offset, value);
-        break;
-    case 16:
-        Inst(Opcode::WriteSharedU16, offset, value);
-        break;
     case 32:
         Inst(Opcode::WriteSharedU32, offset, value);
         break;
@@ -1398,13 +1388,13 @@ F32 IREmitter::ImageSampleDrefExplicitLod(const Value& handle, const Value& body
 }
 
 Value IREmitter::ImageGather(const Value& handle, const Value& coords, const Value& offset,
-                             const Value& offset2, TextureInstInfo info) {
+                             TextureInstInfo info) {
-    return Inst(Opcode::ImageGather, Flags{info}, handle, coords, offset, offset2);
+    return Inst(Opcode::ImageGather, Flags{info}, handle, coords, offset);
 }
 
 Value IREmitter::ImageGatherDref(const Value& handle, const Value& coords, const Value& offset,
-                                 const Value& offset2, const F32& dref, TextureInstInfo info) {
+                                 const F32& dref, TextureInstInfo info) {
-    return Inst(Opcode::ImageGatherDref, Flags{info}, handle, coords, offset, offset2, dref);
+    return Inst(Opcode::ImageGatherDref, Flags{info}, handle, coords, offset, dref);
 }
 
 Value IREmitter::ImageFetch(const Value& handle, const Value& coords, const Value& offset,

src/shader_recompiler/ir/ir_emitter.h

@@ -256,18 +256,17 @@ public:
                                                  const F32& dref, const U32& offset,
                                                  TextureInstInfo info);
 
-    [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod,
+    [[nodiscard]] Value ImageQueryDimension(const Value& handle, const U32& lod,
-                                            const IR::U1& skip_mips);
+                                            const U1& skip_mips);
-    [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod,
+    [[nodiscard]] Value ImageQueryDimension(const Value& handle, const U32& lod,
-                                            const IR::U1& skip_mips, TextureInstInfo info);
+                                            const U1& skip_mips, TextureInstInfo info);
 
     [[nodiscard]] Value ImageQueryLod(const Value& handle, const Value& coords,
                                       TextureInstInfo info);
     [[nodiscard]] Value ImageGather(const Value& handle, const Value& coords, const Value& offset,
-                                    const Value& offset2, TextureInstInfo info);
+                                    TextureInstInfo info);
     [[nodiscard]] Value ImageGatherDref(const Value& handle, const Value& coords,
-                                        const Value& offset, const Value& offset2, const F32& dref,
+                                        const Value& offset, const F32& dref, TextureInstInfo info);
-                                        TextureInstInfo info);
     [[nodiscard]] Value ImageFetch(const Value& handle, const Value& coords, const Value& offset,
                                    const U32& lod, const U32& multisampling, TextureInstInfo info);
     [[nodiscard]] Value ImageGradient(const Value& handle, const Value& coords,

src/shader_recompiler/ir/microinstruction.cpp

@@ -59,8 +59,6 @@ bool Inst::MayHaveSideEffects() const noexcept {
     case Opcode::WriteSharedU128:
     case Opcode::WriteSharedU64:
     case Opcode::WriteSharedU32:
-    case Opcode::WriteSharedU16:
-    case Opcode::WriteSharedU8:
     case Opcode::ImageWrite:
     case Opcode::ImageAtomicIAdd32:
     case Opcode::ImageAtomicSMin32:

src/shader_recompiler/ir/opcodes.inc

@@ -26,15 +26,9 @@ OPCODE(WorkgroupMemoryBarrier,                              Void,
 OPCODE(DeviceMemoryBarrier,                                 Void,                                                                                           )
 
 // Shared memory operations
-OPCODE(LoadSharedU8,                                        U32,            U32,                                                                            )
-OPCODE(LoadSharedS8,                                        U32,            U32,                                                                            )
-OPCODE(LoadSharedU16,                                       U32,            U32,                                                                            )
-OPCODE(LoadSharedS16,                                       U32,            U32,                                                                            )
 OPCODE(LoadSharedU32,                                       U32,            U32,                                                                            )
 OPCODE(LoadSharedU64,                                       U32x2,          U32,                                                                            )
 OPCODE(LoadSharedU128,                                      U32x4,          U32,                                                                            )
-OPCODE(WriteSharedU8,                                       Void,           U32,            U32,                                                            )
-OPCODE(WriteSharedU16,                                      Void,           U32,            U32,                                                            )
 OPCODE(WriteSharedU32,                                      Void,           U32,            U32,                                                            )
 OPCODE(WriteSharedU64,                                      Void,           U32,            U32x2,                                                          )
 OPCODE(WriteSharedU128,                                     Void,           U32,            U32x4,                                                          )
@@ -298,12 +292,12 @@ OPCODE(ConvertU16U32,                                       U16,            U32,
 OPCODE(ConvertU32U16,                                       U32,            U16,                                                                            )
 
 // Image operations
-OPCODE(ImageSampleImplicitLod,                              F32x4,          Opaque,         Opaque,         F32,            U32,                            )
+OPCODE(ImageSampleImplicitLod,                              F32x4,          Opaque,         Opaque,         F32,            Opaque,                         )
-OPCODE(ImageSampleExplicitLod,                              F32x4,          Opaque,         Opaque,         U32,            U32,                            )
+OPCODE(ImageSampleExplicitLod,                              F32x4,          Opaque,         Opaque,         U32,            Opaque,                         )
-OPCODE(ImageSampleDrefImplicitLod,                          F32,            Opaque,         Opaque,         Opaque,         F32,            U32,            )
+OPCODE(ImageSampleDrefImplicitLod,                          F32,            Opaque,         Opaque,         Opaque,         F32,            Opaque,         )
-OPCODE(ImageSampleDrefExplicitLod,                          F32,            Opaque,         Opaque,         Opaque,         U32,            U32,            )
+OPCODE(ImageSampleDrefExplicitLod,                          F32,            Opaque,         Opaque,         Opaque,         U32,            Opaque,         )
-OPCODE(ImageGather,                                         F32x4,          Opaque,         Opaque,         Opaque,         Opaque,                         )
+OPCODE(ImageGather,                                         F32x4,          Opaque,         Opaque,         Opaque,                                         )
-OPCODE(ImageGatherDref,                                     F32x4,          Opaque,         Opaque,         Opaque,         Opaque,         F32,            )
+OPCODE(ImageGatherDref,                                     F32x4,          Opaque,         Opaque,         Opaque,         F32,                            )
 OPCODE(ImageFetch,                                          F32x4,          Opaque,         Opaque,         Opaque,         U32,            Opaque,         )
 OPCODE(ImageQueryDimensions,                                U32x4,          Opaque,         U32,            U1,                                             )
 OPCODE(ImageQueryLod,                                       F32x4,          Opaque,         Opaque,                                                         )

src/shader_recompiler/ir/passes/ir_passes.h

@@ -14,5 +14,6 @@ void DeadCodeEliminationPass(IR::Program& program);
 void ConstantPropagationPass(IR::BlockList& program);
 void ResourceTrackingPass(IR::Program& program);
 void CollectShaderInfoPass(IR::Program& program);
+void LowerSharedMemToRegisters(IR::Program& program);
 
 } // namespace Shader::Optimization

src/shader_recompiler/ir/passes/lower_shared_mem_to_registers.cpp

@@ -0,0 +1,39 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <boost/container/small_vector.hpp>
+#include "shader_recompiler/ir/program.h"
+
+namespace Shader::Optimization {
+
+void LowerSharedMemToRegisters(IR::Program& program) {
+    boost::container::small_vector<IR::Inst*, 8> ds_writes;
+    Info& info{program.info};
+    for (IR::Block* const block : program.blocks) {
+        for (IR::Inst& inst : block->Instructions()) {
+            const auto opcode = inst.GetOpcode();
+            if (opcode == IR::Opcode::WriteSharedU32 || opcode == IR::Opcode::WriteSharedU64) {
+                ds_writes.emplace_back(&inst);
+                continue;
+            }
+            if (opcode == IR::Opcode::LoadSharedU32 || opcode == IR::Opcode::LoadSharedU64) {
+                // Search for write instruction with same offset
+                const IR::Inst* prod = inst.Arg(0).InstRecursive();
+                const auto it = std::ranges::find_if(ds_writes, [&](const IR::Inst* write) {
+                    const IR::Inst* write_prod = write->Arg(0).InstRecursive();
+                    return write_prod->Arg(1).U32() == prod->Arg(1).U32() &&
+                           write_prod->Arg(0) == prod->Arg(0);
+                });
+                ASSERT(it != ds_writes.end());
+                // Replace data read with value written.
+                inst.ReplaceUsesWith((*it)->Arg(1));
+            }
+        }
+    }
+    // We should have eliminated everything. Invalidate data write instructions.
+    for (const auto inst : ds_writes) {
+        inst->Invalidate();
+    }
+}
+
+} // namespace Shader::Optimization

src/shader_recompiler/ir/passes/resource_tracking_pass.cpp

@@ -171,6 +171,22 @@ bool IsImageStorageInstruction(const IR::Inst& inst) {
     }
 }
 
+u32 ImageOffsetArgumentPosition(const IR::Inst& inst) {
+    switch (inst.GetOpcode()) {
+    case IR::Opcode::ImageGather:
+    case IR::Opcode::ImageGatherDref:
+        return 2;
+    case IR::Opcode::ImageSampleExplicitLod:
+    case IR::Opcode::ImageSampleImplicitLod:
+        return 3;
+    case IR::Opcode::ImageSampleDrefExplicitLod:
+    case IR::Opcode::ImageSampleDrefImplicitLod:
+        return 4;
+    default:
+        UNREACHABLE();
+    }
+}
+
 class Descriptors {
 public:
     explicit Descriptors(Info& info_)
@@ -574,33 +590,29 @@ void PatchImageInstruction(IR::Block& block, IR::Inst& inst, Info& info, Descrip
 
     if (inst_info.has_offset) {
         // The offsets are six-bit signed integers: X=[5:0], Y=[13:8], and Z=[21:16].
-        const u32 arg_pos = [&]() -> u32 {
+        const u32 arg_pos = ImageOffsetArgumentPosition(inst);
-            switch (inst.GetOpcode()) {
-            case IR::Opcode::ImageGather:
-            case IR::Opcode::ImageGatherDref:
-                return 2;
-            case IR::Opcode::ImageSampleExplicitLod:
-            case IR::Opcode::ImageSampleImplicitLod:
-                return 3;
-            case IR::Opcode::ImageSampleDrefExplicitLod:
-            case IR::Opcode::ImageSampleDrefImplicitLod:
-                return 4;
-            default:
-                break;
-            }
-            return inst_info.is_depth ? 4 : 3;
-        }();
         const IR::Value arg = inst.Arg(arg_pos);
         ASSERT_MSG(arg.Type() == IR::Type::U32, "Unexpected offset type");
-        const auto f = [&](IR::Value value, u32 offset) -> auto {
+
+        const auto read = [&](u32 offset) -> auto {
             return ir.BitFieldExtract(IR::U32{arg}, ir.Imm32(offset), ir.Imm32(6), true);
         };
 
-        const auto x = f(arg, 0);
+        switch (image.GetType()) {
-        const auto y = f(arg, 8);
+        case AmdGpu::ImageType::Color1D:
-        const auto z = f(arg, 16);
+        case AmdGpu::ImageType::Color1DArray:
-        const IR::Value value = ir.CompositeConstruct(x, y, z);
+            inst.SetArg(arg_pos, read(0));
-        inst.SetArg(arg_pos, value);
+            break;
+        case AmdGpu::ImageType::Color2D:
+        case AmdGpu::ImageType::Color2DArray:
+            inst.SetArg(arg_pos, ir.CompositeConstruct(read(0), read(8)));
+            break;
+        case AmdGpu::ImageType::Color3D:
+            inst.SetArg(arg_pos, ir.CompositeConstruct(read(0), read(8), read(16)));
+            break;
+        default:
+            UNREACHABLE();
+        }
     }
 
     if (inst_info.has_lod_clamp) {

src/shader_recompiler/ir/passes/shader_info_collection_pass.cpp

@@ -16,18 +16,6 @@ void Visit(Info& info, IR::Inst& inst) {
         info.stores.Set(inst.Arg(0).Attribute(), inst.Arg(2).U32());
         break;
     }
-    case IR::Opcode::LoadSharedS8:
-    case IR::Opcode::LoadSharedU8:
-    case IR::Opcode::WriteSharedU8:
-        info.uses_shared_u8 = true;
-        info.uses_shared = true;
-        break;
-    case IR::Opcode::LoadSharedS16:
-    case IR::Opcode::LoadSharedU16:
-    case IR::Opcode::WriteSharedU16:
-        info.uses_shared_u16 = true;
-        info.uses_shared = true;
-        break;
     case IR::Opcode::LoadSharedU32:
     case IR::Opcode::LoadSharedU64:
     case IR::Opcode::WriteSharedU32:

src/shader_recompiler/recompiler.cpp

@@ -58,6 +58,9 @@ IR::Program TranslateProgram(Common::ObjectPool<IR::Inst>& inst_pool,
     Shader::Optimization::SsaRewritePass(program.post_order_blocks);
     Shader::Optimization::ResourceTrackingPass(program);
     Shader::Optimization::ConstantPropagationPass(program.post_order_blocks);
+    if (program.info.stage != Stage::Compute) {
+        Shader::Optimization::LowerSharedMemToRegisters(program);
+    }
     Shader::Optimization::IdentityRemovalPass(program.blocks);
     Shader::Optimization::DeadCodeEliminationPass(program);
     Shader::Optimization::CollectShaderInfoPass(program);

src/shader_recompiler/runtime_info.h

@@ -195,8 +195,6 @@ struct Info {
     bool has_image_query{};
     bool uses_group_quad{};
     bool uses_shared{};
-    bool uses_shared_u8{};
-    bool uses_shared_u16{};
     bool uses_fp16{};
     bool uses_step_rates{};
     bool translation_failed{}; // indicates that shader has unsupported instructions