Implement MUBUF instructions for shorts/bytes (#2856)

* implement loads/store instructions for types smaller than dwords

* initialize s16/s8 types

* set profile for int8/16/64

* also need to zero extend u8/u16 to u32 result

* document unrelated bugs with atomic fmin/max

* remove profile checks and simple emit for added opcodes

---------

Co-authored-by: georgemoralis <giorgosmrls@gmail.com>

src/shader_recompiler/backend/spirv/emit_spirv_atomic.cpp

@@ -268,6 +268,7 @@ Id EmitBufferAtomicFMin32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id addre
     const auto sign_bit_set =
         ctx.OpBitFieldUExtract(ctx.U32[1], u32_value, ctx.ConstU32(31u), ctx.ConstU32(1u));
 
+    // FIXME this needs control flow because it currently executes both atomics
     const auto result = ctx.OpSelect(
         ctx.F32[1], sign_bit_set,
         EmitBitCastF32U32(ctx, EmitBufferAtomicUMax32(ctx, inst, handle, address, u32_value)),
@@ -302,6 +303,7 @@ Id EmitBufferAtomicFMax32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id addre
     const auto sign_bit_set =
         ctx.OpBitFieldUExtract(ctx.U32[1], u32_value, ctx.ConstU32(31u), ctx.ConstU32(1u));
 
+    // FIXME this needs control flow because it currently executes both atomics
     const auto result = ctx.OpSelect(
         ctx.F32[1], sign_bit_set,
         EmitBitCastF32U32(ctx, EmitBufferAtomicUMin32(ctx, inst, handle, address, u32_value)),

src/shader_recompiler/backend/spirv/emit_spirv_convert.cpp

@@ -7,60 +7,32 @@
 namespace Shader::Backend::SPIRV {
 namespace {
 Id ExtractU16(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int16) {
     return ctx.OpUConvert(ctx.U16, value);
-    } else {
-        return ctx.OpBitFieldUExtract(ctx.U32[1], value, ctx.u32_zero_value, ctx.ConstU32(16u));
-    }
 }
 
 Id ExtractS16(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int16) {
     return ctx.OpSConvert(ctx.S16, value);
-    } else {
-        return ctx.OpBitFieldSExtract(ctx.U32[1], value, ctx.u32_zero_value, ctx.ConstU32(16u));
-    }
 }
 
 Id ExtractU8(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int8) {
     return ctx.OpUConvert(ctx.U8, value);
-    } else {
-        return ctx.OpBitFieldUExtract(ctx.U32[1], value, ctx.u32_zero_value, ctx.ConstU32(8u));
-    }
 }
 
 Id ExtractS8(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int8) {
     return ctx.OpSConvert(ctx.S8, value);
-    } else {
-        return ctx.OpBitFieldSExtract(ctx.U32[1], value, ctx.u32_zero_value, ctx.ConstU32(8u));
-    }
 }
 } // Anonymous namespace
 
 Id EmitConvertS16F16(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int16) {
     return ctx.OpSConvert(ctx.U32[1], ctx.OpConvertFToS(ctx.U16, value));
-    } else {
-        return ExtractS16(ctx, ctx.OpConvertFToS(ctx.U32[1], value));
-    }
 }
 
 Id EmitConvertS16F32(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int16) {
     return ctx.OpSConvert(ctx.U32[1], ctx.OpConvertFToS(ctx.U16, value));
-    } else {
-        return ExtractS16(ctx, ctx.OpConvertFToS(ctx.U32[1], value));
-    }
 }
 
 Id EmitConvertS16F64(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int16) {
     return ctx.OpSConvert(ctx.U32[1], ctx.OpConvertFToS(ctx.U16, value));
-    } else {
-        return ExtractS16(ctx, ctx.OpConvertFToS(ctx.U32[1], value));
-    }
 }
 
 Id EmitConvertS32F16(EmitContext& ctx, Id value) {
@@ -88,27 +60,15 @@ Id EmitConvertS64F64(EmitContext& ctx, Id value) {
 }
 
 Id EmitConvertU16F16(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int16) {
     return ctx.OpUConvert(ctx.U32[1], ctx.OpConvertFToU(ctx.U16, value));
-    } else {
-        return ExtractU16(ctx, ctx.OpConvertFToU(ctx.U32[1], value));
-    }
 }
 
 Id EmitConvertU16F32(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int16) {
     return ctx.OpUConvert(ctx.U32[1], ctx.OpConvertFToU(ctx.U16, value));
-    } else {
-        return ExtractU16(ctx, ctx.OpConvertFToU(ctx.U32[1], value));
-    }
 }
 
 Id EmitConvertU16F64(EmitContext& ctx, Id value) {
-    if (ctx.profile.support_int16) {
     return ctx.OpUConvert(ctx.U32[1], ctx.OpConvertFToU(ctx.U16, value));
-    } else {
-        return ExtractU16(ctx, ctx.OpConvertFToU(ctx.U32[1], value));
-    }
 }
 
 Id EmitConvertU32F16(EmitContext& ctx, Id value) {
@@ -271,4 +231,12 @@ Id EmitConvertU32U8(EmitContext& ctx, Id value) {
     return ctx.OpUConvert(ctx.U32[1], value);
 }
 
+Id EmitConvertS32S8(EmitContext& ctx, Id value) {
+    return ctx.OpSConvert(ctx.U32[1], value);
+}
+
+Id EmitConvertS32S16(EmitContext& ctx, Id value) {
+    return ctx.OpSConvert(ctx.U32[1], value);
+}
+
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -488,6 +488,8 @@ Id EmitConvertU16U32(EmitContext& ctx, Id value);
 Id EmitConvertU32U16(EmitContext& ctx, Id value);
 Id EmitConvertU8U32(EmitContext& ctx, Id value);
 Id EmitConvertU32U8(EmitContext& ctx, Id value);
+Id EmitConvertS32S8(EmitContext& ctx, Id value);
+Id EmitConvertS32S16(EmitContext& ctx, Id value);
 
 Id EmitImageSampleRaw(EmitContext& ctx, IR::Inst* inst, u32 handle, Id address1, Id address2,
                       Id address3, Id address4);

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -117,7 +117,9 @@ void EmitContext::DefineArithmeticTypes() {
     void_id = Name(TypeVoid(), "void_id");
     U1[1] = Name(TypeBool(), "bool_id");
     U8 = Name(TypeUInt(8), "u8_id");
+    S8 = Name(TypeSInt(8), "i8_id");
     U16 = Name(TypeUInt(16), "u16_id");
+    S16 = Name(TypeSInt(16), "i16_id");
     if (info.uses_fp16) {
         F16[1] = Name(TypeFloat(16), "f16_id");
         U16 = Name(TypeUInt(16), "u16_id");

src/shader_recompiler/frontend/translate/translate.h

@@ -281,9 +281,10 @@ public:
 
     // Buffer Memory
     // MUBUF / MTBUF
-    void BUFFER_LOAD(u32 num_dwords, bool is_inst_typed, bool is_buffer_typed, const GcnInst& inst);
+    void BUFFER_LOAD(u32 num_dwords, bool is_inst_typed, bool is_buffer_typed, const GcnInst& inst,
-    void BUFFER_STORE(u32 num_dwords, bool is_inst_typed, bool is_buffer_typed,
+                     u32 scalar_width = 32, bool is_signed = false);
-                      const GcnInst& inst);
+    void BUFFER_STORE(u32 num_dwords, bool is_inst_typed, bool is_buffer_typed, const GcnInst& inst,
+                      u32 scalar_width = 32);
     template <typename T = IR::U32>
     void BUFFER_ATOMIC(AtomicOp op, const GcnInst& inst);
 

src/shader_recompiler/frontend/translate/vector_memory.cpp

@@ -28,6 +28,15 @@ void Translator::EmitVectorMemory(const GcnInst& inst) {
     case Opcode::BUFFER_LOAD_FORMAT_XYZW:
         return BUFFER_LOAD(4, false, true, inst);
 
+    case Opcode::BUFFER_LOAD_UBYTE:
+        return BUFFER_LOAD(1, false, false, inst, 8, false);
+    case Opcode::BUFFER_LOAD_SBYTE:
+        return BUFFER_LOAD(1, false, false, inst, 8, true);
+    case Opcode::BUFFER_LOAD_USHORT:
+        return BUFFER_LOAD(1, false, false, inst, 16, false);
+    case Opcode::BUFFER_LOAD_SSHORT:
+        return BUFFER_LOAD(1, false, false, inst, 16, true);
+
     case Opcode::BUFFER_LOAD_DWORD:
         return BUFFER_LOAD(1, false, false, inst);
     case Opcode::BUFFER_LOAD_DWORDX2:
@@ -56,6 +65,11 @@ void Translator::EmitVectorMemory(const GcnInst& inst) {
     case Opcode::TBUFFER_STORE_FORMAT_XYZW:
         return BUFFER_STORE(4, true, false, inst);
 
+    case Opcode::BUFFER_STORE_BYTE:
+        return BUFFER_STORE(1, false, false, inst, 8);
+    case Opcode::BUFFER_STORE_SHORT:
+        return BUFFER_STORE(1, false, false, inst, 16);
+
     case Opcode::BUFFER_STORE_DWORD:
         return BUFFER_STORE(1, false, false, inst);
     case Opcode::BUFFER_STORE_DWORDX2:
@@ -186,7 +200,7 @@ void Translator::EmitVectorMemory(const GcnInst& inst) {
 }
 
 void Translator::BUFFER_LOAD(u32 num_dwords, bool is_inst_typed, bool is_buffer_typed,
-                             const GcnInst& inst) {
+                             const GcnInst& inst, u32 scalar_width, bool is_signed) {
     const auto& mubuf = inst.control.mubuf;
     const bool is_ring = mubuf.glc && mubuf.slc;
     const IR::VectorReg vaddr{inst.src[0].code};
@@ -242,7 +256,26 @@ void Translator::BUFFER_LOAD(u32 num_dwords, bool is_inst_typed, bool is_buffer_
             ir.SetVectorReg(dst_reg + i, IR::F32{ir.CompositeExtract(value, i)});
         }
     } else {
-        const IR::Value value = ir.LoadBufferU32(num_dwords, handle, address, buffer_info);
+        IR::Value value;
+        switch (scalar_width) {
+        case 8: {
+            IR::U8 byte_val = ir.LoadBufferU8(handle, address, buffer_info);
+            value = is_signed ? ir.SConvert(32, byte_val) : ir.UConvert(32, byte_val);
+            break;
+        }
+        case 16: {
+            IR::U16 short_val = ir.LoadBufferU16(handle, address, buffer_info);
+            value = is_signed ? ir.SConvert(32, short_val) : ir.UConvert(32, short_val);
+            break;
+        }
+        case 32:
+            value = ir.LoadBufferU32(num_dwords, handle, address, buffer_info);
+            break;
+
+        default:
+            UNREACHABLE();
+        }
+
         if (num_dwords == 1) {
             ir.SetVectorReg(dst_reg, IR::U32{value});
             return;
@@ -254,7 +287,7 @@ void Translator::BUFFER_LOAD(u32 num_dwords, bool is_inst_typed, bool is_buffer_
 }
 
 void Translator::BUFFER_STORE(u32 num_dwords, bool is_inst_typed, bool is_buffer_typed,
-                              const GcnInst& inst) {
+                              const GcnInst& inst, u32 scalar_width) {
     const auto& mubuf = inst.control.mubuf;
     const bool is_ring = mubuf.glc && mubuf.slc;
     const IR::VectorReg vaddr{inst.src[0].code};
@@ -314,8 +347,23 @@ void Translator::BUFFER_STORE(u32 num_dwords, bool is_inst_typed, bool is_buffer
         }
         ir.StoreBufferFormat(handle, address, ir.CompositeConstruct(comps), buffer_info);
     } else {
-        const auto value = num_dwords == 1 ? comps[0] : ir.CompositeConstruct(comps);
+        IR::Value value = num_dwords == 1 ? comps[0] : ir.CompositeConstruct(comps);
+        if (scalar_width != 32) {
+            value = ir.UConvert(scalar_width, IR::U32{value});
+        }
+        switch (scalar_width) {
+        case 8:
+            ir.StoreBufferU8(handle, address, IR::U8{value}, buffer_info);
+            break;
+        case 16:
+            ir.StoreBufferU16(handle, address, IR::U16{value}, buffer_info);
+            break;
+        case 32:
             ir.StoreBufferU32(num_dwords, handle, address, value, buffer_info);
+            break;
+        default:
+            UNREACHABLE();
+        }
     }
 }
 

src/shader_recompiler/ir/ir_emitter.cpp

@@ -1979,6 +1979,24 @@ U8U16U32U64 IREmitter::UConvert(size_t result_bitsize, const U8U16U32U64& value)
     throw NotImplementedException("Conversion from {} to {} bits", value.Type(), result_bitsize);
 }
 
+U8U16U32U64 IR::IREmitter::SConvert(size_t result_bitsize, const U8U16U32U64& value) {
+    switch (result_bitsize) {
+    case 32:
+        switch (value.Type()) {
+        case Type::U8:
+            return Inst<U32>(Opcode::ConvertS32S8, value);
+        case Type::U16:
+            return Inst<U32>(Opcode::ConvertS32S16, value);
+        default:
+            break;
+        }
+    default:
+        break;
+    }
+    throw NotImplementedException("Signed Conversion from {} to {} bits", value.Type(),
+                                  result_bitsize);
+}
+
 F16F32F64 IREmitter::FPConvert(size_t result_bitsize, const F16F32F64& value) {
     switch (result_bitsize) {
     case 16:

src/shader_recompiler/ir/ir_emitter.h

@@ -325,6 +325,7 @@ public:
                                      const Value& value);
 
     [[nodiscard]] U8U16U32U64 UConvert(size_t result_bitsize, const U8U16U32U64& value);
+    [[nodiscard]] U8U16U32U64 SConvert(size_t result_bitsize, const U8U16U32U64& value);
     [[nodiscard]] F16F32F64 FPConvert(size_t result_bitsize, const F16F32F64& value);
 
     [[nodiscard]] Value ImageAtomicIAdd(const Value& handle, const Value& coords,

src/shader_recompiler/ir/opcodes.inc

@@ -432,6 +432,8 @@ OPCODE(ConvertU16U32,                                       U16,            U32,
 OPCODE(ConvertU32U16,                                       U32,            U16,                                                                            )
 OPCODE(ConvertU8U32,                                        U8,             U32,                                                                            )
 OPCODE(ConvertU32U8,                                        U32,            U8,                                                                             )
+OPCODE(ConvertS32S8,                                        U32,            U8,                                                                             )
+OPCODE(ConvertS32S16,                                       U32,            U16,                                                                            )
 
 // Image operations
 OPCODE(ImageSampleRaw,                                      F32x4,          Opaque,         F32x4,          F32x4,          F32x4,          F32,            Opaque,      )

src/video_core/renderer_vulkan/vk_instance.h

@@ -94,6 +94,21 @@ public:
         return features.shaderFloat64;
     }
 
+    /// Returns true if 64-bit ints are supported in shaders
+    bool IsShaderInt64Supported() const {
+        return features.shaderInt64;
+    }
+
+    /// Returns true if 16-bit ints are supported in shaders
+    bool IsShaderInt16Supported() const {
+        return features.shaderInt16;
+    }
+
+    /// Returns true if 8-bit ints are supported in shaders
+    bool IsShaderInt8Supported() const {
+        return vk12_features.shaderInt8;
+    }
+
     /// Returns true when VK_EXT_custom_border_color is supported
     bool IsCustomBorderColorSupported() const {
         return custom_border_color;

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -203,6 +203,9 @@ PipelineCache::PipelineCache(const Instance& instance_, Scheduler& scheduler_,
     profile = Shader::Profile{
         .supported_spirv = SpirvVersion1_6,
         .subgroup_size = instance.SubgroupSize(),
+        .support_int8 = instance.IsShaderInt8Supported(),
+        .support_int16 = instance.IsShaderInt16Supported(),
+        .support_int64 = instance.IsShaderInt64Supported(),
         .support_float64 = instance.IsShaderFloat64Supported(),
         .support_fp32_denorm_preserve = bool(vk12_props.shaderDenormPreserveFloat32),
         .support_fp32_denorm_flush = bool(vk12_props.shaderDenormFlushToZeroFloat32),