summaryrefslogtreecommitdiff
path: root/src/shader_recompiler/frontend
diff options
context:
space:
mode:
Diffstat (limited to 'src/shader_recompiler/frontend')
-rw-r--r--src/shader_recompiler/frontend/ir/attribute.cpp6
-rw-r--r--src/shader_recompiler/frontend/ir/attribute.h5
-rw-r--r--src/shader_recompiler/frontend/ir/ir_emitter.cpp22
-rw-r--r--src/shader_recompiler/frontend/ir/ir_emitter.h10
-rw-r--r--src/shader_recompiler/frontend/ir/opcodes.h1
-rw-r--r--src/shader_recompiler/frontend/ir/opcodes.inc7
-rw-r--r--src/shader_recompiler/frontend/ir/type.h31
-rw-r--r--src/shader_recompiler/frontend/ir/value.cpp3
-rw-r--r--src/shader_recompiler/frontend/ir/value.h14
-rw-r--r--src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp8
-rw-r--r--src/shader_recompiler/frontend/maxwell/translate_program.cpp118
11 files changed, 135 insertions, 90 deletions
diff --git a/src/shader_recompiler/frontend/ir/attribute.cpp b/src/shader_recompiler/frontend/ir/attribute.cpp
index 7d3d882e4..1bf9db935 100644
--- a/src/shader_recompiler/frontend/ir/attribute.cpp
+++ b/src/shader_recompiler/frontend/ir/attribute.cpp
@@ -446,6 +446,12 @@ std::string NameOf(Attribute attribute) {
446 return "ViewportMask"; 446 return "ViewportMask";
447 case Attribute::FrontFace: 447 case Attribute::FrontFace:
448 return "FrontFace"; 448 return "FrontFace";
449 case Attribute::BaseInstance:
450 return "BaseInstance";
451 case Attribute::BaseVertex:
452 return "BaseVertex";
453 case Attribute::DrawID:
454 return "DrawID";
449 } 455 }
450 return fmt::format("<reserved attribute {}>", static_cast<int>(attribute)); 456 return fmt::format("<reserved attribute {}>", static_cast<int>(attribute));
451} 457}
diff --git a/src/shader_recompiler/frontend/ir/attribute.h b/src/shader_recompiler/frontend/ir/attribute.h
index 6ee3947b1..5f039b6f6 100644
--- a/src/shader_recompiler/frontend/ir/attribute.h
+++ b/src/shader_recompiler/frontend/ir/attribute.h
@@ -219,6 +219,11 @@ enum class Attribute : u64 {
219 FixedFncTexture9Q = 231, 219 FixedFncTexture9Q = 231,
220 ViewportMask = 232, 220 ViewportMask = 232,
221 FrontFace = 255, 221 FrontFace = 255,
222
223 // Implementation attributes
224 BaseInstance = 256,
225 BaseVertex = 257,
226 DrawID = 258,
222}; 227};
223 228
224constexpr size_t NUM_GENERICS = 32; 229constexpr size_t NUM_GENERICS = 32;
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.cpp b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
index 0cdac0eff..b7caa4246 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.cpp
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
@@ -294,6 +294,14 @@ F32 IREmitter::GetAttribute(IR::Attribute attribute, const U32& vertex) {
294 return Inst<F32>(Opcode::GetAttribute, attribute, vertex); 294 return Inst<F32>(Opcode::GetAttribute, attribute, vertex);
295} 295}
296 296
297U32 IREmitter::GetAttributeU32(IR::Attribute attribute) {
298 return GetAttributeU32(attribute, Imm32(0));
299}
300
301U32 IREmitter::GetAttributeU32(IR::Attribute attribute, const U32& vertex) {
302 return Inst<U32>(Opcode::GetAttributeU32, attribute, vertex);
303}
304
297void IREmitter::SetAttribute(IR::Attribute attribute, const F32& value, const U32& vertex) { 305void IREmitter::SetAttribute(IR::Attribute attribute, const F32& value, const U32& vertex) {
298 Inst(Opcode::SetAttribute, attribute, value, vertex); 306 Inst(Opcode::SetAttribute, attribute, value, vertex);
299} 307}
@@ -696,11 +704,6 @@ IR::U32 IREmitter::BitCast<IR::U32, IR::F32>(const IR::F32& value) {
696} 704}
697 705
698template <> 706template <>
699IR::S32 IREmitter::BitCast<IR::S32, IR::F32>(const IR::F32& value) {
700 return Inst<IR::S32>(Opcode::BitCastS32F32, value);
701}
702
703template <>
704IR::F32 IREmitter::BitCast<IR::F32, IR::U32>(const IR::U32& value) { 707IR::F32 IREmitter::BitCast<IR::F32, IR::U32>(const IR::U32& value) {
705 return Inst<IR::F32>(Opcode::BitCastF32U32, value); 708 return Inst<IR::F32>(Opcode::BitCastF32U32, value);
706} 709}
@@ -1843,15 +1846,16 @@ Value IREmitter::ImageFetch(const Value& handle, const Value& coords, const Valu
1843 return Inst(op, Flags{info}, handle, coords, offset, lod, multisampling); 1846 return Inst(op, Flags{info}, handle, coords, offset, lod, multisampling);
1844} 1847}
1845 1848
1846Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod) { 1849Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod,
1850 const IR::U1& skip_mips) {
1847 const Opcode op{handle.IsImmediate() ? Opcode::BoundImageQueryDimensions 1851 const Opcode op{handle.IsImmediate() ? Opcode::BoundImageQueryDimensions
1848 : Opcode::BindlessImageQueryDimensions}; 1852 : Opcode::BindlessImageQueryDimensions};
1849 return Inst(op, handle, lod); 1853 return Inst(op, handle, lod, skip_mips);
1850} 1854}
1851 1855
1852Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod, 1856Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod,
1853 TextureInstInfo info) { 1857 const IR::U1& skip_mips, TextureInstInfo info) {
1854 return Inst(Opcode::ImageQueryDimensions, Flags{info}, handle, lod); 1858 return Inst(Opcode::ImageQueryDimensions, Flags{info}, handle, lod, skip_mips);
1855} 1859}
1856 1860
1857Value IREmitter::ImageQueryLod(const Value& handle, const Value& coords, TextureInstInfo info) { 1861Value IREmitter::ImageQueryLod(const Value& handle, const Value& coords, TextureInstInfo info) {
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.h b/src/shader_recompiler/frontend/ir/ir_emitter.h
index 2df992feb..f3c81dbe1 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.h
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.h
@@ -74,6 +74,8 @@ public:
74 74
75 [[nodiscard]] F32 GetAttribute(IR::Attribute attribute); 75 [[nodiscard]] F32 GetAttribute(IR::Attribute attribute);
76 [[nodiscard]] F32 GetAttribute(IR::Attribute attribute, const U32& vertex); 76 [[nodiscard]] F32 GetAttribute(IR::Attribute attribute, const U32& vertex);
77 [[nodiscard]] U32 GetAttributeU32(IR::Attribute attribute);
78 [[nodiscard]] U32 GetAttributeU32(IR::Attribute attribute, const U32& vertex);
77 void SetAttribute(IR::Attribute attribute, const F32& value, const U32& vertex); 79 void SetAttribute(IR::Attribute attribute, const F32& value, const U32& vertex);
78 80
79 [[nodiscard]] F32 GetAttributeIndexed(const U32& phys_address); 81 [[nodiscard]] F32 GetAttributeIndexed(const U32& phys_address);
@@ -318,9 +320,10 @@ public:
318 [[nodiscard]] F32 ImageSampleDrefExplicitLod(const Value& handle, const Value& coords, 320 [[nodiscard]] F32 ImageSampleDrefExplicitLod(const Value& handle, const Value& coords,
319 const F32& dref, const F32& lod, 321 const F32& dref, const F32& lod,
320 const Value& offset, TextureInstInfo info); 322 const Value& offset, TextureInstInfo info);
321 [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod);
322 [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod, 323 [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod,
323 TextureInstInfo info); 324 const IR::U1& skip_mips);
325 [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod,
326 const IR::U1& skip_mips, TextureInstInfo info);
324 327
325 [[nodiscard]] Value ImageQueryLod(const Value& handle, const Value& coords, 328 [[nodiscard]] Value ImageQueryLod(const Value& handle, const Value& coords,
326 TextureInstInfo info); 329 TextureInstInfo info);
@@ -406,7 +409,8 @@ private:
406 } 409 }
407 410
408 template <typename T> 411 template <typename T>
409 requires(sizeof(T) <= sizeof(u32) && std::is_trivially_copyable_v<T>) struct Flags { 412 requires(sizeof(T) <= sizeof(u32) && std::is_trivially_copyable_v<T>)
413 struct Flags {
410 Flags() = default; 414 Flags() = default;
411 Flags(T proxy_) : proxy{proxy_} {} 415 Flags(T proxy_) : proxy{proxy_} {}
412 416
diff --git a/src/shader_recompiler/frontend/ir/opcodes.h b/src/shader_recompiler/frontend/ir/opcodes.h
index d155afd0f..e300714f3 100644
--- a/src/shader_recompiler/frontend/ir/opcodes.h
+++ b/src/shader_recompiler/frontend/ir/opcodes.h
@@ -38,7 +38,6 @@ constexpr Type U8{Type::U8};
38constexpr Type U16{Type::U16}; 38constexpr Type U16{Type::U16};
39constexpr Type U32{Type::U32}; 39constexpr Type U32{Type::U32};
40constexpr Type U64{Type::U64}; 40constexpr Type U64{Type::U64};
41constexpr Type S32{Type::S32};
42constexpr Type F16{Type::F16}; 41constexpr Type F16{Type::F16};
43constexpr Type F32{Type::F32}; 42constexpr Type F32{Type::F32};
44constexpr Type F64{Type::F64}; 43constexpr Type F64{Type::F64};
diff --git a/src/shader_recompiler/frontend/ir/opcodes.inc b/src/shader_recompiler/frontend/ir/opcodes.inc
index 1fe3749cc..4447d67b0 100644
--- a/src/shader_recompiler/frontend/ir/opcodes.inc
+++ b/src/shader_recompiler/frontend/ir/opcodes.inc
@@ -175,7 +175,6 @@ OPCODE(SelectF64, F64, U1,
175OPCODE(BitCastU16F16, U16, F16, ) 175OPCODE(BitCastU16F16, U16, F16, )
176OPCODE(BitCastU32F32, U32, F32, ) 176OPCODE(BitCastU32F32, U32, F32, )
177OPCODE(BitCastU64F64, U64, F64, ) 177OPCODE(BitCastU64F64, U64, F64, )
178OPCODE(BitCastS32F32, S32, F32, )
179OPCODE(BitCastF16U16, F16, U16, ) 178OPCODE(BitCastF16U16, F16, U16, )
180OPCODE(BitCastF32U32, F32, U32, ) 179OPCODE(BitCastF32U32, F32, U32, )
181OPCODE(BitCastF64U64, F64, U64, ) 180OPCODE(BitCastF64U64, F64, U64, )
@@ -483,7 +482,7 @@ OPCODE(BindlessImageSampleDrefExplicitLod, F32, U32,
483OPCODE(BindlessImageGather, F32x4, U32, Opaque, Opaque, Opaque, ) 482OPCODE(BindlessImageGather, F32x4, U32, Opaque, Opaque, Opaque, )
484OPCODE(BindlessImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, ) 483OPCODE(BindlessImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, )
485OPCODE(BindlessImageFetch, F32x4, U32, Opaque, Opaque, U32, Opaque, ) 484OPCODE(BindlessImageFetch, F32x4, U32, Opaque, Opaque, U32, Opaque, )
486OPCODE(BindlessImageQueryDimensions, U32x4, U32, U32, ) 485OPCODE(BindlessImageQueryDimensions, U32x4, U32, U32, U1, )
487OPCODE(BindlessImageQueryLod, F32x4, U32, Opaque, ) 486OPCODE(BindlessImageQueryLod, F32x4, U32, Opaque, )
488OPCODE(BindlessImageGradient, F32x4, U32, Opaque, Opaque, Opaque, Opaque, ) 487OPCODE(BindlessImageGradient, F32x4, U32, Opaque, Opaque, Opaque, Opaque, )
489OPCODE(BindlessImageRead, U32x4, U32, Opaque, ) 488OPCODE(BindlessImageRead, U32x4, U32, Opaque, )
@@ -496,7 +495,7 @@ OPCODE(BoundImageSampleDrefExplicitLod, F32, U32,
496OPCODE(BoundImageGather, F32x4, U32, Opaque, Opaque, Opaque, ) 495OPCODE(BoundImageGather, F32x4, U32, Opaque, Opaque, Opaque, )
497OPCODE(BoundImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, ) 496OPCODE(BoundImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, )
498OPCODE(BoundImageFetch, F32x4, U32, Opaque, Opaque, U32, Opaque, ) 497OPCODE(BoundImageFetch, F32x4, U32, Opaque, Opaque, U32, Opaque, )
499OPCODE(BoundImageQueryDimensions, U32x4, U32, U32, ) 498OPCODE(BoundImageQueryDimensions, U32x4, U32, U32, U1, )
500OPCODE(BoundImageQueryLod, F32x4, U32, Opaque, ) 499OPCODE(BoundImageQueryLod, F32x4, U32, Opaque, )
501OPCODE(BoundImageGradient, F32x4, U32, Opaque, Opaque, Opaque, Opaque, ) 500OPCODE(BoundImageGradient, F32x4, U32, Opaque, Opaque, Opaque, Opaque, )
502OPCODE(BoundImageRead, U32x4, U32, Opaque, ) 501OPCODE(BoundImageRead, U32x4, U32, Opaque, )
@@ -509,7 +508,7 @@ OPCODE(ImageSampleDrefExplicitLod, F32, Opaq
509OPCODE(ImageGather, F32x4, Opaque, Opaque, Opaque, Opaque, ) 508OPCODE(ImageGather, F32x4, Opaque, Opaque, Opaque, Opaque, )
510OPCODE(ImageGatherDref, F32x4, Opaque, Opaque, Opaque, Opaque, F32, ) 509OPCODE(ImageGatherDref, F32x4, Opaque, Opaque, Opaque, Opaque, F32, )
511OPCODE(ImageFetch, F32x4, Opaque, Opaque, Opaque, U32, Opaque, ) 510OPCODE(ImageFetch, F32x4, Opaque, Opaque, Opaque, U32, Opaque, )
512OPCODE(ImageQueryDimensions, U32x4, Opaque, U32, ) 511OPCODE(ImageQueryDimensions, U32x4, Opaque, U32, U1, )
513OPCODE(ImageQueryLod, F32x4, Opaque, Opaque, ) 512OPCODE(ImageQueryLod, F32x4, Opaque, Opaque, )
514OPCODE(ImageGradient, F32x4, Opaque, Opaque, Opaque, Opaque, Opaque, ) 513OPCODE(ImageGradient, F32x4, Opaque, Opaque, Opaque, Opaque, Opaque, )
515OPCODE(ImageRead, U32x4, Opaque, Opaque, ) 514OPCODE(ImageRead, U32x4, Opaque, Opaque, )
diff --git a/src/shader_recompiler/frontend/ir/type.h b/src/shader_recompiler/frontend/ir/type.h
index 5a7c706ad..04c8c4ddb 100644
--- a/src/shader_recompiler/frontend/ir/type.h
+++ b/src/shader_recompiler/frontend/ir/type.h
@@ -24,22 +24,21 @@ enum class Type {
24 U16 = 1 << 7, 24 U16 = 1 << 7,
25 U32 = 1 << 8, 25 U32 = 1 << 8,
26 U64 = 1 << 9, 26 U64 = 1 << 9,
27 S32 = 1 << 10, 27 F16 = 1 << 10,
28 F16 = 1 << 11, 28 F32 = 1 << 11,
29 F32 = 1 << 12, 29 F64 = 1 << 12,
30 F64 = 1 << 13, 30 U32x2 = 1 << 13,
31 U32x2 = 1 << 14, 31 U32x3 = 1 << 14,
32 U32x3 = 1 << 15, 32 U32x4 = 1 << 15,
33 U32x4 = 1 << 16, 33 F16x2 = 1 << 16,
34 F16x2 = 1 << 17, 34 F16x3 = 1 << 17,
35 F16x3 = 1 << 18, 35 F16x4 = 1 << 18,
36 F16x4 = 1 << 19, 36 F32x2 = 1 << 19,
37 F32x2 = 1 << 20, 37 F32x3 = 1 << 20,
38 F32x3 = 1 << 21, 38 F32x4 = 1 << 21,
39 F32x4 = 1 << 22, 39 F64x2 = 1 << 22,
40 F64x2 = 1 << 23, 40 F64x3 = 1 << 23,
41 F64x3 = 1 << 24, 41 F64x4 = 1 << 24,
42 F64x4 = 1 << 25,
43}; 42};
44DECLARE_ENUM_FLAG_OPERATORS(Type) 43DECLARE_ENUM_FLAG_OPERATORS(Type)
45 44
diff --git a/src/shader_recompiler/frontend/ir/value.cpp b/src/shader_recompiler/frontend/ir/value.cpp
index 30ba12316..346169328 100644
--- a/src/shader_recompiler/frontend/ir/value.cpp
+++ b/src/shader_recompiler/frontend/ir/value.cpp
@@ -23,8 +23,6 @@ Value::Value(u16 value) noexcept : type{Type::U16}, imm_u16{value} {}
23 23
24Value::Value(u32 value) noexcept : type{Type::U32}, imm_u32{value} {} 24Value::Value(u32 value) noexcept : type{Type::U32}, imm_u32{value} {}
25 25
26Value::Value(s32 value) noexcept : type{Type::S32}, imm_s32{value} {}
27
28Value::Value(f32 value) noexcept : type{Type::F32}, imm_f32{value} {} 26Value::Value(f32 value) noexcept : type{Type::F32}, imm_f32{value} {}
29 27
30Value::Value(u64 value) noexcept : type{Type::U64}, imm_u64{value} {} 28Value::Value(u64 value) noexcept : type{Type::U64}, imm_u64{value} {}
@@ -71,7 +69,6 @@ bool Value::operator==(const Value& other) const {
71 return imm_u16 == other.imm_u16; 69 return imm_u16 == other.imm_u16;
72 case Type::U32: 70 case Type::U32:
73 case Type::F32: 71 case Type::F32:
74 case Type::S32:
75 return imm_u32 == other.imm_u32; 72 return imm_u32 == other.imm_u32;
76 case Type::U64: 73 case Type::U64:
77 case Type::F64: 74 case Type::F64:
diff --git a/src/shader_recompiler/frontend/ir/value.h b/src/shader_recompiler/frontend/ir/value.h
index 8b34356fd..22e89dd1b 100644
--- a/src/shader_recompiler/frontend/ir/value.h
+++ b/src/shader_recompiler/frontend/ir/value.h
@@ -101,9 +101,8 @@ public:
101 TypedValue() = default; 101 TypedValue() = default;
102 102
103 template <IR::Type other_type> 103 template <IR::Type other_type>
104 requires((other_type & type_) != IR::Type::Void) explicit(false) 104 requires((other_type & type_) != IR::Type::Void)
105 TypedValue(const TypedValue<other_type>& value) 105 explicit(false) TypedValue(const TypedValue<other_type>& value) : Value(value) {}
106 : Value(value) {}
107 106
108 explicit TypedValue(const Value& value) : Value(value) { 107 explicit TypedValue(const Value& value) : Value(value) {
109 if ((value.Type() & type_) == IR::Type::Void) { 108 if ((value.Type() & type_) == IR::Type::Void) {
@@ -194,16 +193,16 @@ public:
194 void ReplaceOpcode(IR::Opcode opcode); 193 void ReplaceOpcode(IR::Opcode opcode);
195 194
196 template <typename FlagsType> 195 template <typename FlagsType>
197 requires(sizeof(FlagsType) <= sizeof(u32) && std::is_trivially_copyable_v<FlagsType>) 196 requires(sizeof(FlagsType) <= sizeof(u32) && std::is_trivially_copyable_v<FlagsType>)
198 [[nodiscard]] FlagsType Flags() const noexcept { 197 [[nodiscard]] FlagsType Flags() const noexcept {
199 FlagsType ret; 198 FlagsType ret;
200 std::memcpy(reinterpret_cast<char*>(&ret), &flags, sizeof(ret)); 199 std::memcpy(reinterpret_cast<char*>(&ret), &flags, sizeof(ret));
201 return ret; 200 return ret;
202 } 201 }
203 202
204 template <typename FlagsType> 203 template <typename FlagsType>
205 requires(sizeof(FlagsType) <= sizeof(u32) && 204 requires(sizeof(FlagsType) <= sizeof(u32) && std::is_trivially_copyable_v<FlagsType>)
206 std::is_trivially_copyable_v<FlagsType>) void SetFlags(FlagsType value) noexcept { 205 void SetFlags(FlagsType value) noexcept {
207 std::memcpy(&flags, &value, sizeof(value)); 206 std::memcpy(&flags, &value, sizeof(value));
208 } 207 }
209 208
@@ -268,7 +267,6 @@ using U8 = TypedValue<Type::U8>;
268using U16 = TypedValue<Type::U16>; 267using U16 = TypedValue<Type::U16>;
269using U32 = TypedValue<Type::U32>; 268using U32 = TypedValue<Type::U32>;
270using U64 = TypedValue<Type::U64>; 269using U64 = TypedValue<Type::U64>;
271using S32 = TypedValue<Type::S32>;
272using F16 = TypedValue<Type::F16>; 270using F16 = TypedValue<Type::F16>;
273using F32 = TypedValue<Type::F32>; 271using F32 = TypedValue<Type::F32>;
274using F64 = TypedValue<Type::F64>; 272using F64 = TypedValue<Type::F64>;
diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp
index f8cfd4ab6..39af62559 100644
--- a/src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp
@@ -15,11 +15,13 @@ enum class Mode : u64 {
15 SamplePos = 5, 15 SamplePos = 5,
16}; 16};
17 17
18IR::Value Query(TranslatorVisitor& v, const IR::U32& handle, Mode mode, IR::Reg src_reg) { 18IR::Value Query(TranslatorVisitor& v, const IR::U32& handle, Mode mode, IR::Reg src_reg, u64 mask) {
19 switch (mode) { 19 switch (mode) {
20 case Mode::Dimension: { 20 case Mode::Dimension: {
21 const bool needs_num_mips{((mask >> 3) & 1) != 0};
22 const IR::U1 skip_mips{v.ir.Imm1(!needs_num_mips)};
21 const IR::U32 lod{v.X(src_reg)}; 23 const IR::U32 lod{v.X(src_reg)};
22 return v.ir.ImageQueryDimension(handle, lod); 24 return v.ir.ImageQueryDimension(handle, lod, skip_mips);
23 } 25 }
24 case Mode::TextureType: 26 case Mode::TextureType:
25 case Mode::SamplePos: 27 case Mode::SamplePos:
@@ -46,7 +48,7 @@ void Impl(TranslatorVisitor& v, u64 insn, std::optional<u32> cbuf_offset) {
46 handle = v.X(src_reg); 48 handle = v.X(src_reg);
47 ++src_reg; 49 ++src_reg;
48 } 50 }
49 const IR::Value query{Query(v, handle, txq.mode, src_reg)}; 51 const IR::Value query{Query(v, handle, txq.mode, src_reg, txq.mask)};
50 IR::Reg dest_reg{txq.dest_reg}; 52 IR::Reg dest_reg{txq.dest_reg};
51 for (int element = 0; element < 4; ++element) { 53 for (int element = 0; element < 4; ++element) {
52 if (((txq.mask >> element) & 1) == 0) { 54 if (((txq.mask >> element) & 1) == 0) {
diff --git a/src/shader_recompiler/frontend/maxwell/translate_program.cpp b/src/shader_recompiler/frontend/maxwell/translate_program.cpp
index 3adbd2b16..a42453e90 100644
--- a/src/shader_recompiler/frontend/maxwell/translate_program.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate_program.cpp
@@ -171,6 +171,70 @@ std::map<IR::Attribute, IR::Attribute> GenerateLegacyToGenericMappings(
171 } 171 }
172 return mapping; 172 return mapping;
173} 173}
174
175void EmitGeometryPassthrough(IR::IREmitter& ir, const IR::Program& program,
176 const Shader::VaryingState& passthrough_mask,
177 bool passthrough_position,
178 std::optional<IR::Attribute> passthrough_layer_attr) {
179 for (u32 i = 0; i < program.output_vertices; i++) {
180 // Assign generics from input
181 for (u32 j = 0; j < 32; j++) {
182 if (!passthrough_mask.Generic(j)) {
183 continue;
184 }
185
186 const IR::Attribute attr = IR::Attribute::Generic0X + (j * 4);
187 ir.SetAttribute(attr + 0, ir.GetAttribute(attr + 0, ir.Imm32(i)), ir.Imm32(0));
188 ir.SetAttribute(attr + 1, ir.GetAttribute(attr + 1, ir.Imm32(i)), ir.Imm32(0));
189 ir.SetAttribute(attr + 2, ir.GetAttribute(attr + 2, ir.Imm32(i)), ir.Imm32(0));
190 ir.SetAttribute(attr + 3, ir.GetAttribute(attr + 3, ir.Imm32(i)), ir.Imm32(0));
191 }
192
193 if (passthrough_position) {
194 // Assign position from input
195 const IR::Attribute attr = IR::Attribute::PositionX;
196 ir.SetAttribute(attr + 0, ir.GetAttribute(attr + 0, ir.Imm32(i)), ir.Imm32(0));
197 ir.SetAttribute(attr + 1, ir.GetAttribute(attr + 1, ir.Imm32(i)), ir.Imm32(0));
198 ir.SetAttribute(attr + 2, ir.GetAttribute(attr + 2, ir.Imm32(i)), ir.Imm32(0));
199 ir.SetAttribute(attr + 3, ir.GetAttribute(attr + 3, ir.Imm32(i)), ir.Imm32(0));
200 }
201
202 if (passthrough_layer_attr) {
203 // Assign layer
204 ir.SetAttribute(IR::Attribute::Layer, ir.GetAttribute(*passthrough_layer_attr),
205 ir.Imm32(0));
206 }
207
208 // Emit vertex
209 ir.EmitVertex(ir.Imm32(0));
210 }
211 ir.EndPrimitive(ir.Imm32(0));
212}
213
214u32 GetOutputTopologyVertices(OutputTopology output_topology) {
215 switch (output_topology) {
216 case OutputTopology::PointList:
217 return 1;
218 case OutputTopology::LineStrip:
219 return 2;
220 default:
221 return 3;
222 }
223}
224
225void LowerGeometryPassthrough(const IR::Program& program, const HostTranslateInfo& host_info) {
226 for (IR::Block* const block : program.blocks) {
227 for (IR::Inst& inst : block->Instructions()) {
228 if (inst.GetOpcode() == IR::Opcode::Epilogue) {
229 IR::IREmitter ir{*block, IR::Block::InstructionList::s_iterator_to(inst)};
230 EmitGeometryPassthrough(
231 ir, program, program.info.passthrough,
232 program.info.passthrough.AnyComponent(IR::Attribute::PositionX), {});
233 }
234 }
235 }
236}
237
174} // Anonymous namespace 238} // Anonymous namespace
175 239
176IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Block>& block_pool, 240IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Block>& block_pool,
@@ -195,9 +259,14 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
195 program.is_geometry_passthrough = sph.common0.geometry_passthrough != 0; 259 program.is_geometry_passthrough = sph.common0.geometry_passthrough != 0;
196 if (program.is_geometry_passthrough) { 260 if (program.is_geometry_passthrough) {
197 const auto& mask{env.GpPassthroughMask()}; 261 const auto& mask{env.GpPassthroughMask()};
198 for (size_t i = 0; i < program.info.passthrough.mask.size(); ++i) { 262 for (size_t i = 0; i < mask.size() * 32; ++i) {
199 program.info.passthrough.mask[i] = ((mask[i / 32] >> (i % 32)) & 1) == 0; 263 program.info.passthrough.mask[i] = ((mask[i / 32] >> (i % 32)) & 1) == 0;
200 } 264 }
265
266 if (!host_info.support_geometry_shader_passthrough) {
267 program.output_vertices = GetOutputTopologyVertices(program.output_topology);
268 LowerGeometryPassthrough(program, host_info);
269 }
201 } 270 }
202 break; 271 break;
203 } 272 }
@@ -219,11 +288,11 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
219 } 288 }
220 Optimization::SsaRewritePass(program); 289 Optimization::SsaRewritePass(program);
221 290
222 Optimization::ConstantPropagationPass(program); 291 Optimization::ConstantPropagationPass(env, program);
223 292
224 Optimization::PositionPass(env, program); 293 Optimization::PositionPass(env, program);
225 294
226 Optimization::GlobalMemoryToStorageBufferPass(program); 295 Optimization::GlobalMemoryToStorageBufferPass(program, host_info);
227 Optimization::TexturePass(env, program, host_info); 296 Optimization::TexturePass(env, program, host_info);
228 297
229 if (Settings::values.resolution_info.active) { 298 if (Settings::values.resolution_info.active) {
@@ -342,17 +411,7 @@ IR::Program GenerateGeometryPassthrough(ObjectPool<IR::Inst>& inst_pool,
342 IR::Program program; 411 IR::Program program;
343 program.stage = Stage::Geometry; 412 program.stage = Stage::Geometry;
344 program.output_topology = output_topology; 413 program.output_topology = output_topology;
345 switch (output_topology) { 414 program.output_vertices = GetOutputTopologyVertices(output_topology);
346 case OutputTopology::PointList:
347 program.output_vertices = 1;
348 break;
349 case OutputTopology::LineStrip:
350 program.output_vertices = 2;
351 break;
352 default:
353 program.output_vertices = 3;
354 break;
355 }
356 415
357 program.is_geometry_passthrough = false; 416 program.is_geometry_passthrough = false;
358 program.info.loads.mask = source_program.info.stores.mask; 417 program.info.loads.mask = source_program.info.stores.mask;
@@ -366,35 +425,8 @@ IR::Program GenerateGeometryPassthrough(ObjectPool<IR::Inst>& inst_pool,
366 node.data.block = current_block; 425 node.data.block = current_block;
367 426
368 IR::IREmitter ir{*current_block}; 427 IR::IREmitter ir{*current_block};
369 for (u32 i = 0; i < program.output_vertices; i++) { 428 EmitGeometryPassthrough(ir, program, program.info.stores, true,
370 // Assign generics from input 429 source_program.info.emulated_layer);
371 for (u32 j = 0; j < 32; j++) {
372 if (!program.info.stores.Generic(j)) {
373 continue;
374 }
375
376 const IR::Attribute attr = IR::Attribute::Generic0X + (j * 4);
377 ir.SetAttribute(attr + 0, ir.GetAttribute(attr + 0, ir.Imm32(i)), ir.Imm32(0));
378 ir.SetAttribute(attr + 1, ir.GetAttribute(attr + 1, ir.Imm32(i)), ir.Imm32(0));
379 ir.SetAttribute(attr + 2, ir.GetAttribute(attr + 2, ir.Imm32(i)), ir.Imm32(0));
380 ir.SetAttribute(attr + 3, ir.GetAttribute(attr + 3, ir.Imm32(i)), ir.Imm32(0));
381 }
382
383 // Assign position from input
384 const IR::Attribute attr = IR::Attribute::PositionX;
385 ir.SetAttribute(attr + 0, ir.GetAttribute(attr + 0, ir.Imm32(i)), ir.Imm32(0));
386 ir.SetAttribute(attr + 1, ir.GetAttribute(attr + 1, ir.Imm32(i)), ir.Imm32(0));
387 ir.SetAttribute(attr + 2, ir.GetAttribute(attr + 2, ir.Imm32(i)), ir.Imm32(0));
388 ir.SetAttribute(attr + 3, ir.GetAttribute(attr + 3, ir.Imm32(i)), ir.Imm32(0));
389
390 // Assign layer
391 ir.SetAttribute(IR::Attribute::Layer, ir.GetAttribute(source_program.info.emulated_layer),
392 ir.Imm32(0));
393
394 // Emit vertex
395 ir.EmitVertex(ir.Imm32(0));
396 }
397 ir.EndPrimitive(ir.Imm32(0));
398 430
399 IR::Block* return_block{block_pool.Create(inst_pool)}; 431 IR::Block* return_block{block_pool.Create(inst_pool)};
400 IR::IREmitter{*return_block}.Epilogue(); 432 IR::IREmitter{*return_block}.Epilogue();
HIC SVNT DRACONES