From c7c518e280d1ac04adb08d45145690fd06ac7b18 Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Wed, 24 Mar 2021 23:41:55 +0100
Subject: shader: Implement TLD4 and TLD4_B
---
src/shader_recompiler/frontend/ir/ir_emitter.cpp | 13 +++++++++++++
src/shader_recompiler/frontend/ir/ir_emitter.h | 6 ++++++
src/shader_recompiler/frontend/ir/modifiers.h | 1 +
src/shader_recompiler/frontend/ir/opcodes.inc | 6 ++++++
4 files changed, 26 insertions(+)
(limited to 'src/shader_recompiler/frontend/ir')
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.cpp b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
index 418b7f5aca..b365a8a6eb 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.cpp
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
@@ -1474,6 +1474,19 @@ F32 IREmitter::ImageSampleDrefExplicitLod(const Value& handle, const Value& coor
return Inst<F32>(op, Flags{info}, handle, coords, dref, lod_lc, offset);
}
+Value IREmitter::ImageGather(const Value& handle, const Value& coords, const Value& offset,
+ const Value& offset2, TextureInstInfo info) {
+ const Opcode op{handle.IsImmediate() ? Opcode::BoundImageGather : Opcode::BindlessImageGather};
+ return Inst(op, Flags{info}, handle, coords, offset, offset2);
+}
+
+Value IREmitter::ImageGatherDref(const Value& handle, const Value& coords, const Value& offset,
+ const Value& offset2, const F32& dref, TextureInstInfo info) {
+ const Opcode op{handle.IsImmediate() ? Opcode::BoundImageGatherDref
+ : Opcode::BindlessImageGatherDref};
+ return Inst(op, Flags{info}, handle, coords, offset, offset2, dref);
+}
+
U1 IREmitter::VoteAll(const U1& value) {
return Inst<U1>(Opcode::VoteAll, value);
}
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.h b/src/shader_recompiler/frontend/ir/ir_emitter.h
index 64738735e6..04b43197f1 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.h
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.h
@@ -240,6 +240,12 @@ public:
const Value& offset, const F32& lod_clamp,
TextureInstInfo info);
+ [[nodiscard]] Value ImageGather(const Value& handle, const Value& coords, const Value& offset,
+ const Value& offset2, TextureInstInfo info);
+
+ [[nodiscard]] Value ImageGatherDref(const Value& handle, const Value& coords, const Value& offset,
+ const Value& offset2, const F32& dref, TextureInstInfo info);
+
[[nodiscard]] U1 VoteAll(const U1& value);
[[nodiscard]] U1 VoteAny(const U1& value);
[[nodiscard]] U1 VoteEqual(const U1& value);
diff --git a/src/shader_recompiler/frontend/ir/modifiers.h b/src/shader_recompiler/frontend/ir/modifiers.h
index 308c001533..4f09a4b39e 100644
--- a/src/shader_recompiler/frontend/ir/modifiers.h
+++ b/src/shader_recompiler/frontend/ir/modifiers.h
@@ -38,6 +38,7 @@ union TextureInstInfo {
BitField<8, 1, u32> has_bias;
BitField<9, 1, u32> has_lod_clamp;
BitField<10, 1, u32> relaxed_precision;
+ BitField<11, 2, u32> gather_component;
};
static_assert(sizeof(TextureInstInfo) <= sizeof(u32));
diff --git a/src/shader_recompiler/frontend/ir/opcodes.inc b/src/shader_recompiler/frontend/ir/opcodes.inc
index a2479c46ac..60a0bc9802 100644
--- a/src/shader_recompiler/frontend/ir/opcodes.inc
+++ b/src/shader_recompiler/frontend/ir/opcodes.inc
@@ -353,16 +353,22 @@ OPCODE(BindlessImageSampleImplicitLod, F32x4, U32,
OPCODE(BindlessImageSampleExplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(BindlessImageSampleDrefImplicitLod, F32, U32, Opaque, F32, Opaque, Opaque, )
OPCODE(BindlessImageSampleDrefExplicitLod, F32, U32, Opaque, F32, Opaque, Opaque, )
+OPCODE(BindlessImageGather, F32x4, U32, Opaque, Opaque, Opaque, )
+OPCODE(BindlessImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, )
OPCODE(BoundImageSampleImplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(BoundImageSampleExplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(BoundImageSampleDrefImplicitLod, F32, U32, Opaque, F32, Opaque, Opaque, )
OPCODE(BoundImageSampleDrefExplicitLod, F32, U32, Opaque, F32, Opaque, Opaque, )
+OPCODE(BoundImageGather, F32x4, U32, Opaque, Opaque, Opaque, )
+OPCODE(BoundImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, )
OPCODE(ImageSampleImplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(ImageSampleExplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(ImageSampleDrefImplicitLod, F32, U32, Opaque, F32, Opaque, Opaque, )
OPCODE(ImageSampleDrefExplicitLod, F32, U32, Opaque, F32, Opaque, Opaque, )
+OPCODE(ImageGather, F32x4, U32, Opaque, Opaque, Opaque, )
+OPCODE(ImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, )
// Warp operations
OPCODE(VoteAll, U1, U1, )
--
cgit v1.2.3-70-g09d2