From 2091bc56513fd072ae8477b51c9a20be7817e618 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Marcin=20Miko=C5=82ajczyk?= <marcinmikolajcz@gmail.com>
Date: Fri, 30 May 2025 01:56:24 +0200
Subject: [PATCH] Handle R128 bit in MIMG instructions (#3010)

---
 src/shader_recompiler/frontend/decode.cpp                | 1 -
 .../frontend/translate/vector_memory.cpp                 | 5 +++++
 src/shader_recompiler/info.h                             | 9 ++++++++-
 .../ir/passes/resource_tracking_pass.cpp                 | 1 +
 src/shader_recompiler/ir/reg.h                           | 1 +
 5 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/src/shader_recompiler/frontend/decode.cpp b/src/shader_recompiler/frontend/decode.cpp
index 20b78e869..37e8a0973 100644
--- a/src/shader_recompiler/frontend/decode.cpp
+++ b/src/shader_recompiler/frontend/decode.cpp
@@ -1032,7 +1032,6 @@ void GcnDecodeContext::decodeInstructionMIMG(uint64_t hexInstruction) {
 
     m_instruction.control.mimg = *reinterpret_cast<InstControlMIMG*>(&hexInstruction);
     m_instruction.control.mimg.mod = getMimgModifier(m_instruction.opcode);
-    ASSERT(m_instruction.control.mimg.r128 == 0);
 }
 
 void GcnDecodeContext::decodeInstructionDS(uint64_t hexInstruction) {
diff --git a/src/shader_recompiler/frontend/translate/vector_memory.cpp b/src/shader_recompiler/frontend/translate/vector_memory.cpp
index 5639bc56a..8c1946390 100644
--- a/src/shader_recompiler/frontend/translate/vector_memory.cpp
+++ b/src/shader_recompiler/frontend/translate/vector_memory.cpp
@@ -377,6 +377,7 @@ void Translator::IMAGE_LOAD(bool has_mip, const GcnInst& inst) {
     IR::TextureInstInfo info{};
     info.has_lod.Assign(has_mip);
     info.is_array.Assign(mimg.da);
+    info.is_r128.Assign(mimg.r128);
     const IR::Value texel = ir.ImageRead(handle, body, {}, {}, info);
 
     for (u32 i = 0; i < 4; i++) {
@@ -426,6 +427,7 @@ void Translator::IMAGE_GET_RESINFO(const GcnInst& inst) {
 
     IR::TextureInstInfo info{};
     info.is_array.Assign(mimg.da);
+    info.is_r128.Assign(mimg.r128);
 
     const IR::Value size = ir.ImageQueryDimension(tsharp, lod, ir.Imm1(has_mips), info);
 
@@ -451,6 +453,7 @@ void Translator::IMAGE_ATOMIC(AtomicOp op, const GcnInst& inst) {
 
     IR::TextureInstInfo info{};
     info.is_array.Assign(mimg.da);
+    info.is_r128.Assign(mimg.r128);
 
     const IR::Value value = ir.GetVectorReg(val_reg);
     const IR::Value handle = ir.GetScalarReg(tsharp_reg);
@@ -509,6 +512,7 @@ IR::Value EmitImageSample(IR::IREmitter& ir, const GcnInst& inst, const IR::Scal
     info.has_lod.Assign(flags.any(MimgModifier::Lod));
     info.is_array.Assign(mimg.da);
     info.is_unnormalized.Assign(mimg.unrm);
+    info.is_r128.Assign(mimg.r128);
 
     if (gather) {
         info.gather_comp.Assign(std::bit_width(mimg.dmask) - 1);
@@ -617,6 +621,7 @@ void Translator::IMAGE_GET_LOD(const GcnInst& inst) {
 
     IR::TextureInstInfo info{};
     info.is_array.Assign(mimg.da);
+    info.is_r128.Assign(mimg.r128);
 
     const IR::Value handle = ir.GetScalarReg(tsharp_reg);
     const IR::Value body = ir.CompositeConstruct(
diff --git a/src/shader_recompiler/info.h b/src/shader_recompiler/info.h
index d349d7827..24e0741c1 100644
--- a/src/shader_recompiler/info.h
+++ b/src/shader_recompiler/info.h
@@ -84,6 +84,7 @@ struct ImageResource {
     bool is_atomic{};
     bool is_array{};
     bool is_written{};
+    bool is_r128{};
 
     [[nodiscard]] constexpr AmdGpu::Image GetSharp(const Info& info) const noexcept;
 };
@@ -293,7 +294,13 @@ constexpr AmdGpu::Buffer BufferResource::GetSharp(const Info& info) const noexce
 }
 
 constexpr AmdGpu::Image ImageResource::GetSharp(const Info& info) const noexcept {
-    const auto image = info.ReadUdSharp<AmdGpu::Image>(sharp_idx);
+    AmdGpu::Image image{0};
+    if (!is_r128) {
+        image = info.ReadUdSharp<AmdGpu::Image>(sharp_idx);
+    } else {
+        AmdGpu::Buffer buf = info.ReadUdSharp<AmdGpu::Buffer>(sharp_idx);
+        memcpy(&image, &buf, sizeof(buf));
+    }
     if (!image.Valid()) {
         // Fall back to null image if unbound.
         return AmdGpu::Image::Null();
diff --git a/src/shader_recompiler/ir/passes/resource_tracking_pass.cpp b/src/shader_recompiler/ir/passes/resource_tracking_pass.cpp
index cc0bf83d3..18c77e600 100644
--- a/src/shader_recompiler/ir/passes/resource_tracking_pass.cpp
+++ b/src/shader_recompiler/ir/passes/resource_tracking_pass.cpp
@@ -411,6 +411,7 @@ void PatchImageSharp(IR::Block& block, IR::Inst& inst, Info& info, Descriptors&
         .is_atomic = IsImageAtomicInstruction(inst),
         .is_array = bool(inst_info.is_array),
         .is_written = is_written,
+        .is_r128 = bool(inst_info.is_r128),
     });
 
     IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
diff --git a/src/shader_recompiler/ir/reg.h b/src/shader_recompiler/ir/reg.h
index 622190cf0..82aa436a7 100644
--- a/src/shader_recompiler/ir/reg.h
+++ b/src/shader_recompiler/ir/reg.h
@@ -44,6 +44,7 @@ union TextureInstInfo {
     BitField<9, 1, u32> is_array;
     BitField<10, 1, u32> is_unnormalized;
     BitField<11, 1, u32> is_gather;
+    BitField<12, 1, u32> is_r128;
 };
 
 union BufferInstInfo {