1 files changed, 101 insertions, 40 deletions

diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp
index f3ef37873..0b29c213b 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp
@@ -13,132 +13,193 @@
 namespace Shader::Backend::GLSL {
 namespace {
 static constexpr std::string_view cas_loop{R"(
-{} {};
+uint {};
 for (;;){{
-    {} old_value={};
-    {} = atomicCompSwap({},old_value,{}({},{}));
+    uint old_value={};
+    {}=atomicCompSwap({},old_value,{}({},{}));
     if ({}==old_value){{break;}}
 }})"};
 
-void CasFunction(EmitContext& ctx, IR::Inst& inst, std::string_view ssbo, std::string_view value,
-                 std::string_view type, std::string_view function) {
+void CasFunction(EmitContext& ctx, std::string_view ret, std::string_view ssbo,
+                 std::string_view value, std::string_view function) {
+    ctx.Add(cas_loop.data(), ret, ssbo, ret, ssbo, function, ssbo, value, ret);
+}
+
+void CasFunctionInt32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
+                      const IR::Value& offset, std::string_view value, std::string_view function) {
+    const auto ret{ctx.reg_alloc.Define(inst)};
+    const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())};
+    CasFunction(ctx, ret, ssbo, value, function);
+}
+
+void CasFunctionF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
+                    const IR::Value& offset, std::string_view value, std::string_view function) {
+    const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())};
+    const std::string u32_value{fmt::format("floatBitsToUint({})", value)};
+    const auto ret{ctx.reg_alloc.Define(inst)};
+    const auto ret_32{ret + "_u32"};
+    CasFunction(ctx, ret_32, ssbo, u32_value, function);
+    ctx.Add("float {}=uintBitsToFloat({});", ret, ret_32);
+}
+
+void CasFunctionF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
+                      const IR::Value& offset, std::string_view value, std::string_view function) {
+    const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())};
+    const std::string u32_value{fmt::format("packHalf2x16({})", value)};
     const auto ret{ctx.reg_alloc.Define(inst)};
-    ctx.Add(cas_loop.data(), type, ret, type, ssbo, ret, ssbo, function, ssbo, value, ret);
+    CasFunction(ctx, ret, ssbo, u32_value, function);
 }
 } // namespace
 
 void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddU32("{}=atomicAdd(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU32("{}=atomicAdd(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicSMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddS32("{}=atomicMin(ssbo{}_s32[{}],int({}));", inst, binding.U32(), offset.U32(), value);
+    const std::string u32_value{fmt::format("uint({})", value)};
+    CasFunctionInt32(ctx, inst, binding, offset, u32_value, "CasMinS32");
 }
 
 void EmitStorageAtomicUMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddU32("{}=atomicMin(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU32("{}=atomicMin(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicSMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddS32("{}=atomicMax(ssbo{}_s32[{}],int({}));", inst, binding.U32(), offset.U32(), value);
+    const std::string u32_value{fmt::format("uint({})", value)};
+    CasFunctionInt32(ctx, inst, binding, offset, u32_value, "CasMaxS32");
 }
 
 void EmitStorageAtomicUMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddU32("{}=atomicMax(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU32("{}=atomicMax(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicInc32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
-                            [[maybe_unused]] const IR::Value& offset, std::string_view value) {
-    // const auto ret{ctx.reg_alloc.Define(inst)};
-    // const auto type{"uint"};
-    // ctx.Add(cas_loop.data(), type, ret, type, ssbo, ret, ssbo, "CasIncrement", ssbo, value, ret);
-    const std::string ssbo{fmt::format("ssbo{}_u32[{}]", binding.U32(), offset.U32())};
-    CasFunction(ctx, inst, ssbo, value, "uint", "CasIncrement");
+                            const IR::Value& offset, std::string_view value) {
+    CasFunctionInt32(ctx, inst, binding, offset, value, "CasIncrement");
 }
 
 void EmitStorageAtomicDec32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                             const IR::Value& offset, std::string_view value) {
-    const std::string ssbo{fmt::format("ssbo{}_u32[{}]", binding.U32(), offset.U32())};
-    CasFunction(ctx, inst, ssbo, value, "uint", "CasDecrement");
+    CasFunctionInt32(ctx, inst, binding, offset, value, "CasDecrement");
 }
 
 void EmitStorageAtomicAnd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                             const IR::Value& offset, std::string_view value) {
-    ctx.AddU32("{}=atomicAnd(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU32("{}=atomicAnd(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicOr32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                            const IR::Value& offset, std::string_view value) {
-    ctx.AddU32("{}=atomicOr(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU32("{}=atomicOr(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicXor32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                             const IR::Value& offset, std::string_view value) {
-    ctx.AddU32("{}=atomicXor(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU32("{}=atomicXor(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicExchange32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                                  const IR::Value& offset, std::string_view value) {
-    ctx.AddU32("{}=atomicExchange(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU32("{}=atomicExchange(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicIAdd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    // ctx.AddU64("{}=atomicAdd(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
-    ctx.AddU64("{}=ssbo{}_u64[{}];", inst, binding.U32(), offset.U32());
-    ctx.Add("ssbo{}_u64[{}]+={};", binding.U32(), offset.U32(), value);
+    // LOG_WARNING(..., "Op falling to non-atomic");
+    ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
+               binding.U32(), offset.U32() + 1);
+    ctx.Add("ssbo{}[{}]+=unpackUint2x32({}).x;ssbo{}[{}]+=unpackUint2x32({}).y;", binding.U32(),
+            offset.U32(), value, binding.U32(), offset.U32() + 1, value);
 }
 
 void EmitStorageAtomicSMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddS64("{}=atomicMin(int64_t(ssbo{}_u64[{}]),int64_t({}));", inst, binding.U32(),
-               offset.U32(), value);
+    // LOG_WARNING(..., "Op falling to non-atomic");
+    ctx.AddS64("{}=int64_t(ivec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
+               binding.U32(), offset.U32() + 1);
+    ctx.Add(R"(
+for(int i=0;i<2;++i){{
+ssbo{}[{}+i]=uint(min(int(ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));
+}}
+)",
+            binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicUMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddU64("{}=atomicMin(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
+    // LOG_WARNING(..., "Op falling to non-atomic");
+    ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
+               binding.U32(), offset.U32() + 1);
+    ctx.Add(R"(
+for(int i=0;i<2;++i){{
+ssbo{}[{}+i]=min(ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]);
+}}
+)",
+            binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicSMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddS64("{}=atomicMax(int64_t(ssbo{}_u64[{}]),int64_t({}));", inst, binding.U32(),
-               offset.U32(), value);
+    // LOG_WARNING(..., "Op falling to non-atomic");
+    ctx.AddS64("{}=int64_t(ivec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
+               binding.U32(), offset.U32() + 1);
+    ctx.Add(R"(
+for(int i=0;i<2;++i){{
+ssbo{}[{}+i]=uint(max(int(ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));
+}}
+)",
+            binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicUMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddU64("{}=atomicMax(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
+    // LOG_WARNING(..., "Op falling to non-atomic");
+    ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
+               binding.U32(), offset.U32() + 1);
+    ctx.Add(R"(
+for(int i=0;i<2;++i){{
+ssbo{}[{}+i]=max(ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]);
+}}
+)",
+            binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
 }
 
 void EmitStorageAtomicAnd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                             const IR::Value& offset, std::string_view value) {
-    ctx.AddU64("{}=atomicAnd(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU64("{}=uint64_t(uvec2(atomicAnd(ssbo{}[{}],unpackUint2x32({}).x),atomicAnd(ssbo{}[{}],"
+               "unpackUint2x32({}).y)));",
+               inst, binding.U32(), offset.U32(), value, binding.U32(), offset.U32() + 1, value);
 }
 
 void EmitStorageAtomicOr64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                            const IR::Value& offset, std::string_view value) {
-    ctx.AddU64("{}=atomicOr(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU64("{}=uint64_t(uvec2(atomicOr(ssbo{}[{}],unpackUint2x32({}).x),atomicOr(ssbo{}[{}],"
+               "unpackUint2x32({}).y)));",
+               inst, binding.U32(), offset.U32(), value, binding.U32(), offset.U32() + 1, value);
 }
 
 void EmitStorageAtomicXor64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                             const IR::Value& offset, std::string_view value) {
-    ctx.AddU64("{}=atomicXor(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU64("{}=uint64_t(uvec2(atomicXor(ssbo{}[{}],unpackUint2x32({}).x),atomicXor(ssbo{}[{}],"
+               "unpackUint2x32({}).y)));",
+               inst, binding.U32(), offset.U32(), value, binding.U32(), offset.U32() + 1, value);
 }
 
 void EmitStorageAtomicExchange64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                                  const IR::Value& offset, std::string_view value) {
-    ctx.AddU64("{}=atomicExchange(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
+    ctx.AddU64("{}=uint64_t(uvec2(atomicExchange(ssbo{}[{}],unpackUint2x32({}).x),atomicExchange("
+               "ssbo{}[{}],unpackUint2x32({}).y)));",
+               inst, binding.U32(), offset.U32(), value, binding.U32(), offset.U32() + 1, value);
 }
 
 void EmitStorageAtomicAddF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
                              const IR::Value& offset, std::string_view value) {
-    ctx.AddF32("{}=atomicAdd(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
+    CasFunctionF32(ctx, inst, binding, offset, value, "CasFloatAdd");
 }
 
 void EmitStorageAtomicAddF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
@@ -152,7 +213,7 @@ void EmitStorageAtomicAddF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused
                                [[maybe_unused]] const IR::Value& binding,
                                [[maybe_unused]] const IR::Value& offset,
                                [[maybe_unused]] std::string_view value) {
-    throw NotImplementedException("GLSL Instrucion");
+    CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatAdd32x2");
 }
 
 void EmitStorageAtomicMinF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
@@ -166,7 +227,7 @@ void EmitStorageAtomicMinF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused
                                [[maybe_unused]] const IR::Value& binding,
                                [[maybe_unused]] const IR::Value& offset,
                                [[maybe_unused]] std::string_view value) {
-    throw NotImplementedException("GLSL Instrucion");
+    CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMin32x2");
 }
 
 void EmitStorageAtomicMaxF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
@@ -180,7 +241,7 @@ void EmitStorageAtomicMaxF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused
                                [[maybe_unused]] const IR::Value& binding,
                                [[maybe_unused]] const IR::Value& offset,
                                [[maybe_unused]] std::string_view value) {
-    throw NotImplementedException("GLSL Instrucion");
+    CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMax32x2");
 }
 
 void EmitGlobalAtomicIAdd32(EmitContext&) {