glsl: F16x2 storage atomics

7 files changed, 64 insertions, 58 deletions

diff --git a/src/shader_recompiler/backend/glsl/emit_context.cpp b/src/shader_recompiler/backend/glsl/emit_context.cpp
index 7986bf78f..a413219e3 100644
--- a/src/shader_recompiler/backend/glsl/emit_context.cpp
+++ b/src/shader_recompiler/backend/glsl/emit_context.cpp
@@ -39,6 +39,10 @@ void EmitContext::SetupExtensions(std::string& header) {
     if (info.uses_atomic_f16x2_add || info.uses_atomic_f16x2_min || info.uses_atomic_f16x2_max) {
         header += "#extension NV_shader_atomic_fp16_vector : enable\n";
     }
+    if (info.uses_fp16) {
+        // TODO: AMD
+        header += "#extension GL_NV_gpu_shader5 : enable\n";
+    }
 }
 
 void EmitContext::DefineConstantBuffers() {
@@ -89,6 +93,18 @@ void EmitContext::DefineHelperFunctions() {
         code += "uint CasFloatMax32x2(uint op_a,uint op_b){return "
                 "packHalf2x16(max(unpackHalf2x16(op_a),unpackHalf2x16(op_b)));}\n";
     }
+    if (info.uses_atomic_f16x2_add) {
+        code += "uint CasFloatAdd16x2(uint op_a,uint op_b){return "
+                "packFloat2x16(unpackFloat2x16(op_a)+unpackFloat2x16(op_b));}\n";
+    }
+    if (info.uses_atomic_f16x2_min) {
+        code += "uint CasFloatMin16x2(uint op_a,uint op_b){return "
+                "packFloat2x16(min(unpackFloat2x16(op_a),unpackFloat2x16(op_b)));}\n";
+    }
+    if (info.uses_atomic_f16x2_max) {
+        code += "uint CasFloatMax16x2(uint op_a,uint op_b){return "
+                "packFloat2x16(max(unpackFloat2x16(op_a),unpackFloat2x16(op_b)));}\n";
+    }
     // TODO: Track this usage
     code += "uint CasMinS32(uint op_a,uint op_b){return uint(min(int(op_a),int(op_b)));}";
     code += "uint CasMaxS32(uint op_a,uint op_b){return uint(max(int(op_a),int(op_b)));}";
diff --git a/src/shader_recompiler/backend/glsl/emit_context.h b/src/shader_recompiler/backend/glsl/emit_context.h
index ca5657fe7..7f8857fa7 100644
--- a/src/shader_recompiler/backend/glsl/emit_context.h
+++ b/src/shader_recompiler/backend/glsl/emit_context.h
@@ -44,6 +44,11 @@ public:
     }
 
     template <typename... Args>
+    void AddF16x2(const char* format_str, IR::Inst& inst, Args&&... args) {
+        Add<Type::F16x2>(format_str, inst, args...);
+    }
+
+    template <typename... Args>
     void AddU32(const char* format_str, IR::Inst& inst, Args&&... args) {
         Add<Type::U32>(format_str, inst, args...);
     }
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp
index 0b29c213b..b6b326762 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp
@@ -12,8 +12,7 @@
 
 namespace Shader::Backend::GLSL {
 namespace {
-static constexpr std::string_view cas_loop{R"(
-uint {};
+static constexpr std::string_view cas_loop{R"(uint {};
 for (;;){{
     uint old_value={};
     {}=atomicCompSwap({},old_value,{}({},{}));
@@ -49,6 +48,14 @@ void CasFunctionF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding
     const auto ret{ctx.reg_alloc.Define(inst)};
     CasFunction(ctx, ret, ssbo, u32_value, function);
 }
+
+void CasFunctionF16x2(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("packFloat2x16({})", value)};
+    const auto ret{ctx.reg_alloc.Define(inst)};
+    CasFunction(ctx, ret, ssbo, u32_value, function);
+}
 } // namespace
 
 void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@@ -122,11 +129,8 @@ void EmitStorageAtomicSMin64(EmitContext& ctx, IR::Inst& inst, const IR::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]));
-}}
-)",
+    ctx.Add("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);
 }
 
@@ -135,12 +139,9 @@ void EmitStorageAtomicUMin64(EmitContext& ctx, IR::Inst& inst, const IR::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);
+    ctx.Add(
+        "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,
@@ -148,11 +149,8 @@ void EmitStorageAtomicSMax64(EmitContext& ctx, IR::Inst& inst, const IR::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]));
-}}
-)",
+    ctx.Add("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);
 }
 
@@ -161,12 +159,9 @@ void EmitStorageAtomicUMax64(EmitContext& ctx, IR::Inst& inst, const IR::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);
+    ctx.Add(
+        "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,
@@ -202,45 +197,33 @@ void EmitStorageAtomicAddF32(EmitContext& ctx, IR::Inst& inst, const IR::Value&
     CasFunctionF32(ctx, inst, binding, offset, value, "CasFloatAdd");
 }
 
-void EmitStorageAtomicAddF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
-                               [[maybe_unused]] const IR::Value& binding,
-                               [[maybe_unused]] const IR::Value& offset,
-                               [[maybe_unused]] std::string_view value) {
-    throw NotImplementedException("GLSL Instrucion");
+void EmitStorageAtomicAddF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
+                               const IR::Value& offset, std::string_view value) {
+    CasFunctionF16x2(ctx, inst, binding, offset, value, "CasFloatAdd16x2");
 }
 
-void EmitStorageAtomicAddF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
-                               [[maybe_unused]] const IR::Value& binding,
-                               [[maybe_unused]] const IR::Value& offset,
-                               [[maybe_unused]] std::string_view value) {
+void EmitStorageAtomicAddF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
+                               const IR::Value& offset, std::string_view value) {
     CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatAdd32x2");
 }
 
-void EmitStorageAtomicMinF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
-                               [[maybe_unused]] const IR::Value& binding,
-                               [[maybe_unused]] const IR::Value& offset,
-                               [[maybe_unused]] std::string_view value) {
-    throw NotImplementedException("GLSL Instrucion");
+void EmitStorageAtomicMinF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
+                               const IR::Value& offset, std::string_view value) {
+    CasFunctionF16x2(ctx, inst, binding, offset, value, "CasFloatMin16x2");
 }
 
-void EmitStorageAtomicMinF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
-                               [[maybe_unused]] const IR::Value& binding,
-                               [[maybe_unused]] const IR::Value& offset,
-                               [[maybe_unused]] std::string_view value) {
+void EmitStorageAtomicMinF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
+                               const IR::Value& offset, std::string_view value) {
     CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMin32x2");
 }
 
-void EmitStorageAtomicMaxF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
-                               [[maybe_unused]] const IR::Value& binding,
-                               [[maybe_unused]] const IR::Value& offset,
-                               [[maybe_unused]] std::string_view value) {
-    throw NotImplementedException("GLSL Instrucion");
+void EmitStorageAtomicMaxF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
+                               const IR::Value& offset, std::string_view value) {
+    CasFunctionF16x2(ctx, inst, binding, offset, value, "CasFloatMax16x2");
 }
 
-void EmitStorageAtomicMaxF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
-                               [[maybe_unused]] const IR::Value& binding,
-                               [[maybe_unused]] const IR::Value& offset,
-                               [[maybe_unused]] std::string_view value) {
+void EmitStorageAtomicMaxF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
+                               const IR::Value& offset, std::string_view value) {
     CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMax32x2");
 }
 
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_bitwise_conversion.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_bitwise_conversion.cpp
index a1e97b4cb..742f394d4 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_bitwise_conversion.cpp
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_bitwise_conversion.cpp
@@ -62,13 +62,12 @@ void EmitUnpackUint2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value
     ctx.AddU32x2("{}=unpackUint2x32({});", inst, value);
 }
 
-void EmitPackFloat2x16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] std::string_view value) {
-    throw NotImplementedException("GLSL Instruction");
+void EmitPackFloat2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value) {
+    ctx.AddU32("{}=packFloat2x16({});", inst, value);
 }
 
-void EmitUnpackFloat2x16([[maybe_unused]] EmitContext& ctx,
-                         [[maybe_unused]] std::string_view value) {
-    throw NotImplementedException("GLSL Instruction");
+void EmitUnpackFloat2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value) {
+    ctx.AddF16x2("{}=unpackFloat2x16({});", inst, value);
 }
 
 void EmitPackHalf2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value) {
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
index 56b812d84..9f32070b0 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
@@ -224,8 +224,8 @@ void EmitBitCastF32U32(EmitContext& ctx, IR::Inst& inst, std::string_view value)
 void EmitBitCastF64U64(EmitContext& ctx, IR::Inst& inst, std::string_view value);
 void EmitPackUint2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value);
 void EmitUnpackUint2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value);
-void EmitPackFloat2x16(EmitContext& ctx, std::string_view value);
-void EmitUnpackFloat2x16(EmitContext& ctx, std::string_view value);
+void EmitPackFloat2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value);
+void EmitUnpackFloat2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value);
 void EmitPackHalf2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value);
 void EmitUnpackHalf2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value);
 void EmitPackDouble2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value);
diff --git a/src/shader_recompiler/backend/glsl/reg_alloc.cpp b/src/shader_recompiler/backend/glsl/reg_alloc.cpp
index 58c2c408e..3de4c1172 100644
--- a/src/shader_recompiler/backend/glsl/reg_alloc.cpp
+++ b/src/shader_recompiler/backend/glsl/reg_alloc.cpp
@@ -113,6 +113,8 @@ std::string RegAlloc::GetType(Type type, u32 index) {
     switch (type) {
     case Type::U1:
         return "bool ";
+    case Type::F16x2:
+        return "f16vec2 ";
     case Type::U32:
         return "uint ";
     case Type::S32:
diff --git a/src/shader_recompiler/backend/glsl/reg_alloc.h b/src/shader_recompiler/backend/glsl/reg_alloc.h
index 581954e44..9f2ff8718 100644
--- a/src/shader_recompiler/backend/glsl/reg_alloc.h
+++ b/src/shader_recompiler/backend/glsl/reg_alloc.h
@@ -18,6 +18,7 @@ enum class Type;
 namespace Shader::Backend::GLSL {
 enum class Type : u32 {
     U1,
+    F16x2,
     S32,
     U32,
     F32,