Implement extension GL_NV_shader_atomic_int64

SPIRV/GlslangToSpv.cpp

@@ -4749,12 +4749,12 @@ spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv
     case glslang::EOpAtomicMin:
     case glslang::EOpImageAtomicMin:
     case glslang::EOpAtomicCounterMin:
-        opCode = typeProxy == glslang::EbtUint ? spv::OpAtomicUMin : spv::OpAtomicSMin;
+        opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMin : spv::OpAtomicSMin;
         break;
     case glslang::EOpAtomicMax:
     case glslang::EOpImageAtomicMax:
     case glslang::EOpAtomicCounterMax:
-        opCode = typeProxy == glslang::EbtUint ? spv::OpAtomicUMax : spv::OpAtomicSMax;
+        opCode = (typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64) ? spv::OpAtomicUMax : spv::OpAtomicSMax;
         break;
     case glslang::EOpAtomicAnd:
     case glslang::EOpImageAtomicAnd:
@@ -4795,6 +4795,9 @@ spv::Id TGlslangToSpvTraverser::createAtomicOperation(glslang::TOperator op, spv
         break;
     }
 
+    if (typeProxy == glslang::EbtInt64 || typeProxy == glslang::EbtUint64)
+        builder.addCapability(spv::CapabilityInt64Atomics);
+
     // Sort out the operands
     //  - mapping from glslang -> SPV
     //  - there are extra SPV operands with no glslang source

Test/spv.atomicInt64.comp

+#version 450 core
+
+#extension GL_ARB_gpu_shader_int64: enable
+#extension GL_NV_shader_atomic_int64: enable
+
+layout(local_size_x = 16, local_size_y = 16) in;
+
+layout(binding = 0) buffer Buffer
+{
+    int64_t  i64;
+    uint64_t u64;
+} buf;
+
+struct Struct
+{
+    int64_t  i64;
+    uint64_t u64;
+};
+
+shared Struct s;
+
+void main()
+{
+    const int64_t  i64c = -24;
+    const uint64_t u64c = 0xF00000000F; 
+
+    // Test shader storage block
+    int64_t  i64 = 0;
+    uint64_t u64 = 0;
+
+    i64 += atomicMin(buf.i64, i64c);
+    u64 += atomicMin(buf.u64, u64c);
+
+    i64 += atomicMax(buf.i64, i64c);
+    u64 += atomicMax(buf.u64, u64c);
+
+    i64 += atomicAnd(buf.i64, i64c);
+    u64 += atomicAnd(buf.u64, u64c);
+
+    i64 += atomicOr(buf.i64, i64c);
+    u64 += atomicOr(buf.u64, u64c);
+
+    i64 += atomicXor(buf.i64, i64c);
+    u64 += atomicXor(buf.u64, u64c);
+
+    i64 += atomicAdd(buf.i64, i64c);
+    i64 += atomicExchange(buf.i64, i64c);
+    i64 += atomicCompSwap(buf.i64, i64c, i64);
+
+    buf.i64 = i64;
+    buf.u64 = u64;
+
+    // Test shared variable
+    i64 = 0;
+    u64 = 0;
+
+    i64 += atomicMin(s.i64, i64c);
+    u64 += atomicMin(s.u64, u64c);
+
+    i64 += atomicMax(s.i64, i64c);
+    u64 += atomicMax(s.u64, u64c);
+
+    i64 += atomicAnd(s.i64, i64c);
+    u64 += atomicAnd(s.u64, u64c);
+
+    i64 += atomicOr(s.i64, i64c);
+    u64 += atomicOr(s.u64, u64c);
+
+    i64 += atomicXor(s.i64, i64c);
+    u64 += atomicXor(s.u64, u64c);
+
+    i64 += atomicAdd(s.i64, i64c);
+    i64 += atomicExchange(s.i64, i64c);
+    i64 += atomicCompSwap(s.i64, i64c, i64);
+
+    s.i64 = i64;
+    s.u64 = u64;
+}
+

glslang/MachineIndependent/Initialize.cpp

@@ -923,6 +923,32 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
             "\n");
     }
 
+#ifdef NV_EXTENSIONS
+    if (profile != EEsProfile && version >= 440) {
+        commonBuiltins.append(
+            "uint64_t atomicMin(coherent volatile inout uint64_t, uint64_t);"
+            " int64_t atomicMin(coherent volatile inout  int64_t,  int64_t);"
+
+            "uint64_t atomicMax(coherent volatile inout uint64_t, uint64_t);"
+            " int64_t atomicMax(coherent volatile inout  int64_t,  int64_t);"
+
+            "uint64_t atomicAnd(coherent volatile inout uint64_t, uint64_t);"
+            " int64_t atomicAnd(coherent volatile inout  int64_t,  int64_t);"
+
+            "uint64_t atomicOr (coherent volatile inout uint64_t, uint64_t);"
+            " int64_t atomicOr (coherent volatile inout  int64_t,  int64_t);"
+
+            "uint64_t atomicXor(coherent volatile inout uint64_t, uint64_t);"
+            " int64_t atomicXor(coherent volatile inout  int64_t,  int64_t);"
+
+            " int64_t atomicAdd(coherent volatile inout int64_t, int64_t);"
+            " int64_t atomicExchange(coherent volatile inout int64_t, int64_t);"
+            " int64_t atomicCompSwap(coherent volatile inout int64_t, int64_t, int64_t);"
+
+            "\n");
+    }
+#endif
+
     if ((profile == EEsProfile && version >= 310) ||
         (profile != EEsProfile && version >= 450)) {
         commonBuiltins.append(

glslang/MachineIndependent/ParseHelper.cpp

@@ -1551,6 +1551,23 @@ void TParseContext::builtInOpCheck(const TSourceLoc& loc, const TFunction& fnCan
         break;
     }
 
+#ifdef NV_EXTENSIONS
+    case EOpAtomicAdd:
+    case EOpAtomicMin:
+    case EOpAtomicMax:
+    case EOpAtomicAnd:
+    case EOpAtomicOr:
+    case EOpAtomicXor:
+    case EOpAtomicExchange:
+    case EOpAtomicCompSwap:
+    {
+        if (arg0->getType().getBasicType() == EbtInt64 || arg0->getType().getBasicType() == EbtUint64)
+            requireExtensions(loc, 1, &E_GL_NV_shader_atomic_int64, fnCandidate.getName().c_str());
+
+        break;
+    }
+#endif
+
     case EOpInterpolateAtCentroid:
     case EOpInterpolateAtSample:
     case EOpInterpolateAtOffset:

glslang/MachineIndependent/Versions.cpp

@@ -211,6 +211,7 @@ void TParseVersions::initializeExtensionBehavior()
     extensionBehavior[E_GL_NV_viewport_array2]                       = EBhDisable;
     extensionBehavior[E_GL_NV_stereo_view_rendering]                 = EBhDisable;
     extensionBehavior[E_GL_NVX_multiview_per_view_attributes]        = EBhDisable;
+    extensionBehavior[E_GL_NV_shader_atomic_int64]                   = EBhDisable;
 #endif
 
     // AEP
@@ -343,6 +344,7 @@ void TParseVersions::getPreamble(std::string& preamble)
             "#define GL_NV_sample_mask_override_coverage 1\n"
             "#define GL_NV_geometry_shader_passthrough 1\n"
             "#define GL_NV_viewport_array2 1\n"
+            "#define GL_NV_shader_atomic_int64 1\n"
 #endif
             ;
 

glslang/MachineIndependent/Versions.h

@@ -182,6 +182,7 @@ const char* const E_SPV_NV_geometry_shader_passthrough          = "GL_NV_geometr
 const char* const E_GL_NV_viewport_array2                       = "GL_NV_viewport_array2";
 const char* const E_GL_NV_stereo_view_rendering                 = "GL_NV_stereo_view_rendering";
 const char* const E_GL_NVX_multiview_per_view_attributes        = "GL_NVX_multiview_per_view_attributes";
+const char* const E_GL_NV_shader_atomic_int64                   = "GL_NV_shader_atomic_int64";
 
 // Arrays of extensions for the above viewportEXTs duplications
 

gtests/Spv.FromFile.cpp

@@ -410,7 +410,7 @@ INSTANTIATE_TEST_CASE_P(
         "spv.int16.frag",
         "spv.shaderBallotAMD.comp",
         "spv.shaderFragMaskAMD.frag",
-        "spv.textureGatherBiasLod.frag"
+        "spv.textureGatherBiasLod.frag",
     })),
     FileNameAsCustomTestSuffix
 );
@@ -428,6 +428,7 @@ INSTANTIATE_TEST_CASE_P(
     "spv.stereoViewRendering.tesc",
     "spv.multiviewPerViewAttributes.vert",
     "spv.multiviewPerViewAttributes.tesc",
+    "spv.atomicInt64.comp",
 })),
 FileNameAsCustomTestSuffix
 );