Atomics+Wave ops intrinsics fixes. (#3542)

* Fix atomics intrinsics, increase kMaxDescriptorSets.

* Add SPIRVASM to known non-differentiable insts.

* Support fp16 wave ops when targeting glsl.

* Fixes.

* Fix vk validation errors.

* Fix.

* Add to allowed failures.

12 files changed, 157 insertions, 32 deletions

diff --git a/source/slang/core.meta.slang b/source/slang/core.meta.slang
index 3df11fffe..3ebb77f03 100644
--- a/source/slang/core.meta.slang
+++ b/source/slang/core.meta.slang
@@ -217,6 +217,9 @@ attribute_syntax [Differentiable(order:int = 0)] : BackwardDifferentiableAttribu
 __intrinsic_op($(kIROp_RequirePrelude))
 void __requirePrelude(constexpr String preludeText);
 
+__intrinsic_op($(kIROp_RequireGLSLExtension))
+void __requireGLSLExtension(constexpr String preludeText);
+
 /// Interface to denote types as differentiable.
 /// Allows for user-specified differential types as
 /// well as automatic generation, for when the associated type
@@ -2255,6 +2258,10 @@ __intrinsic_op($(kIROp_IsFloat))
 bool __isFloat_impl(T t);
 
 __generic<T>
+__intrinsic_op($(kIROp_IsHalf))
+bool __isHalf_impl(T t);
+
+__generic<T>
 [__unsafeForceInlineEarly]
 bool __isFloat()
 {
@@ -2262,6 +2269,13 @@ bool __isFloat()
 }
 
 __generic<T>
+[__unsafeForceInlineEarly]
+bool __isHalf()
+{
+    return __isHalf_impl(__declVal<T>());
+}
+
+__generic<T>
 __intrinsic_op($(kIROp_IsUnsignedInt))
 bool __isUnsignedInt_impl(T t);
 
diff --git a/source/slang/hlsl.meta.slang b/source/slang/hlsl.meta.slang
index 2900d6ea0..0b60bda0d 100644
--- a/source/slang/hlsl.meta.slang
+++ b/source/slang/hlsl.meta.slang
@@ -1778,8 +1778,8 @@ float __atomicAdd(__ref float value, float amount)
 }
 
 __glsl_version(430)
-__glsl_extension(GL_EXT_shader_atomic_float2)
-half __atomicAdd(__ref half value, half amount)
+__glsl_extension(GL_NV_shader_atomic_fp16_vector)
+half2 __atomicAdd(__ref half2 value, half2 amount)
 {
     __target_switch
     {
@@ -1787,9 +1787,9 @@ half __atomicAdd(__ref half value, half amount)
     case spirv:
         return spirv_asm
         {
-            OpExtension "SPV_EXT_shader_atomic_float16_add";
-            OpCapability AtomicFloat16AddEXT;
-            result:$$half = OpAtomicFAddEXT &value Device None $amount
+            OpExtension "SPV_EXT_shader_atomic_float_add";
+            OpCapability AtomicFloat32AddEXT;
+            result:$$half2 = OpAtomicFAddEXT &value Device None $amount
         };
     }
 }
@@ -2337,7 +2337,7 @@ ${{{{
         __target_switch
         {
         case hlsl:
-            __intrinsic_asm "NvInterlockedAddFp32($0, $1, $2))";
+            __intrinsic_asm "NvInterlockedAddFp16x2($0, $1, $2))";
         }
     }
 
@@ -2364,8 +2364,15 @@ ${{{{
         case glsl:
         case spirv:
             {
-                let buf = __getEquivalentStructuredBuffer<half>(this);
-                originalValue = __atomicAdd(buf[byteAddress / 2], value);
+                let buf = __getEquivalentStructuredBuffer<half2>(this);
+                if ((byteAddress & 2) == 0)
+                {
+                    originalValue = __atomicAdd(buf[byteAddress/4], half2(value, half(0.0))).x;
+                }
+                else
+                {
+                    originalValue = __atomicAdd(buf[byteAddress/4], half2(half(0.0), value)).y;
+                }
                 return;
             }
         }
@@ -7555,6 +7562,9 @@ __target_intrinsic(cuda, "_waveProductMultiple($0, $1)")
 __target_intrinsic(hlsl, "WaveActiveProduct($1)")
 matrix<T,N,M> WaveMaskProduct(WaveMask mask, matrix<T,N,M> expr);
 
+__intrinsic_op($(kIROp_RequireGLSLExtension))
+void __requireGLSLExtension(String extensionName);
+
 __generic<T : __BuiltinArithmeticType>
 __glsl_extension(GL_KHR_shader_subgroup_arithmetic)
 __spirv_version(1.3)
@@ -7562,7 +7572,9 @@ T WaveMaskSum(WaveMask mask, T expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupAdd($1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupAdd($1)";
     case cuda: __intrinsic_asm "_waveSum($0, $1)";
     case hlsl: __intrinsic_asm "WaveActiveSum($1)";
     case spirv:
@@ -7591,7 +7603,9 @@ vector<T,N> WaveMaskSum(WaveMask mask, vector<T,N> expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupAdd($1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupAdd($1)";
     case cuda: __intrinsic_asm "_waveSumMultiple($0, $1)";
     case hlsl: __intrinsic_asm "WaveActiveSum($1)";
     case spirv:
@@ -7627,6 +7641,7 @@ bool WaveMaskAllEqual(WaveMask mask, T value)
     __target_switch
     {
     case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
         __intrinsic_asm "subgroupAllEqual($1)";
     case hlsl:
         __intrinsic_asm "WaveActiveAllEqual($1)";
@@ -7651,6 +7666,7 @@ bool WaveMaskAllEqual(WaveMask mask, vector<T,N> value)
     __target_switch
     {
     case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
         __intrinsic_asm "subgroupAllEqual($1)";
     case hlsl:
         __intrinsic_asm "WaveActiveAllEqual($1)";
@@ -7681,7 +7697,9 @@ T WaveMaskPrefixProduct(WaveMask mask, T expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupExclusiveMul($1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupExclusiveMul($1)";
     case cuda: __intrinsic_asm "_wavePrefixProduct($0, $1)";
     case hlsl: __intrinsic_asm "WavePrefixProduct($1)";
     case spirv:
@@ -7710,7 +7728,9 @@ vector<T,N> WaveMaskPrefixProduct(WaveMask mask, vector<T,N> expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupExclusiveMul($1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupExclusiveMul($1)";
     case cuda: __intrinsic_asm "_wavePrefixProductMultiple($0, $1)";
     case hlsl: __intrinsic_asm "WavePrefixProduct($1)";
     case spirv:
@@ -7744,7 +7764,9 @@ T WaveMaskPrefixSum(WaveMask mask, T expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupExclusiveAdd($1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupExclusiveAdd($1)";
     case cuda: __intrinsic_asm "_wavePrefixSum($0, $1)";
     case hlsl: __intrinsic_asm "WavePrefixSum($1)";
     case spirv:
@@ -7774,7 +7796,9 @@ vector<T,N> WaveMaskPrefixSum(WaveMask mask, vector<T,N> expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupExclusiveAdd($1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupExclusiveAdd($1)";
     case cuda: __intrinsic_asm "_wavePrefixSumMultiple($0, $1)";
     case hlsl: __intrinsic_asm "WavePrefixSum($1)";
     case spirv:
@@ -8281,7 +8305,9 @@ T WaveActive$(opName.hlslName)(T expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroup$(opName.glslName)($0)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroup$(opName.glslName)($0)";
     case hlsl: __intrinsic_asm "WaveActive$(opName.hlslName)";
     case spirv:
         if (__isFloat<T>())
@@ -8320,7 +8346,9 @@ vector<T,N> WaveActive$(opName.hlslName)(vector<T,N> expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroup$(opName.glslName)($0)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroup$(opName.glslName)($0)";
     case hlsl: __intrinsic_asm "WaveActive$(opName.hlslName)";
     case spirv:
         if (__isFloat<T>())
@@ -8574,7 +8602,9 @@ T WavePrefixProduct(T expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupExclusiveMul($0)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupExclusiveMul($0)";
     case hlsl: __intrinsic_asm "WavePrefixProduct";
     case spirv:
         if (__isFloat<T>())
@@ -8609,7 +8639,9 @@ vector<T,N> WavePrefixProduct(vector<T,N> expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupExclusiveMul($0)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupExclusiveMul($0)";
     case hlsl: __intrinsic_asm "WavePrefixProduct";
     case spirv:
         if (__isFloat<T>())
@@ -8647,7 +8679,9 @@ T WavePrefixSum(T expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupExclusiveAdd($0)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupExclusiveAdd($0)";
     case hlsl: __intrinsic_asm "WavePrefixSum";
     case spirv:
         if (__isFloat<T>())
@@ -8678,7 +8712,9 @@ vector<T,N> WavePrefixSum(vector<T,N> expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupExclusiveAdd($0)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupExclusiveAdd($0)";
     case hlsl: __intrinsic_asm "WavePrefixSum";
     case spirv:
         if (__isFloat<T>())
@@ -8716,7 +8752,9 @@ T WaveReadLaneFirst(T expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupBroadcastFirst($0)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupBroadcastFirst($0)";
     case hlsl: __intrinsic_asm "WaveReadLaneFirst";
     case spirv:
         return spirv_asm {OpCapability GroupNonUniformBallot; OpGroupNonUniformBroadcastFirst $$T result Subgroup $expr};
@@ -8732,7 +8770,9 @@ vector<T,N> WaveReadLaneFirst(vector<T,N> expr)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupBroadcastFirst($0)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupBroadcastFirst($0)";
     case hlsl: __intrinsic_asm "WaveReadLaneFirst";
     case spirv:
         return spirv_asm {OpCapability GroupNonUniformBallot; OpGroupNonUniformBroadcastFirst $$vector<T,N> result Subgroup $expr};
@@ -8761,7 +8801,9 @@ T WaveBroadcastLaneAt(T value, constexpr int lane)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupBroadcast($0, $1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupBroadcast($0, $1)";
     case hlsl: __intrinsic_asm "WaveReadLaneAt";
     case spirv:
         let ulane = uint(lane);
@@ -8778,7 +8820,9 @@ vector<T,N> WaveBroadcastLaneAt(vector<T,N> value, constexpr int lane)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupBroadcast($0, $1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupBroadcast($0, $1)";
     case hlsl: __intrinsic_asm "WaveReadLaneAt";
     case spirv:
         let ulane = uint(lane);
@@ -8805,7 +8849,9 @@ T WaveReadLaneAt(T value, int lane)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupShuffle($0, $1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupShuffle($0, $1)";
     case hlsl: __intrinsic_asm "WaveReadLaneAt";
     case spirv:
         let ulane = uint(lane);
@@ -8822,7 +8868,9 @@ vector<T,N> WaveReadLaneAt(vector<T,N> value, int lane)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupShuffle($0, $1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupShuffle($0, $1)";
     case hlsl: __intrinsic_asm "WaveReadLaneAt";
     case spirv:
         let ulane = uint(lane);
@@ -8850,7 +8898,9 @@ T WaveShuffle(T value, int lane)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupShuffle($0, $1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupShuffle($0, $1)";
     case hlsl: __intrinsic_asm "WaveReadLaneAt";
     case spirv:
         let ulane = uint(lane);
@@ -8867,7 +8917,9 @@ vector<T,N> WaveShuffle(vector<T,N> value, int lane)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupShuffle($0, $1)";
+    case glsl:
+        if (__isHalf<T>()) __requireGLSLExtension("GL_EXT_shader_subgroup_extended_types_float16");
+        __intrinsic_asm "subgroupShuffle($0, $1)";
     case hlsl: __intrinsic_asm "WaveReadLaneAt";
     case spirv:
         let ulane = uint(lane);
@@ -8890,7 +8942,8 @@ uint WavePrefixCountBits(bool value)
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupBallotExclusiveBitCount(subgroupBallot($0))";
+    case glsl:
+        __intrinsic_asm "subgroupBallotExclusiveBitCount(subgroupBallot($0))";
     case hlsl: __intrinsic_asm "WavePrefixCountBits($0)";
     case spirv:
         return spirv_asm
@@ -8910,7 +8963,8 @@ uint4 WaveGetConvergedMulti()
 {
     __target_switch
     {
-    case glsl: __intrinsic_asm "subgroupBallot(true)";
+    case glsl:
+        __intrinsic_asm "subgroupBallot(true)";
     case hlsl: __intrinsic_asm "WaveActiveBallot(true)";
     case cuda: __intrinsic_asm "make_uint4(__activemask(), 0, 0, 0)";
     case spirv:
diff --git a/source/slang/slang-compiler.cpp b/source/slang/slang-compiler.cpp
index 491655643..fbdb91ac7 100644
--- a/source/slang/slang-compiler.cpp
+++ b/source/slang/slang-compiler.cpp
@@ -2329,6 +2329,17 @@ namespace Slang
         return false;
     }
 
+    bool CodeGenContext::shouldSkipSPIRVValidation()
+    {
+        if (auto endToEndReq = isEndToEndCompile())
+        {
+            if (endToEndReq->m_skipSPIRVValidation)
+                return true;
+        }
+
+        return false;
+    }
+
     bool CodeGenContext::shouldDumpIR()
     {
         if (getTargetReq()->getTargetFlags() & SLANG_TARGET_FLAG_DUMP_IR)
diff --git a/source/slang/slang-compiler.h b/source/slang/slang-compiler.h
index a95b5a6bc..18a91f78a 100755
--- a/source/slang/slang-compiler.h
+++ b/source/slang/slang-compiler.h
@@ -2562,6 +2562,8 @@ namespace Slang
 
         bool isSpecializationDisabled();
 
+        bool shouldSkipSPIRVValidation();
+
         SlangResult requireTranslationUnitSourceFiles();
 
         //
@@ -2707,7 +2709,8 @@ namespace Slang
         virtual SLANG_NO_THROW void SLANG_MCALL setDebugInfoFormat(SlangDebugInfoFormat format) SLANG_OVERRIDE;
         virtual SLANG_NO_THROW void SLANG_MCALL setReportDownstreamTime(bool value) SLANG_OVERRIDE;
         virtual SLANG_NO_THROW void SLANG_MCALL setReportPerfBenchmark(bool value) SLANG_OVERRIDE;
-        
+        virtual SLANG_NO_THROW void SLANG_MCALL setSkipSPIRVValidation(bool value) SLANG_OVERRIDE;
+
         void setHLSLToVulkanLayoutOptions(int targetIndex, HLSLToVulkanLayoutOptions* vulkanLayoutOptions);
 
         EndToEndCompileRequest(
@@ -2759,6 +2762,8 @@ namespace Slang
 
         // If set, will print out compiler performance benchmark results.
         bool m_reportPerfBenchmark = false;
+
+        bool m_skipSPIRVValidation = false;
         
         String m_diagnosticOutput;
 
diff --git a/source/slang/slang-emit-glsl.cpp b/source/slang/slang-emit-glsl.cpp
index 3d3f9687e..a0151133d 100644
--- a/source/slang/slang-emit-glsl.cpp
+++ b/source/slang/slang-emit-glsl.cpp
@@ -2114,6 +2114,21 @@ void GLSLSourceEmitter::handleRequiredCapabilitiesImpl(IRInst* inst)
 
         }
     }
+
+    // The function may have IRRequireGLSLExtensionInst in its body. We also need to look for them.
+    auto func = as<IRFunc>(inst);
+    if (!func)
+        return;
+    auto block = func->getFirstBlock();
+    if (!block)
+        return;
+    for (auto childInst : block->getChildren())
+    {
+        if (auto requireGLSLExt = as<IRRequireGLSLExtension>(childInst))
+        {
+            _requireGLSLExtension(requireGLSLExt->getExtensionName());
+        }
+    }
 }
 
 static Index _getGLSLVersion(ProfileVersion profile)
diff --git a/source/slang/slang-emit-spirv.cpp b/source/slang/slang-emit-spirv.cpp
index 4e11a7d86..192a40f54 100644
--- a/source/slang/slang-emit-spirv.cpp
+++ b/source/slang/slang-emit-spirv.cpp
@@ -5142,7 +5142,8 @@ SlangResult emitSPIRVFromIR(
 
     StringBuilder runSpirvValEnvVar;
     PlatformUtil::getEnvironmentVariable(UnownedStringSlice("SLANG_RUN_SPIRV_VALIDATION"), runSpirvValEnvVar);
-    if (runSpirvValEnvVar.getUnownedSlice() == "1")
+    if (runSpirvValEnvVar.getUnownedSlice() == "1"
+        && !codeGenContext->shouldSkipSPIRVValidation())
     {
         const auto validationResult = debugValidateSPIRV(spirvOut);
         // If validation isn't available, don't say it failed, it's just a debug
diff --git a/source/slang/slang-ir-autodiff-fwd.cpp b/source/slang/slang-ir-autodiff-fwd.cpp
index 9d13dafae..632f99480 100644
--- a/source/slang/slang-ir-autodiff-fwd.cpp
+++ b/source/slang/slang-ir-autodiff-fwd.cpp
@@ -1906,6 +1906,7 @@ InstPair ForwardDiffTranscriber::transcribeInstImpl(IRBuilder* builder, IRInst*
     case kIROp_DetachDerivative:
     case kIROp_GetSequentialID:
     case kIROp_GetStringHash:
+    case kIROp_SPIRVAsm:
         return transcribeNonDiffInst(builder, origInst);
 
         // A call to createDynamicObject<T>(arbitraryData) cannot provide a diff value,
diff --git a/source/slang/slang-ir-inst-defs.h b/source/slang/slang-ir-inst-defs.h
index ff0b815d4..23fcb3a92 100644
--- a/source/slang/slang-ir-inst-defs.h
+++ b/source/slang/slang-ir-inst-defs.h
@@ -572,6 +572,7 @@ INST_RANGE(Unreachable, MissingReturn, Unreachable)
 INST_RANGE(TerminatorInst, Return, Unreachable)
 
 INST(RequirePrelude, RequirePrelude, 1, 0)
+INST(RequireGLSLExtension, RequireGLSLExtension, 1, 0)
 
 // TODO: We should consider splitting the basic arithmetic/comparison
 // ops into cases for signed integers, unsigned integers, and floating-point
@@ -1014,6 +1015,7 @@ INST(TypeEquals, TypeEquals, 2, 0)
 INST(IsInt, IsInt, 1, 0)
 INST(IsBool, IsBool, 1, 0)
 INST(IsFloat, IsFloat, 1, 0)
+INST(IsHalf, IsHalf, 1, 0)
 INST(IsUnsignedInt, IsUnsignedInt, 1, 0)
 INST(IsSignedInt, IsSignedInt, 1, 0)
 INST(IsVector, IsVector, 1, 0)
diff --git a/source/slang/slang-ir-insts.h b/source/slang/slang-ir-insts.h
index c8309e3df..6e3821c18 100644
--- a/source/slang/slang-ir-insts.h
+++ b/source/slang/slang-ir-insts.h
@@ -3080,6 +3080,12 @@ struct IRRequirePrelude : IRInst
     UnownedStringSlice getPrelude() { return as<IRStringLit>(getOperand(0))->getStringSlice(); }
 };
 
+struct IRRequireGLSLExtension : IRInst
+{
+    IR_LEAF_ISA(RequireGLSLExtension)
+    UnownedStringSlice getExtensionName() { return as<IRStringLit>(getOperand(0))->getStringSlice(); }
+};
+
 struct IRBuilderSourceLocRAII;
 
 struct IRBuilder
diff --git a/source/slang/slang-ir-peephole.cpp b/source/slang/slang-ir-peephole.cpp
index 22bcafdf4..fb67c6842 100644
--- a/source/slang/slang-ir-peephole.cpp
+++ b/source/slang/slang-ir-peephole.cpp
@@ -983,6 +983,7 @@ struct PeepholeContext : InstPassBase
         }
         case kIROp_IsInt:
         case kIROp_IsFloat:
+        case kIROp_IsHalf:
         case kIROp_IsUnsignedInt:
         case kIROp_IsSignedInt:
         case kIROp_IsBool:
@@ -1009,6 +1010,9 @@ struct PeepholeContext : InstPassBase
                     case kIROp_IsFloat:
                         result = isFloatingType(type);
                         break;
+                    case kIROp_IsHalf:
+                        result = type->getOp() == kIROp_HalfType;
+                        break;
                     case kIROp_IsUnsignedInt:
                         result = isIntegralType(type) && !getIntTypeInfo(type).isSigned;
                         break;
diff --git a/source/slang/slang-options.cpp b/source/slang/slang-options.cpp
index ae6195700..3edc7dfb5 100644
--- a/source/slang/slang-options.cpp
+++ b/source/slang/slang-options.cpp
@@ -74,6 +74,7 @@ enum class OptionKind
     EmitIr,
     ReportDownstreamTime,
     ReportPerfBenchmark,
+    SkipSPIRVValidation,
 
     SourceEmbedStyle,
     SourceEmbedName,
@@ -446,6 +447,7 @@ void initCommandOptions(CommandOptions& options)
         { OptionKind::InputFilesRemain, "--", nullptr, "Treat the rest of the command line as input files."},
         { OptionKind::ReportDownstreamTime, "-report-downstream-time", nullptr, "Reports the time spent in the downstream compiler." },
         { OptionKind::ReportPerfBenchmark, "-report-perf-benchmark", nullptr, "Reports compiler performance benchmark results." },
+        { OptionKind::SkipSPIRVValidation, "-skip-spirv-validation", nullptr, "Skips spirv validation." },
         { OptionKind::SourceEmbedStyle, "-source-embed-style", "-source-embed-style <source-embed-style>",
         "If source embedding is enabled, defines the style used. When enabled (with any style other than `none`), "
         "will write compile results into embeddable source for the target language. "
@@ -1930,6 +1932,11 @@ SlangResult OptionsParser::_parse(
                 m_compileRequest->setReportPerfBenchmark(true);
                 break;
             }
+            case OptionKind::SkipSPIRVValidation:
+            {
+                m_compileRequest->setSkipSPIRVValidation(true);
+                break;
+            }
             case OptionKind::ModuleName:
             {
                 CommandLineArg moduleName;
diff --git a/source/slang/slang.cpp b/source/slang/slang.cpp
index 4ce78ee96..b95b21bb5 100644
--- a/source/slang/slang.cpp
+++ b/source/slang/slang.cpp
@@ -5143,6 +5143,11 @@ void EndToEndCompileRequest::setReportPerfBenchmark(bool value)
     m_reportPerfBenchmark = value;
 }
 
+void EndToEndCompileRequest::setSkipSPIRVValidation(bool value)
+{
+    m_skipSPIRVValidation = value;
+}
+
 void EndToEndCompileRequest::setDiagnosticCallback(SlangDiagnosticCallback callback, void const* userData)
 {
     ComPtr<ISlangWriter> writer(new CallbackWriter(callback, userData, WriterFlag::IsConsole));