[Backport to 15] Support for SPV_INTEL_shader_atomic_bfloat16 extension (KhronosGroup#3343)

Dmitry Sidorov · Sidorov, Dmitry · commit ef4acf6d23fc · 2025-09-22T04:20:28.000-07:00
Spec is available here: intel/llvm#20009 Author: "Ratajewski, Andrzej" <andrzej.ratajewski@intel.com> Signed-off-by: Sidorov, Dmitry <dmitry.sidorov@intel.com>
diff --git a/include/LLVMSPIRVExtensions.inc b/include/LLVMSPIRVExtensions.inc
@@ -78,3 +78,4 @@ EXT(SPV_INTEL_bfloat16_arithmetic)
 EXT(SPV_INTEL_ternary_bitwise_function)
 EXT(SPV_INTEL_int4)
 EXT(SPV_INTEL_function_variants)
+EXT(SPV_INTEL_shader_atomic_bfloat16)
diff --git a/lib/SPIRV/libSPIRV/SPIRVInstruction.h b/lib/SPIRV/libSPIRV/SPIRVInstruction.h
@@ -2862,40 +2862,48 @@ class SPIRVAtomicFAddEXTInst : public SPIRVAtomicInstBase {
 public:
   llvm::Optional<ExtensionID> getRequiredExtension() const override {
     assert(hasType());
+    if (getType()->isTypeFloat(16, FPEncodingBFloat16KHR))
+      return ExtensionID::SPV_INTEL_shader_atomic_bfloat16;
     if (getType()->isTypeFloat(16))
       return ExtensionID::SPV_EXT_shader_atomic_float16_add;
     return ExtensionID::SPV_EXT_shader_atomic_float_add;
   }
 
   SPIRVCapVec getRequiredCapability() const override {
     assert(hasType());
+    if (getType()->isTypeFloat(16, FPEncodingBFloat16KHR))
+      return {internal::CapabilityAtomicBFloat16AddINTEL};
     if (getType()->isTypeFloat(16))
       return {CapabilityAtomicFloat16AddEXT};
     if (getType()->isTypeFloat(32))
       return {CapabilityAtomicFloat32AddEXT};
     if (getType()->isTypeFloat(64))
       return {CapabilityAtomicFloat64AddEXT};
     llvm_unreachable(
-        "AtomicFAddEXT can only be generated for f16, f32, f64 types");
+        "AtomicFAddEXT can only be generated for bf16, f16, f32, f64 types");
   }
 };
 
 class SPIRVAtomicFMinMaxEXTBase : public SPIRVAtomicInstBase {
 public:
   llvm::Optional<ExtensionID> getRequiredExtension() const override {
+    if (getType()->isTypeFloat(16, FPEncodingBFloat16KHR))
+      return ExtensionID::SPV_INTEL_shader_atomic_bfloat16;
     return ExtensionID::SPV_EXT_shader_atomic_float_min_max;
   }
 
   SPIRVCapVec getRequiredCapability() const override {
     assert(hasType());
+    if (getType()->isTypeFloat(16, FPEncodingBFloat16KHR))
+      return {internal::CapabilityAtomicBFloat16MinMaxINTEL};
     if (getType()->isTypeFloat(16))
       return {CapabilityAtomicFloat16MinMaxEXT};
     if (getType()->isTypeFloat(32))
       return {CapabilityAtomicFloat32MinMaxEXT};
     if (getType()->isTypeFloat(64))
       return {CapabilityAtomicFloat64MinMaxEXT};
-    llvm_unreachable(
-        "AtomicF(Min|Max)EXT can only be generated for f16, f32, f64 types");
+    llvm_unreachable("AtomicF(Min|Max)EXT can only be generated for bf16, f16, "
+                     "f32, f64 types");
   }
 };
 
diff --git a/lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h b/lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h
@@ -603,6 +603,9 @@ template <> inline void SPIRVMap<Capability, std::string>::init() {
   add(CapabilityLongCompositesINTEL, "LongCompositesINTEL");
   add(CapabilityOptNoneEXT, "OptNoneEXT");
   add(CapabilityAtomicFloat16AddEXT, "AtomicFloat16AddEXT");
+  add(internal::CapabilityAtomicBFloat16AddINTEL, "AtomicBFloat16AddINTEL");
+  add(internal::CapabilityAtomicBFloat16MinMaxINTEL,
+      "AtomicBFloat16MinMaxINTEL");
   add(CapabilityDebugInfoModuleINTEL, "DebugInfoModuleINTEL");
   add(CapabilitySplitBarrierINTEL, "SplitBarrierINTEL");
   add(CapabilityGlobalVariableFPGADecorationsINTEL,
diff --git a/lib/SPIRV/libSPIRV/spirv_internal.hpp b/lib/SPIRV/libSPIRV/spirv_internal.hpp
@@ -110,6 +110,8 @@ enum InternalCapability {
   ICapGlobalVariableDecorationsINTEL = 6146,
   ICapabilityCooperativeMatrixCheckedInstructionsINTEL = 6192,
   ICapabilityBFloat16ArithmeticINTEL = 6226,
+  ICapabilityAtomicBFloat16AddINTEL = 6255,
+  ICapabilityAtomicBFloat16MinMaxINTEL = 6256,
   ICapabilityCooperativeMatrixPrefetchINTEL = 6411,
   ICapabilityComplexFloatMulDivINTEL = 6414,
   ICapabilityTensorFloat32RoundingINTEL = 6425,
@@ -182,6 +184,9 @@ _SPIRV_OP(Capability, BindlessImagesINTEL)
 _SPIRV_OP(Op, ConvertHandleToImageINTEL)
 _SPIRV_OP(Op, ConvertHandleToSamplerINTEL)
 _SPIRV_OP(Op, ConvertHandleToSampledImageINTEL)
+
+_SPIRV_OP(Capability, AtomicBFloat16AddINTEL)
+_SPIRV_OP(Capability, AtomicBFloat16MinMaxINTEL)
 #undef _SPIRV_OP
 
 constexpr SourceLanguage SourceLanguagePython =
diff --git a/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/AtomicFAddEXT.ll b/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/AtomicFAddEXT.ll
@@ -0,0 +1,30 @@
+; RUN: llvm-as %s -o %t.bc
+; RUN: llvm-spirv %t.bc --spirv-ext=+SPV_INTEL_shader_atomic_bfloat16,+SPV_KHR_bfloat16 -o %t.spv
+; RUN: llvm-spirv -to-text %t.spv -o %t.spt
+; RUN: FileCheck < %t.spt %s --check-prefix=CHECK-SPIRV
+
+; RUN: llvm-spirv --spirv-target-env=SPV-IR -r %t.spv -o %t.rev.bc
+; RUN: llvm-dis %t.rev.bc -o - | FileCheck %s --check-prefixes=CHECK-LLVM-SPV
+
+target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-n8:16:32:64"
+target triple = "spir64-unknown-unknown"
+
+; CHECK-SPIRV-DAG: Capability AtomicBFloat16AddINTEL
+; CHECK-SPIRV-DAG: Capability BFloat16TypeKHR
+; CHECK-SPIRV-DAG: Extension "SPV_INTEL_shader_atomic_bfloat16"
+; CHECK-SPIRV-DAG: Extension "SPV_KHR_bfloat16"
+
+; CHECK-SPIRV: TypeFloat [[BFLOAT:[0-9]+]] 16 0
+
+; Function Attrs: convergent norecurse nounwind
+define dso_local spir_func bfloat @test_AtomicFAddEXT_bfloat(ptr addrspace(4) align 2 dereferenceable(4) %Arg) {
+entry:
+  %0 = addrspacecast ptr addrspace(4) %Arg to ptr addrspace(1)
+  ; CHECK-SPIRV: AtomicFAddEXT [[BFLOAT]]
+  ; CHECK-LLVM-SPV: call spir_func bfloat @_Z21__spirv_AtomicFAddEXTPU3AS1u6__bf16iiu6__bf16({{.*}}bfloat
+  %ret = tail call spir_func bfloat @_Z21__spirv_AtomicFAddEXTPU3AS1u6__bf16iiu6__bf16(ptr addrspace(1) %0, i32 1, i32 896, bfloat 1.000000e+00)
+  ret bfloat %ret
+}
+
+; Function Attrs: convergent
+declare dso_local spir_func bfloat @_Z21__spirv_AtomicFAddEXTPU3AS1u6__bf16iiu6__bf16(ptr addrspace(1), i32, i32, bfloat)
diff --git a/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/AtomicFMaxEXT.ll b/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/AtomicFMaxEXT.ll
@@ -0,0 +1,30 @@
+; RUN: llvm-as %s -o %t.bc
+; RUN: llvm-spirv %t.bc --spirv-ext=+SPV_INTEL_shader_atomic_bfloat16,+SPV_KHR_bfloat16 -o %t.spv
+; RUN: llvm-spirv -to-text %t.spv -o %t.spt
+; RUN: FileCheck < %t.spt %s --check-prefix=CHECK-SPIRV
+
+; RUN: llvm-spirv --spirv-target-env=SPV-IR -r %t.spv -o %t.rev.bc
+; RUN: llvm-dis %t.rev.bc -o - | FileCheck %s --check-prefixes=CHECK-LLVM-SPV
+
+target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-n8:16:32:64"
+target triple = "spir64-unknown-unknown"
+
+; CHECK-SPIRV-DAG: Capability AtomicBFloat16MinMaxINTEL
+; CHECK-SPIRV-DAG: Capability BFloat16TypeKHR
+; CHECK-SPIRV-DAG: Extension "SPV_INTEL_shader_atomic_bfloat16"
+; CHECK-SPIRV-DAG: Extension "SPV_KHR_bfloat16"
+
+; CHECK-SPIRV: TypeFloat [[BFLOAT:[0-9]+]] 16 0
+
+; Function Attrs: convergent norecurse nounwind
+define dso_local spir_func bfloat @test_AtomicFMaxEXT_bfloat(ptr addrspace(4) align 2 dereferenceable(4) %Arg) {
+entry:
+  %0 = addrspacecast ptr addrspace(4) %Arg to ptr addrspace(1)
+  ; CHECK-SPIRV: AtomicFMaxEXT [[BFLOAT]]
+  ; CHECK-LLVM-SPV: call spir_func bfloat @_Z21__spirv_AtomicFMaxEXTPU3AS1u6__bf16iiu6__bf16({{.*}}bfloat
+  %ret = tail call spir_func bfloat @_Z21__spirv_AtomicFMaxEXTPU3AS1u6__bf16iiu6__bf16(ptr addrspace(1) %0, i32 1, i32 896, bfloat 1.000000e+00)
+  ret bfloat %ret
+}
+
+; Function Attrs: convergent
+declare dso_local spir_func bfloat @_Z21__spirv_AtomicFMaxEXTPU3AS1u6__bf16iiu6__bf16(ptr addrspace(1), i32, i32, bfloat)
diff --git a/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/AtomicFMinEXT.ll b/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/AtomicFMinEXT.ll
@@ -0,0 +1,30 @@
+; RUN: llvm-as %s -o %t.bc
+; RUN: llvm-spirv %t.bc --spirv-ext=+SPV_INTEL_shader_atomic_bfloat16,+SPV_KHR_bfloat16 -o %t.spv
+; RUN: llvm-spirv -to-text %t.spv -o %t.spt
+; RUN: FileCheck < %t.spt %s --check-prefix=CHECK-SPIRV
+
+; RUN: llvm-spirv --spirv-target-env=SPV-IR -r %t.spv -o %t.rev.bc
+; RUN: llvm-dis %t.rev.bc -o - | FileCheck %s --check-prefixes=CHECK-LLVM-SPV
+
+target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-n8:16:32:64"
+target triple = "spir64-unknown-unknown"
+
+; CHECK-SPIRV-DAG: Capability AtomicBFloat16MinMaxINTEL
+; CHECK-SPIRV-DAG: Capability BFloat16TypeKHR
+; CHECK-SPIRV-DAG: Extension "SPV_INTEL_shader_atomic_bfloat16"
+; CHECK-SPIRV-DAG: Extension "SPV_KHR_bfloat16"
+
+; CHECK-SPIRV: TypeFloat [[BFLOAT:[0-9]+]] 16 0
+
+; Function Attrs: convergent norecurse nounwind
+define dso_local spir_func bfloat @test_AtomicFMinEXT_bfloat(ptr addrspace(4) align 2 dereferenceable(4) %Arg) {
+entry:
+  %0 = addrspacecast ptr addrspace(4) %Arg to ptr addrspace(1)
+  ; CHECK-SPIRV: AtomicFMinEXT [[BFLOAT]]
+  ; CHECK-LLVM-SPV: call spir_func bfloat @_Z21__spirv_AtomicFMinEXTPU3AS1u6__bf16iiu6__bf16({{.*}}bfloat
+  %ret = tail call spir_func bfloat @_Z21__spirv_AtomicFMinEXTPU3AS1u6__bf16iiu6__bf16(ptr addrspace(1) %0, i32 1, i32 896, bfloat 1.000000e+00)
+  ret bfloat %ret
+}
+
+; Function Attrs: convergent
+declare dso_local spir_func bfloat @_Z21__spirv_AtomicFMinEXTPU3AS1u6__bf16iiu6__bf16(ptr addrspace(1), i32, i32, bfloat)
diff --git a/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/atomicrmw_fadd.ll b/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/atomicrmw_fadd.ll
@@ -0,0 +1,34 @@
+; RUN: llvm-as < %s -o %t.bc
+; RUN: llvm-spirv --spirv-ext=+SPV_INTEL_shader_atomic_bfloat16,+SPV_KHR_bfloat16 %t.bc -o %t.spv
+; RUN: llvm-spirv -to-text %t.spv -o - | FileCheck %s
+
+; CHECK-DAG: Extension "SPV_INTEL_shader_atomic_bfloat16"
+; CHECK-DAG: Extension "SPV_KHR_bfloat16"
+; CHECK-DAG: Capability AtomicBFloat16AddINTEL
+; CHECK-DAG: Capability BFloat16TypeKHR
+; CHECK: TypeInt [[Int:[0-9]+]] 32 0
+; CHECK-DAG: Constant [[Int]] [[Device:[0-9]+]] 1 {{$}}
+; CHECK-DAG: Constant [[Int]] [[MemSem_SequentiallyConsistent:[0-9]+]] 16
+; CHECK: TypeFloat [[BFloat:[0-9]+]] 16 0
+; CHECK: Variable {{[0-9]+}} [[BFloatPointer:[0-9]+]]
+; CHECK: Constant [[BFloat]] [[BFloatValue:[0-9]+]] 16936
+
+target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
+target triple = "spir64"
+
+@f = common dso_local local_unnamed_addr addrspace(1) global bfloat 0.000000e+00, align 8
+
+; Function Attrs: nounwind
+define dso_local spir_func void @test_atomicrmw_fadd() local_unnamed_addr #0 {
+entry:
+ %0 = atomicrmw fadd ptr addrspace(1) @f, bfloat 42.000000e+00 seq_cst
+; CHECK: AtomicFAddEXT [[BFloat]] {{[0-9]+}} [[BFloatPointer]] [[Device]] [[MemSem_SequentiallyConsistent]] [[BFloatValue]]
+
+  ret void
+}
+
+attributes #0 = { nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="all" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
+
+!llvm.module.flags = !{!0}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
diff --git a/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/atomicrmw_fminfmax.ll b/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/atomicrmw_fminfmax.ll
@@ -0,0 +1,36 @@
+; RUN: llvm-as < %s -o %t.bc
+; RUN: llvm-spirv --spirv-ext=+SPV_INTEL_shader_atomic_bfloat16,+SPV_KHR_bfloat16 %t.bc -o %t.spv
+; RUN: llvm-spirv -to-text %t.spv -o - | FileCheck %s
+
+; CHECK-DAG: Extension "SPV_INTEL_shader_atomic_bfloat16"
+; CHECK-DAG: Extension "SPV_KHR_bfloat16"
+; CHECK-DAG: AtomicBFloat16MinMaxINTEL
+; CHECK-DAG: Capability BFloat16TypeKHR
+; CHECK: TypeInt [[Int:[0-9]+]] 32 0
+; CHECK-DAG: Constant [[Int]] [[Device:[0-9]+]] 1 {{$}}
+; CHECK-DAG: Constant [[Int]] [[MemSem_SequentiallyConsistent:[0-9]+]] 16
+; CHECK: TypeFloat [[BFloat:[0-9]+]] 16 0
+; CHECK: Variable {{[0-9]+}} [[BFloatPointer:[0-9]+]]
+; CHECK: Constant [[BFloat]] [[BFloatValue:[0-9]+]] 16936
+
+target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
+target triple = "spir64"
+
+@f = common dso_local local_unnamed_addr addrspace(1) global bfloat 0.000000e+00, align 4
+
+; Function Attrs: nounwind
+define dso_local spir_func void @test_atomicrmw_fadd() local_unnamed_addr #0 {
+entry:
+ %0 = atomicrmw fmin ptr addrspace(1) @f, bfloat 42.000000e+00 seq_cst
+; CHECK: AtomicFMinEXT [[BFloat]] {{[0-9]+}} [[BFloatPointer]] [[Device]] [[MemSem_SequentiallyConsistent]] [[BFloatValue]]
+ %1 = atomicrmw fmax ptr addrspace(1) @f, bfloat 42.000000e+00 seq_cst
+; CHECK: AtomicFMaxEXT [[BFloat]] {{[0-9]+}} [[BFloatPointer]] [[Device]] [[MemSem_SequentiallyConsistent]] [[BFloatValue]]
+
+  ret void
+}
+
+attributes #0 = { nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="all" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
+
+!llvm.module.flags = !{!0}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
diff --git a/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/negative.ll b/test/extensions/INTEL/SPV_INTEL_shader_atomic_bfloat16/negative.ll
@@ -0,0 +1,29 @@
+; RUN: llvm-as < %s -o %t.bc
+; RUN: not llvm-spirv --spirv-ext=+SPV_INTEL_shader_atomic_bfloat16 %t.bc 2>&1 | FileCheck %s --check-prefix=CHECK-NO-BF
+; RUN: not llvm-spirv --spirv-ext=+SPV_KHR_bfloat16 %t.bc 2>&1 | FileCheck %s --check-prefix=CHECK-NO-ATOM
+
+; CHECK-NO-BF: RequiresExtension: Feature requires the following SPIR-V extension:
+; CHECK-NO-BF-NEXT: SPV_KHR_bfloat16
+; CHECK-NO-BF-NEXT: NOTE: LLVM module contains bfloat type, translation of which requires this extension
+
+; CHECK-NO-ATOM: RequiresExtension: Feature requires the following SPIR-V extension:
+; CHECK-NO-ATOM-NEXT: SPV_INTEL_shader_atomic_bfloat16
+
+target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
+target triple = "spir64"
+
+@f = common dso_local local_unnamed_addr addrspace(1) global bfloat 0.000000e+00, align 8
+
+; Function Attrs: nounwind
+define dso_local spir_func void @test_atomicrmw_fadd() local_unnamed_addr #0 {
+entry:
+ %0 = atomicrmw fadd ptr addrspace(1) @f, bfloat 42.000000e+00 seq_cst
+
+  ret void
+}
+
+attributes #0 = { nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="all" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
+
+!llvm.module.flags = !{!0}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
