[Backport to llvm_release_180] Add FP4/FP8 operand support for SubgroupMatrixMultiplyAccumulateINTEL (KhronosGroup#3609) (KhronosGroup#3630)

KanclerzPiotr · vmaksimo · web-flow · commit d53bd97f4917 · 2026-03-09T16:46:05.000+01:00
Extend SubgroupMatrixMultiplyAccumulateINTEL to support packed 4-bit and 8-bit floating-point matrix operands by implementing extensions: - SPV_INTEL_subgroup_matrix_multiply_accumulate_float4 - SPV_INTEL_subgroup_matrix_multiply_accumulate_float8 These extensions add operand flags that interpret packed integer data as FP4/FP8 without requiring actual FP4/FP8 type support added by SPV_INTEL_float4 or SPV_EXT_float8. FP4 operands: `MatrixAPackedFloat4E2M1INTEL` (0x40000) / `MatrixBPackedFloat4E2M1INTEL` (0x80000) FP8 operands: `MatrixAPackedFloat8E4M3INTEL` (0x4000) / `MatrixBPackedFloat8E4M3INTEL` (0x8000) `MatrixAPackedFloat8E5M2INTEL` (0x10000) / `MatrixBPackedFloat8E5M2INTEL` (0x20000) Specs: https://github.com/intel/llvm/blob/sycl/sycl/doc/design/spirv-extensions/SPV_INTEL_subgroup_matrix_multiply_accumulate_float4.asciidoc https://github.com/intel/llvm/blob/sycl/sycl/doc/design/spirv-extensions/SPV_INTEL_subgroup_matrix_multiply_accumulate_float8.asciidoc Co-authored-by: Viktoria Maximova <viktoria.maksimova@intel.com>
diff --git a/include/LLVMSPIRVExtensions.inc b/include/LLVMSPIRVExtensions.inc
@@ -77,6 +77,8 @@ EXT(SPV_INTEL_maximum_registers)
 EXT(SPV_INTEL_bindless_images)
 EXT(SPV_INTEL_2d_block_io)
 EXT(SPV_INTEL_subgroup_matrix_multiply_accumulate)
+EXT(SPV_INTEL_subgroup_matrix_multiply_accumulate_float4)
+EXT(SPV_INTEL_subgroup_matrix_multiply_accumulate_float8)
 EXT(SPV_KHR_bfloat16)
 EXT(SPV_INTEL_bfloat16_arithmetic)
 EXT(SPV_INTEL_ternary_bitwise_function)
diff --git a/lib/SPIRV/libSPIRV/SPIRVInstruction.h b/lib/SPIRV/libSPIRV/SPIRVInstruction.h
@@ -4155,6 +4155,66 @@ class SPIRVSubgroupMatrixMultiplyAccumulateINTELInst
   std::optional<ExtensionID> getRequiredExtension() const override {
     return ExtensionID::SPV_INTEL_subgroup_matrix_multiply_accumulate;
   }
+
+protected:
+  void validate() const override {
+    SPIRVInstTemplateBase::validate();
+
+    // Check if FP4 or FP8 matrix operands are used
+    // Operands parameter is the last operand (index 4)
+    auto *NonConstThis =
+        const_cast<SPIRVSubgroupMatrixMultiplyAccumulateINTELInst *>(this);
+    if (NonConstThis->getOperands().size() > 4) {
+      const SPIRVConstant *OperandsConst =
+          static_cast<const SPIRVConstant *>(NonConstThis->getOperand(4));
+      uint64_t OperandsMask = OperandsConst->getZExtIntValue();
+
+      // FP4 operand bits
+      constexpr uint64_t FP4Mask =
+          spv::internal::
+              IMatrixMultiplyAccumulateOperandsMatrixAPackedFloat4E2M1INTELMask |
+          spv::internal::
+              IMatrixMultiplyAccumulateOperandsMatrixBPackedFloat4E2M1INTELMask;
+
+      // FP8 operand bits
+      constexpr uint64_t FP8Mask =
+          spv::internal::
+              IMatrixMultiplyAccumulateOperandsMatrixAPackedFloat8E4M3INTELMask |
+          spv::internal::
+              IMatrixMultiplyAccumulateOperandsMatrixBPackedFloat8E4M3INTELMask |
+          spv::internal::
+              IMatrixMultiplyAccumulateOperandsMatrixAPackedFloat8E5M2INTELMask |
+          spv::internal::
+              IMatrixMultiplyAccumulateOperandsMatrixBPackedFloat8E5M2INTELMask;
+
+      if ((OperandsMask & FP4Mask) != 0) {
+        getModule()->getErrorLog().checkError(
+            getModule()->isAllowedToUseExtension(
+                ExtensionID::
+                    SPV_INTEL_subgroup_matrix_multiply_accumulate_float4),
+            SPIRVEC_RequiresExtension,
+            "SPV_INTEL_subgroup_matrix_multiply_accumulate_float4\n"
+            "SubgroupMatrixMultiplyAccumulateINTEL with FP4 operand flags "
+            "requires this extension");
+        getModule()->addExtension(
+            ExtensionID::SPV_INTEL_subgroup_matrix_multiply_accumulate_float4);
+      }
+
+      if ((OperandsMask & FP8Mask) != 0) {
+        getModule()->getErrorLog().checkError(
+            getModule()->isAllowedToUseExtension(
+                ExtensionID::
+                    SPV_INTEL_subgroup_matrix_multiply_accumulate_float8),
+            SPIRVEC_RequiresExtension,
+            "SPV_INTEL_subgroup_matrix_multiply_accumulate_float8\n"
+            "SubgroupMatrixMultiplyAccumulateINTEL with FP8 operand flags "
+            "requires this extension");
+        getModule()->addExtension(
+            ExtensionID::SPV_INTEL_subgroup_matrix_multiply_accumulate_float8);
+      }
+    }
+  }
+
   SPIRVCapVec getRequiredCapability() const override {
     return getVec(CapabilitySubgroupMatrixMultiplyAccumulateINTEL);
   }
diff --git a/lib/SPIRV/libSPIRV/spirv_internal.hpp b/lib/SPIRV/libSPIRV/spirv_internal.hpp
@@ -171,6 +171,17 @@ enum InternalBuiltIn {
   IBuiltInDeviceBarrierValidINTEL = 6186,
 };
 
+enum InternalMatrixMultiplyAccumulateOperandsMask {
+  // FP8 matrix operands
+  IMatrixMultiplyAccumulateOperandsMatrixAPackedFloat8E4M3INTELMask = 0x4000,
+  IMatrixMultiplyAccumulateOperandsMatrixBPackedFloat8E4M3INTELMask = 0x8000,
+  IMatrixMultiplyAccumulateOperandsMatrixAPackedFloat8E5M2INTELMask = 0x10000,
+  IMatrixMultiplyAccumulateOperandsMatrixBPackedFloat8E5M2INTELMask = 0x20000,
+  // FP4 matrix operands
+  IMatrixMultiplyAccumulateOperandsMatrixAPackedFloat4E2M1INTELMask = 0x40000,
+  IMatrixMultiplyAccumulateOperandsMatrixBPackedFloat4E2M1INTELMask = 0x80000,
+};
+
 #define _SPIRV_OP(x, y) constexpr x x##y = static_cast<x>(I##x##y);
 _SPIRV_OP(Capability, JointMatrixINTEL)
 _SPIRV_OP(Capability, JointMatrixWIInstructionsINTEL)
diff --git a/test/extensions/INTEL/SPV_INTEL_subgroup_matrix_multiply_accumulate/subgroup_matrix_multiply_accumulate_fp4.ll b/test/extensions/INTEL/SPV_INTEL_subgroup_matrix_multiply_accumulate/subgroup_matrix_multiply_accumulate_fp4.ll
@@ -0,0 +1,53 @@
+; This test checks that SubgroupMatrixMultiplyAccumulateINTEL with FP4 operand flags
+; requires the SPV_INTEL_subgroup_matrix_multiply_accumulate_float4 extension.
+
+; RUN: llvm-as %s -o %t.bc
+; RUN: llvm-spirv %t.bc -o %t.spv --spirv-ext=+SPV_INTEL_subgroup_matrix_multiply_accumulate,+SPV_INTEL_subgroup_matrix_multiply_accumulate_float4
+; RUN: llvm-spirv %t.spv -o %t.spt --to-text
+; RUN: FileCheck < %t.spt %s --check-prefix=CHECK-SPIRV
+
+; RUN: not llvm-spirv %t.bc -o %t.spv --spirv-ext=+SPV_INTEL_subgroup_matrix_multiply_accumulate 2>&1 | FileCheck %s --check-prefix=CHECK-ERROR
+
+; CHECK-ERROR: RequiresExtension: Feature requires the following SPIR-V extension:
+; CHECK-ERROR: SPV_INTEL_subgroup_matrix_multiply_accumulate_float4
+
+; CHECK-SPIRV-DAG: Capability SubgroupMatrixMultiplyAccumulateINTEL
+; CHECK-SPIRV-DAG: Extension "SPV_INTEL_subgroup_matrix_multiply_accumulate"
+; CHECK-SPIRV-DAG: Extension "SPV_INTEL_subgroup_matrix_multiply_accumulate_float4"
+; CHECK-SPIRV-DAG: SubgroupMatrixMultiplyAccumulateINTEL {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} 262144
+; CHECK-SPIRV-DAG: SubgroupMatrixMultiplyAccumulateINTEL {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} 524288
+; CHECK-SPIRV-DAG: SubgroupMatrixMultiplyAccumulateINTEL {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} 786432
+
+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-unknown-unknown"
+
+; Test MatrixAPackedFloat4E2M1INTEL operand (0x40000 = 262144)
+define spir_func <4 x float> @test_fp4_matrix_a(<4 x float> %c, <4 x i8> %a, <8 x i8> %b) {
+entry:
+  %result = call spir_func <4 x float> @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv4_hDv8_hDv4_fi(i32 8, <4 x i8> %a, <8 x i8> %b, <4 x float> %c, i32 262144)
+  ret <4 x float> %result
+}
+
+; Test MatrixBPackedFloat4E2M1INTEL operand (0x80000 = 524288)
+define spir_func <4 x float> @test_fp4_matrix_b(<4 x float> %c, <4 x i8> %a, <8 x i8> %b) {
+entry:
+  %result = call spir_func <4 x float> @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv4_hDv8_hDv4_fi(i32 8, <4 x i8> %a, <8 x i8> %b, <4 x float> %c, i32 524288)
+  ret <4 x float> %result
+}
+
+; Test both FP4 operands (0xC0000 = 786432)
+define spir_func <4 x float> @test_fp4_matrix_both(<4 x float> %c, <4 x i8> %a, <8 x i8> %b) {
+entry:
+  %result = call spir_func <4 x float> @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv4_hDv8_hDv4_fi(i32 8, <4 x i8> %a, <8 x i8> %b, <4 x float> %c, i32 786432)
+  ret <4 x float> %result
+}
+
+declare spir_func <4 x float> @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv4_hDv8_hDv4_fi(i32, <4 x i8>, <8 x i8>, <4 x float>, i32)
+
+!opencl.spir.version = !{!0}
+!spirv.Source = !{!1}
+!llvm.ident = !{!2}
+
+!0 = !{i32 1, i32 0}
+!1 = !{i32 4, i32 100000}
+!2 = !{!"clang version 17.0.0"}
diff --git a/test/extensions/INTEL/SPV_INTEL_subgroup_matrix_multiply_accumulate/subgroup_matrix_multiply_accumulate_fp8.ll b/test/extensions/INTEL/SPV_INTEL_subgroup_matrix_multiply_accumulate/subgroup_matrix_multiply_accumulate_fp8.ll
@@ -0,0 +1,61 @@
+; This test checks that SubgroupMatrixMultiplyAccumulateINTEL with FP8 operand flags
+; requires the SPV_INTEL_subgroup_matrix_multiply_accumulate_float8 extension.
+
+; RUN: llvm-as %s -o %t.bc
+; RUN: llvm-spirv %t.bc -o %t.spv --spirv-ext=+SPV_INTEL_subgroup_matrix_multiply_accumulate,+SPV_INTEL_subgroup_matrix_multiply_accumulate_float8
+; RUN: llvm-spirv %t.spv -o %t.spt --to-text
+; RUN: FileCheck < %t.spt %s --check-prefix=CHECK-SPIRV
+
+; RUN: not llvm-spirv %t.bc -o %t.spv --spirv-ext=+SPV_INTEL_subgroup_matrix_multiply_accumulate 2>&1 | FileCheck %s --check-prefix=CHECK-ERROR
+
+; CHECK-ERROR: RequiresExtension: Feature requires the following SPIR-V extension:
+; CHECK-ERROR: SPV_INTEL_subgroup_matrix_multiply_accumulate_float8
+
+; CHECK-SPIRV-DAG: Capability SubgroupMatrixMultiplyAccumulateINTEL
+; CHECK-SPIRV-DAG: Extension "SPV_INTEL_subgroup_matrix_multiply_accumulate"
+; CHECK-SPIRV-DAG: Extension "SPV_INTEL_subgroup_matrix_multiply_accumulate_float8"
+; CHECK-SPIRV-DAG: SubgroupMatrixMultiplyAccumulateINTEL {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} 16384
+; CHECK-SPIRV-DAG: SubgroupMatrixMultiplyAccumulateINTEL {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} 32768
+; CHECK-SPIRV-DAG: SubgroupMatrixMultiplyAccumulateINTEL {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} 65536
+; CHECK-SPIRV-DAG: SubgroupMatrixMultiplyAccumulateINTEL {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} {{[0-9]+}} 131072
+
+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-unknown-unknown"
+
+; Test MatrixAPackedFloat8E4M3INTEL operand (0x4000 = 16384)
+define spir_func <4 x float> @test_fp8_e4m3_matrix_a(<4 x float> %c, <4 x i8> %a, <8 x i8> %b) {
+entry:
+  %result = call spir_func <4 x float> @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv4_hDv8_hDv4_fi(i32 8, <4 x i8> %a, <8 x i8> %b, <4 x float> %c, i32 16384)
+  ret <4 x float> %result
+}
+
+; Test MatrixBPackedFloat8E4M3INTEL operand (0x8000 = 32768)
+define spir_func <4 x float> @test_fp8_e4m3_matrix_b(<4 x float> %c, <4 x i8> %a, <8 x i8> %b) {
+entry:
+  %result = call spir_func <4 x float> @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv4_hDv8_hDv4_fi(i32 8, <4 x i8> %a, <8 x i8> %b, <4 x float> %c, i32 32768)
+  ret <4 x float> %result
+}
+
+; Test MatrixAPackedFloat8E5M2INTEL operand (0x10000 = 65536)
+define spir_func <4 x float> @test_fp8_e5m2_matrix_a(<4 x float> %c, <4 x i8> %a, <8 x i8> %b) {
+entry:
+  %result = call spir_func <4 x float> @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv4_hDv8_hDv4_fi(i32 8, <4 x i8> %a, <8 x i8> %b, <4 x float> %c, i32 65536)
+  ret <4 x float> %result
+}
+
+; Test MatrixBPackedFloat8E5M2INTEL operand (0x20000 = 131072)
+define spir_func <4 x float> @test_fp8_e5m2_matrix_b(<4 x float> %c, <4 x i8> %a, <8 x i8> %b) {
+entry:
+  %result = call spir_func <4 x float> @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv4_hDv8_hDv4_fi(i32 8, <4 x i8> %a, <8 x i8> %b, <4 x float> %c, i32 131072)
+  ret <4 x float> %result
+}
+
+declare spir_func <4 x float> @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv4_hDv8_hDv4_fi(i32, <4 x i8>, <8 x i8>, <4 x float>, i32)
+
+!opencl.spir.version = !{!0}
+!spirv.Source = !{!1}
+!llvm.ident = !{!2}
+
+!0 = !{i32 1, i32 0}
+!1 = !{i32 4, i32 100000}
+!2 = !{!"clang version 17.0.0"}
