Represent FP4 E2M1 with SPV_EXT_ocp_microscaling_types

docs/OCPTypesRepresentationInLLVM.rst

@@ -40,8 +40,8 @@ SPIR-V conversion instructions
 Most conversions will be represented by standard SPIR-V conversion instructions (*OpFConvert*, *OpConvertSToF*, *OpConvertFToS*,
 *OpConvertUToF*, *OpConvertFToU*, *OpSConvert*), which don't carry information about floating-point value's width and encoding.
 This document adds a new set of external function calls, each of which has a name that is formed from encoding a specific conversion
-that it performs. This name has a *__builtin_spirv_* prefix and a postfix indicating the extension (e.g., *EXT* from SPV_EXT_float8,
-*INTEL* from SPV_INTEL_int4/SPV_INTEL_float4/SPV_INTEL_fp_conversions). These calls will be translated to SPIR-V conversion
+that it performs. This name has a *__builtin_spirv_* prefix and a postfix indicating the extension (e.g., *EXT* from SPV_EXT_float8
+and SPV_EXT_ocp_microscaling_types, *INTEL* from SPV_INTEL_int4, SPV_INTEL_float4 and SPV_INTEL_fp_conversions). These calls will be translated to SPIR-V conversion
 instructions operating over the appropriate types. These functions are expected to be mangled following Itanium C++ ABI. SPIR-V consumer
 will apply Itanium mangling during translation to LLVM IR as well.
 
@@ -82,16 +82,19 @@ SPV_INTEL_int4 Conversions
 
   __builtin_spirv_ConvertInt4ToInt8INTEL
 
-SPV_INTEL_float4 Conversions
------------------------------
+SPV_EXT_ocp_microscaling_types Conversions
+-------------------------------------------
 
 **Translated to OpFConvert:**
 
 .. code-block:: C
 
-  __builtin_spirv_ConvertE2M1ToE4M3INTEL, __builtin_spirv_ConvertE2M1ToE5M2INTEL,
-  __builtin_spirv_ConvertE2M1ToFP16INTEL, __builtin_spirv_ConvertE2M1ToBF16INTEL,
-  __builtin_spirv_ConvertFP16ToE2M1INTEL, __builtin_spirv_ConvertBF16ToE2M1INTEL
+  __builtin_spirv_ConvertE2M1ToE4M3EXT, __builtin_spirv_ConvertE2M1ToE5M2EXT,
+  __builtin_spirv_ConvertE2M1ToFP16EXT, __builtin_spirv_ConvertE2M1ToBF16EXT,
+  __builtin_spirv_ConvertFP16ToE2M1EXT, __builtin_spirv_ConvertBF16ToE2M1EXT
+
+For backward compatibility the equivalent INTEL-postfix builtins (e.g. *__builtin_spirv_ConvertE2M1ToFP16INTEL*)
+are also accepted and map to SPV_INTEL_float4; new producers should use the EXT form above.
 
 SPV_INTEL_fp_conversions
 -------------------------
@@ -107,8 +110,9 @@ This extension provides conversions with specialized rounding modes for improved
 
 The result is decorated with *SaturatedToLargestFloat8NormalConversionEXT* (SPV_EXT_float8).
 
-ClampConvert*ToE2M1INTEL builtins are not provided: fp4 (E2M1) saturation is unconditional,
-so they collapse into the plain *Convert\*ToE2M1INTEL* form listed under SPV_INTEL_float4.
+ClampConvert*ToE2M1 builtins are not provided: fp4 (E2M1) saturation is unconditional,
+so they collapse into the plain *Convert\*ToE2M1EXT* / *Convert\*ToE2M1INTEL* form listed under
+SPV_EXT_ocp_microscaling_types / SPV_INTEL_float4.
 
 **Translated to OpClampConvertFToSINTEL (clamp rounding to signed integer):**
 

include/LLVMSPIRVExtensions.inc

@@ -95,6 +95,7 @@ EXT(SPV_EXT_float8)
 EXT(SPV_INTEL_predicated_io)
 EXT(SPV_INTEL_sigmoid)
 EXT(SPV_INTEL_float4)
+EXT(SPV_EXT_ocp_microscaling_types)
 EXT(SPV_INTEL_fp_conversions)
 EXT(SPV_KHR_float_controls2)
 EXT(SPV_INTEL_rounded_divide_sqrt)

lib/SPIRV/SPIRVInternal.h

@@ -1035,7 +1035,8 @@ enum FPEncodingWrap {
   BF16 = FPEncoding::FPEncodingBFloat16KHR,
   E4M3 = FPEncoding::FPEncodingFloat8E4M3EXT,
   E5M2 = FPEncoding::FPEncodingFloat8E5M2EXT,
-  E2M1 = internal::FPEncodingFloat4E2M1INTEL,
+  E2M1 = FPEncoding::FPEncodingFloat4E2M1EXT,
+  E2M1INTEL = internal::FPEncodingFloat4E2M1INTEL,
 };
 
 // Structure describing non-trivial conversions (FP8, FP4 and int4)
@@ -1072,14 +1073,22 @@ typedef SPIRVMap<llvm::StringRef, FPConversionDesc> FPConvertToEncodingMap;
 // clang-format off
 template <> inline void FPConvertToEncodingMap::init() {
   // 4-bit conversions
-  add("ConvertE2M1ToE4M3INTEL",
+  add("ConvertE2M1ToE4M3EXT",
       {FPEncodingWrap::E2M1,      FPEncodingWrap::E4M3,         OpFConvert});
-  add("ConvertE2M1ToE5M2INTEL",
+  add("ConvertE2M1ToE5M2EXT",
       {FPEncodingWrap::E2M1,      FPEncodingWrap::E5M2,         OpFConvert});
-  add("ConvertE2M1ToFP16INTEL",
+  add("ConvertE2M1ToFP16EXT",
       {FPEncodingWrap::E2M1,      FPEncodingWrap::IEEE754,      OpFConvert});
-  add("ConvertE2M1ToBF16INTEL",
+  add("ConvertE2M1ToBF16EXT",
       {FPEncodingWrap::E2M1,      FPEncodingWrap::BF16,         OpFConvert});
+  add("ConvertE2M1ToE4M3INTEL",
+      {FPEncodingWrap::E2M1INTEL, FPEncodingWrap::E4M3,         OpFConvert});
+  add("ConvertE2M1ToE5M2INTEL",
+      {FPEncodingWrap::E2M1INTEL, FPEncodingWrap::E5M2,         OpFConvert});
+  add("ConvertE2M1ToFP16INTEL",
+      {FPEncodingWrap::E2M1INTEL, FPEncodingWrap::IEEE754,      OpFConvert});
+  add("ConvertE2M1ToBF16INTEL",
+      {FPEncodingWrap::E2M1INTEL, FPEncodingWrap::BF16,         OpFConvert});
 
   add("ConvertInt4ToE4M3INTEL",
       {FPEncodingWrap::Integer,      FPEncodingWrap::E4M3,      OpConvertSToF});
@@ -1092,10 +1101,14 @@ template <> inline void FPConvertToEncodingMap::init() {
   add("ConvertInt4ToInt8INTEL",
       {FPEncodingWrap::Integer,      FPEncodingWrap::Integer,   OpSConvert});
 
-  add("ConvertFP16ToE2M1INTEL",
+  add("ConvertFP16ToE2M1EXT",
       {FPEncodingWrap::IEEE754,      FPEncodingWrap::E2M1,      OpFConvert});
-  add("ConvertBF16ToE2M1INTEL",
+  add("ConvertBF16ToE2M1EXT",
       {FPEncodingWrap::BF16,         FPEncodingWrap::E2M1,      OpFConvert});
+  add("ConvertFP16ToE2M1INTEL",
+      {FPEncodingWrap::IEEE754,      FPEncodingWrap::E2M1INTEL, OpFConvert});
+  add("ConvertBF16ToE2M1INTEL",
+      {FPEncodingWrap::BF16,         FPEncodingWrap::E2M1INTEL, OpFConvert});
   add("ConvertFP16ToInt4INTEL",
       {FPEncodingWrap::IEEE754,      FPEncodingWrap::Integer,   OpConvertFToS});
   add("ConvertBF16ToInt4INTEL",

lib/SPIRV/SPIRVReader.cpp

@@ -1042,7 +1042,9 @@ Value *SPIRVToLLVM::transConvertInst(SPIRVValue *BV, Function *F,
 
   auto IsFP4OrFP8Encoding = [](FPEncodingWrap Encoding) -> bool {
     return Encoding == FPEncodingWrap::E4M3 ||
-           Encoding == FPEncodingWrap::E5M2 || Encoding == FPEncodingWrap::E2M1;
+           Encoding == FPEncodingWrap::E5M2 ||
+           Encoding == FPEncodingWrap::E2M1 ||
+           Encoding == FPEncodingWrap::E2M1INTEL;
   };
 
   switch (static_cast<unsigned>(BC->getOpCode())) {
@@ -3145,6 +3147,8 @@ Value *SPIRVToLLVM::transValueWithoutDecoration(SPIRVValue *BV, Function *F,
             OutMatrixElementTy->isTypeFloat(8, FPEncodingFloat8E5M2EXT) ||
             InMatrixElementTy->isTypeFloat(8, FPEncodingFloat8E4M3EXT) ||
             InMatrixElementTy->isTypeFloat(8, FPEncodingFloat8E5M2EXT) ||
+            OutMatrixElementTy->isTypeFloat(4, FPEncodingFloat4E2M1EXT) ||
+            InMatrixElementTy->isTypeFloat(4, FPEncodingFloat4E2M1EXT) ||
             OutMatrixElementTy->isTypeFloat(
                 4, internal::FPEncodingFloat4E2M1INTEL) ||
             InMatrixElementTy->isTypeFloat(4,

lib/SPIRV/SPIRVWriter.cpp

@@ -923,11 +923,14 @@ SPIRVFunction *LLVMToSPIRVBase::transFunctionDecl(Function *F) {
       // generation.
       if (!BM->isAllowedToUseExtension(ExtensionID::SPV_EXT_float8) &&
           !BM->isAllowedToUseExtension(ExtensionID::SPV_INTEL_int4) &&
-          !BM->isAllowedToUseExtension(ExtensionID::SPV_INTEL_float4)) {
-        std::string ErrorStr = "One of the following extensions: "
-                               "SPV_EXT_float8, SPV_INTEL_float4, "
-                               "SPV_INTEL_int4 should be enabled to process "
-                               "conversion builtins";
+          !BM->isAllowedToUseExtension(ExtensionID::SPV_INTEL_float4) &&
+          !BM->isAllowedToUseExtension(
+              ExtensionID::SPV_EXT_ocp_microscaling_types)) {
+        std::string ErrorStr =
+            "One of the following extensions: "
+            "SPV_EXT_float8, SPV_EXT_ocp_microscaling_types, "
+            "SPV_INTEL_float4, SPV_INTEL_int4 should be enabled to process "
+            "conversion builtins";
         getErrorLog().checkError(false, SPIRVEC_RequiresExtension, F, ErrorStr);
       }
       return nullptr;
@@ -5717,8 +5720,9 @@ processMiniFPOrInt4Type(Type *LLVMTy, FPEncodingWrap Encoding,
   unsigned TyWidth = cast<IntegerType>(ScalarTy)->getBitWidth();
   unsigned VecSize = 0;
 
-  bool IsPacked =
-      Encoding == FPEncodingWrap::E2M1 || Encoding == FPEncodingWrap::Integer;
+  bool IsPacked = Encoding == FPEncodingWrap::E2M1 ||
+                  Encoding == FPEncodingWrap::E2M1INTEL ||
+                  Encoding == FPEncodingWrap::Integer;
   if (IsPacked &&
       (TyWidth == 8 || TyWidth == 16 || TyWidth == 32 || TyWidth == 64)) {
     // Int4 or FP4 packed in an integer: each N-bit integer holds N/4 values.

lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h

@@ -705,6 +705,7 @@ template <> inline void SPIRVMap<Capability, std::string>::init() {
   add(internal::CapabilityFloat4E2M1INTEL, "Float4E2M1INTEL");
   add(internal::CapabilityFloat4E2M1CooperativeMatrixINTEL,
       "Float4E2M1CooperativeMatrixINTEL");
+  add(CapabilityFloat4EXT, "Float4EXT");
   add(internal::CapabilityFloatConversionsFtoFINTEL,
       "FloatConversionsFtoFINTEL");
   add(internal::CapabilityFloatConversionsFtoSINTEL,

lib/SPIRV/libSPIRV/SPIRVType.h

@@ -250,6 +250,8 @@ class SPIRVTypeFloat : public SPIRVType {
     if (isTypeFloat(8, FPEncodingFloat8E4M3EXT) ||
         isTypeFloat(8, FPEncodingFloat8E5M2EXT))
       return ExtensionID::SPV_EXT_float8;
+    if (isTypeFloat(4, FPEncodingFloat4E2M1EXT))
+      return ExtensionID::SPV_EXT_ocp_microscaling_types;
     if (isTypeFloat(4, internal::FPEncodingFloat4E2M1INTEL))
       return ExtensionID::SPV_INTEL_float4;
     return {};
@@ -270,6 +272,8 @@ class SPIRVTypeFloat : public SPIRVType {
     } else if (isTypeFloat(8, FPEncodingFloat8E4M3EXT) ||
                isTypeFloat(8, FPEncodingFloat8E5M2EXT)) {
       CV.push_back(CapabilityFloat8EXT);
+    } else if (isTypeFloat(4, FPEncodingFloat4E2M1EXT)) {
+      CV.push_back(CapabilityFloat4EXT);
     } else if (isTypeFloat(4, internal::FPEncodingFloat4E2M1INTEL)) {
       CV.push_back(internal::CapabilityFloat4E2M1INTEL);
     }
@@ -298,14 +302,16 @@ class SPIRVTypeFloat : public SPIRVType {
     assert((BitWidth == 4 || BitWidth == 8 || BitWidth == 16 ||
             BitWidth == 32 || BitWidth == 64) &&
            "Invalid bit width");
-    assert(
-        (FloatingPointEncoding == FPEncodingMax ||
-         (BitWidth == 16 && FloatingPointEncoding == FPEncodingBFloat16KHR) ||
-         (BitWidth == 8 && FloatingPointEncoding == FPEncodingFloat8E4M3EXT) ||
-         (BitWidth == 8 && FloatingPointEncoding == FPEncodingFloat8E5M2EXT) ||
-         (BitWidth == 4 &&
-          FloatingPointEncoding == internal::FPEncodingFloat4E2M1INTEL)) &&
-        "Invalid floating point encoding");
+    bool ValidEncoding =
+        FloatingPointEncoding == FPEncodingMax ||
+        (BitWidth == 16 && FloatingPointEncoding == FPEncodingBFloat16KHR) ||
+        (BitWidth == 8 && FloatingPointEncoding == FPEncodingFloat8E4M3EXT) ||
+        (BitWidth == 8 && FloatingPointEncoding == FPEncodingFloat8E5M2EXT) ||
+        (BitWidth == 4 && FloatingPointEncoding == FPEncodingFloat4E2M1EXT) ||
+        (BitWidth == 4 &&
+         FloatingPointEncoding == internal::FPEncodingFloat4E2M1INTEL);
+    assert(ValidEncoding && "Invalid floating point encoding");
+    (void)ValidEncoding;
   }
 
 private:
@@ -1224,7 +1230,8 @@ class SPIRVTypeCooperativeMatrixKHR : public SPIRVType {
     else if (CompType->isTypeFloat(8, FPEncodingFloat8E4M3EXT) ||
              CompType->isTypeFloat(8, FPEncodingFloat8E5M2EXT))
       CV.push_back(CapabilityFloat8CooperativeMatrixEXT);
-    else if (CompType->isTypeFloat(4, internal::FPEncodingFloat4E2M1INTEL))
+    else if (CompType->isTypeFloat(4, FPEncodingFloat4E2M1EXT) ||
+             CompType->isTypeFloat(4, internal::FPEncodingFloat4E2M1INTEL))
       CV.push_back(internal::CapabilityFloat4E2M1CooperativeMatrixINTEL);
     return CV;
   }

spirv-headers-tag.conf

@@ -1 +1 @@
-5c50cbd25a40f8b60e44a2ccc2f1ba3c9e0d0299
+575b6512579ebde466ed3dfc04e413439d14d95d

test/extensions/INTEL/SPV_INTEL_float4/conversions_intel_encoding.ll

@@ -0,0 +1,61 @@
+; Checks that FP4 (E2M1) conversion builtins with the INTEL postfix are
+; translated using the SPV_INTEL_float4 extension: the Float4E2M1INTEL
+; encoding (6214) and capability, independently of the Float4E2M1EXT encoding
+; from SPV_EXT_ocp_microscaling_types. Round-trips back to the INTEL builtins.
+
+; RUN: llvm-spirv %s -o %t.spv --spirv-ext=+SPV_INTEL_float4,+SPV_INTEL_int4,+SPV_KHR_bfloat16
+; RUN: llvm-spirv %t.spv -o %t.spt --to-text
+; RUN: FileCheck < %t.spt %s --check-prefix=CHECK-SPIRV
+; RUN: llvm-spirv %t.spv -o %t.rev.bc -r --spirv-target-env=SPV-IR
+; RUN: llvm-dis %t.rev.bc -o %t.rev.ll
+; RUN: FileCheck < %t.rev.ll %s --check-prefix=CHECK-LLVM
+
+; CHECK-SPIRV-DAG: Capability Int4TypeINTEL
+; CHECK-SPIRV-DAG: Capability Float4E2M1INTEL
+; CHECK-SPIRV-DAG: Extension "SPV_INTEL_float4"
+; CHECK-SPIRV-DAG: Extension "SPV_INTEL_int4"
+
+; CHECK-SPIRV-DAG: Name [[#fp4e2m1_hf16_scalar:]] "fp4e2m1_hf16_scalar"
+; CHECK-SPIRV-DAG: Name [[#hf16_fp4e2m1_scalar:]] "hf16_fp4e2m1_scalar"
+
+; CHECK-SPIRV-DAG: TypeInt [[#Int4Ty:]] 4 0
+; CHECK-SPIRV-DAG: Constant [[#Int4Ty]] [[#Int4Const:]] 1
+; CHECK-SPIRV-DAG: TypeFloat [[#HFloat16Ty:]] 16 {{$}}
+; CHECK-SPIRV-DAG: TypeFloat [[#E2M1Ty:]] 4 6214
+
+target datalayout = "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
+target triple = "spir-unknown-unknown"
+
+; CHECK-SPIRV: Function [[#]] [[#fp4e2m1_hf16_scalar]] [[#]]
+; CHECK-SPIRV: Bitcast [[#E2M1Ty]] [[#Cast:]] [[#Int4Const]]
+; CHECK-SPIRV: FConvert [[#HFloat16Ty]] [[#Conv:]] [[#Cast]]
+; CHECK-SPIRV: ReturnValue [[#Conv]]
+
+; CHECK-LLVM-LABEL: fp4e2m1_hf16_scalar
+; CHECK-LLVM: %[[#Call:]] = call spir_func half @_Z38__builtin_spirv_ConvertE2M1ToFP16INTELi(i4 1)
+; CHECK-LLVM: ret half %[[#Call]]
+
+define spir_func half @fp4e2m1_hf16_scalar() {
+entry:
+  %0 = call spir_func half @_Z38__builtin_spirv_ConvertE2M1ToFP16INTELi(i4 1)
+  ret half %0
+}
+
+declare dso_local spir_func half @_Z38__builtin_spirv_ConvertE2M1ToFP16INTELi(i4)
+
+; CHECK-SPIRV: Function [[#]] [[#hf16_fp4e2m1_scalar]] [[#]]
+; CHECK-SPIRV: FConvert [[#E2M1Ty]] [[#Conv:]] [[#]]
+; CHECK-SPIRV: Bitcast [[#Int4Ty]] [[#Cast:]] [[#Conv]]
+; CHECK-SPIRV: ReturnValue [[#Cast]]
+
+; CHECK-LLVM-LABEL: hf16_fp4e2m1_scalar
+; CHECK-LLVM: %[[#Call:]] = call spir_func i4 @_Z38__builtin_spirv_ConvertFP16ToE2M1INTELDh(half 1.000000e+00)
+; CHECK-LLVM: ret i4 %[[#Call]]
+
+define spir_func i4 @hf16_fp4e2m1_scalar() {
+entry:
+  %0 = call spir_func i4 @_Z38__builtin_spirv_ConvertFP16ToE2M1INTELDh(half 1.0)
+  ret i4 %0
+}
+
+declare dso_local spir_func i4 @_Z38__builtin_spirv_ConvertFP16ToE2M1INTELDh(half)