FreddyYe created this revision.
Herald added a subscriber: pengfei.
FreddyYe requested review of this revision.
Herald added a project: clang.
Herald added a subscriber: cfe-commits.
Repository:
rG LLVM Github Monorepo
https://reviews.llvm.org/D114059
Files:
clang/lib/Headers/avx512vlbf16intrin.h
clang/test/CodeGen/X86/avx512vlbf16-builtins.c
Index: clang/test/CodeGen/X86/avx512vlbf16-builtins.c
===================================================================
--- clang/test/CodeGen/X86/avx512vlbf16-builtins.c
+++ clang/test/CodeGen/X86/avx512vlbf16-builtins.c
@@ -169,6 +169,15 @@
return _mm_cvtness_sbh(A);
}
+__m128 test_mm_cvtpbh_ps(__m128bh A) {
+ // CHECK-LABEL: @test_mm_cvtpbh_ps
+ // CHECK: sext <4 x i16> %{{.*}} to <4 x i32>
+ // CHECK: @llvm.x86.sse2.pslli.d
+ // CHECK: bitcast <2 x i64> %{{.*}} to <4 x float>
+ // CHECK: ret <4 x float> %{{.*}}
+ return _mm_cvtpbh_ps(A);
+}
+
__m256 test_mm256_cvtpbh_ps(__m128bh A) {
// CHECK-LABEL: @test_mm256_cvtpbh_ps
// CHECK: sext <8 x i16> %{{.*}} to <8 x i32>
@@ -178,6 +187,16 @@
return _mm256_cvtpbh_ps(A);
}
+__m128 test_mm_maskz_cvtpbh_ps(__mmask8 M, __m128bh A) {
+ // CHECK-LABEL: @test_mm_maskz_cvtpbh_ps
+ // CHECK: sext <4 x i16> %{{.*}} to <4 x i32>
+ // CHECK: select <4 x i1> %{{.*}}, <4 x i32> %{{.*}}, <4 x i32> %{{.*}}
+ // CHECK: @llvm.x86.sse2.pslli.d
+ // CHECK: bitcast <2 x i64> %{{.*}} to <4 x float>
+ // CHECK: ret <4 x float> %{{.*}}
+ return _mm_maskz_cvtpbh_ps(M, A);
+}
+
__m256 test_mm256_maskz_cvtpbh_ps(__mmask8 M, __m128bh A) {
// CHECK-LABEL: @test_mm256_maskz_cvtpbh_ps
// CHECK: sext <8 x i16> %{{.*}} to <8 x i32>
@@ -188,6 +207,16 @@
return _mm256_maskz_cvtpbh_ps(M, A);
}
+__m128 test_mm_mask_cvtpbh_ps(__m128 S, __mmask8 M, __m128bh A) {
+ // CHECK-LABEL: @test_mm_mask_cvtpbh_ps
+ // CHECK: sext <4 x i16> %{{.*}} to <4 x i32>
+ // CHECK: @llvm.x86.sse2.pslli.d
+ // CHECK: select <4 x i1> %{{.*}}, <4 x i32> %{{.*}}, <4 x i32> %{{.*}}
+ // CHECK: bitcast <2 x i64> %{{.*}} to <4 x float>
+ // CHECK: ret <4 x float> %{{.*}}
+ return _mm_mask_cvtpbh_ps(S, M, A);
+}
+
__m256 test_mm256_mask_cvtpbh_ps(__m256 S, __mmask8 M, __m128bh A) {
// CHECK-LABEL: @test_mm256_mask_cvtpbh_ps
// CHECK: sext <8 x i16> %{{.*}} to <8 x i32>
Index: clang/lib/Headers/avx512vlbf16intrin.h
===================================================================
--- clang/lib/Headers/avx512vlbf16intrin.h
+++ clang/lib/Headers/avx512vlbf16intrin.h
@@ -420,6 +420,18 @@
return __R[0];
}
+/// Convert Packed BF16 Data to Packed float Data.
+///
+/// \headerfile <x86intrin.h>
+///
+/// \param __A
+/// A 128-bit vector of [4 x bfloat].
+/// \returns A 128-bit vector of [4 x float] come from convertion of __A
+static __inline__ __m128 __DEFAULT_FN_ATTRS128 _mm_cvtpbh_ps(__m128bh __A) {
+ return _mm_castsi128_ps(
+ (__m128i)_mm_slli_epi32((__m128i)_mm_cvtepi16_epi32((__m128i)__A), 16));
+}
+
/// Convert Packed BF16 Data to Packed float Data.
///
/// \headerfile <x86intrin.h>
@@ -432,6 +444,22 @@
(__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16));
}
+/// Convert Packed BF16 Data to Packed float Data using zeroing mask.
+///
+/// \headerfile <x86intrin.h>
+///
+/// \param __U
+/// A 4-bit mask. Elements are zeroed out when the corresponding mask
+/// bit is not set.
+/// \param __A
+/// A 128-bit vector of [4 x bfloat].
+/// \returns A 128-bit vector of [4 x float] come from convertion of __A
+static __inline__ __m128 __DEFAULT_FN_ATTRS128
+_mm_maskz_cvtpbh_ps(__mmask8 __U, __m128bh __A) {
+ return _mm_castsi128_ps((__m128i)_mm_slli_epi32(
+ (__m128i)_mm_maskz_cvtepi16_epi32((__mmask8)__U, (__m128i)__A), 16));
+}
+
/// Convert Packed BF16 Data to Packed float Data using zeroing mask.
///
/// \headerfile <x86intrin.h>
@@ -448,6 +476,26 @@
(__m256i)_mm256_maskz_cvtepi16_epi32((__mmask8)__U, (__m128i)__A), 16));
}
+/// Convert Packed BF16 Data to Packed float Data using merging mask.
+///
+/// \headerfile <x86intrin.h>
+///
+/// \param __S
+/// A 128-bit vector of [4 x float]. Elements are copied from __S when
+/// the corresponding mask bit is not set.
+/// \param __U
+/// A 4-bit mask. Elements are zeroed out when the corresponding mask
+/// bit is not set.
+/// \param __A
+/// A 128-bit vector of [4 x bfloat].
+/// \returns A 128-bit vector of [4 x float] come from convertion of __A
+static __inline__ __m128 __DEFAULT_FN_ATTRS128
+_mm_mask_cvtpbh_ps(__m128 __S, __mmask8 __U, __m128bh __A) {
+ return _mm_castsi128_ps((__m128i)_mm_mask_slli_epi32(
+ (__m128i)__S, (__mmask8)__U, (__m128i)_mm_cvtepi16_epi32((__m128i)__A),
+ 16));
+}
+
/// Convert Packed BF16 Data to Packed float Data using merging mask.
///
/// \headerfile <x86intrin.h>
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
