diff options
| author |  Mark D Ryan <mark.d.ryan@intel.com> | 2018-06-16 15:13:06 -0700 |
|---|---|---|
| committer | Mark D Ryan <mark.d.ryan@intel.com> | 2018-06-16 15:13:06 -0700 |
| commit | 90a53ca6fd8a79d39f926430ec07f626c8260a20 (patch) | |
| tree | 7fc53bc0f5f9202ba9d6287102af5632c15d5c60 /Eigen | |
| parent | 1f54164eca5a8960205b94dd3f3cec87089e4ac6 (diff) | |
Fix the Packet16h version of ptranspose
The AVX512 version of ptranpose for PacketBlock<Packet16h,16> was
reordering the PacketBlock argument incorrectly. This lead to errors in
the multiplication of matrices composed of 16 bit floats on AVX512
machines, if at least of the matrices was using RowMajor order. This
error is responsible for one tensorflow unit test failure on AVX512
machines:
//tensorflow/python/kernel_tests:batch_matmul_op_test
Diffstat (limited to 'Eigen')
| -rw-r--r-- | Eigen/src/Core/arch/CUDA/PacketMathHalf.h | 28 |
1 files changed, 14 insertions, 14 deletions
diff --git a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h index ae9193b0d..9897bd4e5 100644 --- a/Eigen/src/Core/arch/CUDA/PacketMathHalf.h +++ b/Eigen/src/Core/arch/CUDA/PacketMathHalf.h @@ -692,20 +692,20 @@ ptranspose(PacketBlock<Packet16h,16>& kernel) { // NOTE: no unpacklo/hi instr in this case, so using permute instr. __m256i a_p_0 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x20); - __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31); - __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20); - __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31); - __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20); - __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31); - __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20); - __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31); - __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20); - __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31); - __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20); - __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31); - __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20); - __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31); - __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20); + __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20); + __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20); + __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20); + __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20); + __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20); + __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20); + __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20); + __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31); + __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31); + __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31); + __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31); + __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31); + __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31); + __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31); __m256i a_p_f = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x31); kernel.packet[0].x = a_p_0; |