deaeb6c5b8
1. Added bfloat16 based dot as new API: shdot
2. Implemented generic kernel and cooperlake-specific (AVX512-BF16) kernel for shdot
3. Added 4 conversion APIs for bfloat16 data type <=> single/double: shstobf16 shdtobf16 sbf16tos dbf16tod
shstobf16 -- convert single float array to bfloat16 array
shdtobf16 -- convert double float array to bfloat16 array
sbf16tos -- convert bfloat16 array to single float array
dbf16tod -- convert bfloat16 array to double float array
4. Implemented generic kernels for all 4 conversion APIs, and cooperlake-specific kernel for shstobf16 and shdtobf16
5. Update level1 thread facilitate functions and macros to support multi-threading for these new APIs
6. Fix Cooperlake platform detection/specify issue when under dynamic-arch building
7. Change the typedef of bfloat16 from unsigned short to more strict uint16_t
Signed-off-by: Chen, Guobing <guobing.chen@intel.com>
160 lines
8.7 KiB
C
160 lines
8.7 KiB
C
/***************************************************************************
|
|
Copyright (c) 2014, The OpenBLAS Project
|
|
All rights reserved.
|
|
Redistribution and use in source and binary forms, with or without
|
|
modification, are permitted provided that the following conditions are
|
|
met:
|
|
1. Redistributions of source code must retain the above copyright
|
|
notice, this list of conditions and the following disclaimer.
|
|
2. Redistributions in binary form must reproduce the above copyright
|
|
notice, this list of conditions and the following disclaimer in
|
|
the documentation and/or other materials provided with the
|
|
distribution.
|
|
3. Neither the name of the OpenBLAS project nor the names of
|
|
its contributors may be used to endorse or promote products
|
|
derived from this software without specific prior written permission.
|
|
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
|
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
|
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
|
|
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
|
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
|
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
|
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
|
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
|
|
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
*****************************************************************************/
|
|
|
|
/* need a new enough GCC for avx512 support */
|
|
#if (( defined(__GNUC__) && __GNUC__ >= 10 && defined(__AVX512BF16__)) || (defined(__clang__) && __clang_major__ >= 9))
|
|
|
|
#define HAVE_SHDOT_ACCL_KERNEL 1
|
|
#include "common.h"
|
|
#include <immintrin.h>
|
|
|
|
static float shdot_accl_kernel(BLASLONG n, bfloat16 *x, bfloat16 *y)
|
|
{
|
|
__m128 accum128 = _mm_setzero_ps();
|
|
if (n> 127) { /* n range from 128 to inf. */
|
|
long tail_index_32 = n&(~31);
|
|
long tail_index_128 = n&(~127);
|
|
unsigned int tail_mask_uint = (((unsigned int)0xffffffff) >> (32-(n&31)));
|
|
__mmask32 tail_mask = *((__mmask32*) &tail_mask_uint);
|
|
|
|
__m512 accum512_0 = _mm512_setzero_ps();
|
|
__m512 accum512_1 = _mm512_setzero_ps();
|
|
__m512 accum512_2 = _mm512_setzero_ps();
|
|
__m512 accum512_3 = _mm512_setzero_ps();
|
|
|
|
/* Processing the main chunk with 128-elements per round */
|
|
for (long i = 0; i < tail_index_128; i += 128) {
|
|
accum512_0 = _mm512_dpbf16_ps(accum512_0, (__m512bh) _mm512_loadu_si512(&x[i+ 0]), (__m512bh) _mm512_loadu_si512(&y[i+ 0]));
|
|
accum512_1 = _mm512_dpbf16_ps(accum512_1, (__m512bh) _mm512_loadu_si512(&x[i+32]), (__m512bh) _mm512_loadu_si512(&y[i+32]));
|
|
accum512_2 = _mm512_dpbf16_ps(accum512_2, (__m512bh) _mm512_loadu_si512(&x[i+64]), (__m512bh) _mm512_loadu_si512(&y[i+64]));
|
|
accum512_3 = _mm512_dpbf16_ps(accum512_3, (__m512bh) _mm512_loadu_si512(&x[i+96]), (__m512bh) _mm512_loadu_si512(&y[i+96]));
|
|
}
|
|
|
|
/* Processing the remaining <128 chunk with 32-elements per round */
|
|
for (long j = tail_index_128; j < tail_index_32; j += 32) {
|
|
accum512_0 = _mm512_dpbf16_ps(accum512_0, (__m512bh) _mm512_loadu_si512(&x[j]), (__m512bh) _mm512_loadu_si512(&y[j]));
|
|
}
|
|
|
|
/* Processing the remaining <32 chunk with masked 32-elements processing */
|
|
if ((n&31) != 0) {
|
|
accum512_2 = _mm512_dpbf16_ps(accum512_2,
|
|
(__m512bh) _mm512_maskz_loadu_epi16(tail_mask, &x[tail_index_32]),
|
|
(__m512bh) _mm512_maskz_loadu_epi16(tail_mask, &y[tail_index_32]));
|
|
}
|
|
|
|
/* Accumulate the 4 registers into 1 register */
|
|
accum512_0 = _mm512_add_ps(accum512_0, accum512_1);
|
|
accum512_2 = _mm512_add_ps(accum512_2, accum512_3);
|
|
accum512_0 = _mm512_add_ps(accum512_0, accum512_2);
|
|
|
|
__m256 accum256 = _mm256_add_ps(_mm512_castps512_ps256(accum512_0), _mm512_extractf32x8_ps(accum512_0, 1));
|
|
accum128 = _mm_add_ps(_mm256_castps256_ps128(accum256), _mm256_extractf128_ps(accum256, 1));
|
|
} else if (n > 31) { /* n range from 32 to 127 */
|
|
/* Processing <128 chunk with 32-elements per round */
|
|
__m256 accum256 = _mm256_setzero_ps();
|
|
__m256 accum256_1 = _mm256_setzero_ps();
|
|
int tail_index_32 = n&(~31);
|
|
for (int j = 0; j < tail_index_32; j += 32) {
|
|
accum256 = _mm256_dpbf16_ps(accum256, (__m256bh) _mm256_loadu_si256(&x[j+ 0]), (__m256bh) _mm256_loadu_si256(&y[j+ 0]));
|
|
accum256_1 = _mm256_dpbf16_ps(accum256_1, (__m256bh) _mm256_loadu_si256(&x[j+16]), (__m256bh) _mm256_loadu_si256(&y[j+16]));
|
|
}
|
|
accum256 = _mm256_add_ps(accum256, accum256_1);
|
|
|
|
/* Processing the remaining <32 chunk with 16-elements processing */
|
|
if ((n&16) != 0) {
|
|
accum256 = _mm256_dpbf16_ps(accum256, (__m256bh) _mm256_loadu_si256(&x[tail_index_32]), (__m256bh) _mm256_loadu_si256(&y[tail_index_32]));
|
|
}
|
|
accum128 = _mm_add_ps(_mm256_castps256_ps128(accum256), _mm256_extractf128_ps(accum256, 1));
|
|
|
|
/* Processing the remaining <16 chunk with 8-elements processing */
|
|
if ((n&8) != 0) {
|
|
int tail_index_16 = n&(~15);
|
|
accum128 = _mm_dpbf16_ps(accum128, (__m128bh) _mm_loadu_si128(&x[tail_index_16]), (__m128bh) _mm_loadu_si128(&y[tail_index_16]));
|
|
}
|
|
|
|
/* Processing the remaining <8 chunk with masked 8-elements processing */
|
|
if ((n&7) != 0) {
|
|
unsigned char tail_mask_uint = (((unsigned char)0xff) >> (8-(n&7)));
|
|
__mmask8 tail_mask = *((__mmask8*) &tail_mask_uint);
|
|
int tail_index_8 = n&(~7);
|
|
accum128 = _mm_dpbf16_ps(accum128,
|
|
(__m128bh) _mm_maskz_loadu_epi16(tail_mask, &x[tail_index_8]),
|
|
(__m128bh) _mm_maskz_loadu_epi16(tail_mask, &y[tail_index_8]));
|
|
}
|
|
} else if (n > 15) { /* n range from 16 to 31 */
|
|
/* Processing <32 chunk with 16-elements processing */
|
|
__m256 accum256 = _mm256_setzero_ps();
|
|
accum256 = _mm256_dpbf16_ps(accum256, (__m256bh) _mm256_loadu_si256(&x[0]), (__m256bh) _mm256_loadu_si256(&y[0]));
|
|
accum128 += _mm_add_ps(_mm256_castps256_ps128(accum256), _mm256_extractf128_ps(accum256, 1));
|
|
|
|
/* Processing the remaining <16 chunk with 8-elements processing */
|
|
if ((n&8) != 0) {
|
|
int tail_index_16 = n&(~15);
|
|
accum128 = _mm_dpbf16_ps(accum128, (__m128bh) _mm_loadu_si128(&x[tail_index_16]), (__m128bh) _mm_loadu_si128(&y[tail_index_16]));
|
|
}
|
|
|
|
/* Processing the remaining <8 chunk with masked 8-elements processing */
|
|
if ((n&7) != 0) {
|
|
unsigned char tail_mask_uint = (((unsigned char)0xff) >> (8-(n&7)));
|
|
__mmask8 tail_mask = *((__mmask8*) &tail_mask_uint);
|
|
int tail_index_8 = n&(~7);
|
|
accum128 = _mm_dpbf16_ps(accum128,
|
|
(__m128bh) _mm_maskz_loadu_epi16(tail_mask, &x[tail_index_8]),
|
|
(__m128bh) _mm_maskz_loadu_epi16(tail_mask, &y[tail_index_8]));
|
|
}
|
|
} else if (n > 7) { /* n range from 8 to 15 */
|
|
/* Processing <16 chunk with 8-elements processing */
|
|
accum128 = _mm_dpbf16_ps(accum128, (__m128bh) _mm_loadu_si128(&x[0]), (__m128bh) _mm_loadu_si128(&y[0]));
|
|
|
|
/* Processing the remaining <8 chunk with masked 8-elements processing */
|
|
if ((n&7) != 0) {
|
|
unsigned char tail_mask_uint = (((unsigned char)0xff) >> (8-(n&7)));
|
|
__mmask8 tail_mask = *((__mmask8*) &tail_mask_uint);
|
|
int tail_index_8 = n&(~7);
|
|
accum128 = _mm_dpbf16_ps(accum128,
|
|
(__m128bh) _mm_maskz_loadu_epi16(tail_mask, &x[tail_index_8]),
|
|
(__m128bh) _mm_maskz_loadu_epi16(tail_mask, &y[tail_index_8]));
|
|
}
|
|
} else { /* n range from 1 to 7 */
|
|
unsigned char tail_mask_uint = (((unsigned char)0xff) >> (8-(n&7)));
|
|
__mmask8 tail_mask = *((__mmask8*) &tail_mask_uint);
|
|
accum128 = _mm_dpbf16_ps(accum128,
|
|
(__m128bh) _mm_maskz_loadu_epi16(tail_mask, &x[0]),
|
|
(__m128bh) _mm_maskz_loadu_epi16(tail_mask, &y[0]));
|
|
}
|
|
|
|
/* Add up the 4 elements into lowest entry */
|
|
__m128 accum128_1 = _mm_shuffle_ps(accum128, accum128, 14);
|
|
accum128 = _mm_add_ps(accum128, accum128_1);
|
|
accum128_1 = _mm_shuffle_ps(accum128, accum128, 1);
|
|
accum128 = _mm_add_ps(accum128, accum128_1);
|
|
|
|
return accum128[0];
|
|
}
|
|
|
|
#endif
|