| 1 |
// Copyright 2024 Google LLC |
| 2 |
// |
| 3 |
// This source code is licensed under the BSD-style license found in the |
| 4 |
// LICENSE file in the root directory of this source tree. |
| 5 |
|
| 6 |
$assert BATCH_TILE % 16 == 0 |
| 7 |
$assert BATCH_TILE >= 16 |
| 8 |
$SIMD_TILE = BATCH_TILE // 16 |
| 9 |
$assert ACCUMULATORS <= SIMD_TILE |
| 10 |
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" |
| 11 |
#include <assert.h> |
| 12 |
|
| 13 |
#include <immintrin.h> |
| 14 |
|
| 15 |
#include "xnnpack/common.h" |
| 16 |
#include "xnnpack/intrinsics-polyfill.h" |
| 17 |
#include "xnnpack/reduce.h" |
| 18 |
#include "xnnpack/unaligned.h" |
| 19 |
|
| 20 |
|
| 21 |
$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS |
| 22 |
void xnn_f16_f32acc_rsum_ukernel__avx512skx_u${BATCH_TILE}${ACC_SUFFIX}( |
| 23 |
size_t batch, |
| 24 |
const xnn_float16* input, |
| 25 |
float* output, |
| 26 |
const struct xnn_f16_f32acc_scale_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 27 |
{ |
| 28 |
assert(batch != 0); |
| 29 |
assert(batch % sizeof(uint16_t) == 0); |
| 30 |
assert(input != NULL); |
| 31 |
assert(output != NULL); |
| 32 |
|
| 33 |
const uint16_t* i = (const uint16_t*) input; |
| 34 |
$for A in range(ACCUMULATORS): |
| 35 |
__m512 vacc${A} = _mm512_setzero_ps(); |
| 36 |
$if BATCH_TILE > 16: |
| 37 |
for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) { |
| 38 |
const __m512 vt0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i)); |
| 39 |
$for N in range(1, SIMD_TILE): |
| 40 |
const __m512 vt${N} = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + ${N * 16}))); |
| 41 |
i += ${BATCH_TILE}; |
| 42 |
|
| 43 |
$for N in range(SIMD_TILE): |
| 44 |
vacc${N % ACCUMULATORS} = _mm512_add_ps(vacc${N % ACCUMULATORS}, vt${N}); |
| 45 |
} |
| 46 |
$if ACCUMULATORS > 1: |
| 47 |
$ACC_SLICE = 1 |
| 48 |
$while ACC_SLICE < ACCUMULATORS: |
| 49 |
$for A in range(0, ACCUMULATORS, ACC_SLICE * 2): |
| 50 |
$if A + ACC_SLICE < ACCUMULATORS: |
| 51 |
vacc${A} = _mm512_add_ps(vacc${A}, vacc${A + ACC_SLICE}); |
| 52 |
$ACC_SLICE *= 2 |
| 53 |
for (; batch >= 16 * sizeof(uint16_t); batch -= 16 * sizeof(uint16_t)) { |
| 54 |
const __m512 vt = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i)); |
| 55 |
i += 16; |
| 56 |
|
| 57 |
vacc0 = _mm512_add_ps(vacc0, vt); |
| 58 |
} |
| 59 |
if XNN_UNLIKELY(batch != 0) { |
| 60 |
assert(batch >= 1 * sizeof(uint16_t)); |
| 61 |
assert(batch <= 15 * sizeof(uint16_t)); |
| 62 |
|
| 63 |
// Prepare mask for valid elements (depends on batch). |
| 64 |
batch >>= XNN_LOG2_SIZEOF_HALF; |
| 65 |
const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1))); |
| 66 |
|
| 67 |
const __m512 vt = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(vmask, i)); |
| 68 |
|
| 69 |
vacc0 = _mm512_add_ps(vacc0, vt); |
| 70 |
} |
| 71 |
const __m256 vacc256 = _mm256_add_ps(_mm512_castps512_ps256(vacc0), _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(vacc0), 1))); |
| 72 |
__m128 vacc = _mm_add_ps(_mm256_castps256_ps128(vacc256), _mm256_extractf128_ps(vacc256, 1)); |
| 73 |
vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc)); |
| 74 |
vacc = _mm_add_ss(vacc, _mm_movehdup_ps(vacc)); |
| 75 |
vacc = _mm_mul_ss(vacc, _mm_load_ss(¶ms->scalar.scale)); |
| 76 |
|
| 77 |
float vout = _mm_cvtss_f32(vacc); |
| 78 |
*output += vout; |
| 79 |
} |