File size: 3,039 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 |
// Copyright 2023 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
$assert ACCUMULATORS <= SIMD_TILE
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <immintrin.h>
#include <xnnpack/common.h>
#include <xnnpack/reduce.h>
#include <xnnpack/unaligned.h>
$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS
void xnn_f16_f32acc_rsum_ukernel__f16c_x${BATCH_TILE}${ACC_SUFFIX}(
size_t batch,
const void* input,
void* output,
const union xnn_f16_f32acc_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
{
assert(batch != 0);
assert(batch % sizeof(uint16_t) == 0);
assert(input != NULL);
assert(output != NULL);
const uint16_t* i = (const uint16_t*) input;
$for A in range(ACCUMULATORS):
__m256 vacc${A} = _mm256_setzero_ps();
$if BATCH_TILE > 8:
for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
const __m256 vt0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
$for N in range(1, SIMD_TILE):
const __m256 vt${N} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + ${N * 8})));
i += ${BATCH_TILE};
$for N in range(SIMD_TILE):
vacc${N % ACCUMULATORS} = _mm256_add_ps(vacc${N % ACCUMULATORS}, vt${N});
}
$if ACCUMULATORS > 1:
$ACC_SLICE = 1
$while ACC_SLICE < ACCUMULATORS:
$for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
$if A + ACC_SLICE < ACCUMULATORS:
vacc${A} = _mm256_add_ps(vacc${A}, vacc${A + ACC_SLICE});
$ACC_SLICE *= 2
for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
const __m256 vt = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
i += 8;
vacc0 = _mm256_add_ps(vacc0, vt);
}
if XNN_UNLIKELY(batch != 0) {
assert(batch >= 1 * sizeof(uint16_t));
assert(batch <= 7 * sizeof(uint16_t));
const __m128i vmask = _mm_loadu_si128((const __m128i*) ((uintptr_t) ¶ms->avx.mask_table[7] - batch));
const __m128i vh = _mm_castps_si128(_mm_maskload_ps((const float*) i, vmask));
const __m256 vt = _mm256_cvtph_ps(vh);
vacc0 = _mm256_add_ps(vacc0, vt);
i = (const void*) ((uintptr_t) i + batch);
if (batch & (1 * sizeof(uint16_t))) {
const __m128i vh = _mm_insert_epi16(_mm_setzero_si128(), (int) unaligned_load_u16(i - 1), 0);
const __m256 vt = _mm256_zextps128_ps256(_mm_cvtph_ps(vh));
vacc0 = _mm256_add_ps(vacc0, vt);
}
}
__m128 vacc = _mm_add_ps(_mm256_castps256_ps128(vacc0), _mm256_extractf128_ps(vacc0, 1));
vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc));
vacc = _mm_add_ss(vacc, _mm_movehdup_ps(vacc));
vacc = _mm_mul_ss(vacc, _mm_load_ss(¶ms->avx.scale));
const __m128i vout = _mm_cvtps_ph(vacc, _MM_FROUND_TO_NEAREST_INT);
unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout, 0));
}
|