File size: 1,980 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 |
// Copyright 2021 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 % 16 == 0
$assert BATCH_TILE >= 16
$SIMD_TILE = BATCH_TILE // 16
#include <assert.h>
#include <immintrin.h>
#include <xnnpack/common.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/vcvt.h>
void xnn_f16_f32_vcvt_ukernel__avx512skx_x${BATCH_TILE}(
size_t batch,
const void* input,
float* output,
const union xnn_f16_f32_cvt_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;
$if BATCH_TILE > 16:
for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
const __m512 vacc0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i));
$for N in range(1, SIMD_TILE):
const __m512 vacc${N} = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + ${N * 16})));
i += ${BATCH_TILE};
_mm512_storeu_ps(output, vacc0);
$for N in range(1, SIMD_TILE):
_mm512_storeu_ps(output + ${N * 16}, vacc${N});
output += ${BATCH_TILE};
}
for (; batch >= 16 * sizeof(uint16_t); batch -= 16 * sizeof(uint16_t)) {
const __m512 vacc = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i));
i += 16;
_mm512_storeu_ps(output, vacc);
output += 16;
}
if XNN_UNLIKELY(batch != 0) {
assert(batch >= 1 * sizeof(uint16_t));
assert(batch <= 15 * sizeof(uint16_t));
// Prepare mask for valid 32-bit elements (depends on batch).
batch >>= XNN_LOG2_SIZEOF_HALF;
const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << batch) - UINT32_C(1)));
const __m512 vacc = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(vmask, i));
_mm512_mask_storeu_ps(output, vmask, vacc);
}
}
|