File size: 6,069 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 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 |
// Copyright 2022 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 CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert ROW_TILE >= 3
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <immintrin.h>
#include <xnnpack/gavgpool.h>
#include <xnnpack/intrinsics-polyfill.h>
void xnn_f16_gavgpool_minmax_ukernel_${ROW_TILE}x__f16c_c${CHANNEL_TILE}(
size_t rows,
size_t channels,
const void* input,
size_t input_stride,
const void* zero,
void* output,
const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(rows != 0);
assert(rows <= ${ROW_TILE});
assert(channels != 0);
const uint16_t* i0 = input;
$for M in range(1, ROW_TILE):
const uint16_t* i${M} = (const uint16_t*) ((uintptr_t) i${M-1} + input_stride);
$if M % 2 == 1:
if XNN_UNPREDICTABLE(rows < ${M+1}) {
i${M} = (const uint16_t*) zero;
}
$else:
if XNN_UNPREDICTABLE(rows <= ${M}) {
i${M} = (const uint16_t*) zero;
}
uint16_t* o = (uint16_t*) output;
const __m256 vscale = _mm256_load_ps(params->avx.scale);
const __m256 vmin = _mm256_load_ps(params->avx.min);
const __m256 vmax = _mm256_load_ps(params->avx.max);
for (; channels >= ${CHANNEL_TILE}; channels -= ${CHANNEL_TILE}) {
$for M in range(2):
const __m256 vi${M}x${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i${M}));
$for C in range(8, CHANNEL_TILE, 8):
const __m256 vi${M}x${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i${M} + ${C})));
i${M} += ${CHANNEL_TILE};
$for C in range(0, CHANNEL_TILE, 8):
$if C == 0:
const __m256 vi2x${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
$else:
const __m256 vi2x${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + ${C})));
__m128i vacc${ABC[C:C+8]} = _mm256_cvtps_ph(_mm256_add_ps(vi0x${ABC[C:C+8]}, vi1x${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT);
i2 += ${CHANNEL_TILE};
$for M in range(2, ROW_TILE):
$for C in range(0, CHANNEL_TILE, 8):
$if M + 1 != ROW_TILE:
$if C == 0:
const __m256 vi${M+1}x${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i${M+1}));
$else:
const __m256 vi${M+1}x${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i${M+1} + ${C})));
$if C + 8 == CHANNEL_TILE:
i${M+1} += ${CHANNEL_TILE};
vacc${ABC[C:C+8]} = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc${ABC[C:C+8]}), vi${M}x${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT);
$for C in range(0, CHANNEL_TILE, 8):
vacc${ABC[C:C+8]} = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc${ABC[C:C+8]}), vscale), _MM_FROUND_TO_NEAREST_INT);
$for C in range(0, CHANNEL_TILE, 8):
__m256 vout${ABC[C:C+8]} = _mm256_max_ps(_mm256_cvtph_ps(vacc${ABC[C:C+8]}), vmin);
$for C in range(0, CHANNEL_TILE, 8):
vout${ABC[C:C+8]} = _mm256_min_ps(vout${ABC[C:C+8]}, vmax);
_mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vout${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT));
$for C in range(8, CHANNEL_TILE, 8):
_mm_storeu_si128((__m128i*) (o + ${C}), _mm256_cvtps_ph(vout${ABC[C:C+8]}, _MM_FROUND_TO_NEAREST_INT));
o += ${CHANNEL_TILE};
}
if XNN_UNLIKELY(channels != 0) {
${"do " if CHANNEL_TILE > 8 else ""}{
$for M in range(2):
const __m256 vi${M}x${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i${M}));
$if CHANNEL_TILE > 8:
i${M} += 8;
const __m256 vi2x${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
__m128i vacc${ABC[0:8]} = _mm256_cvtps_ph(_mm256_add_ps(vi0x${ABC[0:8]}, vi1x${ABC[0:8]}), _MM_FROUND_TO_NEAREST_INT);
$if CHANNEL_TILE > 8:
i2 += 8;
$for M in range(2, ROW_TILE):
$if M + 1 != ROW_TILE:
const __m256 vi${M+1}x${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i${M+1}));
$if CHANNEL_TILE > 8:
i${M+1} += 8;
vacc${ABC[0:8]} = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vacc${ABC[0:8]}), vi${M}x${ABC[0:8]}), _MM_FROUND_TO_NEAREST_INT);
vacc${ABC[0:8]} = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc${ABC[0:8]}), vscale), _MM_FROUND_TO_NEAREST_INT);
__m256 vout${ABC[0:8]} = _mm256_max_ps(_mm256_cvtph_ps(vacc${ABC[0:8]}), vmin);
vout${ABC[0:8]} = _mm256_min_ps(vout${ABC[0:8]}, vmax);
$if CHANNEL_TILE > 8:
if XNN_LIKELY(channels >= 8) {
_mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vout${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT));
o += 8;
channels -= 8;
} else {
__m128i vh${ABC[0:8]} = _mm256_cvtps_ph(vout${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT);
if (channels & 4) {
_mm_storel_epi64((__m128i*) o, vh${ABC[0:8]});
o += 4;
vh${ABC[0:8]} = _mm_unpackhi_epi64(vh${ABC[0:8]}, vh${ABC[0:8]});
}
if (channels & 2) {
_mm_storeu_si32(o, vh${ABC[0:8]});
o += 2;
vh${ABC[0:8]} = _mm_srli_epi64(vh${ABC[0:8]}, 32);
}
if (channels & 1) {
*o = (uint16_t) _mm_extract_epi16(vh${ABC[0:8]}, 0);
}
channels = 0;
}
$else:
__m128i vh${ABC[0:8]} = _mm256_cvtps_ph(vout${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT);
if (channels & 4) {
_mm_storel_epi64((__m128i*) o, vh${ABC[0:8]});
o += 4;
vh${ABC[0:8]} = _mm_unpackhi_epi64(vh${ABC[0:8]}, vh${ABC[0:8]});
}
if (channels & 2) {
_mm_storeu_si32(o, vh${ABC[0:8]});
o += 2;
vh${ABC[0:8]} = _mm_srli_epi64(vh${ABC[0:8]}, 32);
}
if (channels & 1) {
*o = (uint16_t) _mm_extract_epi16(vh${ABC[0:8]}, 0);
}
}${" while (channels != 0);" if CHANNEL_TILE > 8 else ""}
}
}
|