• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1// Copyright 2021 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$assert DATATYPE in ["QS8", "QU8"]
7$assert BATCH_TILE % 16 == 0
8$assert BATCH_TILE >= 16
9$SIMD_TILE = BATCH_TILE // 4
10$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
11#include <assert.h>
12
13#include <immintrin.h>
14
15#include <xnnpack/common.h>
16#include <xnnpack/intrinsics-polyfill.h>
17#include <xnnpack/vcvt.h>
18
19
20$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
21$_MM512_PACKXS_EPI16 = {"QS8": "_mm512_packs_epi16", "QU8": "_mm512_packus_epi16"}[DATATYPE]
22$_MM256_PACKXS_EPI16 = {"QS8": "_mm256_packs_epi16", "QU8": "_mm256_packus_epi16"}[DATATYPE]
23$_MM_PACKXS_EPI16 = {"QS8": "_mm_packs_epi16", "QU8": "_mm_packus_epi16"}[DATATYPE]
24$_MM512_MAX_EPX8 = {"QS8": "_mm512_max_epi8", "QU8": "_mm512_max_epu8"}[DATATYPE]
25$_MM256_MAX_EPX8 = {"QS8": "_mm256_max_epi8", "QU8": "_mm256_max_epu8"}[DATATYPE]
26$_MM_MAX_EPX8 = {"QS8": "_mm_max_epi8", "QU8": "_mm_max_epu8"}[DATATYPE]
27void xnn_f32_${DATATYPE.lower()}_vcvt_ukernel__avx512skx_x${BATCH_TILE}(
28    size_t n,
29    const float* x,
30    ${XINT8_T}* y,
31    const union xnn_f32_${DATATYPE.lower()}_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
32{
33  assert(n != 0);
34  assert(n % sizeof(float) == 0);
35  assert(x != NULL);
36  assert(y != NULL);
37
38  const __m512 vscale = _mm512_load_ps(params->avx2.scale);
39  const __m512 voutput_max_less_zero_point = _mm512_load_ps(params->avx512.output_max_less_zero_point);
40  const __m512i voutput_zero_point = _mm512_load_si512(params->avx512.output_zero_point);
41  $if SIMD_TILE > 8:
42    const __m512i vshuffle512_mask = _mm512_load_si512(params->avx512.shuffle512_mask);
43  $if SIMD_TILE % 16 != 0:
44    const __m256i vshuffle256_mask = _mm256_load_si256((const __m256i*) params->avx512.shuffle256_mask);
45  $if SIMD_TILE > 8:
46    const __m512i voutput_min = _mm512_load_si512(params->avx512.output_min);
47  $else:
48    const __m256i voutput_min = _mm256_load_si256((const __m256i*) params->avx512.output_min);
49  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
50    __m512 vx0123 = _mm512_loadu_ps(x);
51    $for N in range(4, SIMD_TILE, 4):
52      __m512 vx${ABC[N:N+4]} = _mm512_loadu_ps(x + ${N * 4});
53    x += ${BATCH_TILE};
54
55    $for N in range(0, SIMD_TILE, 4):
56      vx${ABC[N:N+4]} = _mm512_mul_ps(vx${ABC[N:N+4]}, vscale);
57
58    $for N in range(0, SIMD_TILE, 4):
59      vx${ABC[N:N+4]} = _mm512_min_ps(vx${ABC[N:N+4]}, voutput_max_less_zero_point);
60
61    $for N in range(0, SIMD_TILE, 4):
62      const __m512i vacc${ABC[N:N+4]} = _mm512_cvtps_epi32(vx${ABC[N:N+4]});
63
64    $for N in range(0, SIMD_TILE, 8):
65      __m512i vacc${ABC[N]}${ABC[N+4]}${ABC[N+1]}${ABC[N+5]}${ABC[N+2]}${ABC[N+6]}${ABC[N+3]}${ABC[N+7]} = _mm512_packs_epi32(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]});
66
67    $for N in range(0, SIMD_TILE, 8):
68      vacc${ABC[N]}${ABC[N+4]}${ABC[N+1]}${ABC[N+5]}${ABC[N+2]}${ABC[N+6]}${ABC[N+3]}${ABC[N+7]} = _mm512_adds_epi16(vacc${ABC[N]}${ABC[N+4]}${ABC[N+1]}${ABC[N+5]}${ABC[N+2]}${ABC[N+6]}${ABC[N+3]}${ABC[N+7]}, voutput_zero_point);
69
70    $for N in range(0, SIMD_TILE, 16):
71      $if N + 8 < SIMD_TILE:
72        __m512i vy${ABC[N]}${ABC[N+4]}${ABC[N+8]}${ABC[N+12]}${ABC[N+1]}${ABC[N+5]}${ABC[N+9]}${ABC[N+13]}${ABC[N+2]}${ABC[N+6]}${ABC[N+10]}${ABC[N+14]}${ABC[N+3]}${ABC[N+7]}${ABC[N+11]}${ABC[N+15]} = ${_MM512_PACKXS_EPI16}(vacc${ABC[N]}${ABC[N+4]}${ABC[N+1]}${ABC[N+5]}${ABC[N+2]}${ABC[N+6]}${ABC[N+3]}${ABC[N+7]}, vacc${ABC[N+8]}${ABC[N+12]}${ABC[N+9]}${ABC[N+13]}${ABC[N+10]}${ABC[N+14]}${ABC[N+11]}${ABC[N+15]});
73      $else:
74        __m256i vy${ABC[N]}${ABC[N+4]}${ABC[N+2]}${ABC[N+6]}${ABC[N+1]}${ABC[N+5]}${ABC[N+3]}${ABC[N+7]} = ${_MM256_PACKXS_EPI16}(_mm512_castsi512_si256(vacc${ABC[N]}${ABC[N+4]}${ABC[N+1]}${ABC[N+5]}${ABC[N+2]}${ABC[N+6]}${ABC[N+3]}${ABC[N+7]}), _mm512_extracti32x8_epi32(vacc${ABC[N]}${ABC[N+4]}${ABC[N+1]}${ABC[N+5]}${ABC[N+2]}${ABC[N+6]}${ABC[N+3]}${ABC[N+7]}, 1));
75
76    $for N in range(0, SIMD_TILE, 16):
77      $if N + 8 < SIMD_TILE:
78        vy${ABC[N]}${ABC[N+4]}${ABC[N+8]}${ABC[N+12]}${ABC[N+1]}${ABC[N+5]}${ABC[N+9]}${ABC[N+13]}${ABC[N+2]}${ABC[N+6]}${ABC[N+10]}${ABC[N+14]}${ABC[N+3]}${ABC[N+7]}${ABC[N+11]}${ABC[N+15]} = ${_MM512_MAX_EPX8}(vy${ABC[N]}${ABC[N+4]}${ABC[N+8]}${ABC[N+12]}${ABC[N+1]}${ABC[N+5]}${ABC[N+9]}${ABC[N+13]}${ABC[N+2]}${ABC[N+6]}${ABC[N+10]}${ABC[N+14]}${ABC[N+3]}${ABC[N+7]}${ABC[N+11]}${ABC[N+15]}, voutput_min);
79      $elif SIMD_TILE > 8:
80        vy${ABC[N]}${ABC[N+4]}${ABC[N+2]}${ABC[N+6]}${ABC[N+1]}${ABC[N+5]}${ABC[N+3]}${ABC[N+7]} = ${_MM256_MAX_EPX8}(vy${ABC[N]}${ABC[N+4]}${ABC[N+2]}${ABC[N+6]}${ABC[N+1]}${ABC[N+5]}${ABC[N+3]}${ABC[N+7]}, _mm512_castsi512_si256(voutput_min));
81      $else:
82        vy${ABC[N]}${ABC[N+4]}${ABC[N+2]}${ABC[N+6]}${ABC[N+1]}${ABC[N+5]}${ABC[N+3]}${ABC[N+7]} = ${_MM256_MAX_EPX8}(vy${ABC[N]}${ABC[N+4]}${ABC[N+2]}${ABC[N+6]}${ABC[N+1]}${ABC[N+5]}${ABC[N+3]}${ABC[N+7]}, voutput_min);
83
84    $for N in range(0, SIMD_TILE, 16):
85      $if N + 8 < SIMD_TILE:
86        const __m512i vy${ABC[N:N+16]} = _mm512_permutexvar_epi32(vshuffle512_mask, vy${ABC[N]}${ABC[N+4]}${ABC[N+8]}${ABC[N+12]}${ABC[N+1]}${ABC[N+5]}${ABC[N+9]}${ABC[N+13]}${ABC[N+2]}${ABC[N+6]}${ABC[N+10]}${ABC[N+14]}${ABC[N+3]}${ABC[N+7]}${ABC[N+11]}${ABC[N+15]});
87      $else:
88        const __m256i vy${ABC[N:N+8]} = _mm256_permutevar8x32_epi32(vy${ABC[N]}${ABC[N+4]}${ABC[N+2]}${ABC[N+6]}${ABC[N+1]}${ABC[N+5]}${ABC[N+3]}${ABC[N+7]}, vshuffle256_mask);
89
90    $if SIMD_TILE > 8:
91      _mm512_storeu_si512(y, vy${ABC[0:16]});
92    $else:
93      _mm256_storeu_si256((__m256i*) y, vy${ABC[0:8]});
94    $for N in range(16, SIMD_TILE, 16):
95      $if N + 8 < SIMD_TILE:
96        _mm512_storeu_si512(y + ${N * 4}, vy${ABC[N:N+16]});
97      $else:
98        _mm256_storeu_si256((__m256i*) (y + ${N * 4}), vy${ABC[N:N+8]});
99    y += ${BATCH_TILE};
100  }
101  for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
102    __m512 vx0123 = _mm512_loadu_ps(x);
103    vx0123 = _mm512_mul_ps(vx0123, vscale);
104    vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point);
105    x += 16;
106
107    const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123);
108
109    __m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(vacc0123, 1));
110    vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point));
111    const __m128i vy0213 = ${_MM_PACKXS_EPI16}(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213, 1));
112    __m128i vy0123 = _mm_shuffle_epi32(vy0213, _MM_SHUFFLE(3, 1, 2, 0));
113    $if SIMD_TILE > 8:
114      vy0123 = ${_MM_MAX_EPX8}(vy0123, _mm512_castsi512_si128(voutput_min));
115    $else:
116      vy0123 = ${_MM_MAX_EPX8}(vy0123, _mm256_castsi256_si128(voutput_min));
117
118    _mm_storeu_si128((__m128i*) y, vy0123);
119    y += 16;
120  }
121  if XNN_UNLIKELY(n != 0) {
122    assert(n >= 1 * sizeof(float));
123    assert(n <= 15 * sizeof(float));
124
125    // Prepare mask for valid elements (depends on n).
126    n >>= 2 /* log2(sizeof(float)) */;
127    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
128
129    __m512 vx0123 = _mm512_maskz_loadu_ps(vmask, x);
130    vx0123 = _mm512_mul_ps(vx0123, vscale);
131    vx0123 = _mm512_min_ps(vx0123, voutput_max_less_zero_point);
132
133    const __m512i vacc0123 = _mm512_cvtps_epi32(vx0123);
134
135    __m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(vacc0123, 1));
136    vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point));
137    const __m128i vy0213 = ${_MM_PACKXS_EPI16}(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213, 1));
138    __m128i vy0123 = _mm_shuffle_epi32(vy0213, _MM_SHUFFLE(3, 1, 2, 0));
139    $if SIMD_TILE > 8:
140      vy0123 = ${_MM_MAX_EPX8}(vy0123, _mm512_castsi512_si128(voutput_min));
141    $else:
142      vy0123 = ${_MM_MAX_EPX8}(vy0123, _mm256_castsi256_si128(voutput_min));
143
144    _mm_mask_storeu_epi8(y, vmask, vy0123);
145  }
146}
147