xref: /aosp_15_r20/external/XNNPACK/src/qs8-vaddc/avx2-mul32-ld64.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1// Copyright 2020 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 % 8 == 0
8$assert BATCH_TILE >= 8
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11
12#include <immintrin.h>
13
14#include <xnnpack/intrinsics-polyfill.h>
15#include <xnnpack/vadd.h>
16
17
18$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
19$_MM256_CVTEPX8_EPI32 = {"QS8": "_mm256_cvtepi8_epi32", "QU8": "_mm256_cvtepu8_epi32"}[DATATYPE]
20$_MM_PACKXS_EPI16 = {"QS8": "_mm_packs_epi16", "QU8": "_mm_packus_epi16"}[DATATYPE]
21$_MM_MIN_EPX8 = {"QS8": "_mm_min_epi8", "QU8": "_mm_min_epu8"}[DATATYPE]
22$_MM_MAX_EPX8 = {"QS8": "_mm_max_epi8", "QU8": "_mm_max_epu8"}[DATATYPE]
23void xnn_${DATATYPE.lower()}_vaddc_minmax_ukernel__avx2_mul32_ld64_x${BATCH_TILE}(
24    size_t n,
25    const ${XINT8_T}* input_a,
26    const ${XINT8_T}* input_b,
27    ${XINT8_T}* output,
28    const union xnn_${DATATYPE.lower()}_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29{
30  const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier);
31  const __m128i vshift = _mm_load_si128((const __m128i*) params->avx2.shift);
32  $if BATCH_TILE > 8:
33    const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
34  $else:
35    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx2.output_zero_point);
36  const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
37  const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max);
38
39  const __m256i vbias = _mm256_add_epi32(
40    _mm256_broadcastd_epi32(_mm_cvtsi32_si128(params->avx2.b_multiplier[0] * (int32_t) *input_b)),
41    _mm256_load_si256((const __m256i*) params->avx2.bias));
42  for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
43    const __m256i va${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_a));
44    $for N in range(8, BATCH_TILE, 8):
45      const __m256i va${ABC[N:N+8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) (input_a + ${N})));
46    input_a += ${BATCH_TILE};
47
48    $for N in range(0, BATCH_TILE, 8):
49      __m256i vacc${ABC[N:N+8]} = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va${ABC[N:N+8]}, va_multiplier));
50
51    $for N in range(0, BATCH_TILE, 8):
52      vacc${ABC[N:N+8]} = _mm256_sra_epi32(vacc${ABC[N:N+8]}, vshift);
53
54    $for N in range(0, BATCH_TILE, 16):
55      $if N + 8 < BATCH_TILE:
56        __m256i vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]} = _mm256_adds_epi16(_mm256_packs_epi32(vacc${ABC[N:N+8]}, vacc${ABC[N+8:N+16]}), voutput_zero_point);
57      $elif BATCH_TILE > 8:
58        __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]}), _mm256_extracti128_si256(vacc${ABC[N:N+8]}, 1)), _mm256_castsi256_si128(voutput_zero_point));
59      $else:
60        __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]}), _mm256_extracti128_si256(vacc${ABC[N:N+8]}, 1)), voutput_zero_point);
61
62    $for N in range(0, BATCH_TILE, 16):
63      $if N + 8 < BATCH_TILE:
64        __m128i vout${ABC[N:N+16]} = _mm_shuffle_epi32(${_MM_PACKXS_EPI16}(_mm256_castsi256_si128(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}), _mm256_extracti128_si256(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}, 1)), _MM_SHUFFLE(3, 1, 2, 0));
65      $else:
66        __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = ${_MM_PACKXS_EPI16}(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
67
68    $for N in range(0, BATCH_TILE, 16):
69      $if N + 8 < BATCH_TILE:
70        vout${ABC[N:N+16]} = ${_MM_MAX_EPX8}(vout${ABC[N:N+16]}, voutput_min);
71      $else:
72        vout${ABC[N:N+8]}${ABC[N:N+8]} = ${_MM_MAX_EPX8}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_min);
73
74    $for N in range(0, BATCH_TILE, 16):
75      $if N + 8 < BATCH_TILE:
76        vout${ABC[N:N+16]} = ${_MM_MIN_EPX8}(vout${ABC[N:N+16]}, voutput_max);
77      $else:
78        vout${ABC[N:N+8]}${ABC[N:N+8]} = ${_MM_MIN_EPX8}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max);
79
80    $if BATCH_TILE >= 16:
81      _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
82    $else:
83      _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
84    $for N in range(16, BATCH_TILE, 16):
85      $if N + 8 < BATCH_TILE:
86        _mm_storeu_si128((__m128i*) (output + ${N}), vout${ABC[N:N+16]});
87      $else:
88        _mm_storel_epi64((__m128i*) (output + ${N}), vout${ABC[N:N+8]}${ABC[N:N+8]});
89    output += ${BATCH_TILE};
90  }
91  if XNN_UNLIKELY(n != 0) {
92    ${"do " if BATCH_TILE > 8 else ""}{
93      const __m256i va${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_a));
94      $if BATCH_TILE > 8:
95        input_a += 8;
96
97      __m256i vacc${ABC[0:8]} = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va${ABC[0:8]}, va_multiplier));
98
99      vacc${ABC[0:8]} = _mm256_sra_epi32(vacc${ABC[0:8]}, vshift);
100
101      $if BATCH_TILE > 8:
102        __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), _mm256_castsi256_si128(voutput_zero_point));
103      $else:
104        __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), voutput_zero_point);
105      __m128i vout${ABC[0:8]}${ABC[0:8]} = ${_MM_PACKXS_EPI16}(vout${ABC[0:8]}, vout${ABC[0:8]});
106      vout${ABC[0:8]}${ABC[0:8]} = ${_MM_MAX_EPX8}(vout${ABC[0:8]}${ABC[0:8]}, voutput_min);
107      vout${ABC[0:8]}${ABC[0:8]} = ${_MM_MIN_EPX8}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);
108
109      $if BATCH_TILE > 8:
110        if XNN_LIKELY(n >= (8 * sizeof(${XINT8_T}))) {
111          _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
112          output += 8;
113          n -= 8 * sizeof(${XINT8_T});
114        } else {
115          if (n & (4 * sizeof(${XINT8_T}))) {
116            _mm_storeu_si32(output, vout${ABC[0:8]}${ABC[0:8]});
117            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
118            output += 4;
119          }
120          if (n & (2 * sizeof(${XINT8_T}))) {
121            _mm_storeu_si16(output, vout${ABC[0:8]}${ABC[0:8]});
122            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
123            output += 2;
124          }
125          if (n & (1 * sizeof(${XINT8_T}))) {
126            *output = (${XINT8_T}) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
127          }
128          n = 0;
129        }
130      $else:
131        if (n & (4 * sizeof(${XINT8_T}))) {
132          _mm_storeu_si32(output, vout${ABC[0:8]}${ABC[0:8]});
133          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
134          output += 4;
135        }
136        if (n & (2 * sizeof(${XINT8_T}))) {
137          _mm_storeu_si16(output, vout${ABC[0:8]}${ABC[0:8]});
138          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
139          output += 2;
140        }
141        if (n & (1 * sizeof(${XINT8_T}))) {
142          *output = (${XINT8_T}) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
143        }
144    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
145  }
146}
147