xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/unipass-avx2-mul32.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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
7$assert REQUANTIZATION == "FP32"
8$assert DATATYPE in ["QC8", "QS8", "QU8"]
9$assert CHANNEL_TILE % 8 == 0
10$assert CHANNEL_TILE >= 8
11$assert KERNEL_TILE >= 2
12#include <assert.h>
13
14#include <immintrin.h>
15
16#include <xnnpack/dwconv.h>
17#include <xnnpack/unaligned.h>
18
19
20$PARAMS_STRUCT = REQUANTIZATION.lower() + "_avx2"
21$PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower()
22$XINT8_T = "uint8_t" if DATATYPE == "QU8" else "int8_t"
23$_MM256_CVTEPX8_EPI32 = "_mm256_cvtepu8_epi32" if DATATYPE == "QU8" else "_mm256_cvtepi8_epi32"
24$_MM_PACKXS_EPI16 = "_mm_packus_epi16" if DATATYPE == "QU8" else "_mm_packs_epi16"
25$_MM_MIN_EPX8 = "_mm_min_epu8" if DATATYPE == "QU8" else "_mm_min_epi8"
26$_MM_MAX_EPX8 = "_mm_max_epu8" if DATATYPE == "QU8" else "_mm_max_epi8"
27void xnn_${DATATYPE.lower()}_dwconv_minmax_${REQUANTIZATION.lower()}_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__avx2_mul32(
28    size_t channels,
29    size_t output_width,
30    const ${XINT8_T}** input,
31    const void* weights,
32    ${XINT8_T}* output,
33    size_t input_stride,
34    size_t output_increment,
35    size_t input_offset,
36    const ${XINT8_T}* zero,
37    const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
38{
39  assert(channels != 0);
40  assert(output_width != 0);
41
42  $if DATATYPE == "QU8":
43    const __m256i vk_zero_point = _mm256_cvtepu16_epi32(_mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.kernel_zero_point));
44  do {
45    $for K in range(KERNEL_TILE):
46      const ${XINT8_T}* i${K} = input[${K}];
47      assert(i${K} != NULL);
48      if XNN_UNPREDICTABLE(i${K} != zero) {
49        i${K} = (const ${XINT8_T}*) ((uintptr_t) i${K} + input_offset);
50      }
51    input = (const ${XINT8_T}**) ((uintptr_t) input + input_stride);
52
53    size_t c = channels;
54    const void* w = weights;
55    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
56      __m256i vacc${ABC[0:8]} = _mm256_loadu_si256((const __m256i*) w);
57      $for C in range(8, CHANNEL_TILE, 8):
58        __m256i vacc${ABC[C:C+8]} = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + ${C}));
59
60      $for K in range(KERNEL_TILE):
61
62        $for C in range(0, CHANNEL_TILE, 8):
63          $if C == 0:
64            const __m256i vi${K}x${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) i${K}));
65          $else:
66            const __m256i vi${K}x${ABC[C:C+8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) (i${K} + ${C})));
67          $if DATATYPE == "QU8":
68            const __m256i vk${K}x${ABC[C:C+8]} = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(${XINT8_T})))), vk_zero_point);
69          $else:
70            const __m256i vk${K}x${ABC[C:C+8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(${XINT8_T}))));
71        i${K} += ${CHANNEL_TILE};
72
73        $for C in range(0, CHANNEL_TILE, 8):
74          vacc${ABC[C:C+8]} = _mm256_add_epi32(vacc${ABC[C:C+8]}, _mm256_mullo_epi32(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}));
75
76      w = (const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(${XINT8_T}));
77
78      $for C in range(0, CHANNEL_TILE, 8):
79        __m256 vscaled${ABC[C:C+8]} = _mm256_cvtepi32_ps(vacc${ABC[C:C+8]});
80
81      $if DATATYPE == "QC8":
82        const __m256 vscale${ABC[0:8]} = _mm256_loadu_ps((const float*) w);
83        $for C in range(8, CHANNEL_TILE, 8):
84          const __m256 vscale${ABC[C:C+8]} = _mm256_loadu_ps((const float*) w + ${C});
85        w = (const void*) ((const float*) w + ${CHANNEL_TILE});
86        $for C in range(0, CHANNEL_TILE, 8):
87          vscaled${ABC[C:C+8]} = _mm256_mul_ps(vscaled${ABC[C:C+8]}, vscale${ABC[C:C+8]});
88      $else:
89        const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
90        $for C in range(0, CHANNEL_TILE, 8):
91          vscaled${ABC[C:C+8]} = _mm256_mul_ps(vscaled${ABC[C:C+8]}, vscale);
92
93      const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->${PARAMS_STRUCT}.output_max_less_zero_point);
94      $for C in range(0, CHANNEL_TILE, 8):
95        vscaled${ABC[C:C+8]} = _mm256_min_ps(vscaled${ABC[C:C+8]}, voutput_max_less_zero_point);
96
97      $for C in range(0, CHANNEL_TILE, 8):
98        vacc${ABC[C:C+8]} = _mm256_cvtps_epi32(vscaled${ABC[C:C+8]});
99
100      $if CHANNEL_TILE > 8:
101        const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->${PARAMS_STRUCT}.output_zero_point);
102      $else:
103        const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_zero_point);
104      $for C in range(0, CHANNEL_TILE, 16):
105        $if C + 8 < CHANNEL_TILE:
106          __m256i vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_adds_epi16(_mm256_packs_epi32(vacc${ABC[C:C+8]}, vacc${ABC[C+8:C+16]}), voutput_zero_point);
107        $elif CHANNEL_TILE > 8:
108          __m128i vout${ABC[C:C+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[C:C+8]}), _mm256_extracti128_si256(vacc${ABC[C:C+8]}, 1)), _mm256_castsi256_si128(voutput_zero_point));
109        $else:
110          __m128i vout${ABC[C:C+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[C:C+8]}), _mm256_extracti128_si256(vacc${ABC[C:C+8]}, 1)), voutput_zero_point);
111
112      $for C in range(0, CHANNEL_TILE, 16):
113        $if C + 8 < CHANNEL_TILE:
114          __m128i vout${ABC[C:C+16]} = _mm_shuffle_epi32(${_MM_PACKXS_EPI16}(_mm256_castsi256_si128(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}), _mm256_extracti128_si256(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}, 1)), _MM_SHUFFLE(3, 1, 2, 0));
115        $else:
116          __m128i vout${ABC[C:C+8]}${ABC[C:C+8]} = ${_MM_PACKXS_EPI16}(vout${ABC[C:C+8]}, vout${ABC[C:C+8]});
117
118      const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
119      $for C in range(0, CHANNEL_TILE, 16):
120        $if C + 8 < CHANNEL_TILE:
121          vout${ABC[C:C+16]} = ${_MM_MAX_EPX8}(vout${ABC[C:C+16]}, voutput_min);
122        $else:
123          vout${ABC[C:C+8]}${ABC[C:C+8]} = ${_MM_MAX_EPX8}(vout${ABC[C:C+8]}${ABC[C:C+8]}, voutput_min);
124
125      $if CHANNEL_TILE > 8:
126        _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
127      $else:
128        _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[C:C+8]});
129      $for C in range(16, CHANNEL_TILE, 16):
130        $if C + 8 < CHANNEL_TILE:
131          _mm_storeu_si128((__m128i*) (output + ${C}), vout${ABC[C:C+16]});
132        $else:
133          _mm_storel_epi64((__m128i*) (output + ${C}), vout${ABC[C:C+8]}${ABC[C:C+8]});
134      output += ${CHANNEL_TILE};
135    }
136    if XNN_UNLIKELY(c != 0) {
137      $if CHANNEL_TILE > 8:
138        const ${XINT8_T}* k = (const ${XINT8_T}*) ((const int32_t*) w + ${CHANNEL_TILE});
139      ${"do " if CHANNEL_TILE > 8 else ""}{
140        __m256i vacc${ABC[0:8]} = _mm256_loadu_si256((const __m256i*) w);
141
142        $for K in range(KERNEL_TILE):
143
144          const __m256i vi${K}x${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) i${K}));
145          $if DATATYPE == "QU8":
146            $if CHANNEL_TILE > 8:
147              $if K == 0:
148                const __m256i vk${K}x${ABC[0:8]} = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) k)), vk_zero_point);
149              $else:
150                const __m256i vk${K}x${ABC[0:8]} = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + ${K * CHANNEL_TILE}))), vk_zero_point);
151            $else:
152              const __m256i vk${K}x${ABC[0:8]} = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(${XINT8_T})))), vk_zero_point);
153          $else:
154            $if CHANNEL_TILE > 8:
155              $if K == 0:
156                const __m256i vk${K}x${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) k));
157              $else:
158                const __m256i vk${K}x${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (k + ${K * CHANNEL_TILE})));
159            $else:
160              const __m256i vk${K}x${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(${XINT8_T}))));
161          $if CHANNEL_TILE > 8:
162            i${K} += 8;
163
164          vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, _mm256_mullo_epi32(vi${K}x${ABC[0:8]}, vk${K}x${ABC[0:8]}));
165
166        $if CHANNEL_TILE > 8:
167          k += 8;
168
169        __m256 vscaled${ABC[0:8]} = _mm256_cvtepi32_ps(vacc${ABC[0:8]});
170        $if DATATYPE == "QC8":
171          const __m256 vscale${ABC[0:8]} = _mm256_loadu_ps((const float*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${CHANNEL_TILE * KERNEL_TILE} * sizeof(${XINT8_T})));
172          vscaled${ABC[0:8]} = _mm256_mul_ps(vscaled${ABC[0:8]}, vscale${ABC[0:8]});
173        $else:
174          vscaled${ABC[0:8]} = _mm256_mul_ps(vscaled${ABC[0:8]}, _mm256_load_ps(params->fp32_avx2.scale));
175        vscaled${ABC[0:8]} = _mm256_min_ps(vscaled${ABC[0:8]}, _mm256_load_ps(params->${PARAMS_STRUCT}.output_max_less_zero_point));
176        vacc${ABC[0:8]} = _mm256_cvtps_epi32(vscaled${ABC[0:8]});
177
178        $if CHANNEL_TILE > 8:
179          w = (const void*) ((const int32_t*) w + 8);
180
181        const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_zero_point);
182        __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), voutput_zero_point);
183
184        __m128i vout${ABC[0:8]}${ABC[0:8]} = ${_MM_PACKXS_EPI16}(vout${ABC[0:8]}, vout${ABC[0:8]});
185
186        const __m128i voutput_min = _mm_load_si128((const __m128i*) params->${PARAMS_STRUCT}.output_min);
187        vout${ABC[0:8]}${ABC[0:8]} = ${_MM_MAX_EPX8}(vout${ABC[0:8]}${ABC[0:8]}, voutput_min);
188
189        $if CHANNEL_TILE > 8:
190          if XNN_LIKELY(c >= 8) {
191            _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
192            output += 8;
193            c -= 8;
194          } else {
195            if (c & 4) {
196              unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]}));
197              vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
198              output += 4;
199            }
200            if (c & 2) {
201              unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0));
202              vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
203              output += 2;
204            }
205            if (c & 1) {
206              *output = (${XINT8_T}) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
207              output += 1;
208            }
209            c = 0;
210          }
211        $else:
212          if (c & 4) {
213            unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]}));
214            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
215            output += 4;
216          }
217          if (c & 2) {
218            unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0));
219            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
220            output += 2;
221          }
222          if (c & 1) {
223            *output = (${XINT8_T}) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
224            output += 1;
225          }
226      }${" while (c != 0);" if CHANNEL_TILE > 8 else ""}
227    }
228
229    output = (${XINT8_T}*) ((uintptr_t) output + output_increment);
230  } while (--output_width != 0);
231}
232