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 CHANNEL_TILE % 16 == 0
8$assert CHANNEL_TILE >= 16
9$assert KERNEL_TILE >= 2
10#include <assert.h>
11
12#include <immintrin.h>
13
14#include <xnnpack/dwconv.h>
15#include <xnnpack/intrinsics-polyfill.h>
16
17
18void xnn_qs8_dwconv_minmax_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__avx512skx_mul32(
19    size_t channels,
20    size_t output_width,
21    const int8_t** input,
22    const void* weights,
23    int8_t* output,
24    size_t input_stride,
25    size_t output_increment,
26    size_t input_offset,
27    const int8_t* zero,
28    const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29{
30  assert(channels != 0);
31  assert(output_width != 0);
32
33  const __mmask16 vblend_mask = _cvtu32_mask16(0xAAAA);
34  const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.multiplier));
35  const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.rounding));
36  const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
37  const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
38  const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
39  $if CHANNEL_TILE > 16:
40    const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
41    const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_min));
42    const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_max));
43    const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 5, 1, 6, 2, 4, 0);
44  $else:
45    const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
46    const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
47    const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
48
49  do {
50    $for K in range(KERNEL_TILE):
51      const int8_t* i${K} = input[${K}];
52      assert(i${K} != NULL);
53      if XNN_UNPREDICTABLE(i${K} != zero) {
54        i${K} = (const int8_t*) ((uintptr_t) i${K} + input_offset);
55      }
56    input = (const int8_t**) ((uintptr_t) input + input_stride);
57
58    size_t c = channels;
59    const void* w = weights;
60    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
61      __m512i vacc${ABC[0:16]} = _mm512_loadu_si512(w);
62      $for C in range(16, CHANNEL_TILE, 16):
63        __m512i vacc${ABC[C:C+16]} = _mm512_loadu_si512((const void*) ((uintptr_t) w + ${C} * sizeof(int32_t)));
64
65      $for K in range(KERNEL_TILE):
66
67        $for C in range(0, CHANNEL_TILE, 16):
68          $if C == 0:
69            const __m512i vi${K}x${ABC[0:16]} = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i${K}));
70          $else:
71            const __m512i vi${K}x${ABC[C:C+16]} = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i${K} + ${C})));
72          const __m512i vk${K}x${ABC[C:C+16]} = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE + C} * sizeof(int8_t))));
73        i${K} += ${CHANNEL_TILE};
74
75        $for C in range(0, CHANNEL_TILE, 16):
76          vacc${ABC[C:C+16]} = _mm512_add_epi32(vacc${ABC[C:C+16]}, _mm512_mullo_epi32(vi${K}x${ABC[C:C+16]}, vk${K}x${ABC[C:C+16]}));
77
78      w = (const void*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${KERNEL_TILE * CHANNEL_TILE} * sizeof(int8_t));
79
80      $for C in range(0, CHANNEL_TILE, 16):
81        const __m512i vacc${ABC[C+1:C+16:2]} = _mm512_shuffle_epi32(vacc${ABC[C:C+16]}, _MM_SHUFFLE(3, 3, 1, 1));
82
83      $for C in range(0, CHANNEL_TILE, 16):
84        const __m512i vprod${ABC[C:C+16:2]} = _mm512_add_epi64(_mm512_mul_epi32(vacc${ABC[C:C+16]}, vmultiplier), vrounding);
85        const __m512i vprod${ABC[C+1:C+16:2]} = _mm512_add_epi64(_mm512_mul_epi32(vacc${ABC[C+1:C+16:2]}, vmultiplier), vrounding);
86
87      $for C in range(0, CHANNEL_TILE, 16):
88        const __m512i vq31prod${ABC[C:C+16:2]} = _mm512_srli_epi64(vprod${ABC[C:C+16:2]}, 31);
89        const __m512i vq31prod${ABC[C+1:C+16:2]} = _mm512_add_epi64(vprod${ABC[C+1:C+16:2]}, vprod${ABC[C+1:C+16:2]});
90
91      $for C in range(0, CHANNEL_TILE, 16):
92        const __m512i vq31prod${ABC[C:C+16]} = _mm512_mask_blend_epi32(vblend_mask, vq31prod${ABC[C:C+16:2]}, vq31prod${ABC[C+1:C+16:2]});
93
94      $for C in range(0, CHANNEL_TILE, 16):
95        const __m512i vrem${ABC[C:C+16]} =
96          _mm512_add_epi32(_mm512_and_epi32(vq31prod${ABC[C:C+16]}, vremainder_mask), _mm512_srai_epi32(vq31prod${ABC[C:C+16]}, 31));
97
98      $for C in range(0, CHANNEL_TILE, 16):
99        vacc${ABC[C:C+16]} = _mm512_sra_epi32(vq31prod${ABC[C:C+16]}, vshift);
100
101      const __m512i vminus_one = _mm512_set1_epi32(-1);
102      $for C in range(0, CHANNEL_TILE, 16):
103        vacc${ABC[C:C+16]} = _mm512_mask_sub_epi32(vacc${ABC[C:C+16]}, _mm512_cmpgt_epi32_mask(vrem${ABC[C:C+16]}, vremainder_threshold), vacc${ABC[C:C+16]}, vminus_one);
104
105      $for C in range(0, CHANNEL_TILE, 16):
106        $if C + 16 < CHANNEL_TILE:
107          __m512i vout${ABC[C:C+4]}${ABC[C+16:C+20]}${ABC[C+4:C+8]}${ABC[C+20:C+24]}${ABC[C+8:C+12]}${ABC[C+24:C+28]}${ABC[C+12:C+16]}${ABC[C+28:C+32]} = _mm512_adds_epi16(_mm512_packs_epi32(vacc${ABC[C:C+16]}, vacc${ABC[C+16:C+32]}), voutput_zero_point);
108        $elif CHANNEL_TILE > 16:
109          __m256i vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc${ABC[C:C+16]}), _mm512_extracti32x8_epi32(vacc${ABC[C:C+16]}, 1)), _mm512_castsi512_si256(voutput_zero_point));
110        $else:
111          __m256i vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc${ABC[C:C+16]}), _mm512_extracti32x8_epi32(vacc${ABC[C:C+16]}, 1)), voutput_zero_point);
112
113      $for C in range(0, CHANNEL_TILE, 16):
114        $if C + 16 < CHANNEL_TILE:
115          vout${ABC[C:C+4]}${ABC[C+16:C+20]}${ABC[C+4:C+8]}${ABC[C+20:C+24]}${ABC[C+8:C+12]}${ABC[C+24:C+28]}${ABC[C+12:C+16]}${ABC[C+28:C+32]} = _mm512_min_epi16(_mm512_max_epi16(vout${ABC[C:C+4]}${ABC[C+16:C+20]}${ABC[C+4:C+8]}${ABC[C+20:C+24]}${ABC[C+8:C+12]}${ABC[C+24:C+28]}${ABC[C+12:C+16]}${ABC[C+28:C+32]}, voutput_min), voutput_max);
116        $elif CHANNEL_TILE > 16:
117          vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_min_epi16(_mm256_max_epi16(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}, _mm512_castsi512_si256(voutput_min)), _mm512_castsi512_si256(voutput_max));
118        $else:
119          vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_min_epi16(_mm256_max_epi16(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]}, voutput_min), voutput_max);
120
121      $for C in range(0, CHANNEL_TILE, 16):
122        $if C + 16 < CHANNEL_TILE:
123          const __m256i vout${ABC[C:C+4]}${ABC[C+16:C+20]}${ABC[C+4:C+8]}${ABC[C+20:C+24]} = _mm512_castsi512_si256(vout${ABC[C:C+4]}${ABC[C+16:C+20]}${ABC[C+4:C+8]}${ABC[C+20:C+24]}${ABC[C+8:C+12]}${ABC[C+24:C+28]}${ABC[C+12:C+16]}${ABC[C+28:C+32]});
124          const __m256i vout${ABC[C+8:C+12]}${ABC[C+24:C+28]}${ABC[C+12:C+16]}${ABC[C+28:C+32]} = _mm512_extracti32x8_epi32(vout${ABC[C:C+4]}${ABC[C+16:C+20]}${ABC[C+4:C+8]}${ABC[C+20:C+24]}${ABC[C+8:C+12]}${ABC[C+24:C+28]}${ABC[C+12:C+16]}${ABC[C+28:C+32]}, 1);
125          const __m256i vout${ABC[C:C+4]}${ABC[C+16:C+20]}${ABC[C+8:C+12]}${ABC[C+24:C+28]}${ABC[C+4:C+8]}${ABC[C+20:C+24]}${ABC[C+12:C+16]}${ABC[C+28:C+32]} = _mm256_packs_epi16(vout${ABC[C:C+4]}${ABC[C+16:C+20]}${ABC[C+4:C+8]}${ABC[C+20:C+24]}, vout${ABC[C+8:C+12]}${ABC[C+24:C+28]}${ABC[C+12:C+16]}${ABC[C+28:C+32]});
126          __m256i vout${ABC[C:C+32]} = _mm256_permutevar8x32_epi32(vout${ABC[C:C+4]}${ABC[C+16:C+20]}${ABC[C+8:C+12]}${ABC[C+24:C+28]}${ABC[C+4:C+8]}${ABC[C+20:C+24]}${ABC[C+12:C+16]}${ABC[C+28:C+32]}, vpermute_mask);
127        $else:
128          const __m128i vout${ABC[C:C+4]}${ABC[C+8:C+12]} = _mm256_castsi256_si128(vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]});
129          const __m128i vout${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]}${ABC[C+12:C+16]}, 1);
130          __m128i vout${ABC[C:C+16]} = _mm_shuffle_epi32(_mm_packs_epi16(vout${ABC[C:C+4]}${ABC[C+8:C+12]}, vout${ABC[C+4:C+8]}${ABC[C+12:C+16]}), _MM_SHUFFLE(3, 1, 2, 0));
131
132      $if CHANNEL_TILE > 16:
133        _mm256_storeu_si256((__m256i*) output, vout${ABC[0:32]});
134      $else:
135        _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
136      $for C in range(16, CHANNEL_TILE, 16):
137        $if C + 16 < CHANNEL_TILE:
138          _mm256_storeu_si256((__m256i*) (output + ${C}), vout${ABC[C:C+32]});
139        $else:
140          _mm_storeu_si128((__m128i*) (output + ${C}), vout${ABC[C:C+16]});
141      output += ${CHANNEL_TILE};
142    }
143    if XNN_UNLIKELY(c != 0) {
144      // Prepare mask for valid 8-bit elements (depends on nc).
145      const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1)));
146      $if CHANNEL_TILE > 16:
147        const int8_t* k = (const int8_t*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t));
148      ${"do " if CHANNEL_TILE > 16 else ""}{
149        __m512i vacc${ABC[0:16]} = _mm512_loadu_si512(w);
150
151        $for K in range(KERNEL_TILE):
152
153          const __m512i vi${K}x${ABC[0:16]} = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i${K}));
154          $if CHANNEL_TILE > 16:
155            $if K == 0:
156              const __m512i vk${K}x${ABC[0:16]} = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) k));
157            $else:
158              const __m512i vk${K}x${ABC[0:16]} = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + ${K * CHANNEL_TILE})));
159          $else:
160            const __m512i vk${K}x${ABC[0:16]} = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + ${CHANNEL_TILE} * sizeof(int32_t) + ${K * CHANNEL_TILE} * sizeof(int8_t))));
161          $if CHANNEL_TILE > 16:
162            i${K} += 16;
163
164          vacc${ABC[0:16]} = _mm512_add_epi32(vacc${ABC[0:16]}, _mm512_mullo_epi32(vi${K}x${ABC[0:16]}, vk${K}x${ABC[0:16]}));
165
166        $if CHANNEL_TILE > 16:
167          w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
168          k += 16;
169
170        const __m512i vacc${ABC[1:16:2]} = _mm512_shuffle_epi32(vacc${ABC[0:16]}, _MM_SHUFFLE(3, 3, 1, 1));
171
172        const __m512i vprod${ABC[0:16:2]} = _mm512_add_epi64(_mm512_mul_epi32(vacc${ABC[0:16]}, vmultiplier), vrounding);
173        const __m512i vprod${ABC[1:16:2]} = _mm512_add_epi64(_mm512_mul_epi32(vacc${ABC[1:16:2]}, vmultiplier), vrounding);
174
175        const __m512i vq31prod${ABC[0:16:2]} = _mm512_srli_epi64(vprod${ABC[0:16:2]}, 31);
176        const __m512i vq31prod${ABC[1:16:2]} = _mm512_add_epi64(vprod${ABC[1:16:2]}, vprod${ABC[1:16:2]});
177
178        const __m512i vq31prod${ABC[0:16]} = _mm512_mask_blend_epi32(vblend_mask, vq31prod${ABC[0:16:2]}, vq31prod${ABC[1:16:2]});
179
180        const __m512i vrem${ABC[0:16]} = _mm512_add_epi32(_mm512_and_epi32(vq31prod${ABC[0:16]}, vremainder_mask), _mm512_srai_epi32(vq31prod${ABC[0:16]}, 31));
181
182        vacc${ABC[0:16]} = _mm512_sra_epi32(vq31prod${ABC[0:16]}, vshift);
183        vacc${ABC[0:16]} = _mm512_mask_sub_epi32(vacc${ABC[0:16]}, _mm512_cmpgt_epi32_mask(vrem${ABC[0:16]}, vremainder_threshold), vacc${ABC[0:16]}, _mm512_set1_epi32(-1));
184
185        $if CHANNEL_TILE > 16:
186          __m256i vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]} = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc${ABC[0:16]}), _mm512_extracti32x8_epi32(vacc${ABC[0:16]}, 1)), _mm512_castsi512_si256(voutput_zero_point));
187        $else:
188          __m256i vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]} = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc${ABC[0:16]}), _mm512_extracti32x8_epi32(vacc${ABC[0:16]}, 1)), voutput_zero_point);
189
190        $if CHANNEL_TILE > 16:
191          vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]} = _mm256_min_epi16(_mm256_max_epi16(vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]}, _mm512_castsi512_si256(voutput_min)), _mm512_castsi512_si256(voutput_max));
192        $else:
193          vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]} = _mm256_min_epi16(_mm256_max_epi16(vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]}, voutput_min), voutput_max);
194
195        const __m128i vout${ABC[0:4]}${ABC[8:12]} = _mm256_castsi256_si128(vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]});
196        const __m128i vout${ABC[4:8]}${ABC[12:16]} = _mm256_extracti128_si256(vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]}, 1);
197        __m128i vout${ABC[0:16]} = _mm_shuffle_epi32(_mm_packs_epi16(vout${ABC[0:4]}${ABC[8:12]}, vout${ABC[4:8]}${ABC[12:16]}), _MM_SHUFFLE(3, 1, 2, 0));
198
199        $if CHANNEL_TILE > 16:
200          if XNN_LIKELY(c >= 16) {
201            _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
202            output += 16;
203            c -= 16;
204          } else {
205            _mm_mask_storeu_epi8(output, vmask, vout${ABC[0:16]});
206            output = (int8_t*) ((uintptr_t) output + c);
207            c = 0;
208          }
209        $else:
210          _mm_mask_storeu_epi8(output, vmask, vout${ABC[0:16]});
211          output = (int8_t*) ((uintptr_t) output + c);
212      }${" while (c != 0);" if CHANNEL_TILE > 16 else ""}
213    }
214
215    output = (int8_t*) ((uintptr_t) output + output_increment);
216  } while (--output_width != 0);
217}
218