xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/neon-mull-addw-dup.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
7$assert NR % 8 == 0
8$assert 8 <= NR <= 16
9$assert REQUANTIZATION == "RNDNU"
10#include <assert.h>
11
12#include <arm_neon.h>
13
14#include <xnnpack/common.h>
15#include <xnnpack/gemm.h>
16
17
18void xnn_qs8_igemm_minmax_rndnu_ukernel_${MR}x${NR}__neon_mull_addw_dup(
19    size_t mr,
20    size_t nc,
21    size_t kc,
22    size_t ks,
23    const int8_t** restrict a,
24    const void* restrict w,
25    int8_t* restrict c,
26    size_t cm_stride,
27    size_t cn_stride,
28    size_t a_offset,
29    const int8_t* zero,
30    const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31{
32  assert(mr != 0);
33  assert(mr <= ${MR});
34  assert(nc != 0);
35  assert(kc != 0);
36  assert(ks != 0);
37  assert(ks % (${MR} * sizeof(void*)) == 0);
38  assert(a_offset % sizeof(int8_t) == 0);
39  assert(a != NULL);
40  assert(w != NULL);
41  assert(c != NULL);
42
43  int8_t* c0 = c;
44  $for M in range(1, MR):
45    int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
46    $if M % 2 == 0:
47      if XNN_UNPREDICTABLE(mr <= ${M}) {
48        c${M} = c${M-1};
49      }
50    $elif M + 1 == MR:
51      if XNN_UNPREDICTABLE(mr != ${M+1}) {
52        c${M} = c${M-1};
53      }
54    $else:
55      if XNN_UNPREDICTABLE(mr < ${M+1}) {
56        c${M} = c${M-1};
57      }
58
59  do {
60    $for N in range(0, NR, 4):
61      int32x4_t vacc0x${ABC[N:N+4]} = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
62    $for M in range(1, MR):
63      $for N in range(0, NR, 4):
64        int32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
65
66    size_t p = ks;
67    do {
68      $for M in range(MR):
69        const int8_t* restrict a${M} = a[${M}];
70        if XNN_UNPREDICTABLE(a${M} != zero) {
71          a${M} = (const int8_t*) ((uintptr_t) a${M} + a_offset);
72        }
73      a += ${MR};
74
75      size_t k = kc;
76      while (k >= 8 * sizeof(int8_t)) {
77        $for M in range(MR):
78          const int8x8_t va${M} = vld1_s8(a${M}); a${M} += 8;
79
80        $for K in range(8):
81          $for N in range(0, NR, 8):
82            const int8x8_t vb${ABC[N:N+8]}c${K} = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83
84            $for M in range(MR):
85              const int16x8_t vprod${M}x${ABC[N:N+8]}c${K} = vmull_s8(vb${ABC[N:N+8]}c${K}, vdup_lane_s8(va${M}, ${K}));
86              vacc${M}x${ABC[N:N+4]} = vaddw_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vprod${M}x${ABC[N:N+8]}c${K}));
87              vacc${M}x${ABC[N+4:N+8]} = vaddw_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vprod${M}x${ABC[N:N+8]}c${K}));
88
89        k -= 8 * sizeof(int8_t);
90      }
91      if XNN_UNLIKELY(k != 0) {
92        $for M in range(MR):
93          const int8x8_t va${M} = vld1_s8(a${M}); a${M} = (const int8_t*) ((uintptr_t) a${M} + k);
94
95        $for N in range(0, NR, 8):
96          const int8x8_t vb${ABC[N:N+8]}c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97
98        $for M in range(MR):
99          $for N in range(0, NR, 8):
100            const int16x8_t vprod${M}x${ABC[N:N+8]}c0 = vmull_s8(vb${ABC[N:N+8]}c0, vdup_lane_s8(va${M}, 0));
101            vacc${M}x${ABC[N:N+4]} = vaddw_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vprod${M}x${ABC[N:N+8]}c0));
102            vacc${M}x${ABC[N+4:N+8]} = vaddw_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vprod${M}x${ABC[N:N+8]}c0));
103
104        if (k >= 2 * sizeof(int8_t)) {
105          $for N in range(0, NR, 8):
106            const int8x8_t vb${ABC[N:N+8]}c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107
108          $for M in range(MR):
109            $for N in range(0, NR, 8):
110              const int16x8_t vprod${M}x${ABC[N:N+8]}c1 = vmull_s8(vb${ABC[N:N+8]}c1, vdup_lane_s8(va${M}, 1));
111              vacc${M}x${ABC[N:N+4]} = vaddw_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vprod${M}x${ABC[N:N+8]}c1));
112              vacc${M}x${ABC[N+4:N+8]} = vaddw_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vprod${M}x${ABC[N:N+8]}c1));
113
114          if (k > 2 * sizeof(int8_t)) {
115            $for N in range(0, NR, 8):
116              const int8x8_t vb${ABC[N:N+8]}c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117
118            $for M in range(MR):
119              $for N in range(0, NR, 8):
120                const int16x8_t vprod${M}x${ABC[N:N+8]}c2 = vmull_s8(vb${ABC[N:N+8]}c2, vdup_lane_s8(va${M}, 2));
121                vacc${M}x${ABC[N:N+4]} = vaddw_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vprod${M}x${ABC[N:N+8]}c2));
122                vacc${M}x${ABC[N+4:N+8]} = vaddw_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vprod${M}x${ABC[N:N+8]}c2));
123
124            if (k >= 4 * sizeof(int8_t)) {
125              $for N in range(0, NR, 8):
126                const int8x8_t vb${ABC[N:N+8]}c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127
128              $for M in range(MR):
129                $for N in range(0, NR, 8):
130                  const int16x8_t vprod${M}x${ABC[N:N+8]}c3 = vmull_s8(vb${ABC[N:N+8]}c3, vdup_lane_s8(va${M}, 3));
131                  vacc${M}x${ABC[N:N+4]} = vaddw_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vprod${M}x${ABC[N:N+8]}c3));
132                  vacc${M}x${ABC[N+4:N+8]} = vaddw_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vprod${M}x${ABC[N:N+8]}c3));
133
134              if (k > 4 * sizeof(int8_t)) {
135                $for N in range(0, NR, 8):
136                  const int8x8_t vb${ABC[N:N+8]}c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137
138                $for M in range(MR):
139                  $for N in range(0, NR, 8):
140                    const int16x8_t vprod${M}x${ABC[N:N+8]}c4 = vmull_s8(vb${ABC[N:N+8]}c4, vdup_lane_s8(va${M}, 4));
141                    vacc${M}x${ABC[N:N+4]} = vaddw_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vprod${M}x${ABC[N:N+8]}c4));
142                    vacc${M}x${ABC[N+4:N+8]} = vaddw_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vprod${M}x${ABC[N:N+8]}c4));
143
144                if (k >= 6 * sizeof(int8_t)) {
145                  $for N in range(0, NR, 8):
146                    const int8x8_t vb${ABC[N:N+8]}c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
147
148                  $for M in range(MR):
149                    $for N in range(0, NR, 8):
150                      const int16x8_t vprod${M}x${ABC[N:N+8]}c5 = vmull_s8(vb${ABC[N:N+8]}c5, vdup_lane_s8(va${M}, 5));
151                      vacc${M}x${ABC[N:N+4]} = vaddw_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vprod${M}x${ABC[N:N+8]}c5));
152                      vacc${M}x${ABC[N+4:N+8]} = vaddw_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vprod${M}x${ABC[N:N+8]}c5));
153
154                  if (k > 6 * sizeof(int8_t)) {
155                    $for N in range(0, NR, 8):
156                      const int8x8_t vb${ABC[N:N+8]}c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157
158                    $for M in range(MR):
159                      $for N in range(0, NR, 8):
160                        const int16x8_t vprod${M}x${ABC[N:N+8]}c6 = vmull_s8(vb${ABC[N:N+8]}c6, vdup_lane_s8(va${M}, 6));
161                        vacc${M}x${ABC[N:N+4]} = vaddw_s16(vacc${M}x${ABC[N:N+4]}, vget_low_s16(vprod${M}x${ABC[N:N+8]}c6));
162                        vacc${M}x${ABC[N+4:N+8]} = vaddw_s16(vacc${M}x${ABC[N+4:N+8]}, vget_high_s16(vprod${M}x${ABC[N:N+8]}c6));
163                  }
164                }
165              }
166            }
167          }
168        }
169      }
170      p -= ${MR} * sizeof(void*);
171    } while (p != 0);
172
173    const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
174    const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
175    const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
176
177    $for M in range(MR):
178      $for N in range(0, NR, 4):
179        vacc${M}x${ABC[N:N+4]} = vqshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_pre_shift);
180
181    $for M in range(MR):
182      $for N in range(0, NR, 4):
183        vacc${M}x${ABC[N:N+4]} = vqdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier);
184
185    $for M in range(MR):
186      $for N in range(0, NR, 4):
187        vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_post_shift);
188
189    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
190#if XNN_ARCH_ARM64
191    $for M in range(MR):
192      $for N in range(0, NR, 8):
193        const int16x8_t vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vacc${M}x${ABC[N+4:N+8]}), voutput_zero_point);
194
195    $for M in range(MR):
196      $for N in range(0, NR, 16):
197        $if N + 8 < NR:
198          int8x16_t vout${M}x${ABC[N:N+16]} = vqmovn_high_s16(vqmovn_s16(vacc${M}x${ABC[N:N+8]}), vacc${M}x${ABC[N+8:N+16]});
199        $elif M % 2 == 1:
200          int8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vqmovn_high_s16(vqmovn_s16(vacc${M-1}x${ABC[N:N+8]}), vacc${M}x${ABC[N:N+8]});
201        $elif M + 1 == MR:
202          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
203#else
204    $for M in range(MR):
205      $for N in range(0, NR, 8):
206        const int16x8_t vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vqmovn_s32(vacc${M}x${ABC[N+4:N+8]})), voutput_zero_point);
207
208    $for M in range(MR):
209      $for N in range(0, NR, 16):
210        $if N + 8 < NR:
211          int8x16_t vout${M}x${ABC[N:N+16]} = vcombine_s8(vqmovn_s16(vacc${M}x${ABC[N:N+8]}), vqmovn_s16(vacc${M}x${ABC[N+8:N+16]}));
212        $elif M % 2 == 1:
213          int8x16_t vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vcombine_s8(vqmovn_s16(vacc${M-1}x${ABC[N:N+8]}), vqmovn_s16(vacc${M}x${ABC[N:N+8]}));
214        $elif M + 1 == MR:
215          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
216#endif
217    $if NR == 8 and MR == 1:
218      const int8x8_t voutput_min = vld1_dup_s8(&params->rndnu_neon.output_min);
219      const int8x8_t voutput_max = vld1_dup_s8(&params->rndnu_neon.output_max);
220    $else:
221      const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
222      const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
223
224    $for M in reversed(range(MR)):
225      $for N in range(0, NR, 16):
226        $if N + 8 < NR:
227          vout${M}x${ABC[N:N+16]} = vmaxq_s8(vout${M}x${ABC[N:N+16]}, voutput_min);
228        $elif M % 2 == 1:
229          vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vmaxq_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_min);
230        $elif M + 1 == MR:
231          $if NR == 8 and MR == 1:
232            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, voutput_min);
233          $else:
234            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_min));
235
236    $for M in reversed(range(MR)):
237      $for N in range(0, NR, 16):
238        $if N + 8 < NR:
239          vout${M}x${ABC[N:N+16]} = vminq_s8(vout${M}x${ABC[N:N+16]}, voutput_max);
240        $elif M % 2 == 1:
241          vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]} = vminq_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}, voutput_max);
242        $elif M + 1 == MR:
243          $if NR == 8 and MR == 1:
244            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, voutput_max);
245          $else:
246            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_max));
247
248    if (nc >= ${NR}) {
249      $for M in reversed(range(MR)):
250        $for N in range(0, NR, 16):
251          $if N + 8 < NR:
252            vst1q_s8(c${M} + ${N}, vout${M}x${ABC[N:N+16]});
253          $elif M % 2 == 1:
254            vst1_s8(c${M} + ${N}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
255            vst1_s8(c${M-1} + ${N}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
256          $elif M + 1 == MR:
257            vst1_s8(c${M} + ${N}, vout${M}x${ABC[N:N+8]});
258
259      $for M in reversed(range(MR)):
260        c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);
261
262      a = (const int8_t**restrict) ((uintptr_t) a - ks);
263
264      nc -= ${NR};
265    } else {
266      $if NR == 16:
267        $for M in reversed(range(MR)):
268          $if M % 2 == 1:
269            int8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_low_s8(vout${M-1}x0123456789ABCDEF), vget_low_s8(vout${M}x0123456789ABCDEF));
270          $elif M + 1 == MR:
271            int8x8_t vout${M}x01234567 = vget_low_s8(vout${M}x0123456789ABCDEF);
272        if (nc & 8) {
273          $for M in reversed(range(MR)):
274            $if M % 2 == 1:
275              vst1_s8(c${M}, vget_high_s8(vout${M-1}x01234567_${M}x01234567)); c${M} += 8;
276              vst1_s8(c${M-1}, vget_low_s8(vout${M-1}x01234567_${M}x01234567)); c${M-1} += 8;
277            $elif M + 1 == MR:
278              vst1_s8(c${M}, vout${M}x01234567); c${M} += 8;
279          $for M in reversed(range(MR)):
280            $if M % 2 == 1:
281              vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_high_s8(vout${M-1}x0123456789ABCDEF), vget_high_s8(vout${M}x0123456789ABCDEF));
282            $elif M + 1 == MR:
283              vout${M}x01234567 = vget_high_s8(vout${M}x0123456789ABCDEF);
284        }
285      if (nc & 4) {
286        $for M in reversed(range(MR)):
287          $if M % 2 == 1:
288            vst1q_lane_u32((void*) c${M}, vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4;
289            vst1q_lane_u32((void*) c${M-1}, vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4;
290          $elif M + 1 == MR:
291            vst1_lane_u32((void*) c${M}, vreinterpret_u32_s8(vout${M}x01234567), 0); c${M} += 4;
292        $for M in reversed(range(MR)):
293          $if M % 2 == 1:
294            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4);
295          $elif M + 1 == MR:
296            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 4);
297      }
298      if (nc & 2) {
299        $for M in reversed(range(MR)):
300          $if M % 2 == 1:
301            vst1q_lane_u16((void*) c${M}, vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2;
302            vst1q_lane_u16((void*) c${M-1}, vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2;
303          $elif M + 1 == MR:
304            vst1_lane_u16((void*) c${M}, vreinterpret_u16_s8(vout${M}x01234567), 0); c${M} += 2;
305        $for M in reversed(range(MR)):
306          $if M % 2 == 1:
307            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2);
308          $elif M + 1 == MR:
309            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 2);
310      }
311      if (nc & 1) {
312        $for M in reversed(range(MR)):
313          $if M % 2 == 1:
314            vst1q_lane_s8(c${M}, vout${M-1}x01234567_${M}x01234567, 8);
315            vst1q_lane_s8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0);
316          $elif M + 1 == MR:
317            vst1_lane_s8(c${M}, vout${M}x01234567, 0);
318      }
319
320      nc = 0;
321    }
322  } while (nc != 0);
323}
324