xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/c4-neondot.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 NR % 8 == 0
8$assert 8 <= NR <= 16
9$assert REQUANTIZATION in ["FP32", "RNDNU"]
10$assert not CHANNELWISE or REQUANTIZATION == "FP32"
11#include <assert.h>
12
13#include <arm_neon.h>
14
15#include <xnnpack/igemm.h>
16$if REQUANTIZATION == "FP32":
17  #include <xnnpack/intrinsics-polyfill.h>
18#include <xnnpack/math.h>
19
20
21$DATATYPE = "qc8" if CHANNELWISE else "qs8"
22$PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + ("neonv8" if REQUANTIZATION == "FP32" else "neon")
23$PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower()
24void xnn_${DATATYPE}_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot(
25    size_t mr,
26    size_t nc,
27    size_t kc,
28    size_t ks,
29    const int8_t** restrict a,
30    const void* restrict w,
31    int8_t* restrict c,
32    size_t cm_stride,
33    size_t cn_stride,
34    size_t a_offset,
35    const int8_t* zero,
36    const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37{
38  assert(mr != 0);
39  assert(mr <= ${MR});
40  assert(nc != 0);
41  assert(kc != 0);
42  assert(ks != 0);
43  assert(ks % (${MR} * sizeof(void*)) == 0);
44  assert(a_offset % sizeof(int8_t) == 0);
45  assert(a != NULL);
46  assert(w != NULL);
47  assert(c != NULL);
48
49  kc = round_up_po2(kc, 4 * sizeof(int8_t));
50  int8_t* c0 = c;
51  $for M in range(1, MR):
52    int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
53    $if M % 2 == 0:
54      if XNN_UNPREDICTABLE(mr <= ${M}) {
55        c${M} = c${M-1};
56      }
57    $elif M + 1 == MR:
58      if XNN_UNPREDICTABLE(mr != ${M+1}) {
59        c${M} = c${M-1};
60      }
61    $else:
62      if XNN_UNPREDICTABLE(mr < ${M+1}) {
63        c${M} = c${M-1};
64      }
65
66  do {
67    $for N in range(0, NR, 4):
68      int32x4_t vacc0x${ABC[N:N+4]} = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
69    $for M in range(1, MR):
70      $for N in range(0, NR, 4):
71        int32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
72
73    size_t p = ks;
74    do {
75      $for M in range(MR):
76        const int8_t* restrict a${M} = a[${M}];
77        if XNN_UNPREDICTABLE(a${M} != zero) {
78          a${M} = (const int8_t*) ((uintptr_t) a${M} + a_offset);
79        }
80      a += ${MR};
81
82      // Inner accumulation loop along the ${NR} columns.
83      size_t k = kc;
84      // 2x partial unrolled loop to load 8 bytes at a time.
85      while (k >= 8 * sizeof(int8_t)) {
86        // Load a ${MR}x8 block of activations.
87        $for M in range(MR):
88          const int8x8_t va${M}x01234567 = vld1_s8(a${M}); a${M} += 8;
89
90        // Load a 8x${NR} block of weights.
91        $for K in range(0, 8, 4):
92          $for N in range(0, NR, 4):
93            const int8x16_t vb${ABC[K:K+4]}x${ABC[N:N+4]} = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
94
95        // Multiply-accumulate: ${MR}x8 * 8x${NR} --> ${MR}x${NR}.
96        $for K in range(0, 8, 4):
97          $for M in range(MR):
98            $for N in range(0, NR, 4):
99              vacc${M}x${ABC[N:N+4]} = vdotq_lane_s32(vacc${M}x${ABC[N:N+4]}, vb${ABC[K:K+4]}x${ABC[N:N+4]}, va${M}x01234567, ${K//4});
100
101        k -= 8 * sizeof(int8_t);
102      }
103      // Handle up to 4 final positions of `k`
104      if XNN_UNLIKELY(k != 0) {
105        // Load a ${MR}x4 block of activations.
106        $for M in range(MR):
107          const int8x8_t va${M}x01234567 = vld1_s8(a${M});
108
109        // Load a 4x${NR} block of weights.
110        $for N in range(0, NR, 4):
111          const int8x16_t vb0123x${ABC[N:N+4]} = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
112
113        // Multiply-accumulate: ${MR}x4 * 4x${NR} --> ${MR}x${NR}.
114        $for M in range(MR):
115          $for N in range(0, NR, 4):
116            vacc${M}x${ABC[N:N+4]} = vdotq_lane_s32(vacc${M}x${ABC[N:N+4]}, vb0123x${ABC[N:N+4]}, va${M}x01234567, 0);
117      }
118      p -= ${MR} * sizeof(void*);
119    } while (p != 0);
120
121    $if REQUANTIZATION == "RNDNU":
122      const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_pre_shift);
123      const int32x4_t vmultiplier = vld1q_dup_s32(&params->${PARAMS_STRUCT}.multiplier);
124      const int32x4_t vright_post_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_post_shift);
125
126      $for M in range(MR):
127        $for N in range(0, NR, 4):
128          vacc${M}x${ABC[N:N+4]} = vshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_pre_shift);
129
130      $for M in range(MR):
131        $for N in range(0, NR, 4):
132          vacc${M}x${ABC[N:N+4]} = vqdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier);
133
134      $for M in range(MR):
135        $for N in range(0, NR, 4):
136          vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_post_shift);
137    $elif REQUANTIZATION == "FP32":
138      $for M in range(MR):
139        $for N in range(0, NR, 4):
140          float32x4_t vfpacc${M}x${ABC[N:N+4]} = vcvtq_f32_s32(vacc${M}x${ABC[N:N+4]});
141
142      $if CHANNELWISE:
143        $for N in range(0, NR, 4):
144          const float32x4_t vscale${ABC[N:N+4]} = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
145          $for M in range(MR):
146            vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale${ABC[N:N+4]});
147      $else:
148        const float32x4_t vscale = vld1q_dup_f32(&params->${PARAMS_STRUCT}.scale);
149        $for M in range(MR):
150          $for N in range(0, NR, 4):
151            vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale);
152
153      $for M in range(MR):
154        $for N in range(0, NR, 4):
155          vacc${M}x${ABC[N:N+4]} = vcvtnq_s32_f32(vfpacc${M}x${ABC[N:N+4]});
156
157    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->${PARAMS_STRUCT}.output_zero_point);
158#if XNN_ARCH_ARM64
159    $for M in range(MR):
160      $for N in range(0, NR, 8):
161        const int16x8_t vacc${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);
162
163    $for M in range(MR):
164      $for N in range(0, NR, 16):
165        $if N + 8 < NR:
166          int8x16_t vout${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]});
167        $elif M % 2 == 1:
168          int8x16_t vout${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]});
169        $elif M + 1 == MR:
170          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
171#else
172    $for M in range(MR):
173      $for N in range(0, NR, 8):
174        const int16x8_t vacc${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);
175
176    $for M in range(MR):
177      $for N in range(0, NR, 16):
178        $if N + 8 < NR:
179          int8x16_t vout${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]}));
180        $elif M % 2 == 1:
181          int8x16_t vout${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]}));
182        $elif M + 1 == MR:
183          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
184#endif
185    $if NR == 8 and MR == 1:
186      const int8x8_t voutput_min = vld1_dup_s8(&params->${PARAMS_STRUCT}.output_min);
187      const int8x8_t voutput_max = vld1_dup_s8(&params->${PARAMS_STRUCT}.output_max);
188    $else:
189      const int8x16_t voutput_min = vld1q_dup_s8(&params->${PARAMS_STRUCT}.output_min);
190      const int8x16_t voutput_max = vld1q_dup_s8(&params->${PARAMS_STRUCT}.output_max);
191
192    $for M in reversed(range(MR)):
193      $for N in range(0, NR, 16):
194        $if N + 8 < NR:
195          vout${M}x${ABC[N:N+16]} = vmaxq_s8(vout${M}x${ABC[N:N+16]}, voutput_min);
196        $elif M % 2 == 1:
197          vout${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);
198        $elif M + 1 == MR:
199          $if NR == 8 and MR == 1:
200            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, voutput_min);
201          $else:
202            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_min));
203
204    $for M in reversed(range(MR)):
205      $for N in range(0, NR, 16):
206        $if N + 8 < NR:
207          vout${M}x${ABC[N:N+16]} = vminq_s8(vout${M}x${ABC[N:N+16]}, voutput_max);
208        $elif M % 2 == 1:
209          vout${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);
210        $elif M + 1 == MR:
211          $if NR == 8 and MR == 1:
212            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, voutput_max);
213          $else:
214            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_max));
215
216    if (nc >= ${NR}) {
217      $for M in reversed(range(MR)):
218        $for N in range(0, NR, 16):
219          $if N + 8 < NR:
220            vst1q_s8(c${M} + ${N}, vout${M}x${ABC[N:N+16]});
221          $elif M % 2 == 1:
222            vst1_s8(c${M} + ${N}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
223            vst1_s8(c${M-1} + ${N}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
224          $elif M + 1 == MR:
225            vst1_s8(c${M} + ${N}, vout${M}x${ABC[N:N+8]});
226
227      $for M in reversed(range(MR)):
228        c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);
229
230      a = (const int8_t**restrict) ((uintptr_t) a - ks);
231
232      nc -= ${NR};
233    } else {
234      $if NR == 16:
235        $for M in reversed(range(MR)):
236          $if M % 2 == 1:
237            int8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_low_s8(vout${M-1}x0123456789ABCDEF), vget_low_s8(vout${M}x0123456789ABCDEF));
238          $elif M + 1 == MR:
239            int8x8_t vout${M}x01234567 = vget_low_s8(vout${M}x0123456789ABCDEF);
240        if (nc & 8) {
241          $for M in reversed(range(MR)):
242            $if M % 2 == 1:
243              vst1_s8(c${M}, vget_high_s8(vout${M-1}x01234567_${M}x01234567)); c${M} += 8;
244              vst1_s8(c${M-1}, vget_low_s8(vout${M-1}x01234567_${M}x01234567)); c${M-1} += 8;
245            $elif M + 1 == MR:
246              vst1_s8(c${M}, vout${M}x01234567); c${M} += 8;
247          $for M in reversed(range(MR)):
248            $if M % 2 == 1:
249              vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_high_s8(vout${M-1}x0123456789ABCDEF), vget_high_s8(vout${M}x0123456789ABCDEF));
250            $elif M + 1 == MR:
251              vout${M}x01234567 = vget_high_s8(vout${M}x0123456789ABCDEF);
252        }
253      if (nc & 4) {
254        $for M in reversed(range(MR)):
255          $if M % 2 == 1:
256            vst1q_lane_u32((void*) c${M}, vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4;
257            vst1q_lane_u32((void*) c${M-1}, vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4;
258          $elif M + 1 == MR:
259            vst1_lane_u32((void*) c${M}, vreinterpret_u32_s8(vout${M}x01234567), 0); c${M} += 4;
260        $for M in reversed(range(MR)):
261          $if M % 2 == 1:
262            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4);
263          $elif M + 1 == MR:
264            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 4);
265      }
266      if (nc & 2) {
267        $for M in reversed(range(MR)):
268          $if M % 2 == 1:
269            vst1q_lane_u16((void*) c${M}, vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2;
270            vst1q_lane_u16((void*) c${M-1}, vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2;
271          $elif M + 1 == MR:
272            vst1_lane_u16((void*) c${M}, vreinterpret_u16_s8(vout${M}x01234567), 0); c${M} += 2;
273        $for M in reversed(range(MR)):
274          $if M % 2 == 1:
275            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2);
276          $elif M + 1 == MR:
277            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 2);
278      }
279      if (nc & 1) {
280        $for M in reversed(range(MR)):
281          $if M % 2 == 1:
282            vst1q_lane_s8(c${M}, vout${M-1}x01234567_${M}x01234567, 8);
283            vst1q_lane_s8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0);
284          $elif M + 1 == MR:
285            vst1_lane_s8(c${M}, vout${M}x01234567, 0);
286      }
287
288      nc = 0;
289    }
290  } while (nc != 0);
291}
292