• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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$assert 8 <= NR <= 16
9#include <assert.h>
10
11#include <arm_neon.h>
12
13#include <xnnpack/gemm.h>
14#include <xnnpack/math.h>
15
16
17void xnn_qs8_igemm_minmax_ukernel_${MR}x${NR}c16__neon_mlal_padal(
18    size_t mr,
19    size_t nc,
20    size_t kc,
21    size_t ks,
22    const int8_t** restrict a,
23    const void* restrict w,
24    int8_t* restrict c,
25    size_t cm_stride,
26    size_t cn_stride,
27    size_t a_offset,
28    const int8_t* zero,
29    const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
30{
31  assert(mr != 0);
32  assert(mr <= ${MR});
33  assert(nc != 0);
34  assert(kc != 0);
35  assert(ks != 0);
36  assert(ks % (${MR} * sizeof(void*)) == 0);
37  assert(a_offset % sizeof(int8_t) == 0);
38  assert(a != NULL);
39  assert(w != NULL);
40  assert(c != NULL);
41
42  kc = round_up_po2(kc, 16);
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    $for N in range(NR):
61      int32x4_t vacc0x${N} = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62    $for M in range(1, MR):
63      $for N in range(NR):
64        int32x4_t vacc${M}x${N} = vacc0x${N};
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      // KC loop of 16 with up to 15 remainder
76      size_t k = 0;
77      while (k < kc) {
78        $for M in range(MR):
79          const int8x16_t va${M} = vld1q_s8(a${M}); a${M} += 16;
80
81        $for N in range(NR):
82          const int8x16_t vb${N} = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
83
84        $for N in range(NR):
85          $for M in range(MR):
86            int16x8_t vprod${M}x${N} = vmull_s8(vget_low_s8(vb${N}), vget_low_s8(va${M}));
87          $for M in range(MR):
88            vprod${M}x${N} = vmlal_s8(vprod${M}x${N}, vget_high_s8(vb${N}), vget_high_s8(va${M}));
89          $for M in range(MR):
90            vacc${M}x${N} = vpadalq_s16(vacc${M}x${N}, vprod${M}x${N});
91
92        k += 16 * sizeof(int8_t);
93      }
94
95      p -= ${MR} * sizeof(void*);
96    } while (p != 0);
97
98#if XNN_ARCH_ARM64
99    $for M in range(MR):
100      $for N in range(0, NR, 4):
101        const int32x4_t vsum${M}x${ABC[N:N+2]} = vpaddq_s32(vacc${M}x${N}, vacc${M}x${N+1});
102        const int32x4_t vsum${M}x${ABC[N+2:N+4]} = vpaddq_s32(vacc${M}x${N+2}, vacc${M}x${N+3});
103    $for M in range(MR):
104      $for N in range(0, NR, 4):
105        int32x4_t vacc${M}x${ABC[N:N+4]} = vpaddq_s32(vsum${M}x${ABC[N:N+2]}, vsum${M}x${ABC[N+2:N+4]});
106#else
107    $for M in range(MR):
108      $for N in range(0, NR, 4):
109        const int32x2_t vpsum${M}x${ABC[N]} = vadd_s32(vget_low_s32(vacc${M}x${N}), vget_high_s32(vacc${M}x${N}));
110        const int32x2_t vpsum${M}x${ABC[N+1]} = vadd_s32(vget_low_s32(vacc${M}x${N+1}), vget_high_s32(vacc${M}x${N+1}));
111        const int32x2_t vpsum${M}x${ABC[N+2]} = vadd_s32(vget_low_s32(vacc${M}x${N+2}), vget_high_s32(vacc${M}x${N+2}));
112        const int32x2_t vpsum${M}x${ABC[N+3]} = vadd_s32(vget_low_s32(vacc${M}x${N+3}), vget_high_s32(vacc${M}x${N+3}));
113        const int32x2_t vsum${M}x${ABC[N:N+2]} = vpadd_s32(vpsum${M}x${ABC[N]}, vpsum${M}x${ABC[N+1]});
114        const int32x2_t vsum${M}x${ABC[N+2:N+4]} = vpadd_s32(vpsum${M}x${ABC[N+2]}, vpsum${M}x${ABC[N+3]});
115        int32x4_t vacc${M}x${ABC[N:N+4]} = vcombine_s32(vsum${M}x${ABC[N:N+2]}, vsum${M}x${ABC[N+2:N+4]} );
116#endif
117
118    const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
119    $for M in range(MR):
120      $for N in range(0, NR, 4):
121        vacc${M}x${ABC[N:N+4]} = vqrdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier);
122
123    const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
124    const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
125    $for M in range(MR):
126      $for N in range(0, NR, 4):
127        vacc${M}x${ABC[N:N+4]} = vsraq_n_s32(vacc${M}x${ABC[N:N+4]}, vbicq_s32(vacc${M}x${ABC[N:N+4]}, vzero_shift_mask), 31);
128
129    $for M in range(MR):
130      $for N in range(0, NR, 4):
131        vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_shift);
132
133    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
134#if XNN_ARCH_ARM64
135    $for M in range(MR):
136      $for N in range(0, NR, 8):
137        const int16x8_t vacc${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);
138    $for M in range(MR):
139      $for N in range(0, NR, 16):
140        $if N + 8 < NR:
141          int8x16_t vout${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]});
142        $elif M % 2 == 1:
143          int8x16_t vout${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]});
144        $elif M + 1 == MR:
145          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
146#else
147    $for M in range(MR):
148      $for N in range(0, NR, 8):
149        const int16x8_t vacc${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);
150
151    $for M in range(MR):
152      $for N in range(0, NR, 16):
153        $if N + 8 < NR:
154          int8x16_t vout${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]}));
155        $elif M % 2 == 1:
156          int8x16_t vout${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]}));
157        $elif M + 1 == MR:
158          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
159#endif
160    $if NR == 8 and MR == 1:
161      const int8x8_t voutput_min = vld1_dup_s8(&params->neon.output_min);
162      const int8x8_t voutput_max = vld1_dup_s8(&params->neon.output_max);
163    $else:
164      const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
165      const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
166
167    $for M in reversed(range(MR)):
168      $for N in range(0, NR, 16):
169        $if N + 8 < NR:
170          vout${M}x${ABC[N:N+16]} = vmaxq_s8(vout${M}x${ABC[N:N+16]}, voutput_min);
171        $elif M % 2 == 1:
172          vout${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);
173        $elif M + 1 == MR:
174          $if NR == 8 and MR == 1:
175            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, voutput_min);
176          $else:
177            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_min));
178
179    $for M in reversed(range(MR)):
180      $for N in range(0, NR, 16):
181        $if N + 8 < NR:
182          vout${M}x${ABC[N:N+16]} = vminq_s8(vout${M}x${ABC[N:N+16]}, voutput_max);
183        $elif M % 2 == 1:
184          vout${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);
185        $elif M + 1 == MR:
186          $if NR == 8 and MR == 1:
187            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, voutput_max);
188          $else:
189            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_max));
190
191    if (nc >= ${NR}) {
192      $for M in reversed(range(MR)):
193        $for N in range(0, NR, 16):
194          $if N + 8 < NR:
195            vst1q_s8(c${M} + ${N}, vout${M}x${ABC[N:N+16]});
196          $elif M % 2 == 1:
197            vst1_s8(c${M} + ${N}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
198            vst1_s8(c${M-1} + ${N}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
199          $elif M + 1 == MR:
200            vst1_s8(c${M} + ${N}, vout${M}x${ABC[N:N+8]});
201
202      $for M in reversed(range(MR)):
203        c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);
204
205      a = (const int8_t**restrict) ((uintptr_t) a - ks);
206
207      nc -= ${NR};
208    } else {
209      $if NR == 16:
210        $for M in range(MR):
211          $if M % 2 == 1:
212            int8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_low_s8(vout${M-1}x0123456789ABCDEF), vget_low_s8(vout${M}x0123456789ABCDEF));
213          $elif M + 1 == MR:
214            int8x8_t vout${M}x01234567 = vget_low_s8(vout${M}x0123456789ABCDEF);
215        if (nc & 8) {
216          $for M in reversed(range(MR)):
217            $if M % 2 == 1:
218              vst1_s8(c${M}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); c${M} += 8;
219              vst1_s8(c${M-1}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]})); c${M-1} += 8;
220            $elif M + 1 == MR:
221              vst1_s8(c${M}, vout${M}x${ABC[N:N+8]}); c${M} += 8;
222          $for M in reversed(range(MR)):
223            $if M % 2 == 1:
224              vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_high_s8(vout${M-1}x0123456789ABCDEF), vget_high_s8(vout${M}x0123456789ABCDEF));
225            $elif M + 1 == MR:
226              vout${M}x01234567 = vget_high_s8(vout${M}x0123456789ABCDEF);
227        }
228      if (nc & 4) {
229        $for M in reversed(range(MR)):
230          $if M % 2 == 1:
231            vst1q_lane_u32(__builtin_assume_aligned(c${M}, 1), vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4;
232            vst1q_lane_u32(__builtin_assume_aligned(c${M-1}, 1), vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4;
233          $elif M + 1 == MR:
234            vst1_lane_u32(__builtin_assume_aligned(c${M}, 1), vreinterpret_u32_s8(vout${M}x01234567), 0); c${M} += 4;
235        $for M in reversed(range(MR)):
236          $if M % 2 == 1:
237            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4);
238          $elif M + 1 == MR:
239            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 4);
240      }
241      if (nc & 2) {
242        $for M in reversed(range(MR)):
243          $if M % 2 == 1:
244            vst1q_lane_u16(__builtin_assume_aligned(c${M}, 1), vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2;
245            vst1q_lane_u16(__builtin_assume_aligned(c${M-1}, 1), vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2;
246          $elif M + 1 == MR:
247            vst1_lane_u16(__builtin_assume_aligned(c${M}, 1), vreinterpret_u16_s8(vout${M}x01234567), 0); c${M} += 2;
248        $for M in reversed(range(MR)):
249          $if M % 2 == 1:
250            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2);
251          $elif M + 1 == MR:
252            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 2);
253      }
254      if (nc & 1) {
255        $for M in reversed(range(MR)):
256          $if M % 2 == 1:
257            vst1q_lane_s8(c${M}, vout${M-1}x01234567_${M}x01234567, 8);
258            vst1q_lane_s8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0);
259          $elif M + 1 == MR:
260            vst1_lane_s8(c${M}, vout${M}x01234567, 0);
261      }
262
263      nc = 0;
264    }
265  } while (nc != 0);
266}
267