xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/c4-neon-mull-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 in ["FP32", "RNDNU"]
10$assert not CHANNELWISE or REQUANTIZATION == "FP32"
11$assert DUP in ["DUP", "LD1R", "LD2R", "LD4R"]
12#include <assert.h>
13
14#include <arm_neon.h>
15
16#include <xnnpack/gemm.h>
17$if REQUANTIZATION == "FP32" and ARMV8:
18  #include <xnnpack/intrinsics-polyfill.h>
19#include <xnnpack/math.h>
20
21
22$DATATYPE = "qc8" if CHANNELWISE else "qs8"
23$PARAMS_STRUCT = REQUANTIZATION.lower() + "_" + ("neonv8" if REQUANTIZATION == "FP32" and ARMV8 else "neon")
24$PARAMS_UNION = "xnn_%s_conv_minmax_params" % DATATYPE.lower()
25$ISA = "neonv8" if ARMV8 else "neon"
26void xnn_${DATATYPE}_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__${ISA}_${"mlal" if MLA else "mull"}_${DUP.lower()}(
27    size_t mr,
28    size_t nc,
29    size_t kc,
30    size_t ks,
31    const int8_t** restrict a,
32    const void* restrict w,
33    int8_t* restrict c,
34    size_t cm_stride,
35    size_t cn_stride,
36    size_t a_offset,
37    const int8_t* zero,
38    const union ${PARAMS_UNION} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
39{
40  assert(mr != 0);
41  assert(mr <= ${MR});
42  assert(nc != 0);
43  assert(kc != 0);
44  assert(ks != 0);
45  assert(ks % (${MR} * sizeof(void*)) == 0);
46  assert(a_offset % sizeof(int8_t) == 0);
47  assert(a != NULL);
48  assert(w != NULL);
49  assert(c != NULL);
50
51  kc = round_up_po2(kc, 4 * sizeof(int8_t));
52  int8_t* c0 = c;
53  $for M in range(1, MR):
54    int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
55    $if M % 2 == 0:
56      if XNN_UNPREDICTABLE(mr <= ${M}) {
57        c${M} = c${M-1};
58      }
59    $elif M + 1 == MR:
60      if XNN_UNPREDICTABLE(mr != ${M+1}) {
61        c${M} = c${M-1};
62      }
63    $else:
64      if XNN_UNPREDICTABLE(mr < ${M+1}) {
65        c${M} = c${M-1};
66      }
67
68  do {
69    $for N in range(0, NR, 2):
70      int32x4_t vacc0x${ABC[N:N+2]} = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
71    $for M in range(1, MR):
72      $for N in range(0, NR, 2):
73        int32x4_t vacc${M}x${ABC[N:N+2]} = vacc0x${ABC[N:N+2]};
74
75    size_t p = ks;
76    do {
77      $for M in range(MR):
78        const int8_t* restrict a${M} = a[${M}];
79        if XNN_UNPREDICTABLE(a${M} != zero) {
80          a${M} = (const int8_t*) ((uintptr_t) a${M} + a_offset);
81        }
82      a += ${MR};
83
84      size_t k = kc;
85
86      $if MLA:
87        while (k >= 16 * sizeof(int8_t)) {
88          $for M in range(MR):
89            $if DUP == "LD2R":
90              const int32x2x2_t va${M}x0 = vld2_dup_s32((const void*)a${M}); a${M} += 8;
91              const int32x2x2_t va${M}x1 = vld2_dup_s32((const void*)a${M}); a${M} += 8;
92            $elif DUP == "LD1R":
93              const int32x2_t va${M}0x0 = vld1_dup_s32((const void*)a${M});
94              const int32x2_t va${M}1x0 = vld1_dup_s32((const void*)(a${M} + 4)); a${M} += 8;
95              const int32x2_t va${M}0x1 = vld1_dup_s32((const void*)a${M});
96              const int32x2_t va${M}1x1 = vld1_dup_s32((const void*)(a${M} + 4)); a${M} += 8;
97            $else:
98              const int8x8_t va${M}x0 = vld1_s8(a${M}); a${M} += 8;
99              const int8x8_t va${M}x1 = vld1_s8(a${M}); a${M} += 8;
100
101          $for K in range(2):
102            $for N in range(0, NR, 2):
103              const int8x8_t vb${ABC[N:N+2]}c${K}x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104
105          $for K in range(2):
106            $for M in range(MR):
107              $if DUP == "LD2R":
108                const int8x8_t va${M}c${K}x0 = vreinterpret_s8_s32(va${M}x0.val[${K}]);
109                const int8x8_t va${M}c${K}x1 = vreinterpret_s8_s32(va${M}x1.val[${K}]);
110              $elif DUP == "LD1R":
111                const int8x8_t va${M}c${K}x0 = vreinterpret_s8_s32(va${M}${K}x0);
112                const int8x8_t va${M}c${K}x1 = vreinterpret_s8_s32(va${M}${K}x1);
113              $else:
114                const int8x8_t va${M}c${K}x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va${M}x0), ${K}));
115                const int8x8_t va${M}c${K}x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va${M}x1), ${K}));
116
117            $for N in range(0, NR, 2):
118              $for M in range(MR):
119                int16x8_t vprod${M}x${ABC[N:N+2]}c${K} = vmull_s8(vb${ABC[N:N+2]}c${K}x0, va${M}c${K}x0);
120              const int8x8_t vb${ABC[N:N+2]}c${K}x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121              $for M in range(MR):
122                vprod${M}x${ABC[N:N+2]}c${K} = vmlal_s8(vprod${M}x${ABC[N:N+2]}c${K}, vb${ABC[N:N+2]}c${K}x1, va${M}c${K}x1);
123              $for M in range(MR):
124                vacc${M}x${ABC[N:N+2]} = vpadalq_s16(vacc${M}x${ABC[N:N+2]}, vprod${M}x${ABC[N:N+2]}c${K});
125
126          k -= 16 * sizeof(int8_t);
127        }
128
129      ${"if" if MLA else "while"} (k >= 8 * sizeof(int8_t)) {
130        $for M in range(MR):
131          $if DUP == "LD2R":
132            const int32x2x2_t va${M} = vld2_dup_s32((const void*)a${M}); a${M} += 8;
133          $elif DUP == "LD1R":
134            const int32x2_t va${M}0 = vld1_dup_s32((const void*)a${M});
135            const int32x2_t va${M}1 = vld1_dup_s32((const void*)(a${M} + 4)); a${M} += 8;
136          $else:
137            const int8x8_t va${M} = vld1_s8(a${M}); a${M} += 8;
138
139        $for K in range(2):
140          $for N in range(0, NR, 2):
141            const int8x8_t vb${ABC[N:N+2]}c${K} = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142
143        $for K in range(2):
144          $for M in range(MR):
145            $if DUP == "LD2R":
146              const int8x8_t va${M}c${K} = vreinterpret_s8_s32(va${M}.val[${K}]);
147            $elif DUP == "LD1R":
148              const int8x8_t va${M}c${K} = vreinterpret_s8_s32(va${M}${K});
149            $else:
150              const int8x8_t va${M}c${K} = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va${M}), ${K}));
151
152          $for N in range(0, NR, 2):
153            $for M in range(MR):
154              const int16x8_t vprod${M}x${ABC[N:N+2]}c${K} = vmull_s8(vb${ABC[N:N+2]}c${K}, va${M}c${K});
155            $for M in range(MR):
156              vacc${M}x${ABC[N:N+2]} = vpadalq_s16(vacc${M}x${ABC[N:N+2]}, vprod${M}x${ABC[N:N+2]}c${K});
157
158        k -= 8 * sizeof(int8_t);
159      }
160
161      if XNN_UNLIKELY(k != 0) {
162        $for M in range(MR):
163          const int8x8_t va${M} = vld1_s8(a${M}); a${M} = (const int8_t*) ((uintptr_t) a${M} + k);
164
165        $for N in range(0, NR, 2):
166          const int8x8_t vb${ABC[N:N+2]}c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167
168        $for M in range(MR):
169          const int8x8_t va${M}c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va${M}), 0));
170          $for N in range(0, NR, 2):
171            const int16x8_t vprod${M}x${ABC[N:N+2]}c0 = vmull_s8(vb${ABC[N:N+2]}c0, va${M}c0);
172            vacc${M}x${ABC[N:N+2]} = vpadalq_s16(vacc${M}x${ABC[N:N+2]}, vprod${M}x${ABC[N:N+2]}c0);
173      }
174      p -= ${MR} * sizeof(void*);
175    } while (p != 0);
176
177#if XNN_ARCH_ARM64
178    $for M in range(MR):
179      $for N in range(0, NR, 4):
180        int32x4_t vacc${M}x${ABC[N:N+4]} = vpaddq_s32(vacc${M}x${ABC[N:N+2]}, vacc${M}x${ABC[N+2:N+4]});
181#else
182    $for M in range(MR):
183      $for N in range(0, NR, 4):
184        const int32x2_t vsum${M}x${ABC[N:N+2]} = vpadd_s32(vget_low_s32(vacc${M}x${ABC[N:N+2]}), vget_high_s32(vacc${M}x${ABC[N:N+2]}));
185        const int32x2_t vsum${M}x${ABC[N+2:N+4]} = vpadd_s32(vget_low_s32(vacc${M}x${ABC[N+2:N+4]}), vget_high_s32(vacc${M}x${ABC[N+2:N+4]}));
186        int32x4_t vacc${M}x${ABC[N:N+4]} = vcombine_s32(vsum${M}x${ABC[N:N+2]}, vsum${M}x${ABC[N+2:N+4]});
187#endif
188
189    $if REQUANTIZATION == "RNDNU":
190      const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_pre_shift);
191      const int32x4_t vmultiplier = vld1q_dup_s32(&params->${PARAMS_STRUCT}.multiplier);
192      const int32x4_t vright_post_shift = vld1q_dup_s32(&params->${PARAMS_STRUCT}.right_post_shift);
193
194      $for M in range(MR):
195        $for N in range(0, NR, 4):
196          vacc${M}x${ABC[N:N+4]} = vqshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_pre_shift);
197
198      $for M in range(MR):
199        $for N in range(0, NR, 4):
200          vacc${M}x${ABC[N:N+4]} = vqdmulhq_s32(vacc${M}x${ABC[N:N+4]}, vmultiplier);
201
202      $for M in range(MR):
203        $for N in range(0, NR, 4):
204          vacc${M}x${ABC[N:N+4]} = vrshlq_s32(vacc${M}x${ABC[N:N+4]}, vright_post_shift);
205    $elif REQUANTIZATION == "FP32":
206      $for M in range(MR):
207        $for N in range(0, NR, 4):
208          float32x4_t vfpacc${M}x${ABC[N:N+4]} = vcvtq_f32_s32(vacc${M}x${ABC[N:N+4]});
209
210      $if CHANNELWISE:
211        $for N in range(0, NR, 4):
212          const float32x4_t vscale${ABC[N:N+4]} = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
213          $for M in range(MR):
214            vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale${ABC[N:N+4]});
215      $else:
216        const float32x4_t vscale = vld1q_dup_f32(&params->${PARAMS_STRUCT}.scale);
217        $for M in range(MR):
218          $for N in range(0, NR, 4):
219            vfpacc${M}x${ABC[N:N+4]} = vmulq_f32(vfpacc${M}x${ABC[N:N+4]}, vscale);
220
221      $if ARMV8:
222        $for M in range(MR):
223          $for N in range(0, NR, 4):
224            vacc${M}x${ABC[N:N+4]} = vcvtnq_s32_f32(vfpacc${M}x${ABC[N:N+4]});
225      $else:
226        const float32x4_t vmagic_bias = vld1q_dup_f32(&params->${PARAMS_STRUCT}.magic_bias);
227        $for M in range(MR):
228          $for N in range(0, NR, 4):
229            vacc${M}x${ABC[N:N+4]} = vreinterpretq_s32_f32(vaddq_f32(vfpacc${M}x${ABC[N:N+4]}, vmagic_bias));
230
231        const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
232        $for M in range(MR):
233          $for N in range(0, NR, 4):
234            vacc${M}x${ABC[N:N+4]} = vqsubq_s32(vacc${M}x${ABC[N:N+4]}, vmagic_bias_less_output_zero_point);
235
236    $if REQUANTIZATION != "FP32" or ARMV8:
237      const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->${PARAMS_STRUCT}.output_zero_point);
238#if XNN_ARCH_ARM64
239    $for M in range(MR):
240      $for N in range(0, NR, 8):
241        int16x8_t vacc${M}x${ABC[N:N+8]} = vqmovn_high_s32(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vacc${M}x${ABC[N+4:N+8]});
242
243    $if REQUANTIZATION != "FP32" or ARMV8:
244      $for M in range(MR):
245        $for N in range(0, NR, 8):
246          vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vacc${M}x${ABC[N:N+8]}, voutput_zero_point);
247
248    $for M in range(MR):
249      $for N in range(0, NR, 16):
250        $if N + 8 < NR:
251          int8x16_t vout${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]});
252        $elif M % 2 == 1:
253          int8x16_t vout${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]});
254        $elif M + 1 == MR:
255          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
256#else
257    $for M in range(MR):
258      $for N in range(0, NR, 8):
259        int16x8_t vacc${M}x${ABC[N:N+8]} = vcombine_s16(vqmovn_s32(vacc${M}x${ABC[N:N+4]}), vqmovn_s32(vacc${M}x${ABC[N+4:N+8]}));
260
261    $if REQUANTIZATION != "FP32" or ARMV8:
262      $for M in range(MR):
263        $for N in range(0, NR, 8):
264          vacc${M}x${ABC[N:N+8]} = vqaddq_s16(vacc${M}x${ABC[N:N+8]}, voutput_zero_point);
265
266    $for M in range(MR):
267      $for N in range(0, NR, 16):
268        $if N + 8 < NR:
269          int8x16_t vout${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]}));
270        $elif M % 2 == 1:
271          int8x16_t vout${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]}));
272        $elif M + 1 == MR:
273          int8x8_t vout${M}x${ABC[N:N+8]} = vqmovn_s16(vacc${M}x${ABC[N:N+8]});
274#endif
275
276    $if NR == 8 and MR == 1:
277      const int8x8_t voutput_min = vld1_dup_s8(&params->${PARAMS_STRUCT}.output_min);
278    $else:
279      const int8x16_t voutput_min = vld1q_dup_s8(&params->${PARAMS_STRUCT}.output_min);
280    $for M in range(MR):
281      $for N in range(0, NR, 16):
282        $if N + 8 < NR:
283          vout${M}x${ABC[N:N+16]} = vmaxq_s8(vout${M}x${ABC[N:N+16]}, voutput_min);
284        $elif M % 2 == 1:
285          vout${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);
286        $elif M + 1 == MR:
287          $if NR == 8 and MR == 1:
288            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, voutput_min);
289          $else:
290            vout${M}x${ABC[N:N+8]} = vmax_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_min));
291
292    $if NR == 8 and MR == 1:
293      const int8x8_t voutput_max = vld1_dup_s8(&params->${PARAMS_STRUCT}.output_max);
294    $else:
295      const int8x16_t voutput_max = vld1q_dup_s8(&params->${PARAMS_STRUCT}.output_max);
296    $for M in range(MR):
297      $for N in range(0, NR, 16):
298        $if N + 8 < NR:
299          vout${M}x${ABC[N:N+16]} = vminq_s8(vout${M}x${ABC[N:N+16]}, voutput_max);
300        $elif M % 2 == 1:
301          vout${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);
302        $elif M + 1 == MR:
303          $if NR == 8 and MR == 1:
304            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, voutput_max);
305          $else:
306            vout${M}x${ABC[N:N+8]} = vmin_s8(vout${M}x${ABC[N:N+8]}, vget_low_s8(voutput_max));
307
308    if (nc >= ${NR}) {
309      $for M in reversed(range(MR)):
310        $for N in range(0, NR, 16):
311          $if N + 8 < NR:
312            vst1q_s8(c${M} + ${N}, vout${M}x${ABC[N:N+16]});
313          $elif M % 2 == 1:
314            vst1_s8(c${M} + ${N}, vget_high_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
315            vst1_s8(c${M-1} + ${N}, vget_low_s8(vout${M-1}x${ABC[N:N+8]}_${M}x${ABC[N:N+8]}));
316          $elif M + 1 == MR:
317            vst1_s8(c${M} + ${N}, vout${M}x${ABC[N:N+8]});
318
319      $for M in reversed(range(MR)):
320        c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);
321
322      a = (const int8_t**restrict) ((uintptr_t) a - ks);
323
324      nc -= ${NR};
325    } else {
326      $if NR == 16:
327        $for M in reversed(range(MR)):
328          $if M % 2 == 1:
329            int8x16_t vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_low_s8(vout${M-1}x0123456789ABCDEF), vget_low_s8(vout${M}x0123456789ABCDEF));
330          $elif M + 1 == MR:
331            int8x8_t vout${M}x01234567 = vget_low_s8(vout${M}x0123456789ABCDEF);
332        if (nc & 8) {
333          $for M in reversed(range(MR)):
334            $if M % 2 == 1:
335              vst1_s8(c${M}, vget_high_s8(vout${M-1}x01234567_${M}x01234567)); c${M} += 8;
336              vst1_s8(c${M-1}, vget_low_s8(vout${M-1}x01234567_${M}x01234567)); c${M-1} += 8;
337            $elif M + 1 == MR:
338              vst1_s8(c${M}, vout${M}x01234567); c${M} += 8;
339          $for M in reversed(range(MR)):
340            $if M % 2 == 1:
341              vout${M-1}x01234567_${M}x01234567 = vcombine_s8(vget_high_s8(vout${M-1}x0123456789ABCDEF), vget_high_s8(vout${M}x0123456789ABCDEF));
342            $elif M + 1 == MR:
343              vout${M}x01234567 = vget_high_s8(vout${M}x0123456789ABCDEF);
344        }
345      if (nc & 4) {
346        $for M in reversed(range(MR)):
347          $if M % 2 == 1:
348            vst1q_lane_u32((void*) c${M}, vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 2); c${M} += 4;
349            vst1q_lane_u32((void*) c${M-1}, vreinterpretq_u32_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 4;
350          $elif M + 1 == MR:
351            vst1_lane_u32((void*) c${M}, vreinterpret_u32_s8(vout${M}x01234567), 0); c${M} += 4;
352        $for M in reversed(range(MR)):
353          $if M % 2 == 1:
354            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 4);
355          $elif M + 1 == MR:
356            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 4);
357      }
358      if (nc & 2) {
359        $for M in reversed(range(MR)):
360          $if M % 2 == 1:
361            vst1q_lane_u16((void*) c${M}, vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 4); c${M} += 2;
362            vst1q_lane_u16((void*) c${M-1}, vreinterpretq_u16_s8(vout${M-1}x01234567_${M}x01234567), 0); c${M-1} += 2;
363          $elif M + 1 == MR:
364            vst1_lane_u16((void*) c${M}, vreinterpret_u16_s8(vout${M}x01234567), 0); c${M} += 2;
365        $for M in reversed(range(MR)):
366          $if M % 2 == 1:
367            vout${M-1}x01234567_${M}x01234567 = vextq_s8(vout${M-1}x01234567_${M}x01234567, vout${M-1}x01234567_${M}x01234567, 2);
368          $elif M + 1 == MR:
369            vout${M}x01234567 = vext_s8(vout${M}x01234567, vout${M}x01234567, 2);
370      }
371      if (nc & 1) {
372        $for M in reversed(range(MR)):
373          $if M % 2 == 1:
374            vst1q_lane_s8(c${M}, vout${M-1}x01234567_${M}x01234567, 8);
375            vst1q_lane_s8(c${M-1}, vout${M-1}x01234567_${M}x01234567, 0);
376          $elif M + 1 == MR:
377            vst1_lane_s8(c${M}, vout${M}x01234567, 0);
378      }
379
380      nc = 0;
381    }
382  } while (nc != 0);
383}
384