xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/6x8c4-minmax-fp32-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c4-neondot.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const int8_t** restrict a,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const int8_t* zero,
31     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 6);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (6 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(int8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   kc = round_up_po2(kc, 4 * sizeof(int8_t));
45   int8_t* c0 = c;
46   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47   if XNN_UNPREDICTABLE(mr < 2) {
48     c1 = c0;
49   }
50   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51   if XNN_UNPREDICTABLE(mr <= 2) {
52     c2 = c1;
53   }
54   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr < 4) {
56     c3 = c2;
57   }
58   int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59   if XNN_UNPREDICTABLE(mr <= 4) {
60     c4 = c3;
61   }
62   int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
63   if XNN_UNPREDICTABLE(mr != 6) {
64     c5 = c4;
65   }
66 
67   do {
68     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
69     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
70     int32x4_t vacc1x0123 = vacc0x0123;
71     int32x4_t vacc1x4567 = vacc0x4567;
72     int32x4_t vacc2x0123 = vacc0x0123;
73     int32x4_t vacc2x4567 = vacc0x4567;
74     int32x4_t vacc3x0123 = vacc0x0123;
75     int32x4_t vacc3x4567 = vacc0x4567;
76     int32x4_t vacc4x0123 = vacc0x0123;
77     int32x4_t vacc4x4567 = vacc0x4567;
78     int32x4_t vacc5x0123 = vacc0x0123;
79     int32x4_t vacc5x4567 = vacc0x4567;
80 
81     size_t p = ks;
82     do {
83       const int8_t* restrict a0 = a[0];
84       if XNN_UNPREDICTABLE(a0 != zero) {
85         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
86       }
87       const int8_t* restrict a1 = a[1];
88       if XNN_UNPREDICTABLE(a1 != zero) {
89         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
90       }
91       const int8_t* restrict a2 = a[2];
92       if XNN_UNPREDICTABLE(a2 != zero) {
93         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
94       }
95       const int8_t* restrict a3 = a[3];
96       if XNN_UNPREDICTABLE(a3 != zero) {
97         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
98       }
99       const int8_t* restrict a4 = a[4];
100       if XNN_UNPREDICTABLE(a4 != zero) {
101         a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
102       }
103       const int8_t* restrict a5 = a[5];
104       if XNN_UNPREDICTABLE(a5 != zero) {
105         a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
106       }
107       a += 6;
108 
109       // Inner accumulation loop along the 8 columns.
110       size_t k = kc;
111       // 2x partial unrolled loop to load 8 bytes at a time.
112       while (k >= 8 * sizeof(int8_t)) {
113         // Load a 6x8 block of activations.
114         const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
115         const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
116         const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
117         const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
118         const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
119         const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
120 
121         // Load a 8x8 block of weights.
122         const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
123         const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
124         const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
125         const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
126 
127         // Multiply-accumulate: 6x8 * 8x8 --> 6x8.
128         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
129         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
130         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
131         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
132         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
133         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
134         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
135         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
136         vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
137         vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
138         vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
139         vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
140         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
141         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
142         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
143         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
144         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
145         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
146         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
147         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
148         vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
149         vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
150         vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
151         vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
152 
153         k -= 8 * sizeof(int8_t);
154       }
155       // Handle up to 4 final positions of `k`
156       if XNN_UNLIKELY(k != 0) {
157         // Load a 6x4 block of activations.
158         const int8x8_t va0x01234567 = vld1_s8(a0);
159         const int8x8_t va1x01234567 = vld1_s8(a1);
160         const int8x8_t va2x01234567 = vld1_s8(a2);
161         const int8x8_t va3x01234567 = vld1_s8(a3);
162         const int8x8_t va4x01234567 = vld1_s8(a4);
163         const int8x8_t va5x01234567 = vld1_s8(a5);
164 
165         // Load a 4x8 block of weights.
166         const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
167         const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
168 
169         // Multiply-accumulate: 6x4 * 4x8 --> 6x8.
170         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
171         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
172         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
173         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
174         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
175         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
176         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
177         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
178         vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
179         vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
180         vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
181         vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
182       }
183       p -= 6 * sizeof(void*);
184     } while (p != 0);
185 
186     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
187     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
188     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
189     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
190     float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
191     float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
192     float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
193     float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
194     float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
195     float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
196     float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
197     float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
198 
199     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
200     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
201     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
202     vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
203     vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
204     vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
205     vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
206     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
207     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
208     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
209     vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
210     vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
211     vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
212     vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
213 
214     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
215     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
216     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
217     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
218     vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
219     vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
220     vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
221     vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
222     vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123);
223     vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567);
224     vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123);
225     vacc5x4567 = vcvtnq_s32_f32(vfpacc5x4567);
226 
227     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
228 #if XNN_ARCH_ARM64
229     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
230     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
231     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
232     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
233     const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
234     const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
235 
236     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
237     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
238     int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567);
239 #else
240     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
241     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
242     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
243     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
244     const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
245     const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
246 
247     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
248     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
249     int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567));
250 #endif
251     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
252     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
253 
254     vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min);
255     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
256     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
257 
258     vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max);
259     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
260     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
261 
262     if (nc >= 8) {
263       vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567));
264       vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567));
265       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
266       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
267       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
268       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
269 
270       c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
271       c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
272       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
273       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
274       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
275       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
276 
277       a = (const int8_t**restrict) ((uintptr_t) a - ks);
278 
279       nc -= 8;
280     } else {
281       if (nc & 4) {
282         vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
283         vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
284         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
285         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
286         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
287         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
288         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
289         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
290         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
291       }
292       if (nc & 2) {
293         vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
294         vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
295         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
296         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
297         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
298         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
299         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
300         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
301         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
302       }
303       if (nc & 1) {
304         vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
305         vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
306         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
307         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
308         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
309         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
310       }
311 
312       nc = 0;
313     }
314   } while (nc != 0);
315 }
316