xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/4x8c4-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_4x8c4__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_4x8c4__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 <= 4);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (4 * 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 
59   do {
60     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
61     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
62     int32x4_t vacc1x0123 = vacc0x0123;
63     int32x4_t vacc1x4567 = vacc0x4567;
64     int32x4_t vacc2x0123 = vacc0x0123;
65     int32x4_t vacc2x4567 = vacc0x4567;
66     int32x4_t vacc3x0123 = vacc0x0123;
67     int32x4_t vacc3x4567 = vacc0x4567;
68 
69     size_t p = ks;
70     do {
71       const int8_t* restrict a0 = a[0];
72       if XNN_UNPREDICTABLE(a0 != zero) {
73         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74       }
75       const int8_t* restrict a1 = a[1];
76       if XNN_UNPREDICTABLE(a1 != zero) {
77         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78       }
79       const int8_t* restrict a2 = a[2];
80       if XNN_UNPREDICTABLE(a2 != zero) {
81         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
82       }
83       const int8_t* restrict a3 = a[3];
84       if XNN_UNPREDICTABLE(a3 != zero) {
85         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
86       }
87       a += 4;
88 
89       // Inner accumulation loop along the 8 columns.
90       size_t k = kc;
91       // 2x partial unrolled loop to load 8 bytes at a time.
92       while (k >= 8 * sizeof(int8_t)) {
93         // Load a 4x8 block of activations.
94         const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
95         const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
96         const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
97         const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
98 
99         // Load a 8x8 block of weights.
100         const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
101         const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
102         const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
103         const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
104 
105         // Multiply-accumulate: 4x8 * 8x8 --> 4x8.
106         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
107         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
108         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
109         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
110         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
111         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
112         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
113         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
114         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
115         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
116         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
117         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
118         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
119         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
120         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
121         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
122 
123         k -= 8 * sizeof(int8_t);
124       }
125       // Handle up to 4 final positions of `k`
126       if XNN_UNLIKELY(k != 0) {
127         // Load a 4x4 block of activations.
128         const int8x8_t va0x01234567 = vld1_s8(a0);
129         const int8x8_t va1x01234567 = vld1_s8(a1);
130         const int8x8_t va2x01234567 = vld1_s8(a2);
131         const int8x8_t va3x01234567 = vld1_s8(a3);
132 
133         // Load a 4x8 block of weights.
134         const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
135         const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
136 
137         // Multiply-accumulate: 4x4 * 4x8 --> 4x8.
138         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
139         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
140         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
141         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
142         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
143         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
144         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
145         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
146       }
147       p -= 4 * sizeof(void*);
148     } while (p != 0);
149 
150     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
151     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
152     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
153     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
154     float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
155     float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
156     float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
157     float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
158 
159     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
160     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
161     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
162     vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
163     vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
164     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
165     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
166     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
167     vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
168     vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
169 
170     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
171     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
172     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
173     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
174     vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
175     vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
176     vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
177     vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
178 
179     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
180 #if XNN_ARCH_ARM64
181     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
182     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
183     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
184     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
185 
186     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
187     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
188 #else
189     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
190     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
191     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
192     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
193 
194     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
195     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
196 #endif
197     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
198     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
199 
200     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
201     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
202 
203     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
204     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
205 
206     if (nc >= 8) {
207       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
208       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
209       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
210       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
211 
212       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
213       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
214       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
215       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
216 
217       a = (const int8_t**restrict) ((uintptr_t) a - ks);
218 
219       nc -= 8;
220     } else {
221       if (nc & 4) {
222         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
223         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
224         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
225         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
226         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
227         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
228       }
229       if (nc & 2) {
230         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
231         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
232         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
233         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
234         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
235         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
236       }
237       if (nc & 1) {
238         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
239         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
240         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
241         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
242       }
243 
244       nc = 0;
245     }
246   } while (nc != 0);
247 }
248