Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x8c2s4-minmax-neon-mlal.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c2-neon-mull-shuffle.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_2x8c2s4__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 const int8_t* a0 = a;
41 float* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48
49 kc = round_up_po2(kc, 8 * sizeof(int8_t));
50 do {
51 const int32x4_t vizp01 = vld1q_s32(&quantization_params[0].zero_point);
52 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vizp01), 0);
54 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vizp01), 0);
55 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
56 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vizp01), 0);
57 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vizp01), 0);
58
59 size_t k = kc;
60 while (k >= 16 * sizeof(int8_t)) {
61 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
62 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
63 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
64 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
65
66 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74
75 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
76 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
77 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
78 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
79 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
80 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
81 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
82 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
83 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
84 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
85 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
86 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
87 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
88 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
89 va0x0 = vext_s8(va0x0, va0x0, 2);
90 va0x1 = vext_s8(va0x1, va0x1, 2);
91 va1x0 = vext_s8(va1x0, va1x0, 2);
92 va1x1 = vext_s8(va1x1, va1x1, 2);
93 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
94 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
95 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
96 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
97 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
98 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
99 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
100 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
101 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
102 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
103 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
104 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
105 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
106 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
107 va0x0 = vext_s8(va0x0, va0x0, 2);
108 va0x1 = vext_s8(va0x1, va0x1, 2);
109 va1x0 = vext_s8(va1x0, va1x0, 2);
110 va1x1 = vext_s8(va1x1, va1x1, 2);
111 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
112 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
113 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
114 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
115 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
116 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
117 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
118 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
119 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
120 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
121 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
122 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
123 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
124 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
125 va0x0 = vext_s8(va0x0, va0x0, 2);
126 va0x1 = vext_s8(va0x1, va0x1, 2);
127 va1x0 = vext_s8(va1x0, va1x0, 2);
128 va1x1 = vext_s8(va1x1, va1x1, 2);
129 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
130 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
131 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
132 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
133 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
134 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
135 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
136 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
137 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
138 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
139 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
140 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
141 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
142 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
143
144 k -= 16 * sizeof(int8_t);
145 }
146 if (k != 0) {
147 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
148 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
149
150 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
151 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
152 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
153 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
154 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
155 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
156 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
157 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
158
159 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
160 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
161 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
162 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
163 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
164 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
165 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
166 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
167 va0x0 = vext_s8(va0x0, va0x0, 2);
168 va1x0 = vext_s8(va1x0, va1x0, 2);
169 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
170 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
171 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
172 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
173 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
174 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
175 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
176 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
177 va0x0 = vext_s8(va0x0, va0x0, 2);
178 va1x0 = vext_s8(va1x0, va1x0, 2);
179 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
180 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
181 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
182 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
183 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
184 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
185 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
186 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
187 va0x0 = vext_s8(va0x0, va0x0, 2);
188 va1x0 = vext_s8(va1x0, va1x0, 2);
189 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
190 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
191 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
192 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
193 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
194 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
195 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
196 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
197
198 }
199
200 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
201 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
202 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
203 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
204
205 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
206 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
207 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
208 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
209 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
210
211 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
212 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
213
214 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
215 #if XNN_ARCH_ARM64
216 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
217 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
218 #else
219 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
220 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
221 #endif
222 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
223 #if XNN_ARCH_ARM64
224 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
225 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
226 #else
227 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
228 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
229 #endif
230
231 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
232 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
233 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
234 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
235 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
236
237 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
238 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
239 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
240 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
241 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
242
243 if XNN_LIKELY(nc >= 8) {
244 vst1q_f32(c0, vout0x0123);
245 vst1q_f32(c0 + 0, vout0x0123);
246 vst1q_f32(c0 + 4, vout0x4567);
247 vst1q_f32(c1, vout1x0123);
248 vst1q_f32(c1 + 0, vout1x0123);
249 vst1q_f32(c1 + 4, vout1x4567);
250
251 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
252 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
253
254 c0 = (float*) ((uintptr_t) c0 + cn_stride);
255 c1 = (float*) ((uintptr_t) c1 + cn_stride);
256
257 nc -= 8;
258 } else {
259 if (nc & 4) {
260 vst1q_f32(c0, vout0x0123); c0 += 4;
261 vout0x0123 = vout0x4567;
262 vst1q_f32(c1, vout1x0123); c1 += 4;
263 vout1x0123 = vout1x4567;
264 }
265 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
266 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
267 if (nc & 2) {
268 vst1_f32(c0, vout0x01); c0 += 2;
269 vst1_f32(c1, vout1x01); c1 += 2;
270 vout0x01 = vget_high_f32(vout0x0123);
271 vout1x01 = vget_high_f32(vout1x0123);
272 }
273 if (nc & 1) {
274 vst1_lane_f32(c0, vout0x01, 0);
275 vst1_lane_f32(c1, vout1x01, 0);
276 }
277 nc = 0;
278 }
279 } while (nc != 0);
280 }