Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-2x8-minmax-neon-mlal-lane-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/neon-mlal-lane.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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/igemm.h"
17 #include "xnnpack/prefetch.h"
18
19
20 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_2x8__neon_mlal_lane_prfm(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 2);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (2 * sizeof(void*)) == 0);
42 assert(a_offset % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr != 2) {
50 c1 = c0;
51 }
52
53 do {
54 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
55 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
56 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
57 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
58 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
59 int32x4_t vacc1x0123 = vacc0x0123;
60 int32x4_t vacc1x4567 = vacc0x4567;
61
62 size_t p = ks;
63 do {
64 const int8_t* restrict a0 = a[0];
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
67 } else {
68 a0 = zero_data;
69 }
70 const int8_t* restrict a1 = a[1];
71 if XNN_UNPREDICTABLE(a1 != zero) {
72 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
73 } else {
74 a1 = zero_data;
75 }
76 a += 2;
77
78 size_t k = kc;
79 while (k >= 8 * sizeof(int8_t)) {
80 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
81 const int16x8_t vxa0 = vmovl_s8(va0);
82 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
83 const int16x8_t vxa1 = vmovl_s8(va1);
84
85 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
87
88 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
89 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
90 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
91 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
92 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
94
95 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
96 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
97 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
98 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
99 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
100 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
101
102 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
103 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
104 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
105 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
106 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
107 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
108
109 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
110 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
111 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
112 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
113
114 xnn_prefetch_to_l1((const int8_t*) w + 448);
115
116 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
117 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
118
119 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
120 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
121 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
122 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
123 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
124 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
125
126 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
127 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
128 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
129 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
130 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
131 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
132
133 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
134 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
135 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
136 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
137 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
138 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
139
140 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
141 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
142 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
143 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
144
145 k -= 8 * sizeof(int8_t);
146 }
147 if XNN_UNLIKELY(k != 0) {
148 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
149 const int16x8_t vxa0 = vmovl_s8(va0);
150 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
151 const int16x8_t vxa1 = vmovl_s8(va1);
152
153 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
154 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
155
156 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
157 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
158 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
159 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
160
161 if (k >= 2 * sizeof(int8_t)) {
162 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
163 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
164
165 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
166 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
167 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
168 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
169
170 if (k > 2 * sizeof(int8_t)) {
171 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
172 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
173
174 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
175 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
176 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
177 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
178
179 if (k >= 4 * sizeof(int8_t)) {
180 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
181 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
182
183 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
184 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
185 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
186 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
187
188 if (k > 4 * sizeof(int8_t)) {
189 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
190 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
191
192 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
193 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
194 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
195 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
196
197 if (k >= 6 * sizeof(int8_t)) {
198 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
199 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
200
201 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
202 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
203 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
204 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
205
206 if (k > 6 * sizeof(int8_t)) {
207 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
208 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
209
210 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
211 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
212 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
213 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
214 }
215 }
216 }
217 }
218 }
219 }
220 }
221 p -= 2 * sizeof(void*);
222 } while (p != 0);
223
224 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
225 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
226 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
227 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
228
229 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
230 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
231 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
232 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
233 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
234
235 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
236 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
237
238 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
239 #if XNN_ARCH_ARM64
240 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
241 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
242 #else
243 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
244 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
245 #endif
246 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
247 #if XNN_ARCH_ARM64
248 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
249 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
250 #else
251 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
252 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
253 #endif
254
255 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
256 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
257 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
258 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
259 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
260
261 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
262 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
263 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
264 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
265 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
266
267 if XNN_LIKELY(nc >= 8) {
268 vst1q_f32(&c1[0], vout1x0123);
269 vst1q_f32(&c1[4], vout1x4567);
270 vst1q_f32(&c0[0], vout0x0123);
271 vst1q_f32(&c0[4], vout0x4567);
272
273 a = (const int8_t**restrict) ((uintptr_t) a - ks);
274
275 c0 = (float*) ((uintptr_t) c0 + cn_stride);
276 c1 = (float*) ((uintptr_t) c1 + cn_stride);
277
278 nc -= 8;
279 } else {
280 if (nc & 4) {
281 vst1q_f32(c1, vout1x0123); c1 += 4;
282 vout1x0123 = vout1x4567;
283 vst1q_f32(c0, vout0x0123); c0 += 4;
284 vout0x0123 = vout0x4567;
285 }
286 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
287 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
288 if (nc & 2) {
289 vst1_f32(c1, vout1x01); c1 += 2;
290 vst1_f32(c0, vout0x01); c0 += 2;
291 vout1x01 = vget_high_f32(vout1x0123);
292 vout0x01 = vget_high_f32(vout0x0123);
293 }
294 if (nc & 1) {
295 vst1_lane_f32(c1, vout1x01, 0);
296 vst1_lane_f32(c0, vout0x01, 0);
297 }
298 nc = 0;
299 }
300 } while (nc != 0);
301 }