Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-2x8-minmax-neon-mlal-lane.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
18
19 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_2x8__neon_mlal_lane(
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 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const int8_t* zero_data,
32 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 2);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (2 * sizeof(void*)) == 0);
41 assert(a_offset % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr != 2) {
49 c1 = c0;
50 }
51
52 do {
53 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
54 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
55 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
56 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
57 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
58 int32x4_t vacc1x0123 = vacc0x0123;
59 int32x4_t vacc1x4567 = vacc0x4567;
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 } else {
67 a0 = zero_data;
68 }
69 const int8_t* restrict a1 = a[1];
70 if XNN_UNPREDICTABLE(a1 != zero) {
71 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
72 } else {
73 a1 = zero_data;
74 }
75 a += 2;
76
77 size_t k = kc;
78 while (k >= 8 * sizeof(int8_t)) {
79 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
80 const int16x8_t vxa0 = vmovl_s8(va0);
81 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
82 const int16x8_t vxa1 = vmovl_s8(va1);
83
84 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
85 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
86
87 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
88 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
89 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
90 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
91 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
93
94 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
95 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
96 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
97 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
98 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
99 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
100
101 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
102 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
103 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
104 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
105 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
106 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
107
108 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
109 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
110 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
111 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
112
113
114 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
115 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
116
117 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
118 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
119 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
120 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
121 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
122 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
123
124 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
125 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
126 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
127 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
128 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
129 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
130
131 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
132 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
133 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
134 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
135 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
136 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
137
138 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
139 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
140 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
141 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
142
143 k -= 8 * sizeof(int8_t);
144 }
145 if XNN_UNLIKELY(k != 0) {
146 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
147 const int16x8_t vxa0 = vmovl_s8(va0);
148 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
149 const int16x8_t vxa1 = vmovl_s8(va1);
150
151 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
152 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
153
154 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
155 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
156 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
157 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
158
159 if (k >= 2 * sizeof(int8_t)) {
160 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
162
163 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
164 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
165 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
166 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
167
168 if (k > 2 * sizeof(int8_t)) {
169 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
170 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
171
172 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
173 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
174 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
175 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
176
177 if (k >= 4 * sizeof(int8_t)) {
178 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
179 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
180
181 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
182 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
183 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
184 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
185
186 if (k > 4 * sizeof(int8_t)) {
187 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
188 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
189
190 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
191 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
192 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
193 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
194
195 if (k >= 6 * sizeof(int8_t)) {
196 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
197 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
198
199 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
200 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
201 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
202 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
203
204 if (k > 6 * sizeof(int8_t)) {
205 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
206 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
207
208 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
209 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
210 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
211 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
212 }
213 }
214 }
215 }
216 }
217 }
218 }
219 p -= 2 * sizeof(void*);
220 } while (p != 0);
221
222 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
223 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
224 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
225 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
226
227 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
228 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
229 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
230 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
231 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
232
233 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
234 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
235
236 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
237 #if XNN_ARCH_ARM64
238 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
239 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
240 #else
241 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
242 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
243 #endif
244 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
245 #if XNN_ARCH_ARM64
246 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
247 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
248 #else
249 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
250 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
251 #endif
252
253 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
254 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
255 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
256 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
257 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
258
259 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
260 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
261 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
262 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
263 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
264
265 if XNN_LIKELY(nc >= 8) {
266 vst1q_f32(&c1[0], vout1x0123);
267 vst1q_f32(&c1[4], vout1x4567);
268 vst1q_f32(&c0[0], vout0x0123);
269 vst1q_f32(&c0[4], vout0x4567);
270
271 a = (const int8_t**restrict) ((uintptr_t) a - ks);
272
273 c0 = (float*) ((uintptr_t) c0 + cn_stride);
274 c1 = (float*) ((uintptr_t) c1 + cn_stride);
275
276 nc -= 8;
277 } else {
278 if (nc & 4) {
279 vst1q_f32(c1, vout1x0123); c1 += 4;
280 vout1x0123 = vout1x4567;
281 vst1q_f32(c0, vout0x0123); c0 += 4;
282 vout0x0123 = vout0x4567;
283 }
284 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
285 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
286 if (nc & 2) {
287 vst1_f32(c1, vout1x01); c1 += 2;
288 vst1_f32(c0, vout0x01); c0 += 2;
289 vout1x01 = vget_high_f32(vout1x0123);
290 vout0x01 = vget_high_f32(vout0x0123);
291 }
292 if (nc & 1) {
293 vst1_lane_f32(c1, vout1x01, 0);
294 vst1_lane_f32(c0, vout0x01, 0);
295 }
296 nc = 0;
297 }
298 } while (nc != 0);
299 }