Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-2x8c2s4-minmax-neonfp16arith-mlal.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_f16_qc8w_igemm_minmax_ukernel_2x8c2s4__neonfp16arith(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 2);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (2 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 uint16_t* c0 = (uint16_t*) c;
46 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr != 2) {
48 c1 = c0;
49 }
50
51 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
52 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
53 kc = round_up_po2(kc, 8 * sizeof(int8_t));
54 do {
55 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
56 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
57 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
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 >= 16 * sizeof(int8_t)) {
80 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
81 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
82 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
83 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
84
85 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93
94 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
95 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
96 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
97 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
98 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
99 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
100 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
101 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
102 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
103 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
104 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
105 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
106 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
107 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
108 va0x0 = vext_s8(va0x0, va0x0, 2);
109 va0x1 = vext_s8(va0x1, va0x1, 2);
110 va1x0 = vext_s8(va1x0, va1x0, 2);
111 va1x1 = vext_s8(va1x1, va1x1, 2);
112 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
113 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
114 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
115 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
116 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
117 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
118 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
119 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
120 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
121 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
122 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
123 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
124 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
125 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
126 va0x0 = vext_s8(va0x0, va0x0, 2);
127 va0x1 = vext_s8(va0x1, va0x1, 2);
128 va1x0 = vext_s8(va1x0, va1x0, 2);
129 va1x1 = vext_s8(va1x1, va1x1, 2);
130 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
131 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
132 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
133 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
134 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
135 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
136 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
137 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
138 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
139 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
140 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
141 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
142 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
143 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
144 va0x0 = vext_s8(va0x0, va0x0, 2);
145 va0x1 = vext_s8(va0x1, va0x1, 2);
146 va1x0 = vext_s8(va1x0, va1x0, 2);
147 va1x1 = vext_s8(va1x1, va1x1, 2);
148 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
149 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
150 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
151 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
152 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
153 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
154 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
155 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
156 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
157 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
158 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
159 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
160 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
161 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
162
163 k -= 16 * sizeof(int8_t);
164 }
165 if (k != 0) {
166 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
167 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
168
169 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
170 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
171 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
172 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
173 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
174 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
175 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
176 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
177
178 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
179 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
180 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
181 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
182 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
183 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
184 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
185 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
186 va0x0 = vext_s8(va0x0, va0x0, 2);
187 va1x0 = vext_s8(va1x0, va1x0, 2);
188 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
189 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
190 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
191 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
192 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
193 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
194 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
195 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
196 va0x0 = vext_s8(va0x0, va0x0, 2);
197 va1x0 = vext_s8(va1x0, va1x0, 2);
198 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
199 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
200 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
201 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
202 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
203 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
204 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
205 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
206 va0x0 = vext_s8(va0x0, va0x0, 2);
207 va1x0 = vext_s8(va1x0, va1x0, 2);
208 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
209 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
210 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
211 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
212 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
213 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
214 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
215 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
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 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
228 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
229 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
230 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
231
232 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
233 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
234
235 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
236 #if XNN_ARCH_ARM64
237 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
238 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
239 #else
240 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
241 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
242 #endif
243 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
244 #if XNN_ARCH_ARM64
245 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
246 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
247 #else
248 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
249 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
250 #endif
251
252 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
253 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
254 #if XNN_ARCH_ARM64
255 const uint16x8x2_t voutput_minmax = vld2q_dup_u16((const uint16_t*) &params->scalar.min);
256 const float16x8_t voutput_min = vreinterpretq_f16_u16(voutput_minmax.val[0]);
257 const float16x8_t voutput_max = vreinterpretq_f16_u16(voutput_minmax.val[1]);
258 #else
259 // vld2_dup is to work around aarch32 clang bug with vld1q_dup
260 const uint16x4x2_t vminmax = vld2_dup_u16((const uint16_t*) &params->scalar.min);
261 const float16x8_t voutput_min = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[0], vminmax.val[0]));
262 const float16x8_t voutput_max = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[1], vminmax.val[1]));
263 #endif
264 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
265 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
266 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
267 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
268 if XNN_LIKELY(nc >= 8) {
269 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
270 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
271
272 a = (const int8_t**restrict) ((uintptr_t) a - ks);
273
274 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
275 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
276
277 nc -= 8;
278 } else {
279 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
280 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
281 if (nc & 4) {
282 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
283 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
284 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
285 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
286 }
287 if (nc & 2) {
288 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
289 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
290 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
291 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
292 }
293 if (nc & 1) {
294 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
295 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
296 }
297 nc = 0;
298 }
299 } while (nc != 0);
300 }