Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-1x32c4-minmax-neondotfp16arith.c

raw
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/math.h"
16
17
18 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_1x32c4__neondotfp16arith(
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 <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * 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 kc = round_up_po2(kc, 4 * sizeof(int8_t));
46 uint16_t* c0 = (uint16_t*) c;
47
48 do {
49 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
50 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
52 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
54 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
55 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
56 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
57 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
58 const int32x4_t vksumGHIJ = vld1q_s32(w); w = (const int32_t*) w + 4;
59 int32x4_t vacc0xGHIJ = vmulq_s32(vksumGHIJ, vinput_zero_point);
60 const int32x4_t vksumKLMN = vld1q_s32(w); w = (const int32_t*) w + 4;
61 int32x4_t vacc0xKLMN = vmulq_s32(vksumKLMN, vinput_zero_point);
62 const int32x4_t vksumOPQR = vld1q_s32(w); w = (const int32_t*) w + 4;
63 int32x4_t vacc0xOPQR = vmulq_s32(vksumOPQR, vinput_zero_point);
64 const int32x4_t vksumSTUV = vld1q_s32(w); w = (const int32_t*) w + 4;
65 int32x4_t vacc0xSTUV = vmulq_s32(vksumSTUV, vinput_zero_point);
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72 } else {
73 a0 = zero_data;
74 }
75 a += 1;
76
77 // Inner accumulation loop along the 32 columns.
78 size_t k = kc;
79 // 2x partial unrolled loop to load 8 bytes at a time.
80 while (k >= 8 * sizeof(int8_t)) {
81 // Load a 1x8 block of activations.
82 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
83
84 // Load a 8x32 block of weights.
85 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
86 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
87 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
88 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
89 const int8x16_t vb0123xGHIJ = vld1q_s8(w); w = (const int8_t*) w + 16;
90 const int8x16_t vb0123xKLMN = vld1q_s8(w); w = (const int8_t*) w + 16;
91 const int8x16_t vb0123xOPQR = vld1q_s8(w); w = (const int8_t*) w + 16;
92 const int8x16_t vb0123xSTUV = vld1q_s8(w); w = (const int8_t*) w + 16;
93 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
94 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
95 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
96 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
97 const int8x16_t vb4567xGHIJ = vld1q_s8(w); w = (const int8_t*) w + 16;
98 const int8x16_t vb4567xKLMN = vld1q_s8(w); w = (const int8_t*) w + 16;
99 const int8x16_t vb4567xOPQR = vld1q_s8(w); w = (const int8_t*) w + 16;
100 const int8x16_t vb4567xSTUV = vld1q_s8(w); w = (const int8_t*) w + 16;
101
102 // Multiply-accumulate: 1x8 * 8x32 --> 1x32.
103 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
104 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
105 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
106 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
107 vacc0xGHIJ = vdotq_lane_s32(vacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
108 vacc0xKLMN = vdotq_lane_s32(vacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
109 vacc0xOPQR = vdotq_lane_s32(vacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
110 vacc0xSTUV = vdotq_lane_s32(vacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
111 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
112 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
113 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
114 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
115 vacc0xGHIJ = vdotq_lane_s32(vacc0xGHIJ, vb4567xGHIJ, va0x01234567, 1);
116 vacc0xKLMN = vdotq_lane_s32(vacc0xKLMN, vb4567xKLMN, va0x01234567, 1);
117 vacc0xOPQR = vdotq_lane_s32(vacc0xOPQR, vb4567xOPQR, va0x01234567, 1);
118 vacc0xSTUV = vdotq_lane_s32(vacc0xSTUV, vb4567xSTUV, va0x01234567, 1);
119
120 k -= 8 * sizeof(int8_t);
121 }
122 // Handle up to 4 final positions of `k`
123 if XNN_UNLIKELY(k != 0) {
124 // Load a 1x4 block of activations.
125 const int8x8_t va0x01234567 = vld1_s8(a0);
126
127 // Load a 4x32 block of weights.
128 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
129 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
130 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
131 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
132 const int8x16_t vb0123xGHIJ = vld1q_s8(w); w = (const int8_t*) w + 16;
133 const int8x16_t vb0123xKLMN = vld1q_s8(w); w = (const int8_t*) w + 16;
134 const int8x16_t vb0123xOPQR = vld1q_s8(w); w = (const int8_t*) w + 16;
135 const int8x16_t vb0123xSTUV = vld1q_s8(w); w = (const int8_t*) w + 16;
136
137 // Multiply-accumulate: 1x4 * 4x32 --> 1x32.
138 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
139 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
140 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
141 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
142 vacc0xGHIJ = vdotq_lane_s32(vacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
143 vacc0xKLMN = vdotq_lane_s32(vacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
144 vacc0xOPQR = vdotq_lane_s32(vacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
145 vacc0xSTUV = vdotq_lane_s32(vacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
146 }
147 p -= 1 * sizeof(void*);
148 } while (p != 0);
149
150 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
151 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
152 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
153 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
154 float32x4_t vout0xGHIJ = vcvtq_f32_s32(vacc0xGHIJ);
155 float32x4_t vout0xKLMN = vcvtq_f32_s32(vacc0xKLMN);
156 float32x4_t vout0xOPQR = vcvtq_f32_s32(vacc0xOPQR);
157 float32x4_t vout0xSTUV = vcvtq_f32_s32(vacc0xSTUV);
158 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
159 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
160 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
161 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
162 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
163 vout0xGHIJ = vmulq_f32(vout0xGHIJ, vinput_scale);
164 vout0xKLMN = vmulq_f32(vout0xKLMN, vinput_scale);
165 vout0xOPQR = vmulq_f32(vout0xOPQR, vinput_scale);
166 vout0xSTUV = vmulq_f32(vout0xSTUV, vinput_scale);
167
168 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
169 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
170 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
171 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
172 const float32x4_t vfilter_output_scaleGHIJ = vld1q_f32(w); w = (const float*) w + 4;
173 const float32x4_t vfilter_output_scaleKLMN = vld1q_f32(w); w = (const float*) w + 4;
174 const float32x4_t vfilter_output_scaleOPQR = vld1q_f32(w); w = (const float*) w + 4;
175 const float32x4_t vfilter_output_scaleSTUV = vld1q_f32(w); w = (const float*) w + 4;
176
177 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
178 #if XNN_ARCH_ARM64
179 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
180 #else
181 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
182 #endif
183 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
184 #if XNN_ARCH_ARM64
185 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
186 #else
187 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
188 #endif
189 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
190 #if XNN_ARCH_ARM64
191 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
192 #else
193 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
194 #endif
195 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
196 #if XNN_ARCH_ARM64
197 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
198 #else
199 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
200 #endif
201 const float32x4_t vbiasGHIJ = vld1q_f32(w); w = (const float*) w + 4;
202 #if XNN_ARCH_ARM64
203 vout0xGHIJ = vfmaq_f32(vbiasGHIJ, vout0xGHIJ, vfilter_output_scaleGHIJ);
204 #else
205 vout0xGHIJ = vmlaq_f32(vbiasGHIJ, vout0xGHIJ, vfilter_output_scaleGHIJ);
206 #endif
207 const float32x4_t vbiasKLMN = vld1q_f32(w); w = (const float*) w + 4;
208 #if XNN_ARCH_ARM64
209 vout0xKLMN = vfmaq_f32(vbiasKLMN, vout0xKLMN, vfilter_output_scaleKLMN);
210 #else
211 vout0xKLMN = vmlaq_f32(vbiasKLMN, vout0xKLMN, vfilter_output_scaleKLMN);
212 #endif
213 const float32x4_t vbiasOPQR = vld1q_f32(w); w = (const float*) w + 4;
214 #if XNN_ARCH_ARM64
215 vout0xOPQR = vfmaq_f32(vbiasOPQR, vout0xOPQR, vfilter_output_scaleOPQR);
216 #else
217 vout0xOPQR = vmlaq_f32(vbiasOPQR, vout0xOPQR, vfilter_output_scaleOPQR);
218 #endif
219 const float32x4_t vbiasSTUV = vld1q_f32(w); w = (const float*) w + 4;
220 #if XNN_ARCH_ARM64
221 vout0xSTUV = vfmaq_f32(vbiasSTUV, vout0xSTUV, vfilter_output_scaleSTUV);
222 #else
223 vout0xSTUV = vmlaq_f32(vbiasSTUV, vout0xSTUV, vfilter_output_scaleSTUV);
224 #endif
225
226 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
227 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
228 float16x8_t vfp16out0xGHIJKLMN = vcombine_f16(vcvt_f16_f32(vout0xGHIJ), vcvt_f16_f32(vout0xKLMN));
229 float16x8_t vfp16out0xOPQRSTUV = vcombine_f16(vcvt_f16_f32(vout0xOPQR), vcvt_f16_f32(vout0xSTUV));
230
231 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
232 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
233 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
234 vfp16out0xGHIJKLMN = vmaxq_f16(vfp16out0xGHIJKLMN, voutput_min);
235 vfp16out0xOPQRSTUV = vmaxq_f16(vfp16out0xOPQRSTUV, voutput_min);
236 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
237 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
238 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
239 vfp16out0xGHIJKLMN = vminq_f16(vfp16out0xGHIJKLMN, voutput_max);
240 vfp16out0xOPQRSTUV = vminq_f16(vfp16out0xOPQRSTUV, voutput_max);
241 if XNN_LIKELY(nc >= 32) {
242 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
243 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
244 vst1q_u16(c0 + 16, vreinterpretq_u16_f16(vfp16out0xGHIJKLMN));
245 vst1q_u16(c0 + 24, vreinterpretq_u16_f16(vfp16out0xOPQRSTUV));
246
247 a = (const int8_t**restrict) ((uintptr_t) a - ks);
248
249 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
250
251 nc -= 32;
252 } else {
253 if (nc & 16) {
254 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
255 vfp16out0x01234567 = vfp16out0xGHIJKLMN;
256 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x89ABCDEF)); c0 += 8;
257 vfp16out0x89ABCDEF = vfp16out0xOPQRSTUV;
258 }
259 if (nc & 8) {
260 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
261 vfp16out0x01234567 = vfp16out0x89ABCDEF;
262 }
263 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
264 if (nc & 4) {
265 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
266 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
267 }
268 if (nc & 2) {
269 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
270 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
271 }
272 if (nc & 1) {
273 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
274 }
275 nc = 0;
276 }
277 } while (nc != 0);
278 }