Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x16c4-minmax-neondot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_2x16c4__neondot(
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 kc = round_up_po2(kc, 4 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49
50 // Loop over groups of 16 columns.
51 do {
52 // Initialize accumulators with bias. 16 bias values are loaded from the
53 // weight matrix, at the start of the group of 16 columns.
54 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
55 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
56 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
57 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
58 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
59 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
60 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
61 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
62 int32x4_t vacc0x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point01), 0);
63 int32x4_t vacc1x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point01), 0);
64 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
65 int32x4_t vacc0xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point01), 0);
66 int32x4_t vacc1xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point01), 0);
67
68 // Inner accumulation loop along the 16 columns.
69 size_t k = kc;
70 // 2x partial unrolled loop to load 8 bytes at a time.
71 while (k >= 8 * sizeof(int8_t)) {
72 // Load a 2x8 block of activations.
73 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
74 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
75
76 // Load a 8x16 block of weights.
77 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
78 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
79 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
80 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
81 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
82 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
83 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
84 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
85
86 // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
87 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
88 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
89 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
90 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
91 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
92 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
93 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
94 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
95 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
96 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
97 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
98 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
99 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
100 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
101 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
102 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
103
104 k -= 8 * sizeof(int8_t);
105 }
106 // Handle up to 4 final positions of `k`
107 if XNN_UNLIKELY(k != 0) {
108 // Load a 2x4 block of activations.
109 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
110 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
111
112 // Load a 4x16 block of weights.
113 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
114 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
115 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
116 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
117
118 // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
119 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
120 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
121 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
122 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
123 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
124 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
125 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
126 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
127 }
128
129 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
130 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
131 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
132 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
133 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
134 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
135 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
136 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
137 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
138 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
139 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
140 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
141 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
142 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
143 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
144 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
145 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
146
147 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
148 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
149 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
150 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
151
152 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
153 #if XNN_ARCH_ARM64
154 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
155 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
156 #else
157 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
158 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
159 #endif
160 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
161 #if XNN_ARCH_ARM64
162 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
163 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
164 #else
165 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
166 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
167 #endif
168 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
169 #if XNN_ARCH_ARM64
170 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
171 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
172 #else
173 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
174 vout1x89AB = vmlaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
175 #endif
176 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
177 #if XNN_ARCH_ARM64
178 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
179 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
180 #else
181 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
182 vout1xCDEF = vmlaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
183 #endif
184
185 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
186 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
187 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
188 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
189 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
190 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
191 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
192 vout1x89AB = vmaxq_f32(vout1x89AB, voutput_min);
193 vout1xCDEF = vmaxq_f32(vout1xCDEF, voutput_min);
194
195 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
196 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
197 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
198 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
199 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
200 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
201 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
202 vout1x89AB = vminq_f32(vout1x89AB, voutput_max);
203 vout1xCDEF = vminq_f32(vout1xCDEF, voutput_max);
204
205 if XNN_LIKELY(nc >= 16) {
206 vst1q_f32(c0, vout0x0123);
207 vst1q_f32(c0 + 4, vout0x4567);
208 vst1q_f32(c0 + 8, vout0x89AB);
209 vst1q_f32(c0 + 12, vout0xCDEF);
210 vst1q_f32(c1, vout1x0123);
211 vst1q_f32(c1 + 4, vout1x4567);
212 vst1q_f32(c1 + 8, vout1x89AB);
213 vst1q_f32(c1 + 12, vout1xCDEF);
214
215 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
216 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
217
218 c0 = (float*) ((uintptr_t) c0 + cn_stride);
219 c1 = (float*) ((uintptr_t) c1 + cn_stride);
220
221 nc -= 16;
222 } else {
223 if (nc & 8) {
224 vst1q_f32(c0, vout0x0123); c0 += 4;
225 vout0x0123 = vout0x89AB;
226 vst1q_f32(c1, vout1x0123); c1 += 4;
227 vout1x0123 = vout1x89AB;
228 vst1q_f32(c0, vout0x4567); c0 += 4;
229 vout0x4567 = vout0xCDEF;
230 vst1q_f32(c1, vout1x4567); c1 += 4;
231 vout1x4567 = vout1xCDEF;
232 }
233 if (nc & 4) {
234 vst1q_f32(c0, vout0x0123); c0 += 4;
235 vout0x0123 = vout0x4567;
236 vst1q_f32(c1, vout1x0123); c1 += 4;
237 vout1x0123 = vout1x4567;
238 }
239 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
240 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
241 if (nc & 2) {
242 vst1_f32(c0, vout0x01); c0 += 2;
243 vst1_f32(c1, vout1x01); c1 += 2;
244 vout0x01 = vget_high_f32(vout0x0123);
245 vout1x01 = vget_high_f32(vout1x0123);
246 }
247 if (nc & 1) {
248 vst1_lane_f32(c0, vout0x01, 0);
249 vst1_lane_f32(c1, vout1x01, 0);
250 }
251 nc = 0;
252 }
253 } while (nc != 0);
254 }