Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x16-minmax-neon-mlal-lane-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/prefetch.h"
17
18
19 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_1x16__neon_mlal_lane_prfm(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 const int8_t* a0 = a;
42 float* c0 = c;
43
44 do {
45 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
46 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
47 int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
48 int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
49 const int32x4_t vzp0 = vld1q_dup_s32(&quantization_params[0].zero_point);
50 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vzp0);
51 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vzp0);
52 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vzp0);
53 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vzp0);
54
55 size_t k = kc;
56 while (k >= 8 * sizeof(int8_t)) {
57 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
58 const int16x8_t vxa0 = vmovl_s8(va0);
59
60 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
61 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
62 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
63 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
64 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
74 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
75 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const int8_t*) w + 8;
76
77 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
78 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
79 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
80 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
81 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
82 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
83 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
84 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
85 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
86 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
87 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
88 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
89 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
90 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
91 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
92 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
93
94 xnn_prefetch_to_l1((const int8_t*) w + 448);
95 xnn_prefetch_to_l1((const int8_t*) w + 512);
96
97 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
98 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
99 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
100 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
101 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
102 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
103 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
104 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
105 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
106 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
107 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
108 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
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 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
112 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
113
114 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
115 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
116 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
117 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
118 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
119 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
120 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
121 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
122 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
123 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
124 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
125 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
126 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
127 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
128 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
129 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
130
131 k -= 8 * sizeof(int8_t);
132 }
133 if XNN_UNLIKELY(k != 0) {
134 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
135 const int16x8_t vxa0 = vmovl_s8(va0);
136
137 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
138 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
139
140 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
141 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
142
143 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
144 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
145 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
146 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
147
148 if (k >= 2 * sizeof(int8_t)) {
149 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
150 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
151 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
152 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
153
154 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
155 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
156 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
157 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
158
159 if (k > 2 * sizeof(int8_t)) {
160 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
162 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
163 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
164
165 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
166 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
167 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
168 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
169
170 if (k >= 4 * sizeof(int8_t)) {
171 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
172 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
173 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
174 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
175
176 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
177 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
178 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
179 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
180
181 if (k > 4 * sizeof(int8_t)) {
182 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
183 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
184 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
185 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
186
187 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
188 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
189 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
190 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
191
192 if (k >= 6 * sizeof(int8_t)) {
193 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
194 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
195 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
196 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
197
198 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
199 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
200 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
201 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
202
203 if (k > 6 * sizeof(int8_t)) {
204 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
205 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
206 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
207 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
208
209 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
210 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
211 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
212 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
213 }
214 }
215 }
216 }
217 }
218 }
219 }
220
221 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
222 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
223 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
224 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
225
226 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
227 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
228 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
229 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
230 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
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 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
235 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
236
237 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
238 #if XNN_ARCH_ARM64
239 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
240 #else
241 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, 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 #else
247 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
248 #endif
249 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
250 #if XNN_ARCH_ARM64
251 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
252 #else
253 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
254 #endif
255 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
256 #if XNN_ARCH_ARM64
257 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
258 #else
259 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
260 #endif
261
262 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
263 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
264 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
265 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
266 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
267
268 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
269 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
270 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
271 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
272 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
273
274 if XNN_LIKELY(nc >= 16) {
275 vst1q_f32(&c0[0], vout0x0123);
276 vst1q_f32(&c0[4], vout0x4567);
277 vst1q_f32(&c0[8], vout0x89AB);
278 vst1q_f32(&c0[12], vout0xCDEF);
279
280 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
281
282 c0 = (float*) ((uintptr_t) c0 + cn_stride);
283
284 nc -= 16;
285 } else {
286 if (nc & 8) {
287 vst1q_f32(c0, vout0x0123); c0 += 4;
288 vout0x0123 = vout0x89AB;
289 vst1q_f32(c0, vout0x4567); c0 += 4;
290 vout0x4567 = vout0xCDEF;
291 }
292 if (nc & 4) {
293 vst1q_f32(c0, vout0x0123); c0 += 4;
294 vout0x0123 = vout0x4567;
295 }
296 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
297 if (nc & 2) {
298 vst1_f32(c0, vout0x01); c0 += 2;
299 vout0x01 = vget_high_f32(vout0x0123);
300 }
301 if (nc & 1) {
302 vst1_lane_f32(c0, vout0x01, 0);
303 }
304 nc = 0;
305 }
306 } while (nc != 0);
307 }