Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x16-minmax-neon-mlal-lane.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
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_1x16__neon_mlal_lane(
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 <= 1);
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 const int8_t* a0 = a;
41 float* c0 = c;
42
43 do {
44 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
45 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
46 int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
47 int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
48 const int32x4_t vzp0 = vld1q_dup_s32(&quantization_params[0].zero_point);
49 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vzp0);
50 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vzp0);
51 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vzp0);
52 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vzp0);
53
54 size_t k = kc;
55 while (k >= 8 * sizeof(int8_t)) {
56 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
57 const int16x8_t vxa0 = vmovl_s8(va0);
58
59 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
60 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
61 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
62 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
63 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
64 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
74 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const int8_t*) w + 8;
75
76 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
77 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
78 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
79 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
80 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
81 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
82 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
83 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
84 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
85 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
86 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
87 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
88 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
89 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
90 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
91 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
92
93
94 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
95 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
96 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
97 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
98 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
99 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
100 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
101 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
102 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
103 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
104 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
105 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
106 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
107 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
108 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
109 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
110
111 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
112 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
113 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
114 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
115 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
116 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
117 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
118 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
119 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
120 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
121 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
122 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
123 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
124 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
125 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
126 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
127
128 k -= 8 * sizeof(int8_t);
129 }
130 if XNN_UNLIKELY(k != 0) {
131 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
132 const int16x8_t vxa0 = vmovl_s8(va0);
133
134 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
135 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
136
137 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
138 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
139
140 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
141 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
142 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
143 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
144
145 if (k >= 2 * sizeof(int8_t)) {
146 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
147 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
148 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
149 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
150
151 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
152 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
153 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
154 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
155
156 if (k > 2 * sizeof(int8_t)) {
157 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
158 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
160 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
161
162 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
163 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
164 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
165 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
166
167 if (k >= 4 * sizeof(int8_t)) {
168 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
169 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
170 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
171 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
172
173 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
174 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
175 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
176 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
177
178 if (k > 4 * sizeof(int8_t)) {
179 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
180 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
181 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
182 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
183
184 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
185 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
186 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
187 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
188
189 if (k >= 6 * sizeof(int8_t)) {
190 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
191 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
192 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
193 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
194
195 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
196 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
197 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
198 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
199
200 if (k > 6 * sizeof(int8_t)) {
201 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
202 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
203 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
204 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
205
206 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
207 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
208 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
209 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
210 }
211 }
212 }
213 }
214 }
215 }
216 }
217
218 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
219 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
220 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
221 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
222
223 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
224 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
225 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
226 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
227 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
228
229 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
230 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
231 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
232 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
233
234 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
235 #if XNN_ARCH_ARM64
236 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
237 #else
238 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
239 #endif
240 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
241 #if XNN_ARCH_ARM64
242 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
243 #else
244 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
245 #endif
246 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
247 #if XNN_ARCH_ARM64
248 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
249 #else
250 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
251 #endif
252 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
253 #if XNN_ARCH_ARM64
254 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
255 #else
256 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
257 #endif
258
259 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
260 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
261 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
262 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
263 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
264
265 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
266 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
267 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
268 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
269 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
270
271 if XNN_LIKELY(nc >= 16) {
272 vst1q_f32(&c0[0], vout0x0123);
273 vst1q_f32(&c0[4], vout0x4567);
274 vst1q_f32(&c0[8], vout0x89AB);
275 vst1q_f32(&c0[12], vout0xCDEF);
276
277 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
278
279 c0 = (float*) ((uintptr_t) c0 + cn_stride);
280
281 nc -= 16;
282 } else {
283 if (nc & 8) {
284 vst1q_f32(c0, vout0x0123); c0 += 4;
285 vout0x0123 = vout0x89AB;
286 vst1q_f32(c0, vout0x4567); c0 += 4;
287 vout0x4567 = vout0xCDEF;
288 }
289 if (nc & 4) {
290 vst1q_f32(c0, vout0x0123); c0 += 4;
291 vout0x0123 = vout0x4567;
292 }
293 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
294 if (nc & 2) {
295 vst1_f32(c0, vout0x01); c0 += 2;
296 vout0x01 = vget_high_f32(vout0x0123);
297 }
298 if (nc & 1) {
299 vst1_lane_f32(c0, vout0x01, 0);
300 }
301 nc = 0;
302 }
303 } while (nc != 0);
304 }