Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x4c2-minmax-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-wasmsimd-dot16x2.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 <wasm_simd128.h>
13
14 #include "xnnpack/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_3x4c2__wasmsimd_dot16x2_ld64(
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 float* 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_f32_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 <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (3 * 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, 2 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55
56 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
57 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
58 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
59 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
60 XNN_FORCE_REALIZATION(vmin);
61 XNN_FORCE_REALIZATION(vmax);
62
63
64 do {
65 const v128_t vksum0123 = wasm_v128_load(w);
66 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
67 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
68 v128_t vacc2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
69 w = (const void*) ((const int32_t*) w + 4);
70
71 size_t p = ks;
72 do {
73 const int8_t* restrict a0 = a[0];
74 if XNN_UNPREDICTABLE(a0 != zero) {
75 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
76 } else {
77 a0 = zero_data;
78 }
79 const int8_t* restrict a1 = a[1];
80 if XNN_UNPREDICTABLE(a1 != zero) {
81 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
82 } else {
83 a1 = zero_data;
84 }
85 const int8_t* restrict a2 = a[2];
86 if XNN_UNPREDICTABLE(a2 != zero) {
87 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
88 } else {
89 a2 = zero_data;
90 }
91 a += 3;
92
93 size_t k = kc;
94 while (k >= 8 * sizeof(int8_t)) {
95 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
96 a0 += 8;
97 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
98 a1 += 8;
99 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
100 a2 += 8;
101
102 const v128_t vxb0 = wasm_i16x8_load8x8(w);
103
104 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
105 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
106 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
107 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
108 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
109 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
110 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
111
112 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
113 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
114 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
115 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
116 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
117 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
118 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
119
120 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
121 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
122 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
123 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
124 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
125 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
126 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
127
128 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
129 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
130 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
131 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
132 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
133 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
134
135 w = (const void*) ((const int8_t*) w + 32);
136 k -= 8 * sizeof(int8_t);
137 }
138 if (k != 0) {
139 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
140 a0 = (const int8_t*) ((uintptr_t) a0 + k);
141 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
142 a1 = (const int8_t*) ((uintptr_t) a1 + k);
143 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
144 a2 = (const int8_t*) ((uintptr_t) a2 + k);
145
146 const v128_t vxb0 = wasm_i16x8_load8x8(w);
147 w = (const int8_t*) w + 8;
148
149 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
150 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
151 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
152 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
153 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
154 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
155
156 if (k > 2 * sizeof(int8_t)) {
157 const v128_t vxb1 = wasm_i16x8_load8x8(w);
158 w = (const int8_t*) w + 8;
159
160 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
161 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
162 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
163 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
164 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
165 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
166
167 if (k > 4 * sizeof(int8_t)) {
168 const v128_t vxb2 = wasm_i16x8_load8x8(w);
169 w = (const int8_t*) w + 8;
170
171 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
172 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
173 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
174 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
175 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
176 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
177 }
178 }
179 }
180 p -= 3 * sizeof(void*);
181 } while (p != 0);
182
183 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
184 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
185 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
186
187 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
188 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
189 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
190
191 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
192 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
193 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
194 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
195 w = (const float*) w + 4;
196
197 const v128_t vbias0123 = wasm_v128_load(w);
198 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
199 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
200 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
201 w = (const float*) w + 4;
202
203 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
204 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
205 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
206
207 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
208 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
209 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
210
211 if XNN_LIKELY(nc >= 4) {
212 wasm_v128_store(c2, vacc2x0123);
213 wasm_v128_store(c1, vacc1x0123);
214 wasm_v128_store(c0, vacc0x0123);
215
216 a = (const int8_t**restrict) ((uintptr_t) a - ks);
217
218 c0 = (float*) ((uintptr_t) c0 + cn_stride);
219 c1 = (float*) ((uintptr_t) c1 + cn_stride);
220 c2 = (float*) ((uintptr_t) c2 + cn_stride);
221
222 nc -= 4;
223 } else {
224 if (nc & 2) {
225 wasm_v128_store64_lane(c2, vacc2x0123, 0);
226 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
227 c2 += 2;
228 wasm_v128_store64_lane(c1, vacc1x0123, 0);
229 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
230 c1 += 2;
231 wasm_v128_store64_lane(c0, vacc0x0123, 0);
232 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
233 c0 += 2;
234 }
235 if (nc & 1) {
236 wasm_v128_store32_lane(c2, vacc2x0123, 0);
237 wasm_v128_store32_lane(c1, vacc1x0123, 0);
238 wasm_v128_store32_lane(c0, vacc0x0123, 0);
239 }
240 nc = 0;
241 }
242 } while (nc != 0);
243 }