Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x4c2-minmax-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_3x4c2__wasmsimd_dot16x2_ld128(
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 <= 3);
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, 2 * 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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55
56 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
57 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
58 XNN_FORCE_REALIZATION(vmin);
59 XNN_FORCE_REALIZATION(vmax);
60
61
62 do {
63 const v128_t vksum0123 = wasm_v128_load(w);
64 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
65 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
66 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
67 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
68 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
69 v128_t vacc2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
70 w = (const int32_t*) w + 4;
71
72 size_t k = kc;
73 while (k >= 8 * sizeof(int8_t)) {
74 const v128_t vxa0 = wasm_i16x8_load8x8((const v128_t*) a0);
75 a0 += 8;
76 const v128_t vxa1 = wasm_i16x8_load8x8((const v128_t*) a1);
77 a1 += 8;
78 const v128_t vxa2 = wasm_i16x8_load8x8((const v128_t*) a2);
79 a2 += 8;
80
81 const v128_t vb01 = wasm_v128_load(w);
82 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
83 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
84
85 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
86 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
87 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
88 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
89 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
90 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
91
92 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
93 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
94 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
95 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
96 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
97 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
98 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
99 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
100 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
101
102 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
103 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
104 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
105 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
106 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
107 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
108
109 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
110 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
111 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
112 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
113 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
114 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
115
116 w = (const int8_t*) w + 32;
117 k -= 8 * sizeof(int8_t);
118 }
119 if (k != 0) {
120 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
121 a0 = (const int8_t*) ((uintptr_t) a0 + k);
122 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
123 a1 = (const int8_t*) ((uintptr_t) a1 + k);
124 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
125 a2 = (const int8_t*) ((uintptr_t) a2 + k);
126
127 const v128_t vxb0 = wasm_i16x8_load8x8(w);
128 w = (const int8_t*) w + 8;
129
130 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
131 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
132 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
133 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
134 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
135 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
136
137 if (k > 2 * sizeof(int8_t)) {
138 const v128_t vxb1 = wasm_i16x8_load8x8(w);
139 w = (const int8_t*) w + 8;
140
141 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
142 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
143 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
144 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
145 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
146 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
147
148 if (k > 4 * sizeof(int8_t)) {
149 const v128_t vxb2 = wasm_i16x8_load8x8(w);
150 w = (const int8_t*) w + 8;
151
152 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
153 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
154 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
155 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
156 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
157 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
158 }
159 }
160 }
161
162 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
163 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
164 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
165
166 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
167 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
168 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
169
170 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
171 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
172 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
173
174 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
175 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
176 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
177 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
178 w = (const float*) w + 4;
179
180 const v128_t vbias0123 = wasm_v128_load(w);
181 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
182 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
183 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
184 w = (const float*) w + 4;
185
186 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
187 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
188 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
189
190 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
191 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
192 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
193
194 if XNN_LIKELY(nc >= 4) {
195 wasm_v128_store(c0, vacc0x0123);
196 wasm_v128_store(c1, vacc1x0123);
197 wasm_v128_store(c2, vacc2x0123);
198
199 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
200 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
201 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
202
203 c0 = (float*) ((uintptr_t) c0 + cn_stride);
204 c1 = (float*) ((uintptr_t) c1 + cn_stride);
205 c2 = (float*) ((uintptr_t) c2 + cn_stride);
206
207 nc -= 4;
208 } else {
209 if (nc & 2) {
210 wasm_v128_store64_lane(c0, vacc0x0123, 0);
211 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
212 c0 += 2;
213 wasm_v128_store64_lane(c1, vacc1x0123, 0);
214 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
215 c1 += 2;
216 wasm_v128_store64_lane(c2, vacc2x0123, 0);
217 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
218 c2 += 2;
219 }
220 if (nc & 1) {
221 wasm_v128_store32_lane(c0, vacc0x0123, 0);
222 wasm_v128_store32_lane(c1, vacc1x0123, 0);
223 wasm_v128_store32_lane(c2, vacc2x0123, 0);
224 }
225 nc = 0;
226 }
227 } while (nc != 0);
228 }