Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-3x4c8-minmax-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-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_qc4w_gemm_minmax_ukernel_3x4c8__wasmsimd_dot16x2_ld64(
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 struct xnn_f32_qc4w_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, 8 * 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 const v128_t vmask = wasm_i8x16_const_splat(0xF0);
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 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
68 const v128_t vinit1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
69 const v128_t vinit2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
70
71 const v128_t vzero = wasm_i32x4_const_splat(0);
72 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
73 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
74 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
75 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
76 v128_t vacc1x0 = wasm_v32x4_shuffle(vinit1x0123, vzero, 0, 5, 6, 7);
77 v128_t vacc1x1 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 1, 6, 7);
78 v128_t vacc1x2 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 2, 7);
79 v128_t vacc1x3 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 6, 3);
80 v128_t vacc2x0 = wasm_v32x4_shuffle(vinit2x0123, vzero, 0, 5, 6, 7);
81 v128_t vacc2x1 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 1, 6, 7);
82 v128_t vacc2x2 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 2, 7);
83 v128_t vacc2x3 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 6, 3);
84 w = (const int32_t*) w + 4;
85
86 size_t k = kc;
87 while (k >= 16 * sizeof(int8_t)) {
88 v128_t vxa0;
89 v128_t vxa1;
90 v128_t vxa2;
91 vxa0 = wasm_i16x8_load8x8(a0);
92 a0 += 8;
93 vxa1 = wasm_i16x8_load8x8(a1);
94 a1 += 8;
95 vxa2 = wasm_i16x8_load8x8(a2);
96 a2 += 8;
97 const v128_t vb0 = wasm_v128_load(w);
98 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 16);
99 const v128_t vbm0 = wasm_i8x16_shl(vb0, 4);
100 const v128_t vbm2 = wasm_i8x16_shl(vb2, 4);
101 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vbm0);
102 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vbm0);
103 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vbm2);
104 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vbm2);
105 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
106 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
107 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
108 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
109 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
110 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
111 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
112 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
113 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
114 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
115 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
116 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
117 vxa0 = wasm_i16x8_load8x8(a0);
118 a0 += 8;
119 vxa1 = wasm_i16x8_load8x8(a1);
120 a1 += 8;
121 vxa2 = wasm_i16x8_load8x8(a2);
122 a2 += 8;
123 const v128_t vbm4 = wasm_v128_and(vb0, vmask);
124 const v128_t vbm6 = wasm_v128_and(vb2, vmask);
125 const v128_t vxb4 = wasm_i16x8_extend_low_i8x16(vbm4);
126 const v128_t vxb5 = wasm_i16x8_extend_high_i8x16(vbm4);
127 const v128_t vxb6 = wasm_i16x8_extend_low_i8x16(vbm6);
128 const v128_t vxb7 = wasm_i16x8_extend_high_i8x16(vbm6);
129 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb4));
130 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb4));
131 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb4));
132 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb5));
133 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb5));
134 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb5));
135 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb6));
136 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb6));
137 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb6));
138 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb7));
139 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb7));
140 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb7));
141 w = (const int8_t*) w + 32;
142 k -= 16 * sizeof(int8_t);
143 }
144
145 while (k >= 8 * sizeof(int8_t)) {
146 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
147 a0 += 8;
148 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
149 a1 += 8;
150 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
151 a2 += 8;
152
153 const v128_t vb0 = wasm_v128_load(w);
154 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 16);
155 const v128_t vbm0 = wasm_i8x16_shl(vb0, 4);
156 const v128_t vbm2 = wasm_i8x16_shl(vb2, 4);
157 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vbm0);
158 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vbm0);
159 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vbm2);
160 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vbm2);
161
162 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
163 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
164 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
165
166 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
167 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
168 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
169
170 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
171 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
172 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
173
174 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
175 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
176 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
177
178 w = (const int8_t*) w + 32; // only low nibble used
179 k -= 8 * sizeof(int8_t);
180 };
181
182 const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
183 const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
184 const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
185 const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
186 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
187 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
188
189 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
190 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
191 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
192
193 vacc0x0123 = wasm_i32x4_shr(vacc0x0123, 4);
194 vacc1x0123 = wasm_i32x4_shr(vacc1x0123, 4);
195 vacc2x0123 = wasm_i32x4_shr(vacc2x0123, 4);
196 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
197 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
198 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
199
200 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
201 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
202 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
203
204 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
205 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
206 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
207
208 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
209 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
210 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
211 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
212 w = (const float*) w + 4;
213
214 const v128_t vbias0123 = wasm_v128_load(w);
215 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
216 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
217 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
218 w = (const float*) w + 4;
219
220 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
221 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
222 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
223
224 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
225 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
226 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
227
228 if XNN_LIKELY(nc >= 4) {
229 wasm_v128_store(c0, vacc0x0123);
230 wasm_v128_store(c1, vacc1x0123);
231 wasm_v128_store(c2, vacc2x0123);
232
233 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
234 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
235 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
236
237 c0 = (float*) ((uintptr_t) c0 + cn_stride);
238 c1 = (float*) ((uintptr_t) c1 + cn_stride);
239 c2 = (float*) ((uintptr_t) c2 + cn_stride);
240
241 nc -= 4;
242 } else {
243 if (nc & 2) {
244 wasm_v128_store64_lane(c0, vacc0x0123, 0);
245 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
246 c0 += 2;
247 wasm_v128_store64_lane(c1, vacc1x0123, 0);
248 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
249 c1 += 2;
250 wasm_v128_store64_lane(c2, vacc2x0123, 0);
251 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
252 c2 += 2;
253 }
254 if (nc & 1) {
255 wasm_v128_store32_lane(c0, vacc0x0123, 0);
256 wasm_v128_store32_lane(c1, vacc1x0123, 0);
257 wasm_v128_store32_lane(c2, vacc2x0123, 0);
258 }
259 nc = 0;
260 }
261 } while (nc != 0);
262 }