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
xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64(
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 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
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);
41 const int8_t* a0 = a;
42 int8_t* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 int8_t* c1 = (int8_t*) ((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 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr != 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61
62 do {
63 v128_t vacc0x0123 = wasm_v128_load(w);
64 v128_t vacc1x0123 = vacc0x0123;
65 v128_t vacc2x0123 = vacc0x0123;
66 v128_t vacc3x0123 = vacc0x0123;
67 w = (const void*) ((const int32_t*) w + 4);
68
69 size_t k = kc;
70 while (k >= 8 * sizeof(int8_t)) {
71 const v128_t vxa0 = wasm_i16x8_load8x8((const v128_t*) a0);
72 a0 += 8;
73 const v128_t vxa1 = wasm_i16x8_load8x8((const v128_t*) a1);
74 a1 += 8;
75 const v128_t vxa2 = wasm_i16x8_load8x8((const v128_t*) a2);
76 a2 += 8;
77 const v128_t vxa3 = wasm_i16x8_load8x8((const v128_t*) a3);
78 a3 += 8;
79
80 const v128_t vxb0 = wasm_i16x8_load8x8(w);
81
82 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
83 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
84 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
85 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
86 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
87 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
88 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
89 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
90 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
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 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
99 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
100 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
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 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
109 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
110 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
111
112 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
113 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
114 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
115 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
116 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
117 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
118 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
119 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 3, 3, 3, 3), vxb3));
120
121 w = (const void*) ((const int8_t*) w + 32);
122 k -= 8 * sizeof(int8_t);
123 }
124 if (k != 0) {
125 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
126 a0 = (const int8_t*) ((uintptr_t) a0 + k);
127 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
128 a1 = (const int8_t*) ((uintptr_t) a1 + k);
129 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
130 a2 = (const int8_t*) ((uintptr_t) a2 + k);
131 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
132 a3 = (const int8_t*) ((uintptr_t) a3 + k);
133
134 const v128_t vxb0 = wasm_i16x8_load8x8(w);
135 w = (const void*) ((const int8_t*) w + 8);
136
137 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
138 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
139 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
140 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
141 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
142 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
143 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
144 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
145
146 if (k > 2 * sizeof(int8_t)) {
147 const v128_t vxb1 = wasm_i16x8_load8x8(w);
148 w = (const void*) ((const int8_t*) w + 8);
149
150 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
151 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
152 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
153 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
154 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
155 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
156 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
157 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
158
159 if (k > 4 * sizeof(int8_t)) {
160 const v128_t vxb2 = wasm_i16x8_load8x8(w);
161 w = (const void*) ((const int8_t*) w + 8);
162
163 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
164 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
165 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
166 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
167 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
168 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
169 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
170 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
171 }
172 }
173 }
174
175 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
176 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
177 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
178 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
179
180 const v128_t vscale0123 = wasm_v128_load(w);
181 w = (const void*) ((const float*) w + 4);
182 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
183 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
184 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
185 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale0123);
186
187 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
188 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
189 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
190 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
191 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
192
193 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
194 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
195 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
196 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
197 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
198
199 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
200 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
201 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
202 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
203 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
204
205 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
206 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
207
208 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
209
210 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
211 vout = wasm_i8x16_min(vout, voutput_max);
212
213 if (nc >= 4) {
214 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
215 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
216 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
217 *((float*) c3) = (float) wasm_f32x4_extract_lane(vout, 3);
218
219 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
220 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
221 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
222 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
223
224 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
225 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
226 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
227 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
228
229 nc -= 4;
230 } else {
231 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
232 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
233 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
234 uint32_t vout3 = wasm_i32x4_extract_lane(vout, 3);
235 if (nc & 2) {
236 *((uint16_t*) c0) = (uint16_t) vout0;
237 vout0 >>= 16;
238 c0 += 2;
239 *((uint16_t*) c1) = (uint16_t) vout1;
240 vout1 >>= 16;
241 c1 += 2;
242 *((uint16_t*) c2) = (uint16_t) vout2;
243 vout2 >>= 16;
244 c2 += 2;
245 *((uint16_t*) c3) = (uint16_t) vout3;
246 vout3 >>= 16;
247 c3 += 2;
248 }
249 if (nc & 1) {
250 *c0 = (int8_t) vout0;
251 *c1 = (int8_t) vout1;
252 *c2 = (int8_t) vout2;
253 *c3 = (int8_t) vout3;
254 }
255
256 nc = 0;
257 }
258 } while (nc != 0);
259 }
260