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_qu8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_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(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2);
41 const uint8_t* a0 = a;
42 uint8_t* c0 = c;
43 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
50 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
56 uint8_t* c3 = (uint8_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 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
71 while (k >= 8 * sizeof(uint8_t)) {
72 const v128_t vxa0 = wasm_u16x8_load8x8((const v128_t*) a0);
73 a0 += 8;
74 const v128_t vxa1 = wasm_u16x8_load8x8((const v128_t*) a1);
75 a1 += 8;
76 const v128_t vxa2 = wasm_u16x8_load8x8((const v128_t*) a2);
77 a2 += 8;
78 const v128_t vxa3 = wasm_u16x8_load8x8((const v128_t*) a3);
79 a3 += 8;
80
81 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
82
83 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
84 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
85 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
86 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
87 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
88 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
89 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
90 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
91 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
92
93 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
94 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
95 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
96 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
97 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
98 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
99 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
100 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
101 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
102
103 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
104 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
105 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
106 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
107 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
108 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
109 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
110 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
111 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
112
113 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
114 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
115 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
116 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
117 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
118 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
119 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
120 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 3, 3, 3, 3), vxb3));
121
122 w = (const void*) ((const uint8_t*) w + 32);
123 k -= 8 * sizeof(uint8_t);
124 }
125 if (k != 0) {
126 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
127 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
128 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
129 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
130 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
131 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
132 const v128_t vxa3 = wasm_u16x8_load8x8(a3);
133 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
134
135 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
136 w = (const void*) ((const uint8_t*) w + 8);
137
138 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
139 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
140 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
141 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
142 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
143 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
144 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
145 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
146
147 if (k > 2 * sizeof(uint8_t)) {
148 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
149 w = (const void*) ((const uint8_t*) w + 8);
150
151 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
152 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
153 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
154 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
155 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
156 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
157 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
158 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
159
160 if (k > 4 * sizeof(uint8_t)) {
161 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
162 w = (const void*) ((const uint8_t*) w + 8);
163
164 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
165 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
166 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
167 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
168 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
169 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
170 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
171 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
172 }
173 }
174 }
175
176 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
177 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
178 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
179 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
180
181 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
182 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
183 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
184 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
185 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale);
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_u8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
209
210 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
211 vout = wasm_u8x16_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 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
220 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
221 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
222 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
223
224 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
225 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
226 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
227 a3 = (const uint8_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 = (uint8_t) vout0;
251 *c1 = (uint8_t) vout1;
252 *c2 = (uint8_t) vout2;
253 *c3 = (uint8_t) vout3;
254 }
255
256 nc = 0;
257 }
258 } while (nc != 0);
259 }
260