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
xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2);
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 do {
55 v128_t vacc0x0123 = wasm_v128_load(w);
56 v128_t vacc1x0123 = vacc0x0123;
57 v128_t vacc2x0123 = vacc0x0123;
58 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t p = ks;
61 do {
62 const uint8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const uint8_t* restrict a1 = a[1];
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 const uint8_t* restrict a2 = a[2];
71 if XNN_UNPREDICTABLE(a2 != zero) {
72 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
73 }
74 a += 3;
75
76 size_t k = kc;
77 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
78 while (k >= 8 * sizeof(uint8_t)) {
79 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
80 a0 += 8;
81 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
82 a1 += 8;
83 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
84 a2 += 8;
85
86 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
87
88 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
89 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
90 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
91 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
92 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
93 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
94 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
95
96 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
97 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
98 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
99 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
100 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
101 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
102 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
103
104 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
105 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
106 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
107 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
108 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
109 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
110 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
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
119 w = (const void*) ((const uint8_t*) w + 32);
120 k -= 8 * sizeof(uint8_t);
121 }
122 if (k != 0) {
123 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
124 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
125 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
126 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
127 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
128 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
129
130 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
131 w = (const void*) ((const uint8_t*) w + 8);
132
133 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
134 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
135 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
136 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
137 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
138 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
139
140 if (k > 2 * sizeof(uint8_t)) {
141 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
142 w = (const void*) ((const uint8_t*) w + 8);
143
144 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
145 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
146 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
147 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
148 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
149 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
150
151 if (k > 4 * sizeof(uint8_t)) {
152 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
153 w = (const void*) ((const uint8_t*) w + 8);
154
155 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
156 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
157 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
158 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
159 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
160 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
161 }
162 }
163 }
164 p -= 3 * sizeof(void*);
165 } while (p != 0);
166
167 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
168 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
169 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
170
171 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
172 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
173 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
174 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
175
176 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
177 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
178 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
179 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
180
181 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
182 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
183 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
184 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
185
186 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
187 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
188 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
189 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
190
191 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
192 v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
193
194 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
195
196 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
197 vout = wasm_u8x16_min(vout, voutput_max);
198
199 if (nc >= 4) {
200 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
201 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
202 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
203
204 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
205 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
206 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
207
208 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
209
210 nc -= 4;
211 } else {
212 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
213 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
214 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
215 if (nc & 2) {
216 *((uint16_t*) c2) = (uint16_t) vout2;
217 vout2 >>= 16;
218 c2 += 2;
219 *((uint16_t*) c1) = (uint16_t) vout1;
220 vout1 >>= 16;
221 c1 += 2;
222 *((uint16_t*) c0) = (uint16_t) vout0;
223 vout0 >>= 16;
224 c0 += 2;
225 }
226 if (nc & 1) {
227 *c2 = (uint8_t) vout2;
228 *c1 = (uint8_t) vout1;
229 *c0 = (uint8_t) vout0;
230 }
231
232 nc = 0;
233 }
234 } while (nc != 0);
235 }
236