1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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
xnn_qu8_igemm_minmax_fp32_ukernel_4x4c8__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_4x4c8__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 <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * 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, 8);
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 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
59 do {
60 v128_t vacc0x0 = wasm_v128_load32_zero(w);
61 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
62 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
63 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
64 v128_t vacc1x0 = vacc0x0;
65 v128_t vacc1x1 = vacc0x1;
66 v128_t vacc1x2 = vacc0x2;
67 v128_t vacc1x3 = vacc0x3;
68 v128_t vacc2x0 = vacc0x0;
69 v128_t vacc2x1 = vacc0x1;
70 v128_t vacc2x2 = vacc0x2;
71 v128_t vacc2x3 = vacc0x3;
72 v128_t vacc3x0 = vacc0x0;
73 v128_t vacc3x1 = vacc0x1;
74 v128_t vacc3x2 = vacc0x2;
75 v128_t vacc3x3 = vacc0x3;
76 w = (const void*) ((const int32_t*) w + 4);
77
78 size_t p = ks;
79 do {
80 const uint8_t* restrict a0 = a[0];
81 if XNN_UNPREDICTABLE(a0 != zero) {
82 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
83 }
84 const uint8_t* restrict a1 = a[1];
85 if XNN_UNPREDICTABLE(a1 != zero) {
86 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
87 }
88 const uint8_t* restrict a2 = a[2];
89 if XNN_UNPREDICTABLE(a2 != zero) {
90 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
91 }
92 const uint8_t* restrict a3 = a[3];
93 if XNN_UNPREDICTABLE(a3 != zero) {
94 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
95 }
96 a += 4;
97
98 size_t k = 0;
99 while (k < kc) {
100 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
101 a0 += 8;
102 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
103 a1 += 8;
104 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
105 a2 += 8;
106 const v128_t vxa3 = wasm_u16x8_load8x8(a3);
107 a3 += 8;
108
109 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
110
111 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
112 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
113 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
114 vacc3x0 = wasm_i32x4_add(vacc3x0, wasm_i32x4_dot_i16x8(vxa3, vxb0));
115 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
116
117 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
118 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
119 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
120 vacc3x1 = wasm_i32x4_add(vacc3x1, wasm_i32x4_dot_i16x8(vxa3, vxb1));
121 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
122
123 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
124 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
125 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
126 vacc3x2 = wasm_i32x4_add(vacc3x2, wasm_i32x4_dot_i16x8(vxa3, vxb2));
127 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
128
129 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
130 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
131 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
132 vacc3x3 = wasm_i32x4_add(vacc3x3, wasm_i32x4_dot_i16x8(vxa3, vxb3));
133
134 w = (const void*) ((const uint8_t*) w + 32);
135 k += 8 * sizeof(uint8_t);
136 }
137 p -= 4 * sizeof(void*);
138 } while (p != 0);
139
140 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));
141 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));
142 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));
143 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));
144 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));
145 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));
146 const v128_t vacc3x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x0, vacc3x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x0, vacc3x2, 2, 6, 3, 7));
147 const v128_t vacc3x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x1, vacc3x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x1, vacc3x3, 2, 6, 3, 7));
148
149 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
150 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
151 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
152 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
153
154 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
155 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
156 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
157 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
158
159 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
160 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
161 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
162 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
163 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale);
164
165 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
166 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
167 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
168 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
169 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
170
171 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
172 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
173 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
174 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
175 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
176
177 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
178 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
179 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
180 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
181 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
182
183 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
184 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
185
186 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
187
188 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
189 vout = wasm_u8x16_min(vout, voutput_max);
190
191 if (nc >= 4) {
192 *((float*) c3) = (float) wasm_f32x4_extract_lane(vout, 3);
193 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
194 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
195 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
196
197 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
198 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
199 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
200 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
201
202 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
203
204 nc -= 4;
205 } else {
206 uint32_t vout3 = wasm_i32x4_extract_lane(vout, 3);
207 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
208 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
209 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
210 if (nc & 2) {
211 *((uint16_t*) c3) = (uint16_t) vout3;
212 vout3 >>= 16;
213 c3 += 2;
214 *((uint16_t*) c2) = (uint16_t) vout2;
215 vout2 >>= 16;
216 c2 += 2;
217 *((uint16_t*) c1) = (uint16_t) vout1;
218 vout1 >>= 16;
219 c1 += 2;
220 *((uint16_t*) c0) = (uint16_t) vout0;
221 vout0 >>= 16;
222 c0 += 2;
223 }
224 if (nc & 1) {
225 *c3 = (uint8_t) vout3;
226 *c2 = (uint8_t) vout2;
227 *c1 = (uint8_t) vout1;
228 *c0 = (uint8_t) vout0;
229 }
230
231 nc = 0;
232 }
233 } while (nc != 0);
234 }
235