1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2s4-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_4x4c2s4__wasmsimd_dot16x2_ld128(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_4x4c2s4__wasmsimd_dot16x2_ld128(
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 uint8_t* c0 = c;
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 kc = round_up_po2(kc, 8 * sizeof(uint8_t));
58 do {
59 v128_t vacc0x0123 = wasm_v128_load(w);
60 v128_t vacc1x0123 = vacc0x0123;
61 v128_t vacc2x0123 = vacc0x0123;
62 v128_t vacc3x0123 = vacc0x0123;
63 w = (const void*) ((const int32_t*) w + 4);
64
65 size_t p = ks;
66 do {
67 const uint8_t* restrict a0 = a[0];
68 if XNN_UNPREDICTABLE(a0 != zero) {
69 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
70 }
71 const uint8_t* restrict a1 = a[1];
72 if XNN_UNPREDICTABLE(a1 != zero) {
73 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
74 }
75 const uint8_t* restrict a2 = a[2];
76 if XNN_UNPREDICTABLE(a2 != zero) {
77 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
78 }
79 const uint8_t* restrict a3 = a[3];
80 if XNN_UNPREDICTABLE(a3 != zero) {
81 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
82 }
83 a += 4;
84
85 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
86 size_t k = kc;
87 do {
88 v128_t vxa0 = wasm_u16x8_load8x8(a0);
89 a0 += 8;
90 v128_t vxa1 = wasm_u16x8_load8x8(a1);
91 a1 += 8;
92 v128_t vxa2 = wasm_u16x8_load8x8(a2);
93 a2 += 8;
94 v128_t vxa3 = wasm_u16x8_load8x8(a3);
95 a3 += 8;
96
97 const v128_t vb01 = wasm_v128_load(w);
98 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
99 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
100
101 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
102 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
103 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
104 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
105 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
106 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
107 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb0));
108 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
109
110 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
111 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
112 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
113 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
114 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
115 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
116 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb1));
117 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
118 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
119 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
120 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
121
122 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
123 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
124 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
125 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
126 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
127 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
128 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb2));
129 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
130
131 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
132 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
133 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
134 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb3));
135
136 w = (const uint8_t*) w + 32;
137 k -= 8 * sizeof(uint8_t);
138 } while (k != 0);
139 p -= 4 * sizeof(void*);
140 } while (p != 0);
141
142 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
143 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
144 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
145 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
146
147 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
148 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
149 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
150 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
151 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale);
152
153 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
154 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
155 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
156 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
157 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
158
159 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
160 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
161 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
162 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
163 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
164
165 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
166 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
167 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
168 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
169 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
170
171 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
172 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
173
174 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
175
176 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
177 vout = wasm_u8x16_min(vout, voutput_max);
178
179 if (nc >= 4) {
180 *((float*) c3) = (float) wasm_f32x4_extract_lane(vout, 3);
181 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
182 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
183 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
184
185 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
186 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
187 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
188 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
189
190 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
191
192 nc -= 4;
193 } else {
194 uint32_t vout3 = wasm_i32x4_extract_lane(vout, 3);
195 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
196 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
197 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
198 if (nc & 2) {
199 *((uint16_t*) c3) = (uint16_t) vout3;
200 vout3 >>= 16;
201 c3 += 2;
202 *((uint16_t*) c2) = (uint16_t) vout2;
203 vout2 >>= 16;
204 c2 += 2;
205 *((uint16_t*) c1) = (uint16_t) vout1;
206 vout1 >>= 16;
207 c1 += 2;
208 *((uint16_t*) c0) = (uint16_t) vout0;
209 vout0 >>= 16;
210 c0 += 2;
211 }
212 if (nc & 1) {
213 *c3 = (uint8_t) vout3;
214 *c2 = (uint8_t) vout2;
215 *c1 = (uint8_t) vout1;
216 *c0 = (uint8_t) vout0;
217 }
218
219 nc = 0;
220 }
221 } while (nc != 0);
222 }
223