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_3x4c2__wasmsimd_dot16x2_ld128(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_3x4c2__wasmsimd_dot16x2_ld128(
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 <= 3);
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
56 do {
57 v128_t vacc0x0123 = wasm_v128_load(w);
58 v128_t vacc1x0123 = vacc0x0123;
59 v128_t vacc2x0123 = vacc0x0123;
60 w = (const void*) ((const int32_t*) w + 4);
61
62 size_t k = kc;
63 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
64 while (k >= 8 * sizeof(uint8_t)) {
65 const v128_t vxa0 = wasm_u16x8_load8x8((const v128_t*) a0);
66 a0 += 8;
67 const v128_t vxa1 = wasm_u16x8_load8x8((const v128_t*) a1);
68 a1 += 8;
69 const v128_t vxa2 = wasm_u16x8_load8x8((const v128_t*) a2);
70 a2 += 8;
71
72 const v128_t vb01 = wasm_v128_load(w);
73 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
74 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
75
76 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
77 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
78 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
79 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
80 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
81 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
82
83 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
84 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
85 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
86 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
87 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
88 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
89 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
90 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
91 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
92
93 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
94 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
95 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
96 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
97 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
98 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
99
100 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
101 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
102 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
103 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
104 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
105 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
106
107 w = (const void*) ((const uint8_t*) w + 32);
108 k -= 8 * sizeof(uint8_t);
109 }
110 if (k != 0) {
111 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
112 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
113 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
114 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
115 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
116 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
117
118 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
119 w = (const void*) ((const uint8_t*) w + 8);
120
121 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
122 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
123 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
124 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
125 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
126 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
127
128 if (k > 2 * sizeof(uint8_t)) {
129 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
130 w = (const void*) ((const uint8_t*) w + 8);
131
132 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
133 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
134 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
135 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
136 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
137 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
138
139 if (k > 4 * sizeof(uint8_t)) {
140 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
141 w = (const void*) ((const uint8_t*) w + 8);
142
143 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
144 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
145 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
146 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
147 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
148 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
149 }
150 }
151 }
152
153 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
154 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
155 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
156
157 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
158 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
159 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
160 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
161
162 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
163 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
164 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
165 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
166
167 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
168 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
169 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
170 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
171
172 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
173 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
174 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
175 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
176
177 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
178 v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
179
180 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
181
182 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
183 vout = wasm_u8x16_min(vout, voutput_max);
184
185 if (nc >= 4) {
186 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
187 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
188 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
189
190 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
191 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
192 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
193
194 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
195 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
196 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
197
198 nc -= 4;
199 } else {
200 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
201 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
202 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
203 if (nc & 2) {
204 *((uint16_t*) c0) = (uint16_t) vout0;
205 vout0 >>= 16;
206 c0 += 2;
207 *((uint16_t*) c1) = (uint16_t) vout1;
208 vout1 >>= 16;
209 c1 += 2;
210 *((uint16_t*) c2) = (uint16_t) vout2;
211 vout2 >>= 16;
212 c2 += 2;
213 }
214 if (nc & 1) {
215 *c0 = (uint8_t) vout0;
216 *c1 = (uint8_t) vout1;
217 *c2 = (uint8_t) vout2;
218 }
219
220 nc = 0;
221 }
222 } while (nc != 0);
223 }
224