1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_fp32_ukernel_4x4c8__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)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(uint8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 8);
40 const uint8_t* a0 = a;
41 uint8_t* c0 = c;
42 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
43 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
55 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
62 do {
63 v128_t vacc0x0 = wasm_v128_load32_zero(w);
64 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
65 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
66 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
67 v128_t vacc1x0 = vacc0x0;
68 v128_t vacc1x1 = vacc0x1;
69 v128_t vacc1x2 = vacc0x2;
70 v128_t vacc1x3 = vacc0x3;
71 v128_t vacc2x0 = vacc0x0;
72 v128_t vacc2x1 = vacc0x1;
73 v128_t vacc2x2 = vacc0x2;
74 v128_t vacc2x3 = vacc0x3;
75 v128_t vacc3x0 = vacc0x0;
76 v128_t vacc3x1 = vacc0x1;
77 v128_t vacc3x2 = vacc0x2;
78 v128_t vacc3x3 = vacc0x3;
79 w = (const void*) ((const int32_t*) w + 4);
80
81 size_t k = 0;
82 while (k < kc) {
83 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
84 a0 += 8;
85 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
86 a1 += 8;
87 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
88 a2 += 8;
89 const v128_t vxa3 = wasm_u16x8_load8x8(a3);
90 a3 += 8;
91
92 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
93
94 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
95 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
96 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
97 vacc3x0 = wasm_i32x4_add(vacc3x0, wasm_i32x4_dot_i16x8(vxa3, vxb0));
98 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const int8_t*) w + 8), vb_zero_point);
99
100 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
101 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
102 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
103 vacc3x1 = wasm_i32x4_add(vacc3x1, wasm_i32x4_dot_i16x8(vxa3, vxb1));
104 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const int8_t*) w + 16), vb_zero_point);
105
106 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
107 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
108 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
109 vacc3x2 = wasm_i32x4_add(vacc3x2, wasm_i32x4_dot_i16x8(vxa3, vxb2));
110 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const int8_t*) w + 24), vb_zero_point);
111
112 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
113 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
114 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
115 vacc3x3 = wasm_i32x4_add(vacc3x3, wasm_i32x4_dot_i16x8(vxa3, vxb3));
116
117 w = (const void*) ((const uint8_t*) w + 32);
118 k += 8 * sizeof(uint8_t);
119 }
120
121 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));
122 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));
123 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));
124 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));
125 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));
126 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));
127 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));
128 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));
129
130 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
131 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
132 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
133 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
134
135 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
136 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
137 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
138 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
139
140 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
141 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
142 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
143 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
144 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale);
145
146 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
147 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
148 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
149 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
150 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
151
152 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
153 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
154 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
155 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
156 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
157
158 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
159 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
160 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
161 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
162 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
163
164 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
165 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
166
167 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
168
169 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
170 vout = wasm_u8x16_min(vout, voutput_max);
171
172 if (nc >= 4) {
173 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
174 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
175 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
176 *((float*) c3) = (float) wasm_f32x4_extract_lane(vout, 3);
177
178 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
179 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
180 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
181 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
182
183 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
184 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
185 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
186 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
187
188 nc -= 4;
189 } else {
190 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
191 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
192 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
193 uint32_t vout3 = wasm_i32x4_extract_lane(vout, 3);
194 if (nc & 2) {
195 *((uint16_t*) c0) = (uint16_t) vout0;
196 vout0 >>= 16;
197 c0 += 2;
198 *((uint16_t*) c1) = (uint16_t) vout1;
199 vout1 >>= 16;
200 c1 += 2;
201 *((uint16_t*) c2) = (uint16_t) vout2;
202 vout2 >>= 16;
203 c2 += 2;
204 *((uint16_t*) c3) = (uint16_t) vout3;
205 vout3 >>= 16;
206 c3 += 2;
207 }
208 if (nc & 1) {
209 *c0 = (uint8_t) vout0;
210 *c1 = (uint8_t) vout1;
211 *c2 = (uint8_t) vout2;
212 *c3 = (uint8_t) vout3;
213 }
214
215 nc = 0;
216 }
217 } while (nc != 0);
218 }
219