1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16
17
18
xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__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_4x4c2s4__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 <= 4);
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 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 kc = round_up_po2(kc, 8 * sizeof(uint8_t));
62 do {
63 v128_t vacc0x0123 = wasm_v128_load(w);
64 v128_t vacc1x0123 = vacc0x0123;
65 v128_t vacc2x0123 = vacc0x0123;
66 v128_t vacc3x0123 = vacc0x0123;
67 w = (const void*) ((const int32_t*) w + 4);
68
69 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
70 size_t k = kc;
71 do {
72 v128_t vxa0 = wasm_u16x8_load8x8((const v128_t*) a0);
73 a0 += 8;
74 v128_t vxa1 = wasm_u16x8_load8x8((const v128_t*) a1);
75 a1 += 8;
76 v128_t vxa2 = wasm_u16x8_load8x8((const v128_t*) a2);
77 a2 += 8;
78 v128_t vxa3 = wasm_u16x8_load8x8((const v128_t*) a3);
79 a3 += 8;
80
81 const v128_t vb01 = wasm_v128_load(w);
82 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
83 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
84
85 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
86 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
87 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
88 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
89 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
90 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
91 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb0));
92 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
93
94 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
95 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
96 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
97 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
98 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
99 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
100 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb1));
101 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
102 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
103 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
104 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
105
106 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
107 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
108 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
109 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
110 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
111 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
112 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb2));
113 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
114
115 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
116 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
117 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
118 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb3));
119
120 w = (const uint8_t*) w + 32;
121 k -= 8 * sizeof(uint8_t);
122 } while (k != 0);
123
124 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
125 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
126 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
127 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
128
129 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
130 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
131 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
132 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
133 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale);
134
135 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
136 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
137 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
138 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
139 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
140
141 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
142 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
143 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
144 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
145 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
146
147 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
148 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
149 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
150 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
151 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
152
153 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
154 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
155
156 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
157
158 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
159 vout = wasm_u8x16_min(vout, voutput_max);
160
161 if (nc >= 4) {
162 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
163 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
164 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
165 *((float*) c3) = (float) wasm_f32x4_extract_lane(vout, 3);
166
167 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
168 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
169 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
170 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
171
172 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
173 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
174 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
175 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
176
177 nc -= 4;
178 } else {
179 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
180 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
181 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
182 uint32_t vout3 = wasm_i32x4_extract_lane(vout, 3);
183 if (nc & 2) {
184 *((uint16_t*) c0) = (uint16_t) vout0;
185 vout0 >>= 16;
186 c0 += 2;
187 *((uint16_t*) c1) = (uint16_t) vout1;
188 vout1 >>= 16;
189 c1 += 2;
190 *((uint16_t*) c2) = (uint16_t) vout2;
191 vout2 >>= 16;
192 c2 += 2;
193 *((uint16_t*) c3) = (uint16_t) vout3;
194 vout3 >>= 16;
195 c3 += 2;
196 }
197 if (nc & 1) {
198 *c0 = (uint8_t) vout0;
199 *c1 = (uint8_t) vout1;
200 *c2 = (uint8_t) vout2;
201 *c3 = (uint8_t) vout3;
202 }
203
204 nc = 0;
205 }
206 } while (nc != 0);
207 }
208