1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2);
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 v128_t vacc0x0123 = wasm_v128_load(w);
52 v128_t vacc1x0123 = vacc0x0123;
53 w = (const void*) ((const int32_t*) w + 4);
54
55 size_t p = ks;
56 do {
57 const int8_t* restrict a0 = a[0];
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60 }
61 const int8_t* restrict a1 = a[1];
62 if XNN_UNPREDICTABLE(a1 != zero) {
63 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
64 }
65 a += 2;
66
67 size_t k = kc;
68 while (k >= 8 * sizeof(int8_t)) {
69 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
70 a0 += 8;
71 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
72 a1 += 8;
73
74 const v128_t vxb0 = wasm_i16x8_load8x8(w);
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 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
81
82 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
83 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
84 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
85 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
86 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
87
88 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
89 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
90 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
91 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
92 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
93
94 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
95 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
96 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
97 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
98
99 w = (const void*) ((const int8_t*) w + 32);
100 k -= 8 * sizeof(int8_t);
101 }
102 if (k != 0) {
103 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
104 a0 = (const int8_t*) ((uintptr_t) a0 + k);
105 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
106 a1 = (const int8_t*) ((uintptr_t) a1 + k);
107
108 const v128_t vxb0 = wasm_i16x8_load8x8(w);
109 w = (const void*) ((const int8_t*) w + 8);
110
111 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
112 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
113 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
114 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
115
116 if (k > 2 * sizeof(int8_t)) {
117 const v128_t vxb1 = wasm_i16x8_load8x8(w);
118 w = (const void*) ((const int8_t*) w + 8);
119
120 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
121 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
122 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
123 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
124
125 if (k > 4 * sizeof(int8_t)) {
126 const v128_t vxb2 = wasm_i16x8_load8x8(w);
127 w = (const void*) ((const int8_t*) w + 8);
128
129 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
130 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
131 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
132 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
133 }
134 }
135 }
136 p -= 2 * sizeof(void*);
137 } while (p != 0);
138
139 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
140 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
141
142 const v128_t vscale0123 = wasm_v128_load(w);
143 w = (const void*) ((const float*) w + 4);
144 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
145 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
146
147 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
148 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
149 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
150
151 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
152 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
153 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
154
155 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
156 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
157 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
158
159 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
160
161 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
162
163 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
164 vout = wasm_i8x16_min(vout, voutput_max);
165
166 if (nc >= 4) {
167 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
168 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
169
170 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
171 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
172
173 a = (const int8_t**restrict) ((uintptr_t) a - ks);
174
175 nc -= 4;
176 } else {
177 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
178 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
179 if (nc & 2) {
180 *((uint16_t*) c1) = (uint16_t) vout1;
181 vout1 >>= 16;
182 c1 += 2;
183 *((uint16_t*) c0) = (uint16_t) vout0;
184 vout0 >>= 16;
185 c0 += 2;
186 }
187 if (nc & 1) {
188 *c1 = (int8_t) vout1;
189 *c0 = (int8_t) vout0;
190 }
191
192 nc = 0;
193 }
194 } while (nc != 0);
195 }
196