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_1x4c2__wasmsimd_dot16x2_ld128(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_1x4c2__wasmsimd_dot16x2_ld128(
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 <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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
46 do {
47 v128_t vacc0x0123 = wasm_v128_load(w);
48 w = (const void*) ((const int32_t*) w + 4);
49
50 size_t p = ks;
51 do {
52 const int8_t* restrict a0 = a[0];
53 if XNN_UNPREDICTABLE(a0 != zero) {
54 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
55 }
56 a += 1;
57
58 size_t k = kc;
59 while (k >= 8 * sizeof(int8_t)) {
60 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
61 a0 += 8;
62
63 const v128_t vb01 = wasm_v128_load(w);
64 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
65 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
66
67 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
68 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
69
70 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
71 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
72 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
73 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
74 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
75
76 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
77 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
78
79 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
80 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
81
82 w = (const void*) ((const int8_t*) w + 32);
83 k -= 8 * sizeof(int8_t);
84 }
85 if (k != 0) {
86 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
87 a0 = (const int8_t*) ((uintptr_t) a0 + k);
88
89 const v128_t vxb0 = wasm_i16x8_load8x8(w);
90 w = (const void*) ((const int8_t*) w + 8);
91
92 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
93 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
94
95 if (k > 2 * sizeof(int8_t)) {
96 const v128_t vxb1 = wasm_i16x8_load8x8(w);
97 w = (const void*) ((const int8_t*) w + 8);
98
99 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
100 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
101
102 if (k > 4 * sizeof(int8_t)) {
103 const v128_t vxb2 = wasm_i16x8_load8x8(w);
104 w = (const void*) ((const int8_t*) w + 8);
105
106 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
107 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
108 }
109 }
110 }
111 p -= 1 * sizeof(void*);
112 } while (p != 0);
113
114 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
115
116 const v128_t vscale0123 = wasm_v128_load(w);
117 w = (const void*) ((const float*) w + 4);
118 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
119
120 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
121 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
122
123 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
124 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
125
126 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
127 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
128
129 v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
130
131 v128_t vout = wasm_i8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
132
133 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
134 vout = wasm_i8x16_min(vout, voutput_max);
135
136 if (nc >= 4) {
137 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
138
139 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
140
141 a = (const int8_t**restrict) ((uintptr_t) a - ks);
142
143 nc -= 4;
144 } else {
145 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
146 if (nc & 2) {
147 *((uint16_t*) c0) = (uint16_t) vout0;
148 vout0 >>= 16;
149 c0 += 2;
150 }
151 if (nc & 1) {
152 *c0 = (int8_t) vout0;
153 }
154
155 nc = 0;
156 }
157 } while (nc != 0);
158 }
159