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_qu8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_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(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2);
44 uint8_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 uint8_t* restrict a0 = a[0];
53 if XNN_UNPREDICTABLE(a0 != zero) {
54 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
55 }
56 a += 1;
57
58 size_t k = kc;
59 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
60 while (k >= 8 * sizeof(uint8_t)) {
61 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
62 a0 += 8;
63
64 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
65
66 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
67 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
68 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
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 vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
73
74 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
75 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
76 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
77
78 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
79 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
80
81 w = (const void*) ((const uint8_t*) w + 32);
82 k -= 8 * sizeof(uint8_t);
83 }
84 if (k != 0) {
85 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
86 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
87
88 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
89 w = (const void*) ((const uint8_t*) w + 8);
90
91 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
92 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
93
94 if (k > 2 * sizeof(uint8_t)) {
95 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
96 w = (const void*) ((const uint8_t*) w + 8);
97
98 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
99 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
100
101 if (k > 4 * sizeof(uint8_t)) {
102 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
103 w = (const void*) ((const uint8_t*) w + 8);
104
105 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
106 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
107 }
108 }
109 }
110 p -= 1 * sizeof(void*);
111 } while (p != 0);
112
113 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
114
115 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
116 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
117
118 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
119 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
120
121 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
122 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
123
124 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
125 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
126
127 v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
128
129 v128_t vout = wasm_u8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
130
131 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
132 vout = wasm_u8x16_min(vout, voutput_max);
133
134 if (nc >= 4) {
135 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
136
137 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
138
139 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
140
141 nc -= 4;
142 } else {
143 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
144 if (nc & 2) {
145 *((uint16_t*) c0) = (uint16_t) vout0;
146 vout0 >>= 16;
147 c0 += 2;
148 }
149 if (nc & 1) {
150 *c0 = (uint8_t) vout0;
151 }
152
153 nc = 0;
154 }
155 } while (nc != 0);
156 }
157