1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_fp32_ukernel_1x4c8__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_1x4c8__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, 8);
44 uint8_t* c0 = c;
45
46 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
47 do {
48 v128_t vacc0x0 = wasm_v128_load32_zero(w);
49 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
50 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
51 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
52 w = (const void*) ((const int32_t*) w + 4);
53
54 size_t p = ks;
55 do {
56 const uint8_t* restrict a0 = a[0];
57 if XNN_UNPREDICTABLE(a0 != zero) {
58 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
59 }
60 a += 1;
61
62 size_t k = 0;
63 while (k < kc) {
64 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
65 a0 += 8;
66
67 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
68
69 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
70 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
71
72 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
73 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
74
75 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
76 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
77
78 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
79
80 w = (const void*) ((const uint8_t*) w + 32);
81 k += 8 * sizeof(uint8_t);
82 }
83 p -= 1 * sizeof(void*);
84 } while (p != 0);
85
86 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));
87 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));
88
89 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
90
91 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
92
93 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
94 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
95
96 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
97 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
98
99 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
100 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
101
102 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
103 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
104
105 v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
106
107 v128_t vout = wasm_u8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
108
109 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
110 vout = wasm_u8x16_min(vout, voutput_max);
111
112 if (nc >= 4) {
113 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
114
115 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
116
117 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
118
119 nc -= 4;
120 } else {
121 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
122 if (nc & 2) {
123 *((uint16_t*) c0) = (uint16_t) vout0;
124 vout0 >>= 16;
125 c0 += 2;
126 }
127 if (nc & 1) {
128 *c0 = (uint8_t) vout0;
129 }
130
131 nc = 0;
132 }
133 } while (nc != 0);
134 }
135