1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-wasmsimd-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_qc8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_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_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_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_qs8_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, 8);
44 int8_t* c0 = c;
45
46 do {
47 v128_t vacc0x0 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[0]);
48 v128_t vacc0x1 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[1]);
49 v128_t vacc0x2 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[2]);
50 v128_t vacc0x3 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[3]);
51 w = (const void*) ((const int32_t*) w + 4);
52
53 size_t p = ks;
54 do {
55 const int8_t* restrict a0 = a[0];
56 if XNN_UNPREDICTABLE(a0 != zero) {
57 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
58 }
59 a += 1;
60
61 size_t k = 0;
62 while (k < kc) {
63 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
64 a0 += 8;
65
66 const v128_t vxb0 = wasm_i16x8_load8x8(w);
67
68 const v128_t vprod0x0 = wasm_i16x8_mul(vxa0, vxb0);
69 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_extend_low_i16x8(vprod0x0));
70 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_extend_high_i16x8(vprod0x0));
71 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
72
73 const v128_t vprod0x1 = wasm_i16x8_mul(vxa0, vxb1);
74 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_extend_low_i16x8(vprod0x1));
75 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_extend_high_i16x8(vprod0x1));
76 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
77
78 const v128_t vprod0x2 = wasm_i16x8_mul(vxa0, vxb2);
79 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_extend_low_i16x8(vprod0x2));
80 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_extend_high_i16x8(vprod0x2));
81 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
82
83 const v128_t vprod0x3 = wasm_i16x8_mul(vxa0, vxb3);
84 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_extend_low_i16x8(vprod0x3));
85 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_extend_high_i16x8(vprod0x3));
86
87 w = (const void*) ((const int8_t*) w + 32);
88 k += 8 * sizeof(int8_t);
89 }
90 p -= 1 * sizeof(void*);
91 } while (p != 0);
92
93 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));
94 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));
95
96 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
97
98 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
99
100 const v128_t vscale0123 = wasm_v128_load(w);
101 w = (const void*) ((const float*) w + 4);
102 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
103
104 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd.magic_bias);
105 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
106
107 const v128_t vmagic_min = wasm_v128_load64_splat(params->wasmsimd.magic_min);
108 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
109
110 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->wasmsimd.magic_bias_less_output_zero_point);
111 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
112
113 v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
114
115 v128_t vout = wasm_i8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
116
117 const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd.output_max);
118 vout = wasm_i8x16_min(vout, voutput_max);
119
120 if (nc >= 4) {
121 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
122
123 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
124
125 a = (const int8_t**restrict) ((uintptr_t) a - ks);
126
127 nc -= 4;
128 } else {
129 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
130 if (nc & 2) {
131 *((uint16_t*) c0) = (uint16_t) vout0;
132 vout0 >>= 16;
133 c0 += 2;
134 }
135 if (nc & 1) {
136 *c0 = (int8_t) vout0;
137 }
138
139 nc = 0;
140 }
141 } while (nc != 0);
142 }
143