Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x4c2-minmax-wasmsimd-dot16x2-ld128.c

raw
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
18 void xnn_qd8_f32_qc8w_igemm_minmax_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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 2 * sizeof(int8_t));
46 float* c0 = c;
47
48 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
49 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
50 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
51 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
52 XNN_FORCE_REALIZATION(vmin);
53 XNN_FORCE_REALIZATION(vmax);
54
55
56 do {
57 const v128_t vksum0123 = wasm_v128_load(w);
58 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
59 w = (const void*) ((const int32_t*) w + 4);
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 } else {
67 a0 = zero_data;
68 }
69 a += 1;
70
71 size_t k = kc;
72 while (k >= 8 * sizeof(int8_t)) {
73 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
74 a0 += 8;
75
76 const v128_t vb01 = wasm_v128_load(w);
77 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
78 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
79
80 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
81 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
82
83 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
84 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
85 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
86 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
87 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
88
89 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
90 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
91
92 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
93 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
94
95 w = (const void*) ((const int8_t*) w + 32);
96 k -= 8 * sizeof(int8_t);
97 }
98 if (k != 0) {
99 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
100 a0 = (const int8_t*) ((uintptr_t) a0 + k);
101
102 const v128_t vxb0 = wasm_i16x8_load8x8(w);
103 w = (const int8_t*) w + 8;
104
105 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
106 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
107
108 if (k > 2 * sizeof(int8_t)) {
109 const v128_t vxb1 = wasm_i16x8_load8x8(w);
110 w = (const int8_t*) w + 8;
111
112 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
113 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
114
115 if (k > 4 * sizeof(int8_t)) {
116 const v128_t vxb2 = wasm_i16x8_load8x8(w);
117 w = (const int8_t*) w + 8;
118
119 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
120 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
121 }
122 }
123 }
124 p -= 1 * sizeof(void*);
125 } while (p != 0);
126
127 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
128
129 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
130
131 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
132 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
133 w = (const float*) w + 4;
134
135 const v128_t vbias0123 = wasm_v128_load(w);
136 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
137 w = (const float*) w + 4;
138
139 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
140
141 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
142
143 if XNN_LIKELY(nc >= 4) {
144 wasm_v128_store(c0, vacc0x0123);
145
146 a = (const int8_t**restrict) ((uintptr_t) a - ks);
147
148 c0 = (float*) ((uintptr_t) c0 + cn_stride);
149
150 nc -= 4;
151 } else {
152 if (nc & 2) {
153 wasm_v128_store64_lane(c0, vacc0x0123, 0);
154 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
155 c0 += 2;
156 }
157 if (nc & 1) {
158 wasm_v128_store32_lane(c0, vacc0x0123, 0);
159 }
160 nc = 0;
161 }
162 } while (nc != 0);
163 }