Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x4c2-minmax-fp32-wasmsimd-dot16x2-ld64.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_qs8_qc8w_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 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_qc8w_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 * sizeof(int8_t));
44 int8_t* c0 = c;
45
46
47 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
48 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
49 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
50 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
51 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
52 XNN_FORCE_REALIZATION(vmagic_bias);
53 XNN_FORCE_REALIZATION(vmagic_min);
54 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
55 XNN_FORCE_REALIZATION(voutput_max);
56
57
58 do {
59 v128_t vacc0x0123 = wasm_v128_load(w);
60 w = (const void*) ((const int32_t*) w + 4);
61
62 size_t p = ks;
63 do {
64 const int8_t* restrict a0 = a[0];
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
67 }
68 a += 1;
69
70 size_t k = kc;
71 while (k >= 8 * sizeof(int8_t)) {
72 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
73 a0 += 8;
74
75 const v128_t vxb0 = wasm_i16x8_load8x8(w);
76
77 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
78 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
79 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
80
81 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
82 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
83 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
84
85 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
86 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
87 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
88
89 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
90 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
91
92 w = (const void*) ((const int8_t*) w + 32);
93 k -= 8 * sizeof(int8_t);
94 }
95 if (k != 0) {
96 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
97 a0 = (const int8_t*) ((uintptr_t) a0 + k);
98
99 const v128_t vxb0 = wasm_i16x8_load8x8(w);
100 w = (const int8_t*) w + 8;
101
102 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
103 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
104
105 if (k > 2 * sizeof(int8_t)) {
106 const v128_t vxb1 = wasm_i16x8_load8x8(w);
107 w = (const int8_t*) w + 8;
108
109 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
110 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
111
112 if (k > 4 * sizeof(int8_t)) {
113 const v128_t vxb2 = wasm_i16x8_load8x8(w);
114 w = (const int8_t*) w + 8;
115
116 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
117 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
118 }
119 }
120 }
121 p -= 1 * sizeof(void*);
122 } while (p != 0);
123
124 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
125
126 const v128_t vscale0123 = wasm_v128_load(w);
127 w = (const float*) w + 4;
128 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
129
130 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
131
132 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
133
134 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
135
136 v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
137
138 v128_t vout = wasm_i8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
139
140 vout = wasm_i8x16_min(vout, voutput_max);
141
142 if (nc >= 4) {
143 wasm_v128_store32_lane(c0, vout, 0);
144
145 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
146
147 a = (const int8_t**restrict) ((uintptr_t) a - ks);
148
149 nc -= 4;
150 } else {
151 if (nc & 2) {
152 wasm_v128_store16_lane(c0, vout, 0);
153 c0 += 2;
154
155 vout = wasm_u32x4_shr(vout, 16);
156 }
157 if (nc & 1) {
158 wasm_v128_store8_lane(c0, vout, 0);
159 }
160
161 nc = 0;
162 }
163 } while (nc != 0);
164 }