-
Notifications
You must be signed in to change notification settings - Fork 391
/
Copy pathwasmsimd-mul32-ld64.c.in
168 lines (141 loc) · 7.46 KB
/
wasmsimd-mul32-ld64.c.in
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
// Copyright 2021 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
$assert DATATYPE in ["QS8", "QU8"]
$assert REQUANTIZATION == "FP32"
$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <wasm_simd128.h>
#include <xnnpack/vmul.h>
$PARAMS_STRUCT = REQUANTIZATION.lower() + "_wasmsimd"
$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$WASM_X16X8_LOAD8X8 = "wasm_u16x8_load8x8" if DATATYPE == "QU8" else "wasm_i16x8_load8x8"
$WASM_X32X4_EXTEND_LOW_X16X8 = "wasm_u32x4_extend_low_u16x8" if DATATYPE == "QU8" else "wasm_i32x4_extend_low_i16x8"
$WASM_X32X4_EXTEND_HIGH_X16X8 = "wasm_u32x4_extend_high_u16x8" if DATATYPE == "QU8" else "wasm_i32x4_extend_high_i16x8"
$WASM_X8X16_NARROW_I16X8 = {"QS8": "wasm_i8x16_narrow_i16x8", "QU8": "wasm_u8x16_narrow_i16x8"}[DATATYPE]
$WASM_X8X16_MIN = {"QS8": "wasm_i8x16_min", "QU8": "wasm_u8x16_min"}[DATATYPE]
void xnn_${DATATYPE.lower()}_vmulc_minmax_${REQUANTIZATION.lower()}_ukernel__wasmsimd_mul32_ld64_x${BATCH_TILE}(
size_t batch,
const ${XINT8_T}* input_a,
const ${XINT8_T}* input_b,
${XINT8_T}* output,
const union xnn_${DATATYPE.lower()}_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(${XINT8_T}) == 0);
assert(input_a != NULL);
assert(input_b != NULL);
assert(output != NULL);
const v128_t va_zero_point = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.a_zero_point);
const v128_t vscale = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.scale);
const v128_t vmagic_bias = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias);
const v128_t vmagic_min = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_min);
const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.magic_bias_less_output_zero_point);
const v128_t voutput_max = wasm_v128_load64_splat(params->${PARAMS_STRUCT}.output_max);
const v128_t vxb = wasm_i16x8_sub(
wasm_i16x8_splat((int16_t) *input_b), wasm_v128_load64_splat(params->${PARAMS_STRUCT}.b_zero_point));
for (; batch >= ${BATCH_TILE} * sizeof(${XINT8_T}); batch -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
const v128_t va${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(input_a);
$for N in range(8, BATCH_TILE, 8):
const v128_t va${ABC[N:N+8]} = ${WASM_X16X8_LOAD8X8}(input_a + ${N});
input_a += ${BATCH_TILE};
$for N in range(0, BATCH_TILE, 8):
const v128_t vxa${ABC[N:N+8]} = wasm_i16x8_sub(va${ABC[N:N+8]}, va_zero_point);
$for N in range(0, BATCH_TILE, 8):
v128_t vacc${ABC[N:N+4]} = wasm_i32x4_extmul_low_i16x8(vxa${ABC[N:N+8]}, vxb);
v128_t vacc${ABC[N+4:N+8]} = wasm_i32x4_extmul_high_i16x8(vxa${ABC[N:N+8]}, vxb);
$for N in range(0, BATCH_TILE, 4):
vacc${ABC[N:N+4]} = wasm_f32x4_convert_i32x4(vacc${ABC[N:N+4]});
$for N in range(0, BATCH_TILE, 4):
vacc${ABC[N:N+4]} = wasm_f32x4_mul(vacc${ABC[N:N+4]}, vscale);
$for N in range(0, BATCH_TILE, 4):
vacc${ABC[N:N+4]} = wasm_f32x4_add(vacc${ABC[N:N+4]}, vmagic_bias);
$for N in range(0, BATCH_TILE, 4):
vacc${ABC[N:N+4]} = wasm_i32x4_max(vacc${ABC[N:N+4]}, vmagic_min);
$for N in range(0, BATCH_TILE, 4):
vacc${ABC[N:N+4]} = wasm_i32x4_sub(vacc${ABC[N:N+4]}, vmagic_bias_less_output_zero_point);
$for N in range(0, BATCH_TILE, 8):
v128_t vout${ABC[N:N+8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]});
$for N in range(0, BATCH_TILE, 16):
$if N + 8 < BATCH_TILE:
v128_t vout${ABC[N:N+16]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]});
$else:
v128_t vout${ABC[N:N+8]}${ABC[N:N+8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
$for N in range(0, BATCH_TILE, 16):
$if N + 8 < BATCH_TILE:
vout${ABC[N:N+16]} = ${WASM_X8X16_MIN}(vout${ABC[N:N+16]}, voutput_max);
$else:
vout${ABC[N:N+8]}${ABC[N:N+8]} = ${WASM_X8X16_MIN}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max);
$if BATCH_TILE >= 16:
wasm_v128_store(output, vout${ABC[0:16]});
$else:
wasm_v128_store64_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
$for N in range(16, BATCH_TILE, 16):
$if N + 8 < BATCH_TILE:
wasm_v128_store(output + ${N}, vout${ABC[N:N+16]});
$else:
wasm_v128_store64_lane(output + ${N}, vout${ABC[0:8]}${ABC[0:8]}, 0);
output += ${BATCH_TILE};
}
if XNN_UNLIKELY(batch != 0) {
${"do " if BATCH_TILE > 8 else ""}{
const v128_t va${ABC[0:8]} = ${WASM_X16X8_LOAD8X8}(input_a);
$if BATCH_TILE > 8:
input_a += 8;
const v128_t vxa${ABC[0:8]} = wasm_i16x8_sub(va${ABC[0:8]}, va_zero_point);
v128_t vacc${ABC[0:4]} = wasm_i32x4_extmul_low_i16x8(vxa${ABC[0:8]}, vxb);
v128_t vacc${ABC[4:8]} = wasm_i32x4_extmul_high_i16x8(vxa${ABC[0:8]}, vxb);
vacc${ABC[0:4]} = wasm_f32x4_convert_i32x4(vacc${ABC[0:4]});
vacc${ABC[4:8]} = wasm_f32x4_convert_i32x4(vacc${ABC[4:8]});
vacc${ABC[0:4]} = wasm_f32x4_mul(vacc${ABC[0:4]}, vscale);
vacc${ABC[4:8]} = wasm_f32x4_mul(vacc${ABC[4:8]}, vscale);
vacc${ABC[0:4]} = wasm_f32x4_add(vacc${ABC[0:4]}, vmagic_bias);
vacc${ABC[4:8]} = wasm_f32x4_add(vacc${ABC[4:8]}, vmagic_bias);
vacc${ABC[0:4]} = wasm_i32x4_max(vacc${ABC[0:4]}, vmagic_min);
vacc${ABC[4:8]} = wasm_i32x4_max(vacc${ABC[4:8]}, vmagic_min);
vacc${ABC[0:4]} = wasm_i32x4_sub(vacc${ABC[0:4]}, vmagic_bias_less_output_zero_point);
vacc${ABC[4:8]} = wasm_i32x4_sub(vacc${ABC[4:8]}, vmagic_bias_less_output_zero_point);
v128_t vout${ABC[0:8]} = wasm_i16x8_narrow_i32x4(vacc${ABC[0:4]}, vacc${ABC[4:8]});
v128_t vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_NARROW_I16X8}(vout${ABC[0:8]}, vout${ABC[0:8]});
vout${ABC[0:8]}${ABC[0:8]} = ${WASM_X8X16_MIN}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);
$if BATCH_TILE > 8:
if XNN_LIKELY(batch >= (8 * sizeof(${XINT8_T}))) {
wasm_v128_store64_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
output += 8;
batch -= 8 * sizeof(${XINT8_T});
} else {
if (batch & (4 * sizeof(${XINT8_T}))) {
wasm_v128_store32_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
output += 4;
}
if (batch & (2 * sizeof(${XINT8_T}))) {
wasm_v128_store16_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
output += 2;
}
if (batch & (1 * sizeof(${XINT8_T}))) {
wasm_v128_store8_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
}
batch = 0;
}
$else:
if (batch & (4 * sizeof(${XINT8_T}))) {
wasm_v128_store32_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
vout${ABC[0:8]}${ABC[0:8]} = wasm_u64x2_shr(vout${ABC[0:8]}${ABC[0:8]}, 32);
output += 4;
}
if (batch & (2 * sizeof(${XINT8_T}))) {
wasm_v128_store16_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
vout${ABC[0:8]}${ABC[0:8]} = wasm_u32x4_shr(vout${ABC[0:8]}${ABC[0:8]}, 16);
output += 2;
}
if (batch & (1 * sizeof(${XINT8_T}))) {
wasm_v128_store8_lane(output, vout${ABC[0:8]}${ABC[0:8]}, 0);
}
}${" while (batch != 0);" if BATCH_TILE > 8 else ""}
}
}