| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-vadd/wasmsimd.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/common.h" |
| 15 |
#include "xnnpack/vbinary.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qu8_vadd_minmax_ukernel__wasmsimd_u8( |
| 19 |
size_t batch, |
| 20 |
const uint8_t* input_a, |
| 21 |
const uint8_t* input_b, |
| 22 |
uint8_t* output, |
| 23 |
const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 24 |
{ |
| 25 |
assert(batch != 0); |
| 26 |
assert(batch % sizeof(uint8_t) == 0); |
| 27 |
assert(input_a != NULL); |
| 28 |
assert(input_b != NULL); |
| 29 |
assert(output != NULL); |
| 30 |
|
| 31 |
const v128_t vbias = wasm_v128_load32_splat(¶ms->scalar.bias); |
| 32 |
const v128_t va_multiplier = wasm_v128_load32_splat(¶ms->scalar.a_multiplier); |
| 33 |
const v128_t vb_multiplier = wasm_v128_load32_splat(¶ms->scalar.b_multiplier); |
| 34 |
const uint32_t vshift = params->scalar.shift; |
| 35 |
const v128_t voutput_zero_point = wasm_v128_load16_splat(¶ms->scalar.output_zero_point); |
| 36 |
const v128_t voutput_min = wasm_v128_load8_splat(¶ms->scalar.output_min); |
| 37 |
const v128_t voutput_max = wasm_v128_load8_splat(¶ms->scalar.output_max); |
| 38 |
|
| 39 |
XNN_FORCE_REALIZATION(vbias); |
| 40 |
XNN_FORCE_REALIZATION(va_multiplier); |
| 41 |
XNN_FORCE_REALIZATION(vb_multiplier); |
| 42 |
XNN_FORCE_REALIZATION(voutput_zero_point); |
| 43 |
XNN_FORCE_REALIZATION(voutput_min); |
| 44 |
XNN_FORCE_REALIZATION(voutput_max); |
| 45 |
|
| 46 |
for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) { |
| 47 |
const v128_t va01234567 = wasm_u16x8_load8x8(input_a); |
| 48 |
const v128_t vb01234567 = wasm_u16x8_load8x8(input_b); |
| 49 |
input_a += 8; |
| 50 |
input_b += 8; |
| 51 |
|
| 52 |
v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier)); |
| 53 |
v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier)); |
| 54 |
|
| 55 |
vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vb01234567), vb_multiplier)); |
| 56 |
vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vb01234567), vb_multiplier)); |
| 57 |
|
| 58 |
vacc0123 = wasm_i32x4_shr(vacc0123, vshift); |
| 59 |
vacc4567 = wasm_i32x4_shr(vacc4567, vshift); |
| 60 |
|
| 61 |
v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point); |
| 62 |
|
| 63 |
v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567); |
| 64 |
|
| 65 |
vout0123456701234567 = wasm_u8x16_max(vout0123456701234567, voutput_min); |
| 66 |
|
| 67 |
vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max); |
| 68 |
|
| 69 |
wasm_v128_store64_lane(output, vout0123456701234567, 0); |
| 70 |
output += 8; |
| 71 |
} |
| 72 |
if XNN_UNLIKELY(batch != 0) { |
| 73 |
{ |
| 74 |
const v128_t va01234567 = wasm_u16x8_load8x8(input_a); |
| 75 |
const v128_t vb01234567 = wasm_u16x8_load8x8(input_b); |
| 76 |
|
| 77 |
v128_t vacc0123 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(va01234567), va_multiplier)); |
| 78 |
v128_t vacc4567 = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va01234567), va_multiplier)); |
| 79 |
|
| 80 |
vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vb01234567), vb_multiplier)); |
| 81 |
vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vb01234567), vb_multiplier)); |
| 82 |
|
| 83 |
vacc0123 = wasm_i32x4_shr(vacc0123, vshift); |
| 84 |
vacc4567 = wasm_i32x4_shr(vacc4567, vshift); |
| 85 |
|
| 86 |
v128_t vout01234567 = wasm_i16x8_add_sat(wasm_i16x8_narrow_i32x4(vacc0123, vacc4567), voutput_zero_point); |
| 87 |
|
| 88 |
v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567); |
| 89 |
vout0123456701234567 = wasm_u8x16_max(vout0123456701234567, voutput_min); |
| 90 |
vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max); |
| 91 |
|
| 92 |
if (batch & (4 * sizeof(uint8_t))) { |
| 93 |
wasm_v128_store32_lane(output, vout0123456701234567, 0); |
| 94 |
vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32); |
| 95 |
output += 4; |
| 96 |
} |
| 97 |
if (batch & (2 * sizeof(uint8_t))) { |
| 98 |
wasm_v128_store16_lane(output, vout0123456701234567, 0); |
| 99 |
vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16); |
| 100 |
output += 2; |
| 101 |
} |
| 102 |
if (batch & (1 * sizeof(uint8_t))) { |
| 103 |
wasm_v128_store8_lane(output, vout0123456701234567, 0); |
| 104 |
} |
| 105 |
} |
| 106 |
} |
| 107 |
} |