| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-vcvt/wasmsimd.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2022 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/vcvt.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qu8_vcvt_ukernel__wasmsimd_u8( |
| 19 |
size_t batch, |
| 20 |
const uint8_t* input, |
| 21 |
uint8_t* output, |
| 22 |
const struct xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 23 |
{ |
| 24 |
assert(batch != 0); |
| 25 |
assert(batch % sizeof(uint8_t) == 0); |
| 26 |
assert(input != NULL); |
| 27 |
assert(output != NULL); |
| 28 |
|
| 29 |
const v128_t vinput_zero_point = wasm_v128_load16_splat(¶ms->scalar.input_zero_point); |
| 30 |
const v128_t vmultiplier = wasm_i16x8_splat(-params->scalar.multiplier); |
| 31 |
const v128_t voutput_zero_point = wasm_v128_load16_splat(¶ms->scalar.output_zero_point); |
| 32 |
XNN_FORCE_REALIZATION(vinput_zero_point); |
| 33 |
XNN_FORCE_REALIZATION(vmultiplier); |
| 34 |
XNN_FORCE_REALIZATION(voutput_zero_point); |
| 35 |
for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) { |
| 36 |
v128_t vacc = wasm_u16x8_load8x8(input); |
| 37 |
vacc = wasm_i16x8_sub(vinput_zero_point, vacc); |
| 38 |
vacc = wasm_i16x8_shl(vacc, 7); |
| 39 |
vacc = wasm_i16x8_q15mulr_sat(vacc, vmultiplier); |
| 40 |
vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point); |
| 41 |
input += 8; |
| 42 |
|
| 43 |
const v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc); |
| 44 |
wasm_v128_store64_lane(output, vy, 0); |
| 45 |
output += 8; |
| 46 |
} |
| 47 |
if XNN_UNLIKELY(batch != 0) { |
| 48 |
assert(batch >= 1 * sizeof(uint8_t)); |
| 49 |
assert(batch <= 7 * sizeof(uint8_t)); |
| 50 |
|
| 51 |
v128_t vacc = wasm_u16x8_load8x8(input); |
| 52 |
vacc = wasm_i16x8_sub(vinput_zero_point, vacc); |
| 53 |
vacc = wasm_i16x8_shl(vacc, 7); |
| 54 |
vacc = wasm_i16x8_q15mulr_sat(vacc, vmultiplier); |
| 55 |
vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point); |
| 56 |
|
| 57 |
v128_t vy = wasm_u8x16_narrow_i16x8(vacc, vacc); |
| 58 |
if (batch & (4 * sizeof(uint8_t))) { |
| 59 |
wasm_v128_store32_lane(output, vy, 0); |
| 60 |
vy = wasm_u64x2_shr(vy, 32); |
| 61 |
output += 4; |
| 62 |
} |
| 63 |
if (batch & (2 * sizeof(uint8_t))) { |
| 64 |
wasm_v128_store16_lane(output, vy, 0); |
| 65 |
vy = wasm_u32x4_shr(vy, 16); |
| 66 |
output += 2; |
| 67 |
} |
| 68 |
if (batch & (1 * sizeof(uint8_t))) { |
| 69 |
wasm_v128_store8_lane(output, vy, 0); |
| 70 |
} |
| 71 |
} |
| 72 |
} |