File size: 2,260 Bytes
8b7c501 |
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 |
// Copyright 2022 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 BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP in ["RNDNE", "RNDZ", "RNDU", "RNDD"]
#include <assert.h>
#include <arm_neon.h>
#include <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/vunary.h>
$VRNDQ_F16 = {
$ "RNDNE": "vrndnq_f16",
$ "RNDZ": "vrndq_f16",
$ "RNDU": "vrndpq_f16",
$ "RNDD": "vrndmq_f16",
$}[OP]
void xnn_f16_v${OP.lower()}_ukernel__neonfp16arith_x${BATCH_TILE}(
size_t batch,
const void* input,
void* output,
const union xnn_f16_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(uint16_t) == 0);
assert(input != NULL);
assert(output != NULL);
const uint16_t* i = (const uint16_t*) input;
uint16_t* o = (uint16_t*) output;
$if BATCH_TILE > 8:
for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
$for N in range(SIMD_TILE):
float16x8_t vacc${ABC[N]} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
$for N in range(SIMD_TILE):
vacc${ABC[N]} = ${VRNDQ_F16}(vacc${ABC[N]});
$for N in range(SIMD_TILE):
vst1q_u16(o, vreinterpretq_u16_f16(vacc${ABC[N]})); o += 8;
}
for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
float16x8_t vacc = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
vacc = ${VRNDQ_F16}(vacc);
vst1q_u16(o, vreinterpretq_u16_f16(vacc)); o += 8;
}
if XNN_UNLIKELY(batch != 0) {
float16x8_t vacc = vreinterpretq_f16_u16(vld1q_u16(i));
vacc = ${VRNDQ_F16}(vacc);
float16x4_t vacc_lo = vget_low_f16(vacc);
if (batch & (4 * sizeof(uint16_t))) {
vst1_u16(o, vreinterpret_u16_f16(vacc_lo)); o += 4;
vacc_lo = vget_high_f16(vacc);
}
if (batch & (2 * sizeof(uint16_t))) {
vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2;
vacc_lo = vext_f16(vacc_lo, vacc_lo, 2);
}
if (batch & (1 * sizeof(uint16_t))) {
vst1_lane_u16(o, vreinterpret_u16_f16(vacc_lo), 0);
}
}
}
|