File size: 2,227 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 |
// 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 BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
#include <assert.h>
#include <arm_neon.h>
#include <xnnpack/common.h>
#include <xnnpack/vcvt.h>
void xnn_f16_f32_vcvt_ukernel__neonfp16_x${BATCH_TILE}(
size_t batch,
const void* input,
float* output,
const union xnn_f16_f32_cvt_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;
$if BATCH_TILE > 8:
for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
$for N in range(SIMD_TILE):
const float16x8_t vh${N} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
$for N in range(SIMD_TILE):
const float32x4_t vf${2*N} = vcvt_f32_f16(vget_low_f16(vh${N}));
const float32x4_t vf${2*N+1} = vcvt_f32_f16(vget_high_f16(vh${N}));
$for N in range(2*SIMD_TILE):
vst1q_f32(output, vf${N}); output += 4;
}
for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
const float16x8_t vh = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
const float32x4_t vf_lo = vcvt_f32_f16(vget_low_f16(vh));
const float32x4_t vf_hi = vcvt_f32_f16(vget_high_f16(vh));
vst1q_f32(output, vf_lo); output += 4;
vst1q_f32(output, vf_hi); output += 4;
}
if XNN_UNLIKELY(batch != 0) {
assert(batch >= 1 * sizeof(uint16_t));
assert(batch <= 7 * sizeof(uint16_t));
const float16x8_t vh = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
float32x4_t vf = vcvt_f32_f16(vget_low_f16(vh));
if (batch & (4 * sizeof(uint16_t))) {
vst1q_f32(output, vf); output += 4;
vf = vcvt_f32_f16(vget_high_f16(vh));
}
float32x2_t vf_lo = vget_low_f32(vf);
if (batch & (2 * sizeof(uint16_t))) {
vst1_f32(output, vf_lo); output += 2;
vf_lo = vget_high_f32(vf);
}
if (batch & (1 * sizeof(uint16_t))) {
vst1_lane_f32(output, vf_lo, 0);
}
}
}
|