| 1 |
// Copyright 2023 Google LLC |
| 2 |
// |
| 3 |
// This source code is licensed under the BSD-style license found in the |
| 4 |
// LICENSE file in the root directory of this source tree. |
| 5 |
|
| 6 |
$assert BATCH_TILE % 8 == 0 or BATCH_TILE == 4 |
| 7 |
$assert BATCH_TILE >= 4 |
| 8 |
$SIMD_TILE = BATCH_TILE // 4 |
| 9 |
$assert ACCUMULATORS <= SIMD_TILE |
| 10 |
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" |
| 11 |
#include <assert.h> |
| 12 |
|
| 13 |
#include <arm_neon.h> |
| 14 |
|
| 15 |
#include "xnnpack/common.h" |
| 16 |
#include "xnnpack/reduce.h" |
| 17 |
|
| 18 |
|
| 19 |
$ACC_SUFFIX = "" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS |
| 20 |
void xnn_f16_f32acc_rsum_ukernel__neonfp16arith_u${BATCH_TILE}${ACC_SUFFIX}( |
| 21 |
size_t batch, |
| 22 |
const xnn_float16* input, |
| 23 |
float* output, |
| 24 |
const struct xnn_f16_f32acc_scale_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 25 |
{ |
| 26 |
assert(batch != 0); |
| 27 |
assert(batch % sizeof(uint16_t) == 0); |
| 28 |
assert(input != NULL); |
| 29 |
assert(output != NULL); |
| 30 |
|
| 31 |
const uint16_t* i = (const uint16_t*) input; |
| 32 |
$for A in range(ACCUMULATORS): |
| 33 |
float32x4_t vacc${A} = vmovq_n_f32(0.0f); |
| 34 |
$if BATCH_TILE > 8: |
| 35 |
for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) { |
| 36 |
$for N in range(0, SIMD_TILE, 2): |
| 37 |
const float16x8_t vh${ABC[N:N+2]} = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8; |
| 38 |
|
| 39 |
$for N in range(0, SIMD_TILE, 2): |
| 40 |
const float32x4_t vt${N} = vcvt_f32_f16(vget_low_f16(vh${ABC[N:N+2]})); |
| 41 |
const float32x4_t vt${N+1} = vcvt_f32_f16(vget_high_f16(vh${ABC[N:N+2]})); |
| 42 |
|
| 43 |
$for N in range(SIMD_TILE): |
| 44 |
vacc${N % ACCUMULATORS} = vaddq_f32(vacc${N % ACCUMULATORS}, vt${N}); |
| 45 |
} |
| 46 |
$if ACCUMULATORS > 1: |
| 47 |
$ACC_SLICE = 1 |
| 48 |
$while ACC_SLICE < ACCUMULATORS: |
| 49 |
$for A in range(0, ACCUMULATORS, ACC_SLICE * 2): |
| 50 |
$if A + ACC_SLICE < ACCUMULATORS: |
| 51 |
vacc${A} = vaddq_f32(vacc${A}, vacc${A + ACC_SLICE}); |
| 52 |
$ACC_SLICE *= 2 |
| 53 |
for (; batch >= 4 * sizeof(uint16_t); batch -= 4 * sizeof(uint16_t)) { |
| 54 |
const float16x4_t vh = vreinterpret_f16_u16(vld1_u16(i)); i += 4; |
| 55 |
const float32x4_t vt = vcvt_f32_f16(vh); |
| 56 |
vacc0 = vaddq_f32(vacc0, vt); |
| 57 |
} |
| 58 |
const float32x2_t vscale = vld1_dup_f32(¶ms->scalar.scale); |
| 59 |
float32x2_t vacc = vadd_f32(vget_low_f32(vacc0), vget_high_f32(vacc0)); |
| 60 |
if XNN_UNLIKELY(batch & (2 * sizeof(uint16_t))) { |
| 61 |
const float16x4_t vh = vreinterpret_f16_u32(vld1_dup_u32((const void*) i)); i += 2; |
| 62 |
const float32x4_t vt = vcvt_f32_f16(vh); |
| 63 |
vacc = vadd_f32(vacc, vget_low_f32(vt)); |
| 64 |
} |
| 65 |
vacc = vpadd_f32(vacc, vacc); |
| 66 |
if XNN_UNLIKELY(batch & (1 * sizeof(uint16_t))) { |
| 67 |
const float16x4_t vh = vreinterpret_f16_u16(vld1_dup_u16(i)); |
| 68 |
const float32x4_t vt = vcvt_f32_f16(vh); |
| 69 |
vacc = vadd_f32(vacc, vget_low_f32(vt)); |
| 70 |
} |
| 71 |
vacc = vmul_f32(vacc, vscale); |
| 72 |
|
| 73 |
float vout = vget_lane_f32(vacc, 0); |
| 74 |
*output += vout; |
| 75 |
} |