test / src /f16-gavgpool /multipass-neonfp16arith.c.in
Androidonnxfork's picture
Upload folder using huggingface_hub
8b7c501
raw
history blame
8.71 kB
// Copyright 2020 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 CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert ROW_TILE >= 3
$assert ROW_SUBTILE >= 3
$assert ROW_SUBTILE <= ROW_TILE
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <arm_neon.h>
#include <xnnpack/gavgpool.h>
#include <xnnpack/math.h>
void xnn_f16_gavgpool_minmax_ukernel_${ROW_TILE}p${ROW_SUBTILE}x__neonfp16arith_c${CHANNEL_TILE}(
size_t rows,
size_t channels,
const void* input,
size_t input_stride,
const void* zero,
void* buffer,
void* output,
const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(rows > ${ROW_TILE});
assert(channels != 0);
const uint16_t* i0 = input;
$for M in range(1, ROW_TILE):
const uint16_t* i${M} = (const uint16_t*) ((uintptr_t) i${M-1} + input_stride);
const size_t input_increment = ${ROW_TILE} * input_stride - round_up_po2(channels, 8) * sizeof(uint16_t);
uint16_t* b = buffer;
size_t c = channels;
for (; ${"c >= %d" % CHANNEL_TILE if CHANNEL_TILE > 8 else "c != 0"}; ${("c -= %d" if CHANNEL_TILE > 8 else "c = doz(c, %d)") % CHANNEL_TILE}) {
$for M in range(2):
$for C in range(0, CHANNEL_TILE, 8):
const float16x8_t vi${M}x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i${M})); i${M} += 8;
$for C in range(0, CHANNEL_TILE, 8):
const float16x8_t vi2x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
float16x8_t vacc${ABC[C:C+8]} = vaddq_f16(vi0x${ABC[C:C+8]}, vi1x${ABC[C:C+8]});
$for M in range(2, ROW_TILE):
$for C in range(0, CHANNEL_TILE, 8):
$if M + 1 != ROW_TILE:
const float16x8_t vi${M+1}x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i${M+1})); i${M+1} += 8;
vacc${ABC[C:C+8]} = vaddq_f16(vacc${ABC[C:C+8]}, vi${M}x${ABC[C:C+8]});
$for C in range(0, CHANNEL_TILE, 8):
vst1q_u16(b, vreinterpretq_u16_f16(vacc${ABC[C:C+8]})); b += 8;
}
$if CHANNEL_TILE > 8:
if XNN_UNLIKELY(c != 0) {
do {
$for M in range(3):
const float16x8_t vi${M}x${ABC[0:8]} = vreinterpretq_f16_u16(vld1q_u16(i${M})); i${M} += 8;
float16x8_t vacc${ABC[0:8]} = vaddq_f16(vi0x${ABC[0:8]}, vi1x${ABC[0:8]});
$for M in range(2, ROW_TILE):
$if M + 1 != ROW_TILE:
const float16x8_t vi${M+1}x${ABC[0:8]} = vreinterpretq_f16_u16(vld1q_u16(i${M+1})); i${M+1} += 8;
vacc${ABC[0:8]} = vaddq_f16(vacc${ABC[0:8]}, vi${M}x${ABC[0:8]});
vst1q_u16(b, vreinterpretq_u16_f16(vacc${ABC[0:8]})); b += 8;
c = doz(c, 8);
} while (c != 0);
}
for (rows -= ${ROW_TILE}; rows > ${ROW_SUBTILE}; rows -= ${ROW_SUBTILE}) {
$for M in range(ROW_SUBTILE):
i${M} = (const uint16_t*) ((uintptr_t) i${M + ROW_TILE - ROW_SUBTILE} + input_increment);
uint16_t* b = buffer;
size_t c = channels;
for (; ${"c >= %d" % CHANNEL_TILE if CHANNEL_TILE > 8 else "c != 0"}; ${("c -= %d" if CHANNEL_TILE > 8 else "c = doz(c, %d)") % CHANNEL_TILE}) {
float16x8_t vacc${ABC[0:8]} = vreinterpretq_f16_u16(vld1q_u16(b));
$for C in range(8, CHANNEL_TILE, 8):
float16x8_t vacc${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(b + ${C}));
$for C in range(0, CHANNEL_TILE, 8):
const float16x8_t vi0x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 8):
$if M + 1 != ROW_TILE:
const float16x8_t vi${M+1}x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i${M+1})); i${M+1} += 8;
vacc${ABC[C:C+8]} = vaddq_f16(vacc${ABC[C:C+8]}, vi${M}x${ABC[C:C+8]});
$for C in range(0, CHANNEL_TILE, 8):
vst1q_u16(b, vreinterpretq_u16_f16(vacc${ABC[C:C+8]})); b += 8;
}
$if CHANNEL_TILE > 8:
if XNN_UNLIKELY(c != 0) {
do {
float16x8_t vacc${ABC[0:8]} = vreinterpretq_f16_u16(vld1q_u16(b));
const float16x8_t vi0x${ABC[0:8]} = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
$for M in range(ROW_TILE):
$if M + 1 != ROW_TILE:
const float16x8_t vi${M+1}x${ABC[0:8]} = vreinterpretq_f16_u16(vld1q_u16(i${M+1})); i${M+1} += 8;
vacc${ABC[0:8]} = vaddq_f16(vacc${ABC[0:8]}, vi${M}x${ABC[0:8]});
vst1q_u16(b, vreinterpretq_u16_f16(vacc${ABC[0:8]})); b += 8;
c = doz(c, 8);
} while (c != 0);
}
}
i0 = (const uint16_t*) ((uintptr_t) i${ROW_TILE - ROW_SUBTILE} + input_increment);
$for M in range(1, ROW_SUBTILE):
i${M} = (const uint16_t*) ((uintptr_t) i${M + ROW_TILE - ROW_SUBTILE} + input_increment);
$if M % 2 == 1:
if XNN_UNPREDICTABLE(rows < ${M+1}) {
i${M} = (const uint16_t*) zero;
}
$else:
if XNN_UNPREDICTABLE(rows <= ${M}) {
i${M} = (const uint16_t*) zero;
}
const float16x8_t vscale = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.scale));
const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.min));
const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.max));
for (; channels >= ${CHANNEL_TILE}; channels -= ${CHANNEL_TILE}) {
$for C in range(0, CHANNEL_TILE, 8):
float16x8_t vacc${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(buffer)); buffer = (uint16_t*) buffer + 8;
$for C in range(0, CHANNEL_TILE, 8):
const float16x8_t vi0x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 8):
$if M + 1 != ROW_TILE:
const float16x8_t vi${M+1}x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i${M+1})); i${M+1} += 8;
vacc${ABC[C:C+8]} = vaddq_f16(vacc${ABC[C:C+8]}, vi${M}x${ABC[C:C+8]});
$for C in range(0, CHANNEL_TILE, 8):
vacc${ABC[C:C+8]} = vmulq_f16(vacc${ABC[C:C+8]}, vscale);
$for C in range(0, CHANNEL_TILE, 8):
vacc${ABC[C:C+8]} = vmaxq_f16(vacc${ABC[C:C+8]}, vmin);
$for C in range(0, CHANNEL_TILE, 8):
vacc${ABC[C:C+8]} = vminq_f16(vacc${ABC[C:C+8]}, vmax);
$for C in range(0, CHANNEL_TILE, 8):
vst1q_u16(output, vreinterpretq_u16_f16(vacc${ABC[C:C+8]})); output = (uint16_t*) output + 8;
}
if XNN_UNLIKELY(channels != 0) {
${"do " if CHANNEL_TILE > 8 else ""}{
float16x8_t vacc${ABC[0:8]} = vreinterpretq_f16_u16(vld1q_u16(buffer)); buffer = (uint16_t*) buffer + 8;
const float16x8_t vi0x${ABC[0:8]} = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
$for M in range(ROW_TILE):
$if M + 1 != ROW_TILE:
const float16x8_t vi${M+1}x${ABC[0:8]} = vreinterpretq_f16_u16(vld1q_u16(i${M+1})); i${M+1} += 8;
vacc${ABC[0:8]} = vaddq_f16(vacc${ABC[0:8]}, vi${M}x${ABC[0:8]});
vacc${ABC[0:8]} = vmulq_f16(vacc${ABC[0:8]}, vscale);
vacc${ABC[0:8]} = vmaxq_f16(vacc${ABC[0:8]}, vmin);
vacc${ABC[0:8]} = vminq_f16(vacc${ABC[0:8]}, vmax);
$if CHANNEL_TILE > 8:
if XNN_LIKELY(channels >= 8) {
vst1q_u16(output, vreinterpretq_u16_f16(vacc${ABC[0:8]})); output = (uint16_t*) output + 8;
channels -= 8;
} else {
float16x4_t vacc${ABC[0:4]} = vget_low_f16(vacc${ABC[0:8]});
if (channels & 4) {
vst1_u16(output, vreinterpret_u16_f16(vacc${ABC[0:4]})); output = (uint16_t*) output + 4;
vacc${ABC[0:4]} = vget_high_f16(vacc${ABC[0:8]});
}
if (channels & 2) {
vst1_lane_u32(output, vreinterpret_u32_f16(vacc${ABC[0:4]}), 0); output = (uint16_t*) output + 2;
vacc${ABC[0:4]} = vext_f16(vacc${ABC[0:4]}, vacc${ABC[0:4]}, 2);
}
if (channels & 1) {
vst1_lane_u16(output, vreinterpret_u16_f16(vacc${ABC[0:4]}), 0); output = (uint16_t*) output + 1;
}
channels = 0;
}
$else:
float16x4_t vacc${ABC[0:4]} = vget_low_f16(vacc${ABC[0:8]});
if (channels & 4) {
vst1_u16(output, vreinterpret_u16_f16(vacc${ABC[0:4]})); output = (uint16_t*) output + 4;
vacc${ABC[0:4]} = vget_high_f16(vacc${ABC[0:8]});
}
if (channels & 2) {
vst1_lane_u32(output, vreinterpret_u32_f16(vacc${ABC[0:4]}), 0); output = (uint16_t*) output + 2;
vacc${ABC[0:4]} = vext_f16(vacc${ABC[0:4]}, vacc${ABC[0:4]}, 2);
}
if (channels & 1) {
vst1_lane_u16(output, vreinterpret_u16_f16(vacc${ABC[0:4]}), 0); output = (uint16_t*) output + 1;
}
}${" while (channels != 0);" if CHANNEL_TILE > 8 else ""}
}
}