blob: bc24791c2e7db9b916c93b23ad33b8ba1d2d08e4 [file] [log] [blame]
// Copyright 2019 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 % 4 == 0
$assert CHANNEL_TILE >= 4
$assert ROW_TILE >= 1
$ABC = "0123456789ABCDEFGHIJKLMN"
#include <assert.h>
#include <arm_neon.h>
#include <xnnpack/math.h>
#include <xnnpack/vmulcaddc.h>
void xnn_f32_vmulcaddc_ukernel_c${CHANNEL_TILE}__${"neonfma" if FMA else "neon"}_${ROW_TILE}x(
size_t rows,
size_t channels,
const float*restrict input,
size_t input_stride,
const float*restrict weights,
float*restrict output,
size_t output_stride,
const union xnn_f32_output_params params[restrict static 1])
{
assert(rows != 0);
assert(channels != 0);
assert(channels % sizeof(float) == 0);
const float* i0 = input;
float* o0 = output;
$for M in range(1, ROW_TILE):
const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_stride);
float* o${M} = (float*) ((uintptr_t) o${M-1} + output_stride);
$if M % 2 == 0:
if XNN_UNPREDICTABLE(rows <= ${M}) {
i${M} = i${M-1};
o${M} = o${M-1};
}
$else:
if XNN_UNPREDICTABLE(rows < ${M+1}) {
i${M} = i${M-1};
o${M} = o${M-1};
}
const size_t input_increment = input_stride * ${ROW_TILE} - channels;
const size_t output_increment = output_stride * ${ROW_TILE} - channels;
const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
do {
const float* w = weights;
size_t c = channels;
for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
$for C in range(0, CHANNEL_TILE, 4):
const float32x4_t vscale${ABC[C:C+4]} = vld1q_f32(w); w += 4;
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
float32x4_t vacc${M}x${ABC[C:C+4]} = vld1q_f32(i${M}); i${M} += 4;
$if not FMA:
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vacc${M}x${ABC[C:C+4]} = vmulq_f32(vacc${M}x${ABC[C:C+4]}, vscale${ABC[C:C+4]});
$for C in range(0, CHANNEL_TILE, 4):
const float32x4_t vbias${ABC[C:C+4]} = vld1q_f32(w); w += 4;
$if not FMA:
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vacc${M}x${ABC[C:C+4]} = vaddq_f32(vacc${M}x${ABC[C:C+4]}, vbias${ABC[C:C+4]});
$else:
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vacc${M}x${ABC[C:C+4]} = vfmaq_f32(vbias${ABC[C:C+4]}, vscale${ABC[C:C+4]}, vacc${M}x${ABC[C:C+4]});
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vacc${M}x${ABC[C:C+4]} = vmaxq_f32(vacc${M}x${ABC[C:C+4]}, vmin);
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vacc${M}x${ABC[C:C+4]} = vminq_f32(vacc${M}x${ABC[C:C+4]}, vmax);
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vst1q_f32(o${M}, vacc${M}x${ABC[C:C+4]}); o${M} += 4;
}
$if CHANNEL_TILE > 4:
for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
const float32x4_t vscale0123 = vld1q_f32(w); w += 4;
$for M in range(ROW_TILE):
float32x4_t vacc${M}x0123 = vld1q_f32(i${M}); i${M} += 4;
$if not FMA:
$for M in range(ROW_TILE):
vacc${M}x0123 = vmulq_f32(vacc${M}x0123, vscale0123);
const float32x4_t vbias0123 = vld1q_f32(w + ${CHANNEL_TILE - 4});
$if not FMA:
$for M in range(ROW_TILE):
vacc${M}x0123 = vaddq_f32(vacc${M}x0123, vbias0123);
$else:
$for M in range(ROW_TILE):
vacc${M}x0123 = vfmaq_f32(vbias0123, vscale0123, vacc${M}x0123);
$for M in range(ROW_TILE):
vacc${M}x0123 = vmaxq_f32(vacc${M}x0123, vmin);
$for M in range(ROW_TILE):
vacc${M}x0123 = vminq_f32(vacc${M}x0123, vmax);
$for M in range(ROW_TILE):
vst1q_f32(o${M}, vacc${M}x0123); o${M} += 4;
}
if XNN_UNLIKELY(c != 0) {
const float32x4_t vscale0123 = vld1q_f32(w);
$for M in range(ROW_TILE):
float32x4_t vacc${M}x0123 = vld1q_f32(i${M}); i${M} = (const float*) ((uintptr_t) i${M} + c);
$if not FMA:
$for M in range(ROW_TILE):
vacc${M}x0123 = vmulq_f32(vacc${M}x0123, vscale0123);
const float32x4_t vbias0123 = vld1q_f32(w + ${CHANNEL_TILE});
$if not FMA:
$for M in range(ROW_TILE):
vacc${M}x0123 = vaddq_f32(vacc${M}x0123, vbias0123);
$else:
$for M in range(ROW_TILE):
vacc${M}x0123 = vfmaq_f32(vbias0123, vscale0123, vacc${M}x0123);
$for M in range(ROW_TILE):
vacc${M}x0123 = vmaxq_f32(vacc${M}x0123, vmin);
$for M in range(ROW_TILE):
vacc${M}x0123 = vminq_f32(vacc${M}x0123, vmax);
$for M in range(ROW_TILE):
float32x2_t vacc${M}x01 = vget_low_f32(vacc${M}x0123);
if (c & (2 * sizeof(float))) {
$for M in range(ROW_TILE):
vst1_f32(o${M}, vacc${M}x01); o${M} += 2;
$for M in range(ROW_TILE):
vacc${M}x01 = vget_high_f32(vacc${M}x0123);
}
if (c & (1 * sizeof(float))) {
$for M in range(ROW_TILE):
vst1_lane_f32(o${M}, vacc${M}x01, 0); o${M} += 1;
}
}
$for M in range(ROW_TILE):
i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
o${M} = (float*) ((uintptr_t) o${M} + output_increment);
$if M % 2 == 1:
if XNN_UNPREDICTABLE(rows < ${ROW_TILE + M + 1}) {
i${M} = i${M-1};
o${M} = o${M-1};
}
$elif M != 0:
if XNN_UNPREDICTABLE(rows <= ${ROW_TILE + M}) {
i${M} = i${M-1};
o${M} = o${M-1};
}
rows = doz(rows, ${ROW_TILE});
} while (rows != 0);
}