blob: fdb559fef9e589bf768db56fbf4ce3858c921709 [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
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <xmmintrin.h>
#include <xnnpack/math.h>
#include <xnnpack/vmulcaddc.h>
void xnn_f32_vmulcaddc_ukernel_c${CHANNEL_TILE}__sse_${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_minmax_params params[restrict XNN_MIN_ELEMENTS(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 __m128 vmin = _mm_load_ps(params->sse.min);
const __m128 vmax = _mm_load_ps(params->sse.max);
do {
const float* w = weights;
size_t c = channels;
for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
const __m128 vscale${ABC[0:4]} = _mm_load_ps(w);
$for C in range(4, CHANNEL_TILE, 4):
const __m128 vscale${ABC[C:C+4]} = _mm_load_ps(w + ${C});
$for M in range(ROW_TILE):
__m128 vacc${M}x${ABC[0:4]} = _mm_loadu_ps(i${M});
$for C in range(4, CHANNEL_TILE, 4):
__m128 vacc${M}x${ABC[C:C+4]} = _mm_loadu_ps(i${M} + ${C});
i${M} += ${CHANNEL_TILE};
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vacc${M}x${ABC[C:C+4]} = _mm_mul_ps(vacc${M}x${ABC[C:C+4]}, vscale${ABC[C:C+4]});
$for C in range(0, CHANNEL_TILE, 4):
const __m128 vbias${ABC[C:C+4]} = _mm_load_ps(w + ${C + CHANNEL_TILE});
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vacc${M}x${ABC[C:C+4]} = _mm_add_ps(vacc${M}x${ABC[C:C+4]}, vbias${ABC[C:C+4]});
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vacc${M}x${ABC[C:C+4]} = _mm_max_ps(vacc${M}x${ABC[C:C+4]}, vmin);
$for M in range(ROW_TILE):
$for C in range(0, CHANNEL_TILE, 4):
vacc${M}x${ABC[C:C+4]} = _mm_min_ps(vacc${M}x${ABC[C:C+4]}, vmax);
$for M in range(ROW_TILE):
_mm_storeu_ps(o${M}, vacc${M}x${ABC[0:4]});
$for C in range(4, CHANNEL_TILE, 4):
_mm_storeu_ps(o${M} + ${C}, vacc${M}x${ABC[C:C+4]});
o${M} += ${CHANNEL_TILE};
w += ${CHANNEL_TILE * 2};
}
$if CHANNEL_TILE > 4:
for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
const __m128 vscale0123 = _mm_load_ps(w);
$for M in range(ROW_TILE):
__m128 vacc${M}x0123 = _mm_loadu_ps(i${M});
i${M} += 4;
$for M in range(ROW_TILE):
vacc${M}x0123 = _mm_mul_ps(vacc${M}x0123, vscale0123);
const __m128 vbias0123 = _mm_load_ps(w + ${CHANNEL_TILE});
$for M in range(ROW_TILE):
vacc${M}x0123 = _mm_add_ps(vacc${M}x0123, vbias0123);
$for M in range(ROW_TILE):
vacc${M}x0123 = _mm_max_ps(vacc${M}x0123, vmin);
$for M in range(ROW_TILE):
vacc${M}x0123 = _mm_min_ps(vacc${M}x0123, vmax);
$for M in range(ROW_TILE):
_mm_storeu_ps(o${M}, vacc${M}x0123);
o${M} += 4;
w += 4;
}
if XNN_UNLIKELY(c != 0) {
const __m128 vscale0123 = _mm_load_ps(w);
$for M in range(ROW_TILE):
__m128 vacc${M}x0123 = _mm_loadu_ps(i${M});
i${M} = (const float*) ((uintptr_t) i${M} + c);
$for M in range(ROW_TILE):
vacc${M}x0123 = _mm_mul_ps(vacc${M}x0123, vscale0123);
const __m128 vbias0123 = _mm_load_ps(w + ${CHANNEL_TILE});
$for M in range(ROW_TILE):
vacc${M}x0123 = _mm_add_ps(vacc${M}x0123, vbias0123);
$for M in range(ROW_TILE):
vacc${M}x0123 = _mm_max_ps(vacc${M}x0123, vmin);
$for M in range(ROW_TILE):
vacc${M}x0123 = _mm_min_ps(vacc${M}x0123, vmax);
if (c & (2 * sizeof(float))) {
$for M in range(ROW_TILE):
_mm_storel_pi((__m64*) o${M}, vacc${M}x0123);
$for M in range(ROW_TILE):
vacc${M}x0123 = _mm_movehl_ps(vacc${M}x0123, vacc${M}x0123);
$for M in range(ROW_TILE):
o${M} += 2;
}
if (c & (1 * sizeof(float))) {
$for M in range(ROW_TILE):
_mm_store_ss(o${M}, vacc${M}x0123);
$for M in range(ROW_TILE):
o${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);
}