|
|
|
|
|
|
|
|
|
|
|
$assert BATCH_TILE >= 1 |
|
#include <assert.h> |
|
#include <stddef.h> |
|
#include <stdint.h> |
|
|
|
#include <xnnpack/math.h> |
|
#include <xnnpack/vsquareabs.h> |
|
|
|
|
|
void xnn_cs16_vsquareabs_ukernel__scalar_x${BATCH_TILE}( |
|
size_t batch, |
|
const int16_t* input, |
|
uint32_t* output) |
|
{ |
|
assert(batch != 0); |
|
assert(batch % (sizeof(int16_t) * 2) == 0); |
|
assert(input != NULL); |
|
assert(output != NULL); |
|
|
|
$if BATCH_TILE == 1: |
|
do { |
|
const int32_t vr = (int32_t) input[0]; |
|
const int32_t vi = (int32_t) input[1]; |
|
input += 2; |
|
|
|
uint32_t vacc = (uint32_t) (vr * vr); |
|
vacc += (uint32_t) (vi * vi); |
|
|
|
*output++ = vacc; |
|
batch -= sizeof(int16_t) * 2; |
|
} while (batch != 0); |
|
$else: |
|
for (; batch >= ${BATCH_TILE * 2} * sizeof(int16_t); batch -= ${BATCH_TILE * 2} * sizeof(int16_t)) { |
|
$for C in range(BATCH_TILE): |
|
const int32_t vr${C} = (int32_t) input[${C * 2}]; |
|
const int32_t vi${C} = (int32_t) input[${C * 2 + 1}]; |
|
input += ${BATCH_TILE} * 2; |
|
|
|
$for C in range(BATCH_TILE): |
|
uint32_t vacc${C} = (uint32_t) (vr${C} * vr${C}); |
|
|
|
$for C in range(BATCH_TILE): |
|
vacc${C} += (uint32_t) (vi${C} * vi${C}); |
|
|
|
$for C in range(BATCH_TILE): |
|
output[${C}] = vacc${C}; |
|
output += ${BATCH_TILE}; |
|
} |
|
$if BATCH_TILE == 2: |
|
if XNN_LIKELY(batch != 0) { |
|
assert(batch == 2 * sizeof(int16_t)); |
|
|
|
const int32_t vr = (int32_t) input[0]; |
|
const int32_t vi = (int32_t) input[1]; |
|
|
|
uint32_t vacc = (uint32_t) (vr * vr); |
|
vacc += (uint32_t) (vi * vi); |
|
|
|
*output = vacc; |
|
} |
|
$else: |
|
if XNN_LIKELY(batch != 0) { |
|
do { |
|
const int32_t vr = (int32_t) input[0]; |
|
const int32_t vi = (int32_t) input[1]; |
|
input += 2; |
|
|
|
uint32_t vacc = (uint32_t) (vr * vr); |
|
vacc += (uint32_t) (vi * vi); |
|
|
|
*output++ = vacc; |
|
batch -= sizeof(int16_t) * 2; |
|
} while (batch != 0); |
|
} |
|
} |
|
|