|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#include <assert.h> |
|
|
|
#include <arm_fp16.h> |
|
|
|
#include <xnnpack/common.h> |
|
#include <xnnpack/intrinsics-polyfill.h> |
|
#include <xnnpack/vbinary.h> |
|
|
|
|
|
void xnn_f16_vmin_ukernel__fp16arith_x4( |
|
size_t batch, |
|
const void* restrict input_a, |
|
const void* restrict input_b, |
|
void* restrict output, |
|
const union xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS(1)]) |
|
{ |
|
assert(batch != 0); |
|
assert(batch % sizeof(float16_t) == 0); |
|
assert(input_a != NULL); |
|
assert(input_b != NULL); |
|
assert(output != NULL); |
|
|
|
const float16_t* a = (const float16_t*) input_a; |
|
const float16_t* b = (const float16_t*) input_b; |
|
float16_t* o = (float16_t*) output; |
|
|
|
|
|
for (; batch >= 4 * sizeof(float16_t); batch -= 4 * sizeof(float16_t)) { |
|
const float16_t va0 = *a++; |
|
const float16_t va1 = *a++; |
|
const float16_t va2 = *a++; |
|
const float16_t va3 = *a++; |
|
|
|
const float16_t vb0 = *b++; |
|
const float16_t vb1 = *b++; |
|
const float16_t vb2 = *b++; |
|
const float16_t vb3 = *b++; |
|
|
|
float16_t vacc0 = vminnmh_f16(va0, vb0); |
|
float16_t vacc1 = vminnmh_f16(va1, vb1); |
|
float16_t vacc2 = vminnmh_f16(va2, vb2); |
|
float16_t vacc3 = vminnmh_f16(va3, vb3); |
|
|
|
|
|
|
|
*o++ = vacc0; |
|
*o++ = vacc1; |
|
*o++ = vacc2; |
|
*o++ = vacc3; |
|
} |
|
if XNN_UNLIKELY(batch != 0) { |
|
do { |
|
const float16_t va = *a++; |
|
const float16_t vb = *b++; |
|
float16_t vacc = vminnmh_f16(va, vb); |
|
*o++ = vacc; |
|
batch -= sizeof(float16_t); |
|
} while (batch != 0); |
|
} |
|
} |
|
|