|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#include <assert.h> |
|
|
|
#include <xnnpack/common.h> |
|
#include <xnnpack/math.h> |
|
#include <xnnpack/vcvt.h> |
|
|
|
|
|
void xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x2( |
|
size_t batch, |
|
const float* input, |
|
void* output, |
|
const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) |
|
{ |
|
assert(batch != 0); |
|
assert(batch % sizeof(float) == 0); |
|
assert(input != NULL); |
|
assert(output != NULL); |
|
|
|
const uint32_t vnonsign_mask = params->scalar_bitcast.nonsign_mask; |
|
const uint32_t vexp_bias = params->scalar_bitcast.exp_bias; |
|
const float vscale_to_inf = params->scalar_bitcast.scale_to_inf; |
|
const uint32_t vexpw_max = params->scalar_bitcast.expw_max; |
|
const float vscale_to_zero = params->scalar_bitcast.scale_to_zero; |
|
const uint32_t vbias_min = params->scalar_bitcast.bias_min; |
|
const uint16_t vexph_mask = params->scalar_bitcast.exph_mask; |
|
const uint16_t vmanth_mask = params->scalar_bitcast.manth_mask; |
|
const uint16_t vnanh = params->scalar_bitcast.nanh; |
|
|
|
const uint32_t* i = (const uint32_t*) input; |
|
uint16_t* o = (uint16_t*) output; |
|
for (; batch >= 2 * sizeof(float); batch -= 2 * sizeof(float)) { |
|
const uint32_t vw0 = i[0]; |
|
const uint32_t vw1 = i[1]; |
|
i += 2; |
|
|
|
const uint32_t vnonsignw0 = vw0 & vnonsign_mask; |
|
const uint32_t vnonsignw1 = vw1 & vnonsign_mask; |
|
|
|
float vf0 = uint32_as_float(vnonsignw0); |
|
float vf1 = uint32_as_float(vnonsignw1); |
|
const uint32_t vsignw0 = vw0 ^ vnonsignw0; |
|
const uint32_t vsignw1 = vw1 ^ vnonsignw1; |
|
uint32_t vbias0 = vnonsignw0 + vexp_bias; |
|
uint32_t vbias1 = vnonsignw1 + vexp_bias; |
|
|
|
vf0 *= vscale_to_inf; |
|
vf1 *= vscale_to_inf; |
|
vbias0 &= vexpw_max; |
|
vbias1 &= vexpw_max; |
|
|
|
vf0 *= vscale_to_zero; |
|
vf1 *= vscale_to_zero; |
|
vbias0 = math_max_u32(vbias0, vbias_min); |
|
vbias1 = math_max_u32(vbias1, vbias_min); |
|
|
|
vf0 += uint32_as_float(vbias0); |
|
vf1 += uint32_as_float(vbias1); |
|
|
|
const uint32_t vbits0 = float_as_uint32(vf0); |
|
const uint32_t vbits1 = float_as_uint32(vf1); |
|
|
|
const uint16_t vexph0 = (uint16_t) (vbits0 >> 13) & vexph_mask; |
|
const uint16_t vexph1 = (uint16_t) (vbits1 >> 13) & vexph_mask; |
|
const uint16_t vmanth0 = (uint16_t) vbits0 & vmanth_mask; |
|
const uint16_t vmanth1 = (uint16_t) vbits1 & vmanth_mask; |
|
const uint16_t vsignh0 = (uint16_t) (vsignw0 >> 16); |
|
const uint16_t vsignh1 = (uint16_t) (vsignw1 >> 16); |
|
|
|
uint16_t vh0 = vexph0 + vmanth0; |
|
uint16_t vh1 = vexph1 + vmanth1; |
|
if XNN_UNPREDICTABLE(vnonsignw0 > vexpw_max) { |
|
vh0 = vnanh; |
|
} |
|
if XNN_UNPREDICTABLE(vnonsignw1 > vexpw_max) { |
|
vh1 = vnanh; |
|
} |
|
vh0 |= vsignh0; |
|
vh1 |= vsignh1; |
|
|
|
o[0] = vh0; |
|
o[1] = vh1; |
|
o += 2; |
|
} |
|
if XNN_UNLIKELY(batch != 0) { |
|
const uint32_t vw = *i; |
|
|
|
const uint32_t vnonsignw = vw & vnonsign_mask; |
|
|
|
float vf = uint32_as_float(vnonsignw); |
|
const uint32_t vsignw = vw ^ vnonsignw; |
|
uint32_t vbias = vnonsignw + vexp_bias; |
|
|
|
vf *= vscale_to_inf; |
|
vbias &= vexpw_max; |
|
|
|
vf *= vscale_to_zero; |
|
vbias = math_max_u32(vbias, vbias_min); |
|
|
|
vf += uint32_as_float(vbias); |
|
|
|
const uint32_t vbits = float_as_uint32(vf); |
|
|
|
const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask; |
|
const uint16_t vmanth = (uint16_t) vbits & vmanth_mask; |
|
const uint16_t vsignh = (uint16_t) (vsignw >> 16); |
|
|
|
uint16_t vh = vexph + vmanth; |
|
if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) { |
|
vh = vnanh; |
|
} |
|
vh |= vsignh; |
|
|
|
*o = vh; |
|
} |
|
} |
|
|