void test_vcvtQf32_u32 (void)
{
float32x4_t out_float32x4_t;
uint32x4_t arg0_uint32x4_t;
out_float32x4_t = vcvtq_f32_u32 (arg0_uint32x4_t);
}
void word2float48_neon(const uint8_t *t8, const int pitch, float *p) {
const uint16_t *t = (const uint16_t *)t8;
for (int i = 0; i < 4; i++) {
uint32x4_t u1 = vmovl_u16(vld1_u16(t));
uint32x4_t u2 = vmovl_u16(vld1_u16(t + 4));
uint32x4_t u3 = vmovl_u16(vld1_u16(t + 8));
float32x4_t f1 = vcvtq_f32_u32(u1);
float32x4_t f2 = vcvtq_f32_u32(u2);
float32x4_t f3 = vcvtq_f32_u32(u3);
vst1q_f32(p, f1);
vst1q_f32(p + 4, f2);
vst1q_f32(p + 8, f3);
t += pitch * 2; // it was already halved
p += 12;
}
}
//Kernel function: saxpy
void saxpy_vector(KernelArgs* args) {
//Setup
const float32x4_t MASK_FALSE = vdupq_n_f32(0.f);
const float32x4_t MASK_TRUE = vcvtq_f32_u32(vceqq_f32(MASK_FALSE, MASK_FALSE));
//Uniforms
//Fuses
//Literals
//Stack variables
float32x4_t scale, x, y, result, var060, var061;
//Loop over input
uint64_t index;
for(index = 0; index < args->N; index += 4) {
//Inputs
scale = vld1q_f32(&args->scale[index]);
x = vld1q_f32(&args->x[index]);
y = vld1q_f32(&args->y[index]);
//Begin kernel logic
{
//>>> result = scale * x + y
var061 = vmulq_f32(scale, x);
var060 = vaddq_f32(var061, y);
result = vbslq_f32(vcvtq_u32_f32(MASK_TRUE), var060, result);
}
//End kernel logic
//Outputs
vst1q_f32(&args->result[index], result);
}
}
void byte2float48_neon(const uint8_t *t, const int pitch, float *p) {
uint16x8_t m0, m1, m2, m3, m4, m5;
uint32x2_t temp1, temp4;
m0 = vmovl_u8(vld1_u8(t));
temp1 = vld1_lane_u32((const uint32_t *)(t + 8), temp1, 0);
temp1 = vld1_lane_u32((const uint32_t *)(t + pitch * 2), temp1, 1);
m1 = vmovl_u8(vreinterpret_u8_u32(temp1));
m2 = vmovl_u8(vld1_u8(t + pitch * 2 + 4));
t += pitch * 4;
m3 = vmovl_u8(vld1_u8(t));
temp4 = vld1_lane_u32((const uint32_t *)(t + 8), temp4, 0);
temp4 = vld1_lane_u32((const uint32_t *)(t + pitch * 2), temp4, 1);
m4 = vmovl_u8(vreinterpret_u8_u32(temp4));
m5 = vmovl_u8(vld1_u8(t + pitch * 2 + 4));
vst1q_f32(p, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m0))));
vst1q_f32(p + 4, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m0))));
vst1q_f32(p + 8, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m1))));
vst1q_f32(p + 12, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m1))));
vst1q_f32(p + 16, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m2))));
vst1q_f32(p + 20, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m2))));
vst1q_f32(p + 24, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m3))));
vst1q_f32(p + 28, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m3))));
vst1q_f32(p + 32, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m4))));
vst1q_f32(p + 36, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m4))));
vst1q_f32(p + 40, vcvtq_f32_u32(vmovl_u16(vget_low_u16(m5))));
vst1q_f32(p + 44, vcvtq_f32_u32(vmovl_u16(vget_high_u16(m5))));
}
void meanStdDev(const Size2D &size,
const u16 * srcBase, ptrdiff_t srcStride,
f32 * pMean, f32 * pStdDev)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
size_t blockSize0 = 1 << 10, roiw4 = size.width & ~3;
f64 fsum = 0.0f, fsqsum = 0.0f;
f32 arsum[8];
uint32x4_t v_zero = vdupq_n_u32(0u), v_sum;
float32x4_t v_zero_f = vdupq_n_f32(0.0f), v_sqsum;
for (size_t i = 0; i < size.height; ++i)
{
const u16 * src = internal::getRowPtr(srcBase, srcStride, i);
size_t j = 0u;
while (j < roiw4)
{
size_t blockSize = std::min(roiw4 - j, blockSize0) + j;
v_sum = v_zero;
v_sqsum = v_zero_f;
for ( ; j + 16 < blockSize ; j += 16)
{
internal::prefetch(src + j);
uint16x8_t v_src0 = vld1q_u16(src + j), v_src1 = vld1q_u16(src + j + 8);
// 0
uint32x4_t v_srclo = vmovl_u16(vget_low_u16(v_src0));
uint32x4_t v_srchi = vmovl_u16(vget_high_u16(v_src0));
v_sum = vaddq_u32(v_sum, vaddq_u32(v_srclo, v_srchi));
float32x4_t v_srclo_f = vcvtq_f32_u32(v_srclo);
float32x4_t v_srchi_f = vcvtq_f32_u32(v_srchi);
v_sqsum = vmlaq_f32(v_sqsum, v_srclo_f, v_srclo_f);
v_sqsum = vmlaq_f32(v_sqsum, v_srchi_f, v_srchi_f);
// 1
v_srclo = vmovl_u16(vget_low_u16(v_src1));
v_srchi = vmovl_u16(vget_high_u16(v_src1));
v_sum = vaddq_u32(v_sum, vaddq_u32(v_srclo, v_srchi));
v_srclo_f = vcvtq_f32_u32(v_srclo);
v_srchi_f = vcvtq_f32_u32(v_srchi);
v_sqsum = vmlaq_f32(v_sqsum, v_srclo_f, v_srclo_f);
v_sqsum = vmlaq_f32(v_sqsum, v_srchi_f, v_srchi_f);
}
for ( ; j < blockSize; j += 4)
{
uint32x4_t v_src = vmovl_u16(vld1_u16(src + j));
float32x4_t v_src_f = vcvtq_f32_u32(v_src);
v_sum = vaddq_u32(v_sum, v_src);
v_sqsum = vmlaq_f32(v_sqsum, v_src_f, v_src_f);
}
vst1q_f32(arsum, vcvtq_f32_u32(v_sum));
vst1q_f32(arsum + 4, v_sqsum);
fsum += (f64)arsum[0] + arsum[1] + arsum[2] + arsum[3];
fsqsum += (f64)arsum[4] + arsum[5] + arsum[6] + arsum[7];
}
// collect a few last elements in the current row
for ( ; j < size.width; ++j)
{
f32 srcval = src[j];
fsum += srcval;
fsqsum += srcval * srcval;
}
}
// calc mean and stddev
f64 itotal = 1.0 / size.total();
f64 mean = fsum * itotal;
f64 stddev = sqrt(std::max(fsqsum * itotal - mean * mean, 0.0));
if (pMean)
*pMean = mean;
if (pStdDev)
*pStdDev = stddev;
#else
(void)size;
(void)srcBase;
(void)srcStride;
(void)pMean;
(void)pStdDev;
#endif
}
