inline float32x2_t cv_vrsqrt_f32(float32x2_t val)
{
float32x2_t e = vrsqrte_f32(val);
e = vmul_f32(vrsqrts_f32(vmul_f32(e, e), val), e);
e = vmul_f32(vrsqrts_f32(vmul_f32(e, e), val), e);
return e;
}
inline float32x2_t cv_vrecp_f32(float32x2_t val)
{
float32x2_t reciprocal = vrecpe_f32(val);
reciprocal = vmul_f32(vrecps_f32(val, reciprocal), reciprocal);
reciprocal = vmul_f32(vrecps_f32(val, reciprocal), reciprocal);
return reciprocal;
}
f64 dotProduct(const Size2D &_size,
const f32 * src0Base, ptrdiff_t src0Stride,
const f32 * src1Base, ptrdiff_t src1Stride)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
Size2D size(_size);
if (src0Stride == src1Stride &&
src0Stride == (ptrdiff_t)(size.width * sizeof(f32)))
{
size.width *= size.height;
size.height = 1;
}
#define DOT_FLOAT_BLOCKSIZE (1 << 13)
f64 result = 0.0;
for (size_t row = 0; row < size.height; ++row)
{
const f32 * src0 = internal::getRowPtr(src0Base, src0Stride, row);
const f32 * src1 = internal::getRowPtr(src1Base, src1Stride, row);
size_t i = 0;
while(i + 4 <= size.width)
{
size_t lim = std::min(i + DOT_FLOAT_BLOCKSIZE, size.width) - 4;
float32x4_t v_sum = vdupq_n_f32(0.0f);
for( ; i <= lim; i += 4 )
{
internal::prefetch(src0 + i);
internal::prefetch(src1 + i);
v_sum = vmlaq_f32(v_sum, vld1q_f32(src0 + i), vld1q_f32(src1 + i));
}
float32x2_t vres = vpadd_f32(vget_low_f32(v_sum),vget_high_f32(v_sum));
result += vget_lane_f32(vres, 0) + vget_lane_f32(vres, 1);
}
if(i + 2 <= size.width)
{
float32x2_t vres = vmul_f32(vld1_f32(src0 + i), vld1_f32(src1 + i));
result += vget_lane_f32(vres, 0) + vget_lane_f32(vres, 1);
i += 2;
}
for (; i < size.width; ++i)
result += src0[i] * src1[i];
}
return result;
#else
(void)_size;
(void)src0Base;
(void)src0Stride;
(void)src1Base;
(void)src1Stride;
return 0;
#endif
}
void test_vmulf32 (void)
{
float32x2_t out_float32x2_t;
float32x2_t arg0_float32x2_t;
float32x2_t arg1_float32x2_t;
out_float32x2_t = vmul_f32 (arg0_float32x2_t, arg1_float32x2_t);
}
void dot_loop(const GLfloat *verts, const GLfloat *params, GLfloat *out, GLint count) {
#ifdef __ARM_NEON__
float32x2_t acc;
float32x2x3_t vert;
float32x2x3_t param = vld3_f32((const float32_t *)params);
for (; count != 0; count -= 1) {
vert = vld3_f32((const float32_t *)verts);
acc = vmul_f32(vert.val[0], param.val[0]);
acc = vmla_f32(acc, vert.val[1], param.val[1]);
acc = vmla_f32(acc, vert.val[2], param.val[2]);
vst1_f32((float32_t *)out, acc);
out += 2;
verts += 3;
}
#else
for (int i = 0; i < count; i++) {
out[0] = dot(verts, params);
out += 2;
verts += 3;
}
#endif
}