/*!
Calculate integral images.
\param dst - [Output] Destination block pointer
\param dstr - [Input] Destination block stride
\param src - [Input] Source block pointer
\param tlst - [Input] Temporary line pointer
\param sstr - [Input] Source block stride
\param bw - [Input] Block width
\param bh - [Input] Block height
\param ncu - [Input] Number of CUs (32/64)
*/
void
apu_integral_image(vec32u __cmem* dst, int dstr, const vec08u __cmem* src, vec32u __cmem* tlst, int sstr, int bw, int bh, int ncu)
{
for (int y=0; y<bh; ++y)
{
// Local Integration
const vec08u __cmem* s = src;
vec16u v0 = (vec16u)s[0];
vec16u v1 = (vec16u)s[1];
vec16u v2 = (vec16u)s[2];
vec16u v3 = (vec16u)s[3];
s += 4;
vec16u ltot = 0;
for (int x=4; x<bw; x+=4) chess_loop_range(1,) // at least 8 pixels
{
ltot += v0; v0 = (vec16u)s[0];
ltot += v1; v1 = (vec16u)s[1];
ltot += v2; v2 = (vec16u)s[2];
ltot += v3; v3 = (vec16u)s[3];
s += 4;
}
ltot += v0;
ltot += v1;
ltot += v2;
ltot += v3;
// Reduction
vec32u ttot = (vec32u)ltot;
ltot = vmsr(ltot);
for (int c=1; c<ncu; ++c)
{
ttot += (vec32u)ltot;
ltot = vmsr(ltot);
}
//
s = src;
tlst[bw-1] = dst[bw-1] = ttot + tlst[bw-1]; ttot -= (vec32u)s[bw-1];
for (int x=(bw-2); x>=0; --x)
{
tlst[x] = dst[x] = ttot + tlst[x]; ttot -= (vec32u)s[x];
}
dst += dstr;
src += sstr;
}
}
void neon_init(void) {
uint32_t v;
// F**k this took a long time to discover ...
// First, need to enable access to co-processors c10 and c11 - vfp and neon
v = mrc15("c1, c0, 2");
v |= 0xf<<20;
mcr15("c1, c0, 2", v);
// required apparently
asm volatile("isb");
// Next, just need to enable NEON instructions. The ROM or uboot has already turned it on.
vmsr(FPEXC, 1<<30);
}