61 行
4.5 KiB
Metal
61 行
4.5 KiB
Metal
#include <metal_stdlib>
|
|
using namespace metal;
|
|
struct PyramidInfo {int width[5],height[5],offset[5],levels;};
|
|
kernel void dyTrackPyramid(device float*f[[buffer(0)]],constant PyramidInfo&i[[buffer(1)]],constant int&level[[buffer(2)]],uint2 p[[thread_position_in_grid]]) {
|
|
int w=i.width[level],h=i.height[level];if(p.x>=uint(w)||p.y>=uint(h))return;
|
|
int sw=i.width[level-1],sh=i.height[level-1];float sum=0;
|
|
for(int y=-1;y<=1;y++)for(int x=-1;x<=1;x++)sum+=f[i.offset[level-1]+clamp(int(p.y)*2+y,0,sh-1)*sw+clamp(int(p.x)*2+x,0,sw-1)]*(x==0?2.f:1.f)*(y==0?2.f:1.f);
|
|
f[i.offset[level]+p.y*w+p.x]=sum/16.f;
|
|
}
|
|
inline bool inside(float2 p,int w,int h){return all(p>=12.f)&&p.x<w-13&&p.y<h-13;}
|
|
inline float pixel(device const float*f,int w,float2 p) {
|
|
int2 a=int2(p);float2 d=p-float2(a);int idx=a.y*w+a.x;
|
|
return mix(mix(f[idx],f[idx+1],d.x),mix(f[idx+w],f[idx+w+1],d.x),d.y);
|
|
}
|
|
// One SIMD group per feature: lanes cooperatively reduce the 21x21 patch.
|
|
// Forward and backward checks stay in the same dispatch and reuse the resident pyramids.
|
|
bool trackFlow(device const float*prev,device const float*next,constant PyramidInfo&i,float2 p,float2 guess,thread float2&q,thread float&error,uint lane) {
|
|
int level=i.levels-1;for(;level>0;level--){float s=float(1<<level);if(inside(p/s,i.width[level],i.height[level])&&inside(guess/s,i.width[level],i.height[level]))break;}
|
|
q=guess/float(1<<level);error=0;
|
|
for(;level>=0;level--) {
|
|
float2 pp=p/float(1<<level);int w=i.width[level],h=i.height[level];if(!inside(pp,w,h))return false;
|
|
device const float*a=prev+i.offset[level];device const float*b=next+i.offset[level];
|
|
float ref[14],gx[14],gy[14],mean=0,hxx=0,hxy=0,hyy=0;
|
|
for(uint j=0;j<14;j++) {
|
|
uint k=lane+j*32;ref[j]=gx[j]=gy[j]=0;if(k>=441)continue;
|
|
float2 pos=pp+float2(int(k%21)-10,int(k/21)-10);ref[j]=pixel(a,w,pos);mean+=ref[j];
|
|
gx[j]=.5f*(pixel(a,w,pos+float2(1,0))-pixel(a,w,pos-float2(1,0)));
|
|
gy[j]=.5f*(pixel(a,w,pos+float2(0,1))-pixel(a,w,pos-float2(0,1)));
|
|
hxx+=gx[j]*gx[j];hxy+=gx[j]*gy[j];hyy+=gy[j]*gy[j];
|
|
}
|
|
mean=simd_sum(mean)/441;hxx=simd_sum(hxx);hxy=simd_sum(hxy);hyy=simd_sum(hyy);
|
|
float det=hxx*hyy-hxy*hxy;if(det<1e-8f||hxx+hyy<1e-5f)return false;
|
|
for(int iteration=0;iteration<30;iteration++) {
|
|
if(!inside(q,w,h))return false;float target[14],tm=0;
|
|
for(uint j=0;j<14;j++){uint k=lane+j*32;target[j]=k<441?pixel(b,w,q+float2(int(k%21)-10,int(k/21)-10)):0;tm+=target[j];}
|
|
tm=simd_sum(tm)/441;float bx=0,by=0,sse=0;
|
|
for(uint j=0;j<14;j++)if(lane+j*32<441){float e=ref[j]-mean-target[j]+tm;bx+=gx[j]*e;by+=gy[j]*e;sse+=e*e;}
|
|
bx=simd_sum(bx);by=simd_sum(by);sse=simd_sum(sse);
|
|
float2 delta=float2(hyy*bx-hxy*by,hxx*by-hxy*bx)/det;if(!all(isfinite(delta))||length(delta)>8)return false;
|
|
q+=delta;error=sqrt(sse/441);if(dot(delta,delta)<.0001f)break;
|
|
}
|
|
if(level>0)q*=2;
|
|
}
|
|
return inside(q,i.width[0],i.height[0])&&error<.12f;
|
|
}
|
|
kernel void dyTrackFlow(device const float*prev[[buffer(0)]],device const float*next[[buffer(1)]],device const float2*points[[buffer(2)]],device float4*results[[buffer(3)]],constant PyramidInfo&i[[buffer(4)]],uint gid[[thread_position_in_grid]],uint lane[[thread_index_in_simdgroup]]) {
|
|
uint idx=gid/32;float2 p=points[idx],q=0,back=0;float error=0,backError=0;
|
|
bool ok=trackFlow(prev,next,i,p,p,q,error,lane);
|
|
if(ok)ok=trackFlow(next,prev,i,q,p,back,backError,lane)&&distance(back,p)<=1.1f;
|
|
if(lane==0)results[idx]=float4(q,error,ok?1.f:0.f);
|
|
}
|
|
|
|
// Full host frame -> top-first display luma. Low-pass the source footprint
|
|
// before decimation so 4K fabric detail cannot alias into inconsistent motion.
|
|
// This analysis-only image never changes the full-resolution color render.
|
|
kernel void dyToolCapture(texture2d<float> source [[texture(0)]],device float* pixels [[buffer(0)]],constant uint3& info [[buffer(1)]],uint2 pos [[thread_position_in_grid]]){
|
|
if(pos.x>=info.x||pos.y>=info.y)return;constexpr sampler f(coord::normalized,address::clamp_to_edge,filter::linear);uint2 taps=clamp(uint2(ceil(float2(source.get_width(),source.get_height())/float2(info.xy))),uint2(1),uint2(8));float sum=0;
|
|
for(uint y=0;y<taps.y;y++)for(uint x=0;x<taps.x;x++){float2 uv=(float2(pos)+(float2(x,y)+.5f)/float2(taps))/float2(info.xy);if(info.z)uv.y=1-uv.y;float4 sample=source.sample(f,uv);float3 linear=sample.a>0?sample.rgb/sample.a:float3(0);float3 encoded=select(1.055f*pow(max(linear,float3(0)),float3(1.f/2.4f))-.055f,linear*12.92f,linear<=.0031308f);sum+=dot(clamp(encoded,0.f,1.f),float3(.299f,.587f,.114f));}
|
|
pixels[pos.y*info.x+pos.x]=sum/float(taps.x*taps.y);
|
|
}
|