Kernels:

---

phanerozoic

/

skyrim-vr-upscale

like 0

// perf.cpp — measures the resolution-scaling performance benefit of the SVR
// upscale path, isolated with GPU timestamp queries. No VR runtime involved.
//
// The saving from any upscaler is that the game renders fewer, cheaper pixels;
// the kernel then reconstructs full resolution at a small fixed cost. This
// harness renders an expensive procedural fragment shader (domain-warped fBm,
// octave count = per-pixel shading load) to eye-sized render targets, times
// the scene on the GPU at full resolution and at renderScale R, runs the real
// SVR kernel (svr_cuda.dll) to reconstruct R -> full, and reports the net
// frame speedup against the pixel-count ceiling. The octave sweep spans light
// to heavy shading so the speedup curve is visible; a real title sits on it
// wherever its own per-pixel cost lands.
#include <d3d11.h>
#include <d3dcompiler.h>

#include <cstdint>
#include <cstdio>
#include <vector>

#pragma comment(lib, "d3d11.lib")
#pragma comment(lib, "d3dcompiler.lib")

static const char* kShader = R"(
cbuffer CB : register(b0) { float4 p; }  // x=octaves, yz=size, w=time
struct VSOut { float4 pos : SV_Position; float2 uv : TEXCOORD0; };
VSOut vs(uint id : SV_VertexID) {
  VSOut o; float2 xy = float2((id << 1) & 2, id & 2);
  o.pos = float4(xy * 2.0 - 1.0, 0.0, 1.0); o.uv = xy; return o;
}
float hash(float2 n){ return frac(sin(dot(n, float2(12.9898,78.233)))*43758.5453); }
float vnoise(float2 x){
  float2 i=floor(x), f=frac(x); f=f*f*(3.0-2.0*f);
  float a=hash(i), b=hash(i+float2(1,0)), c=hash(i+float2(0,1)), d=hash(i+float2(1,1));
  return lerp(lerp(a,b,f.x), lerp(c,d,f.x), f.y);
}
float4 ps(VSOut v) : SV_Target {
  float2 uv = v.uv * p.yz * 0.02;
  float2 q = uv; float acc = 0.0, amp = 0.5;
  int N = (int)p.x;
  [loop] for (int i = 0; i < N; ++i) {
    q += float2(vnoise(q*1.7 + p.w), vnoise(q*1.7 - p.w)) * 0.35;
    acc += amp * vnoise(q); amp *= 0.6; q *= 1.9;
  }
  float3 col = 0.5 + 0.5*cos(6.2831*(acc + float3(0.0,0.33,0.67)));
  return float4(col, 1.0);
}
)";

typedef int (*SvrProcessFn)(ID3D11Texture2D*, int, int, int, int, int,
                            ID3D11Texture2D*, int, int, int, int, int,
                            float, float, float, float, float);

static ID3D11Texture2D* mkrt(ID3D11Device* d, int w, int h) {
  D3D11_TEXTURE2D_DESC td = {};
  td.Width = w; td.Height = h; td.MipLevels = 1; td.ArraySize = 1;
  td.Format = DXGI_FORMAT_R8G8B8A8_UNORM; td.SampleDesc.Count = 1;
  td.Usage = D3D11_USAGE_DEFAULT;
  td.BindFlags = D3D11_BIND_RENDER_TARGET | D3D11_BIND_SHADER_RESOURCE;
  ID3D11Texture2D* t = nullptr; d->CreateTexture2D(&td, nullptr, &t); return t;
}

int main() {
  HMODULE svr = LoadLibraryW(L"svr_cuda.dll");
  SvrProcessFn svr_process = svr ? (SvrProcessFn)GetProcAddress(svr, "svr_process") : nullptr;
  if (!svr_process) { printf("svr_cuda.dll svr_process missing\n"); return 1; }

  ID3D11Device* dev = nullptr; ID3D11DeviceContext* ctx = nullptr;
  D3D_FEATURE_LEVEL fl;
  if (FAILED(D3D11CreateDevice(nullptr, D3D_DRIVER_TYPE_HARDWARE, nullptr, 0,
                               nullptr, 0, D3D11_SDK_VERSION, &dev, &fl, &ctx))) {
    printf("device create failed\n"); return 1;
  }

  ID3DBlob *vsb=nullptr,*psb=nullptr,*eb=nullptr;
  D3DCompile(kShader, strlen(kShader), 0,0,0, "vs","vs_5_0",0,0,&vsb,&eb);
  if (FAILED(D3DCompile(kShader, strlen(kShader), 0,0,0, "ps","ps_5_0",0,0,&psb,&eb))) {
    printf("ps: %s\n", eb?(char*)eb->GetBufferPointer():"?"); return 1; }
  ID3D11VertexShader* vs=nullptr; ID3D11PixelShader* ps=nullptr;
  dev->CreateVertexShader(vsb->GetBufferPointer(), vsb->GetBufferSize(),0,&vs);
  dev->CreatePixelShader(psb->GetBufferPointer(), psb->GetBufferSize(),0,&ps);
  D3D11_BUFFER_DESC bd={}; bd.ByteWidth=16; bd.Usage=D3D11_USAGE_DEFAULT;
  bd.BindFlags=D3D11_BIND_CONSTANT_BUFFER; ID3D11Buffer* cb=nullptr;
  dev->CreateBuffer(&bd,nullptr,&cb);
  D3D11_RASTERIZER_DESC rd={}; rd.FillMode=D3D11_FILL_SOLID; rd.CullMode=D3D11_CULL_NONE;
  ID3D11RasterizerState* rs=nullptr; dev->CreateRasterizerState(&rd,&rs);
  ctx->RSSetState(rs);

  ID3D11Query *qd=nullptr,*q0=nullptr,*q1=nullptr;
  D3D11_QUERY_DESC dqd={D3D11_QUERY_TIMESTAMP_DISJOINT,0}; dev->CreateQuery(&dqd,&qd);
  D3D11_QUERY_DESC tqd={D3D11_QUERY_TIMESTAMP,0};
  dev->CreateQuery(&tqd,&q0); dev->CreateQuery(&tqd,&q1);

  auto render_two_eyes = [&](ID3D11RenderTargetView* rL, ID3D11RenderTargetView* rR,
                             int w, int h, int oct, float t) {
    float p[4] = {(float)oct, (float)w, (float)h, t};
    ctx->UpdateSubresource(cb,0,nullptr,p,0,0);
    ctx->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
    ctx->IASetInputLayout(nullptr);
    ctx->VSSetShader(vs,nullptr,0); ctx->VSSetConstantBuffers(0,1,&cb);
    ctx->PSSetShader(ps,nullptr,0); ctx->PSSetConstantBuffers(0,1,&cb);
    ctx->RSSetState(rs);
    D3D11_VIEWPORT vp={0,0,(float)w,(float)h,0,1}; ctx->RSSetViewports(1,&vp);
    ID3D11RenderTargetView* rts[2]={rL,rR};
    for (int e=0;e<2;++e){ ctx->OMSetRenderTargets(1,&rts[e],nullptr); ctx->Draw(3,0); }
  };

  auto gpu_ms = [&](auto&& fn)->double{
    // warm
    for(int i=0;i<3;++i) fn(); ctx->Flush();
    double best=1e30;
    for(int rep=0;rep<12;++rep){
      ctx->Begin(qd); ctx->End(q0); fn(); ctx->End(q1); ctx->End(qd);
      ctx->Flush();
      D3D11_QUERY_DATA_TIMESTAMP_DISJOINT dj; UINT64 t0,t1;
      while(ctx->GetData(qd,&dj,sizeof(dj),0)!=S_OK){}
      while(ctx->GetData(q0,&t0,sizeof(t0),0)!=S_OK){}
      while(ctx->GetData(q1,&t1,sizeof(t1),0)!=S_OK){}
      if(dj.Disjoint) continue;
      double ms = 1000.0*(double)(t1-t0)/(double)dj.Frequency;
      if(ms<best) best=ms;
    }
    return best;
  };

  auto kernel_ms = [&](ID3D11Texture2D* srcL, ID3D11Texture2D* srcR, int iw,int ih,
                       ID3D11Texture2D* dstL, ID3D11Texture2D* dstR, int ow,int oh)->double{
    float r1=0.65f*oh, r0=0.55f*r1;
    auto once=[&]{ svr_process(srcL,0,0,0,iw,ih,dstL,0,0,0,ow,oh,0.46f*ow,0.5f*oh,r0,r1,0.4f);
                  svr_process(srcR,0,0,0,iw,ih,dstR,0,0,0,ow,oh,0.54f*ow,0.5f*oh,r0,r1,0.4f); };
    for(int i=0;i<5;++i) once();               // warm + register
    LARGE_INTEGER f,a,b; QueryPerformanceFrequency(&f);
    QueryPerformanceCounter(&a);
    const int N=60; for(int i=0;i<N;++i) once();
    QueryPerformanceCounter(&b);
    return 1000.0*(double)(b.QuadPart-a.QuadPart)/(double)f.QuadPart/N;
  };

  struct Res { int w,h; const char* name; };
  Res shapes[] = { {1440,1600,"Quest3 low"}, {2016,2240,"Quest3 med"},
                   {2464,2736,"Quest3 high"} };
  float scales[] = { 0.90f, 0.77f, 0.67f, 0.59f };
  int octaves[] = { 4, 8, 16, 24 };

  printf("SVR resolution-scaling benefit (GPU timestamps, per frame = 2 eyes)\n");
  printf("shading load = fBm octave count; a real title sits where its per-pixel cost lands\n\n");

  for (auto& sh : shapes) {
    ID3D11Texture2D* fullL=mkrt(dev,sh.w,sh.h); ID3D11Texture2D* fullR=mkrt(dev,sh.w,sh.h);
    ID3D11RenderTargetView *fLrtv=nullptr,*fRrtv=nullptr;
    dev->CreateRenderTargetView(fullL,nullptr,&fLrtv);
    dev->CreateRenderTargetView(fullR,nullptr,&fRrtv);
    printf("== %s  %dx%d/eye ==\n", sh.name, sh.w, sh.h);
    printf("%7s %8s | %9s %9s %8s %8s | %8s %8s\n",
           "scale","pix%","full ms","scaled ms","kern ms","frame ms","speedup","ceiling");
    for (float sc : scales) {
      int rw=(int)(sh.w*sc+0.5f), rh=(int)(sh.h*sc+0.5f);
      ID3D11Texture2D* redL=mkrt(dev,rw,rh); ID3D11Texture2D* redR=mkrt(dev,rw,rh);
      ID3D11RenderTargetView *rLrtv=nullptr,*rRrtv=nullptr;
      dev->CreateRenderTargetView(redL,nullptr,&rLrtv);
      dev->CreateRenderTargetView(redR,nullptr,&rRrtv);
      double pixr = (double)(rw*rh)/(double)(sh.w*sh.h);
      for (int oct : octaves) {
        double tf = gpu_ms([&]{ render_two_eyes(fLrtv,fRrtv,sh.w,sh.h,oct,1.3f); });
        double ts = gpu_ms([&]{ render_two_eyes(rLrtv,rRrtv,rw,rh,oct,1.3f); });
        double tk = kernel_ms(redL,redR,rw,rh,fullL,fullR,sh.w,sh.h);
        double frame = ts + tk;
        printf("%6.2f  %6.1f%% | %9.3f %9.3f %8.3f %8.3f | %7.2fx %7.2fx  (oct %d)\n",
               sc, 100.0*pixr, tf, ts, tk, frame, tf/frame, 1.0/pixr, oct);
      }
      rLrtv->Release(); rRrtv->Release(); redL->Release(); redR->Release();
    }
    fLrtv->Release(); fRrtv->Release(); fullL->Release(); fullR->Release();
    printf("\n");
  }
  printf("done\n");
  return 0;
}