mrq/FidelityFX-FSR2
1
0
Fork 1
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Force fp32 during accumulation pass on Intel

Browse Source 
fp16 accumulation is creating dark artifacts on Intel GPUs, possibly related to a clamping issue
This commit is contained in:
Daniele Pieroni 2022-10-03 15:30:56 -04:00
parent f8b3d1b2c3
commit 6f185cfd8f
1 changed files with 45 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
src/ffx-fsr2-api/dx12/ffx_fsr2_dx12.cpp
View File

@ -22,6 +22,7 @@
#include <codecvt> // convert string to wstring
#include <DirectXMath.h>
#include <d3d12.h>
#include <dxgi1_3.h>
#include <d3dcompiler.h>
#include <d3d12shader.h>
#include "d3dx12.h"
@ -621,7 +622,6 @@ FfxErrorCode CreateBackendContextDX12(FfxFsr2Interface* backendInterface, FfxDev
dx12Device->AddRef();
backendContext->device = dx12Device;
}
// init resource linked list
backendContext->nextStaticResource = 1;
backendContext->nextDynamicResource = FSR2_MAX_RESOURCE_COUNT - 1;
@ -928,6 +928,43 @@ FfxResourceDescription GetResourceDescriptorDX12(
return resourceDescription;
}
static bool isLuidsSame(LUID luid1, LUID luid2)
{
return memcmp(&luid1, &luid2, sizeof(LUID)) == 0;
}
static bool IsIntelAdapter(ID3D12Device* device)
{
bool isIntel = false;
IDXGIFactory* pFactory = nullptr;
if (SUCCEEDED(CreateDXGIFactory2(0, IID_PPV_ARGS(&pFactory))))
{
IDXGIAdapter* pAdapter = nullptr;
UINT i = 0;
while (pFactory->EnumAdapters(i++, &pAdapter) != DXGI_ERROR_NOT_FOUND)
{
DXGI_ADAPTER_DESC desc{};
if (SUCCEEDED(pAdapter->GetDesc(&desc)))
{
if (isLuidsSame(desc.AdapterLuid, device->GetAdapterLuid()))
{
if (desc.VendorId == 0x8086)
isIntel = true;
}
pAdapter->Release();
}
}
pFactory->Release();
}
return isIntel;
}
FfxErrorCode CreatePipelineDX12(
FfxFsr2Interface* backendInterface,
FfxFsr2Pass pass,
@ -981,6 +1018,13 @@ FfxErrorCode CreatePipelineDX12(
supportedFP16 = !!(d3d12Options.MinPrecisionSupport & D3D12_SHADER_MIN_PRECISION_SUPPORT_16_BIT);
}
if (pass == FFX_FSR2_PASS_ACCUMULATE || pass == FFX_FSR2_PASS_ACCUMULATE_SHARPEN)
{
// Workaround: Disable FP16 path for the accumulate pass on Intel due to a clamping error.
if (IsIntelAdapter(dx12Device))
supportedFP16 = false;
}
// work out what permutation to load.
uint32_t flags = 0;
flags |= (pipelineDescription->contextFlags & FFX_FSR2_ENABLE_HIGH_DYNAMIC_RANGE) ? FSR2_SHADER_PERMUTATION_HDR_COLOR_INPUT : 0;