mirror/Sunshine
1
0
Fork 0
mirror of https://github.com/LizardByte/Sunshine.git synced 2025-01-30 03:32:43 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fix color conversion for SDR FP16 capture formats

Browse Source
This commit is contained in:
Cameron Gutman 2023-04-12 23:43:29 -05:00
parent 010440dbd3
commit b5c8bbaa24
3 changed files with 112 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

57
src/platform/windows/display_vram.cpp
View File

@ -91,9 +91,11 @@ namespace platf::dxgi {
blob_t convert_UV_vs_hlsl;
blob_t convert_UV_ps_hlsl;
blob_t convert_UV_linear_ps_hlsl;
blob_t convert_UV_PQ_ps_hlsl;
blob_t scene_vs_hlsl;
blob_t convert_Y_ps_hlsl;
blob_t convert_Y_linear_ps_hlsl;
blob_t convert_Y_PQ_ps_hlsl;
blob_t scene_ps_hlsl;
blob_t scene_NW_ps_hlsl;
@ -116,6 +118,9 @@ namespace platf::dxgi {
// Unique identifier for this image
uint32_t id = 0;
// DXGI format of this image texture
DXGI_FORMAT format;
virtual ~img_d3d_t() override {
if (encoder_texture_handle) {
CloseHandle(encoder_texture_handle);
@ -338,14 +343,14 @@ namespace platf::dxgi {
device_ctx->OMSetRenderTargets(1, &nv12_Y_rt, nullptr);
device_ctx->VSSetShader(scene_vs.get(), nullptr, 0);
device_ctx->PSSetShader(convert_Y_ps.get(), nullptr, 0);
device_ctx->PSSetShader(img.format == DXGI_FORMAT_R16G16B16A16_FLOAT ? convert_Y_fp16_ps.get() : convert_Y_ps.get(), nullptr, 0);
device_ctx->RSSetViewports(1, &outY_view);
device_ctx->PSSetShaderResources(0, 1, &img_ctx.encoder_input_res);
device_ctx->Draw(3, 0);
device_ctx->OMSetRenderTargets(1, &nv12_UV_rt, nullptr);
device_ctx->VSSetShader(convert_UV_vs.get(), nullptr, 0);
device_ctx->PSSetShader(convert_UV_ps.get(), nullptr, 0);
device_ctx->PSSetShader(img.format == DXGI_FORMAT_R16G16B16A16_FLOAT ? convert_UV_fp16_ps.get() : convert_UV_ps.get(), nullptr, 0);
device_ctx->RSSetViewports(1, &outUV_view);
device_ctx->Draw(3, 0);
@ -577,34 +582,48 @@ namespace platf::dxgi {
}
// If the display is in HDR and we're streaming HDR, we'll be converting scRGB to SMPTE 2084 PQ.
// NB: We can consume scRGB in SDR with our regular shaders because it behaves like UNORM input.
if (format == DXGI_FORMAT_P010 && display->is_hdr()) {
status = device->CreatePixelShader(convert_Y_PQ_ps_hlsl->GetBufferPointer(), convert_Y_PQ_ps_hlsl->GetBufferSize(), nullptr, &convert_Y_ps);
status = device->CreatePixelShader(convert_Y_PQ_ps_hlsl->GetBufferPointer(), convert_Y_PQ_ps_hlsl->GetBufferSize(), nullptr, &convert_Y_fp16_ps);
if (status) {
BOOST_LOG(error) << "Failed to create convertY pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
return -1;
}
status = device->CreatePixelShader(convert_UV_PQ_ps_hlsl->GetBufferPointer(), convert_UV_PQ_ps_hlsl->GetBufferSize(), nullptr, &convert_UV_ps);
status = device->CreatePixelShader(convert_UV_PQ_ps_hlsl->GetBufferPointer(), convert_UV_PQ_ps_hlsl->GetBufferSize(), nullptr, &convert_UV_fp16_ps);
if (status) {
BOOST_LOG(error) << "Failed to create convertUV pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
return -1;
}
}
else {
status = device->CreatePixelShader(convert_Y_ps_hlsl->GetBufferPointer(), convert_Y_ps_hlsl->GetBufferSize(), nullptr, &convert_Y_ps);
// If the display is in Advanced Color mode, the desktop format will be scRGB FP16.
// scRGB uses linear gamma, so we must use our linear to sRGB conversion shaders.
status = device->CreatePixelShader(convert_Y_linear_ps_hlsl->GetBufferPointer(), convert_Y_linear_ps_hlsl->GetBufferSize(), nullptr, &convert_Y_fp16_ps);
if (status) {
BOOST_LOG(error) << "Failed to create convertY pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
return -1;
}
status = device->CreatePixelShader(convert_UV_ps_hlsl->GetBufferPointer(), convert_UV_ps_hlsl->GetBufferSize(), nullptr, &convert_UV_ps);
status = device->CreatePixelShader(convert_UV_linear_ps_hlsl->GetBufferPointer(), convert_UV_linear_ps_hlsl->GetBufferSize(), nullptr, &convert_UV_fp16_ps);
if (status) {
BOOST_LOG(error) << "Failed to create convertUV pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
return -1;
}
}
// These shaders consume standard 8-bit sRGB input
status = device->CreatePixelShader(convert_Y_ps_hlsl->GetBufferPointer(), convert_Y_ps_hlsl->GetBufferSize(), nullptr, &convert_Y_ps);
if (status) {
BOOST_LOG(error) << "Failed to create convertY pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
return -1;
}
status = device->CreatePixelShader(convert_UV_ps_hlsl->GetBufferPointer(), convert_UV_ps_hlsl->GetBufferSize(), nullptr, &convert_UV_ps);
if (status) {
BOOST_LOG(error) << "Failed to create convertUV pixel shader [0x"sv << util::hex(status).to_string_view() << ']';
return -1;
}
color_matrix = make_buffer(device.get(), ::video::colors[0]);
if (!color_matrix) {
BOOST_LOG(error) << "Failed to create color matrix buffer"sv;
@ -750,7 +769,9 @@ namespace platf::dxgi {
vs_t convert_UV_vs;
ps_t convert_UV_ps;
ps_t convert_UV_fp16_ps;
ps_t convert_Y_ps;
ps_t convert_Y_fp16_ps;
vs_t scene_vs;
D3D11_VIEWPORT outY_view;
@ -1118,6 +1139,7 @@ namespace platf::dxgi {
img->pixel_pitch = get_pixel_pitch();
img->row_pitch = img->pixel_pitch * img->width;
img->dummy = dummy;
img->format = (capture_format == DXGI_FORMAT_UNKNOWN) ? DXGI_FORMAT_B8G8R8A8_UNORM : capture_format;
D3D11_TEXTURE2D_DESC t {};
t.Width = img->width;
@ -1126,7 +1148,7 @@ namespace platf::dxgi {
t.ArraySize = 1;
t.SampleDesc.Count = 1;
t.Usage = D3D11_USAGE_DEFAULT;
t.Format = (capture_format == DXGI_FORMAT_UNKNOWN) ? DXGI_FORMAT_B8G8R8A8_UNORM : capture_format;
t.Format = img->format;
t.BindFlags = D3D11_BIND_SHADER_RESOURCE | D3D11_BIND_RENDER_TARGET;
t.MiscFlags = D3D11_RESOURCE_MISC_SHARED_NTHANDLE | D3D11_RESOURCE_MISC_SHARED_KEYEDMUTEX;
@ -1190,8 +1212,9 @@ namespace platf::dxgi {
std::vector<DXGI_FORMAT>
display_vram_t::get_supported_capture_formats() {
return {
// scRGB FP16 is the desired format for HDR content. This will also handle
// 10-bit SDR displays with the increased precision of FP16 vs 8-bit UNORMs.
// scRGB FP16 is the ideal format for Wide Color Gamut and Advanced Color
// displays (both SDR and HDR). This format uses linear gamma, so we will
// use a linear->PQ shader for HDR and a linear->sRGB shader for SDR.
DXGI_FORMAT_R16G16B16A16_FLOAT,
// DXGI_FORMAT_R10G10B10A2_UNORM seems like it might give us frames already
@ -1203,8 +1226,8 @@ namespace platf::dxgi {
// but we avoid it for now.
// We include the 8-bit modes too for when the display is in SDR mode,
// while the client stream is HDR-capable. These UNORM formats behave
// like a degenerate case of scRGB FP16 with values between 0.0f-1.0f.
// while the client stream is HDR-capable. These UNORM formats can
// use our normal pixel shaders that expect sRGB input.
DXGI_FORMAT_B8G8R8A8_UNORM,
DXGI_FORMAT_R8G8B8A8_UNORM,
};
@ -1250,6 +1273,11 @@ namespace platf::dxgi {
return -1;
}
convert_Y_linear_ps_hlsl = compile_pixel_shader(SUNSHINE_SHADERS_DIR "/ConvertYPS_Linear.hlsl");
if (!convert_Y_linear_ps_hlsl) {
return -1;
}
convert_UV_ps_hlsl = compile_pixel_shader(SUNSHINE_SHADERS_DIR "/ConvertUVPS.hlsl");
if (!convert_UV_ps_hlsl) {
return -1;
@ -1260,6 +1288,11 @@ namespace platf::dxgi {
return -1;
}
convert_UV_linear_ps_hlsl = compile_pixel_shader(SUNSHINE_SHADERS_DIR "/ConvertUVPS_Linear.hlsl");
if (!convert_UV_linear_ps_hlsl) {
return -1;
}
convert_UV_vs_hlsl = compile_vertex_shader(SUNSHINE_SHADERS_DIR "/ConvertUVVS.hlsl");
if (!convert_UV_vs_hlsl) {
return -1;

36
src_assets/windows/assets/shaders/directx/ConvertUVPS_Linear.hlsl Normal file
View File

@ -0,0 +1,36 @@
Texture2D image : register(t0);
SamplerState def_sampler : register(s0);
struct FragTexWide {
float3 uuv : TEXCOORD0;
};
cbuffer ColorMatrix : register(b0) {
float4 color_vec_y;
float4 color_vec_u;
float4 color_vec_v;
float2 range_y;
float2 range_uv;
};
// This is a fast sRGB approximation from Microsoft's ColorSpaceUtility.hlsli
float3 ApplySRGBCurve(float3 x)
{
return x < 0.0031308 ? 12.92 * x : 1.13005 * sqrt(x - 0.00228) - 0.13448 * x + 0.005719;
}
float2 main_ps(FragTexWide input) : SV_Target
{
float3 rgb_left = ApplySRGBCurve(saturate(image.Sample(def_sampler, input.uuv.xz)).rgb);
float3 rgb_right = ApplySRGBCurve(saturate(image.Sample(def_sampler, input.uuv.yz)).rgb);
float3 rgb = (rgb_left + rgb_right) * 0.5;
float u = dot(color_vec_u.xyz, rgb) + color_vec_u.w;
float v = dot(color_vec_v.xyz, rgb) + color_vec_v.w;
u = u * range_uv.x + range_uv.y;
v = v * range_uv.x + range_uv.y;
return float2(u, v);
}

31
src_assets/windows/assets/shaders/directx/ConvertYPS_Linear.hlsl Normal file
View File

@ -0,0 +1,31 @@
Texture2D image : register(t0);
SamplerState def_sampler : register(s0);
cbuffer ColorMatrix : register(b0) {
float4 color_vec_y;
float4 color_vec_u;
float4 color_vec_v;
float2 range_y;
float2 range_uv;
};
struct PS_INPUT
{
float4 pos : SV_POSITION;
float2 tex : TEXCOORD;
};
// This is a fast sRGB approximation from Microsoft's ColorSpaceUtility.hlsli
float3 ApplySRGBCurve(float3 x)
{
return x < 0.0031308 ? 12.92 * x : 1.13005 * sqrt(x - 0.00228) - 0.13448 * x + 0.005719;
}
float main_ps(PS_INPUT frag_in) : SV_Target
{
float3 rgb = ApplySRGBCurve(saturate(image.Sample(def_sampler, frag_in.tex, 0)).rgb);
float y = dot(color_vec_y.xyz, rgb) + color_vec_y.w;
return y * range_y.x + range_y.y;
}