From 4b5071f062db47b96eca915cedf2cc075f2ff229 Mon Sep 17 00:00:00 2001
From: kiseido <kiseido@users.noreply.github.com>
Date: Wed, 19 Oct 2022 16:21:29 -0700
Subject: [PATCH] Copying changes from PotatoOfDoom's repo
---
CMakeLists.txt | 2 +
src/ffx-fsr2-api/dx12/ffx_fsr2_dx12.cpp | 4 --
src/ffx-fsr2-api/ffx_fsr2.cpp | 9 +--
src/ffx-fsr2-api/ffx_fsr2.h | 2 +-
src/ffx-fsr2-api/ffx_types.h | 1 -
.../shaders/ffx_fsr2_accumulate_pass.glsl | 1 +
.../shaders/ffx_fsr2_callbacks_glsl.h | 11 ++-
.../shaders/ffx_fsr2_callbacks_hlsl.h | 68 ++++++++++---------
.../shaders/ffx_fsr2_depth_clip.h | 4 ++
src/ffx-fsr2-api/shaders/ffx_fsr2_sample.h | 40 +++++------
src/ffx-fsr2-api/shaders/ffx_fsr2_upsample.h | 6 +-
src/ffx-fsr2-api/vk/ffx_fsr2_vk.cpp | 6 +-
.../vk/shaders/ffx_fsr2_shaders_vk.cpp | 2 +
13 files changed, 81 insertions(+), 75 deletions(-)
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9a5424c..2911401 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -79,6 +79,8 @@ option (FFX_FSR2_API_VK "Build FSR 2.0 Vulkan backend" ${GFX_API_VK})
# reference libs used by both backends
add_subdirectory(libs/cauldron)
add_subdirectory(src/Common)
+
+set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)
add_subdirectory(src/ffx-fsr2-api)
if(GFX_API_VK)
diff --git a/src/ffx-fsr2-api/dx12/ffx_fsr2_dx12.cpp b/src/ffx-fsr2-api/dx12/ffx_fsr2_dx12.cpp
index 8973648..4b0f507 100644
--- a/src/ffx-fsr2-api/dx12/ffx_fsr2_dx12.cpp
+++ b/src/ffx-fsr2-api/dx12/ffx_fsr2_dx12.cpp
@@ -245,8 +245,6 @@ DXGI_FORMAT ffxGetDX12FormatFromSurfaceFormat(FfxSurfaceFormat surfaceFormat)
return DXGI_FORMAT_R32G32B32A32_FLOAT;
case(FFX_SURFACE_FORMAT_R16G16B16A16_FLOAT):
return DXGI_FORMAT_R16G16B16A16_FLOAT;
- case(FFX_SURFACE_FORMAT_R16G16B16A16_UNORM):
- return DXGI_FORMAT_R16G16B16A16_UNORM;
case(FFX_SURFACE_FORMAT_R32G32_FLOAT):
return DXGI_FORMAT_R32G32_FLOAT;
case(FFX_SURFACE_FORMAT_R32_UINT):
@@ -298,8 +296,6 @@ FfxSurfaceFormat ffxGetSurfaceFormatDX12(DXGI_FORMAT format)
return FFX_SURFACE_FORMAT_R32G32B32A32_FLOAT;
case(DXGI_FORMAT_R16G16B16A16_FLOAT):
return FFX_SURFACE_FORMAT_R16G16B16A16_FLOAT;
- case(DXGI_FORMAT_R16G16B16A16_UNORM):
- return FFX_SURFACE_FORMAT_R16G16B16A16_UNORM;
case(DXGI_FORMAT_R32G32_FLOAT):
return FFX_SURFACE_FORMAT_R32G32_FLOAT;
case(DXGI_FORMAT_R32_UINT):
diff --git a/src/ffx-fsr2-api/ffx_fsr2.cpp b/src/ffx-fsr2-api/ffx_fsr2.cpp
index c7a342f..260a688 100644
--- a/src/ffx-fsr2-api/ffx_fsr2.cpp
+++ b/src/ffx-fsr2-api/ffx_fsr2.cpp
@@ -29,6 +29,7 @@
#include "shaders/ffx_fsr1.h"
#include "shaders/ffx_spd.h"
#include "shaders/ffx_fsr2_callbacks_hlsl.h"
+#include <cmath>
#include "ffx_fsr2_maximum_bias.h"
@@ -161,7 +162,7 @@ FfxConstantBuffer globalFsr2ConstantBuffers[3] = {
// Lanczos
static float lanczos2(float value)
{
- return abs(value) < FFX_EPSILON ? 1.f : (sinf(FFX_PI * value) / (FFX_PI * value)) * (sinf(0.5f * FFX_PI * value) / (0.5f * FFX_PI * value));
+ return std::abs(value) < FFX_EPSILON ? 1.f : (sinf(FFX_PI * value) / (FFX_PI * value)) * (sinf(0.5f * FFX_PI * value) / (0.5f * FFX_PI * value));
}
// Calculate halton number for index and base.
@@ -331,7 +332,7 @@ static FfxErrorCode fsr2Create(FfxFsr2Context_Private* context, const FfxFsr2Con
const Fsr2ResourceDescription internalSurfaceDesc[] = {
{ FFX_FSR2_RESOURCE_IDENTIFIER_PREPARED_INPUT_COLOR, L"FSR2_PreparedInputColor", FFX_RESOURCE_USAGE_UAV,
- FFX_SURFACE_FORMAT_R16G16B16A16_UNORM, contextDescription->maxRenderSize.width, contextDescription->maxRenderSize.height, 1, FFX_RESOURCE_FLAGS_ALIASABLE },
+ FFX_SURFACE_FORMAT_R16G16B16A16_FLOAT, contextDescription->maxRenderSize.width, contextDescription->maxRenderSize.height, 1, FFX_RESOURCE_FLAGS_ALIASABLE },
{ FFX_FSR2_RESOURCE_IDENTIFIER_RECONSTRUCTED_PREVIOUS_NEAREST_DEPTH, L"FSR2_ReconstructedPrevNearestDepth", FFX_RESOURCE_USAGE_UAV,
FFX_SURFACE_FORMAT_R32_UINT, contextDescription->maxRenderSize.width, contextDescription->maxRenderSize.height, 1, FFX_RESOURCE_FLAGS_ALIASABLE },
@@ -662,8 +663,8 @@ static FfxErrorCode fsr2Dispatch(FfxFsr2Context_Private* context, const FfxFsr2D
// compute jitter cancellation
if (context->contextDescription.flags & FFX_FSR2_ENABLE_MOTION_VECTORS_JITTER_CANCELLATION) {
- context->constants.motionVectorJitterCancellation[0] = (context->previousJitterOffset[0] - context->constants.jitterOffset[0]) / motionVectorsTargetSize[0];
- context->constants.motionVectorJitterCancellation[1] = (context->previousJitterOffset[1] - context->constants.jitterOffset[1]) / motionVectorsTargetSize[1];
+ context->constants.motionVectorJitterCancellation[0] = (context->previousJitterOffset[0] - context->constants.jitterOffset[0]) / motionVectorsTargetSize[0] * context->constants.motionVectorScale[0];
+ context->constants.motionVectorJitterCancellation[1] = (context->previousJitterOffset[1] - context->constants.jitterOffset[1]) / motionVectorsTargetSize[1] * context->constants.motionVectorScale[1];
context->previousJitterOffset[0] = context->constants.jitterOffset[0];
context->previousJitterOffset[1] = context->constants.jitterOffset[1];
diff --git a/src/ffx-fsr2-api/ffx_fsr2.h b/src/ffx-fsr2-api/ffx_fsr2.h
index ee2ff7d..c8b0fa5 100644
--- a/src/ffx-fsr2-api/ffx_fsr2.h
+++ b/src/ffx-fsr2-api/ffx_fsr2.h
@@ -40,7 +40,7 @@
/// FidelityFX Super Resolution 2 patch version.
///
/// @ingroup FSR2
-#define FFX_FSR2_VERSION_PATCH (2)
+#define FFX_FSR2_VERSION_PATCH (1)
/// The size of the context specified in 32bit values.
///
diff --git a/src/ffx-fsr2-api/ffx_types.h b/src/ffx-fsr2-api/ffx_types.h
index 0079572..75fb0e8 100644
--- a/src/ffx-fsr2-api/ffx_types.h
+++ b/src/ffx-fsr2-api/ffx_types.h
@@ -57,7 +57,6 @@ typedef enum FfxSurfaceFormat {
FFX_SURFACE_FORMAT_R32G32B32A32_TYPELESS, ///< 32 bit per channel, 4 channel typeless format
FFX_SURFACE_FORMAT_R32G32B32A32_FLOAT, ///< 32 bit per channel, 4 channel float format
FFX_SURFACE_FORMAT_R16G16B16A16_FLOAT, ///< 16 bit per channel, 4 channel float format
- FFX_SURFACE_FORMAT_R16G16B16A16_UNORM, ///< 16 bit per channel, 4 channel unsigned normalized format
FFX_SURFACE_FORMAT_R32G32_FLOAT, ///< 32 bit per channel, 2 channel float format
FFX_SURFACE_FORMAT_R32_UINT, ///< 32 bit per channel, 1 channel float format
FFX_SURFACE_FORMAT_R8G8B8A8_TYPELESS, ///< 8 bit per channel, 4 channel float format
diff --git a/src/ffx-fsr2-api/shaders/ffx_fsr2_accumulate_pass.glsl b/src/ffx-fsr2-api/shaders/ffx_fsr2_accumulate_pass.glsl
index e1ee116..4f0d694 100644
--- a/src/ffx-fsr2-api/shaders/ffx_fsr2_accumulate_pass.glsl
+++ b/src/ffx-fsr2-api/shaders/ffx_fsr2_accumulate_pass.glsl
@@ -42,6 +42,7 @@
#extension GL_GOOGLE_include_directive : require
#extension GL_EXT_samplerless_texture_functions : require
+#extension GL_EXT_shader_image_load_formatted : require
#define FSR2_BIND_SRV_EXPOSURE 0
#define FSR2_BIND_SRV_DILATED_REACTIVE_MASKS 1
diff --git a/src/ffx-fsr2-api/shaders/ffx_fsr2_callbacks_glsl.h b/src/ffx-fsr2-api/shaders/ffx_fsr2_callbacks_glsl.h
index d598250..0d5519b 100644
--- a/src/ffx-fsr2-api/shaders/ffx_fsr2_callbacks_glsl.h
+++ b/src/ffx-fsr2-api/shaders/ffx_fsr2_callbacks_glsl.h
@@ -221,7 +221,7 @@ layout (set = 0, binding = 1) uniform sampler s_LinearClamp;
layout (set = 1, binding = FSR2_BIND_UAV_DEPTH_CLIP, r32f) uniform image2D rw_depth_clip;
#endif
#if defined FSR2_BIND_UAV_PREPARED_INPUT_COLOR
- layout (set = 1, binding = FSR2_BIND_UAV_PREPARED_INPUT_COLOR, rgba16) uniform image2D rw_prepared_input_color;
+ layout (set = 1, binding = FSR2_BIND_UAV_PREPARED_INPUT_COLOR, rgba32f) uniform image2D rw_prepared_input_color;
#endif
#if defined FSR2_BIND_UAV_LUMA_HISTORY
layout (set = 1, binding = FSR2_BIND_UAV_LUMA_HISTORY, rgba32f) uniform image2D rw_luma_history;
@@ -591,6 +591,9 @@ void SetReconstructedDepth(FfxInt32x2 iPxSample, FfxUInt32 uValue)
void StoreDilatedDepth(FFX_PARAMETER_IN FfxInt32x2 iPxPos, FFX_PARAMETER_IN FfxFloat32 fDepth)
{
#if defined(FSR2_BIND_UAV_DILATED_DEPTH)
+#if !FFX_FSR2_OPTION_INVERTED_DEPTH
+ fDepth = 1.0 - fDepth; // Preserve precision as well as we can in FP16.
+#endif
//FfxUInt32 uDepth = f32tof16(fDepth);
imageStore(rw_dilatedDepth, iPxPos, vec4(fDepth, 0.0f, 0.0f, 0.0f));
#endif
@@ -625,7 +628,11 @@ FfxFloat32x2 SampleDilatedMotionVector(FfxFloat32x2 fUV)
FfxFloat32 LoadDilatedDepth(FfxInt32x2 iPxInput)
{
#if defined(FSR2_BIND_SRV_DILATED_DEPTH)
- return texelFetch(r_dilatedDepth, iPxInput, 0).r;
+ FfxFloat32 fDepth = texelFetch(r_dilatedDepth, iPxInput, 0).r;
+#if !FFX_FSR2_OPTION_INVERTED_DEPTH
+ fDepth = 1.0 - fDepth; // Reconstruct from FP16.
+#endif
+ return fDepth;
#else
return 0.f;
#endif
diff --git a/src/ffx-fsr2-api/shaders/ffx_fsr2_callbacks_hlsl.h b/src/ffx-fsr2-api/shaders/ffx_fsr2_callbacks_hlsl.h
index 4641927..f6412f7 100644
--- a/src/ffx-fsr2-api/shaders/ffx_fsr2_callbacks_hlsl.h
+++ b/src/ffx-fsr2-api/shaders/ffx_fsr2_callbacks_hlsl.h
@@ -234,26 +234,26 @@ SamplerState s_LinearClamp : register(s1);
// SRVs
#if defined(FFX_INTERNAL)
- Texture2D<FfxFloat32x4> r_input_color_jittered : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_COLOR);
- Texture2D<FfxFloat32x4> r_motion_vectors : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_MOTION_VECTORS);
- Texture2D<FfxFloat32> r_depth : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_DEPTH);
- Texture2D<FfxFloat32x2> r_exposure : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_EXPOSURE);
- Texture2D<FfxFloat32> r_reactive_mask : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_REACTIVE_MASK);
- Texture2D<FfxFloat32> r_transparency_and_composition_mask : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_TRANSPARENCY_AND_COMPOSITION_MASK);
- Texture2D<FfxUInt32> r_reconstructed_previous_nearest_depth : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_RECONSTRUCTED_PREVIOUS_NEAREST_DEPTH);
- Texture2D<FfxFloat32x2> r_dilated_motion_vectors : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_MOTION_VECTORS);
- Texture2D<FfxFloat32> r_dilatedDepth : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_DEPTH);
- Texture2D<FfxFloat32x4> r_internal_upscaled_color : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INTERNAL_UPSCALED_COLOR);
- Texture2D<FfxFloat32x3> r_lock_status : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_LOCK_STATUS);
- Texture2D<FfxFloat32> r_depth_clip : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DEPTH_CLIP);
- Texture2D<unorm FfxFloat32x4> r_prepared_input_color : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_PREPARED_INPUT_COLOR);
- Texture2D<unorm FfxFloat32x4> r_luma_history : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_LUMA_HISTORY);
- Texture2D<FfxFloat32x4> r_rcas_input : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_RCAS_INPUT);
- Texture2D<FfxFloat32> r_lanczos_lut : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_LANCZOS_LUT);
- Texture2D<FfxFloat32> r_imgMips : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_AUTO_EXPOSURE);
- Texture2D<FfxFloat32> r_upsample_maximum_bias_lut : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTITIER_UPSAMPLE_MAXIMUM_BIAS_LUT);
- Texture2D<FfxFloat32x2> r_dilated_reactive_masks : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_REACTIVE_MASKS);
- Texture2D<FfxFloat32x4> r_debug_out : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DEBUG_OUTPUT);
+ Texture2D<FfxFloat32x4> r_input_color_jittered : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_COLOR);
+ Texture2D<FfxFloat32x4> r_motion_vectors : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_MOTION_VECTORS);
+ Texture2D<FfxFloat32> r_depth : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_DEPTH);
+ Texture2D<FfxFloat32x2> r_exposure : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_EXPOSURE);
+ Texture2D<FfxFloat32> r_reactive_mask : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_REACTIVE_MASK);
+ Texture2D<FfxFloat32> r_transparency_and_composition_mask : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INPUT_TRANSPARENCY_AND_COMPOSITION_MASK);
+ Texture2D<FfxUInt32> r_reconstructed_previous_nearest_depth : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_RECONSTRUCTED_PREVIOUS_NEAREST_DEPTH);
+ Texture2D<FfxFloat32x2> r_dilated_motion_vectors : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_MOTION_VECTORS);
+ Texture2D<FfxFloat32> r_dilatedDepth : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_DEPTH);
+ Texture2D<FfxFloat32x4> r_internal_upscaled_color : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_INTERNAL_UPSCALED_COLOR);
+ Texture2D<FfxFloat32x3> r_lock_status : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_LOCK_STATUS);
+ Texture2D<FfxFloat32> r_depth_clip : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DEPTH_CLIP);
+ Texture2D<FfxFloat32x4> r_prepared_input_color : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_PREPARED_INPUT_COLOR);
+ Texture2D<unorm FfxFloat32x4> r_luma_history : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_LUMA_HISTORY);
+ Texture2D<FfxFloat32x4> r_rcas_input : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_RCAS_INPUT);
+ Texture2D<FfxFloat32> r_lanczos_lut : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_LANCZOS_LUT);
+ Texture2D<FfxFloat32> r_imgMips : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_AUTO_EXPOSURE);
+ Texture2D<FfxFloat32> r_upsample_maximum_bias_lut : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTITIER_UPSAMPLE_MAXIMUM_BIAS_LUT);
+ Texture2D<FfxFloat32x2> r_dilated_reactive_masks : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_REACTIVE_MASKS);
+ Texture2D<FfxFloat32x4> r_debug_out : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_DEBUG_OUTPUT);
// declarations not current form, no accessor functions
Texture2D<FfxFloat32x4> r_transparency_mask : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_TRANSPARENCY_MASK);
@@ -275,15 +275,15 @@ SamplerState s_LinearClamp : register(s1);
Texture2D<FfxFloat32x4> r_motion_vectors_reflection : FFX_FSR2_DECLARE_SRV(FFX_FSR2_RESOURCE_IDENTIFIER_VELOCITY_REFLECTION);
// UAV declarations
- RWTexture2D<FfxUInt32> rw_reconstructed_previous_nearest_depth : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_RECONSTRUCTED_PREVIOUS_NEAREST_DEPTH);
- RWTexture2D<FfxFloat32x2> rw_dilated_motion_vectors : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_MOTION_VECTORS);
- RWTexture2D<FfxFloat32> rw_dilatedDepth : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_DEPTH);
- RWTexture2D<FfxFloat32x4> rw_internal_upscaled_color : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_INTERNAL_UPSCALED_COLOR);
- RWTexture2D<FfxFloat32x3> rw_lock_status : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_LOCK_STATUS);
- RWTexture2D<FfxFloat32> rw_depth_clip : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_DEPTH_CLIP);
- RWTexture2D<unorm FfxFloat32x4> rw_prepared_input_color : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_PREPARED_INPUT_COLOR);
- RWTexture2D<unorm FfxFloat32x4> rw_luma_history : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_LUMA_HISTORY);
- RWTexture2D<FfxFloat32x4> rw_upscaled_output : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_UPSCALED_OUTPUT);
+ RWTexture2D<FfxUInt32> rw_reconstructed_previous_nearest_depth : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_RECONSTRUCTED_PREVIOUS_NEAREST_DEPTH);
+ RWTexture2D<FfxFloat32x2> rw_dilated_motion_vectors : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_MOTION_VECTORS);
+ RWTexture2D<FfxFloat32> rw_dilatedDepth : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_DILATED_DEPTH);
+ RWTexture2D<FfxFloat32x4> rw_internal_upscaled_color : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_INTERNAL_UPSCALED_COLOR);
+ RWTexture2D<FfxFloat32x3> rw_lock_status : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_LOCK_STATUS);
+ RWTexture2D<FfxFloat32> rw_depth_clip : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_DEPTH_CLIP);
+ RWTexture2D<FfxFloat32x4> rw_prepared_input_color : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_PREPARED_INPUT_COLOR);
+ RWTexture2D<unorm FfxFloat32x4> rw_luma_history : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_LUMA_HISTORY);
+ RWTexture2D<FfxFloat32x4> rw_upscaled_output : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_UPSCALED_OUTPUT);
globallycoherent RWTexture2D<FfxFloat32> rw_img_mip_shading_change : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_AUTO_EXPOSURE_MIPMAP_SHADING_CHANGE);
globallycoherent RWTexture2D<FfxFloat32> rw_img_mip_5 : FFX_FSR2_DECLARE_UAV(FFX_FSR2_RESOURCE_IDENTIFIER_AUTO_EXPOSURE_MIPMAP_5);
@@ -330,7 +330,7 @@ SamplerState s_LinearClamp : register(s1);
Texture2D<FfxFloat32> r_depth_clip : FFX_FSR2_DECLARE_SRV(FSR2_BIND_SRV_DEPTH_CLIP);
#endif
#if defined FSR2_BIND_SRV_PREPARED_INPUT_COLOR
- Texture2D<unorm FfxFloat32x4> r_prepared_input_color : FFX_FSR2_DECLARE_SRV(FSR2_BIND_SRV_PREPARED_INPUT_COLOR);
+ Texture2D<FfxFloat32x4> r_prepared_input_color : FFX_FSR2_DECLARE_SRV(FSR2_BIND_SRV_PREPARED_INPUT_COLOR);
#endif
#if defined FSR2_BIND_SRV_LUMA_HISTORY
Texture2D<unorm FfxFloat32x4> r_luma_history : FFX_FSR2_DECLARE_SRV(FSR2_BIND_SRV_LUMA_HISTORY);
@@ -371,7 +371,7 @@ SamplerState s_LinearClamp : register(s1);
RWTexture2D<FfxFloat32> rw_depth_clip : FFX_FSR2_DECLARE_UAV(FSR2_BIND_UAV_DEPTH_CLIP);
#endif
#if defined FSR2_BIND_UAV_PREPARED_INPUT_COLOR
- RWTexture2D<unorm FfxFloat32x4> rw_prepared_input_color : FFX_FSR2_DECLARE_UAV(FSR2_BIND_UAV_PREPARED_INPUT_COLOR);
+ RWTexture2D<FfxFloat32x4> rw_prepared_input_color : FFX_FSR2_DECLARE_UAV(FSR2_BIND_UAV_PREPARED_INPUT_COLOR);
#endif
#if defined FSR2_BIND_UAV_LUMA_HISTORY
RWTexture2D<unorm FfxFloat32x4> rw_luma_history : FFX_FSR2_DECLARE_UAV(FSR2_BIND_UAV_LUMA_HISTORY);
@@ -778,7 +778,11 @@ FfxFloat32x2 SampleDilatedMotionVector(FfxFloat32x2 fUV)
FfxFloat32 LoadDilatedDepth(FfxUInt32x2 iPxInput)
{
#if defined(FSR2_BIND_SRV_DILATED_DEPTH) || defined(FFX_INTERNAL)
- return r_dilatedDepth[iPxInput];
+ FfxFloat32 fDepth = r_dilatedDepth[iPxInput];
+#if !FFX_FSR2_OPTION_INVERTED_DEPTH
+ fDepth = 1.0 - fDepth; // Reconstruct from FP16.
+#endif
+ return fDepth;
#else
return 0.f;
#endif
diff --git a/src/ffx-fsr2-api/shaders/ffx_fsr2_depth_clip.h b/src/ffx-fsr2-api/shaders/ffx_fsr2_depth_clip.h
index 81db737..4748c96 100644
--- a/src/ffx-fsr2-api/shaders/ffx_fsr2_depth_clip.h
+++ b/src/ffx-fsr2-api/shaders/ffx_fsr2_depth_clip.h
@@ -34,8 +34,12 @@ FfxFloat32 ComputeSampleDepthClip(FfxInt32x2 iPxSamplePos, FfxFloat32 fPreviousD
const FfxFloat32 fHalfViewportWidth = RenderSize().x * 0.5f;
FfxFloat32 fDepthThreshold = ffxMin(fCurrentDepthViewSpace, fPrevNearestDepthViewSpace);
+#if !FFX_FSR2_OPTION_INVERTED_DEPTH
+ const FfxFloat32 Ksep = 4.0f * 1.37e-05f; // Arbitrary hack to make normal depth work better.
+#else
// WARNING: Ksep only works with reversed-z with infinite projection.
const FfxFloat32 Ksep = 1.37e-05f;
+#endif
FfxFloat32 fRequiredDepthSeparation = Ksep * fDepthThreshold * TanHalfFoV() * fHalfViewportWidth;
FfxFloat32 fDepthDiff = fCurrentDepthViewSpace - fPrevNearestDepthViewSpace;
diff --git a/src/ffx-fsr2-api/shaders/ffx_fsr2_sample.h b/src/ffx-fsr2-api/shaders/ffx_fsr2_sample.h
index cfa9db8..f697d70 100644
--- a/src/ffx-fsr2-api/shaders/ffx_fsr2_sample.h
+++ b/src/ffx-fsr2-api/shaders/ffx_fsr2_sample.h
@@ -563,32 +563,24 @@ FFX_MIN16_I2 ClampCoord(FFX_MIN16_I2 iPxSample, FFX_MIN16_I2 iPxOffset, FFX_MIN1
// BE CAREFUL: there is some precision issues and (3253, 125) leading to (3252.9989778, 125.001102)
// is common, so iPxSample can "jitter"
-#define DeclareCustomTextureSample(Name, InterpolateSamples, FetchSamples) \
- FfxFloat32x4 Name(FfxFloat32x2 fUvSample, FfxInt32x2 iTextureSize) \
- { \
- FfxFloat32x2 fPxSample = fUvSample * FfxFloat32x2(iTextureSize) - FfxFloat32x2(0.5f, 0.5f); \
- FfxInt32x2 iPxSample = FfxInt32x2(floor(fPxSample)); \
- /* Clamp base coords */ \
- iPxSample.x = ffxMax(0, ffxMin(iPxSample.x, iTextureSize.x - 1)); \
- iPxSample.y = ffxMax(0, ffxMin(iPxSample.y, iTextureSize.y - 1)); \
- /* */ \
- FfxFloat32x2 fPxFrac = ffxFract(fPxSample); \
- FfxFloat32x4 fColorXY = FfxFloat32x4(InterpolateSamples(FetchSamples(iPxSample, iTextureSize), fPxFrac)); \
- return fColorXY; \
+#define DeclareCustomTextureSample(Name, InterpolateSamples, FetchSamples) \
+ FfxFloat32x4 Name(FfxFloat32x2 fUvSample, FfxInt32x2 iTextureSize) \
+ { \
+ FfxFloat32x2 fPxSample = fUvSample * FfxFloat32x2(iTextureSize) - FfxFloat32x2(0.5f, 0.5f); \
+ FfxInt32x2 iPxSample = FfxInt32x2(floor(fPxSample)); \
+ FfxFloat32x2 fPxFrac = ffxFract(fPxSample); \
+ FfxFloat32x4 fColorXY = FfxFloat32x4(InterpolateSamples(FetchSamples(iPxSample, iTextureSize), fPxFrac)); \
+ return fColorXY; \
}
-#define DeclareCustomTextureSampleMin16(Name, InterpolateSamples, FetchSamples) \
- FFX_MIN16_F4 Name(FfxFloat32x2 fUvSample, FfxInt32x2 iTextureSize) \
- { \
- FfxFloat32x2 fPxSample = fUvSample * FfxFloat32x2(iTextureSize) - FfxFloat32x2(0.5f, 0.5f); \
- FfxInt32x2 iPxSample = FfxInt32x2(floor(fPxSample)); \
- /* Clamp base coords */ \
- iPxSample.x = ffxMax(0, ffxMin(iPxSample.x, iTextureSize.x - 1)); \
- iPxSample.y = ffxMax(0, ffxMin(iPxSample.y, iTextureSize.y - 1)); \
- /* */ \
- FFX_MIN16_F2 fPxFrac = FFX_MIN16_F2(ffxFract(fPxSample)); \
- FFX_MIN16_F4 fColorXY = FFX_MIN16_F4(InterpolateSamples(FetchSamples(iPxSample, iTextureSize), fPxFrac)); \
- return fColorXY; \
+#define DeclareCustomTextureSampleMin16(Name, InterpolateSamples, FetchSamples) \
+ FFX_MIN16_F4 Name(FfxFloat32x2 fUvSample, FfxInt32x2 iTextureSize) \
+ { \
+ FfxFloat32x2 fPxSample = fUvSample * FfxFloat32x2(iTextureSize) - FfxFloat32x2(0.5f, 0.5f); \
+ FfxInt32x2 iPxSample = FfxInt32x2(floor(fPxSample)); \
+ FFX_MIN16_F2 fPxFrac = FFX_MIN16_F2(ffxFract(fPxSample)); \
+ FFX_MIN16_F4 fColorXY = FFX_MIN16_F4(InterpolateSamples(FetchSamples(iPxSample, iTextureSize), fPxFrac)); \
+ return fColorXY; \
}
#define FFX_FSR2_CONCAT_ID(x, y) x ## y
diff --git a/src/ffx-fsr2-api/shaders/ffx_fsr2_upsample.h b/src/ffx-fsr2-api/shaders/ffx_fsr2_upsample.h
index a66a5be..80524d4 100644
--- a/src/ffx-fsr2-api/shaders/ffx_fsr2_upsample.h
+++ b/src/ffx-fsr2-api/shaders/ffx_fsr2_upsample.h
@@ -46,10 +46,10 @@ FfxFloat32 GetUpsampleLanczosWeight(FfxFloat32x2 fSrcSampleOffset, FfxFloat32x2
FfxFloat32x2 fSrcSampleOffsetBiased = fSrcSampleOffset * fKernelWeight;
#if FFX_FSR2_OPTION_UPSAMPLE_USE_LANCZOS_TYPE == 0 // LANCZOS_TYPE_REFERENCE
FfxFloat32 fSampleWeight = Lanczos2(length(fSrcSampleOffsetBiased));
-#elif FFX_FSR2_OPTION_UPSAMPLE_USE_LANCZOS_TYPE == 1 // LANCZOS_TYPE_APPROXIMATE
- FfxFloat32 fSampleWeight = Lanczos2ApproxSq(dot(fSrcSampleOffsetBiased, fSrcSampleOffsetBiased));
-#elif FFX_FSR2_OPTION_UPSAMPLE_USE_LANCZOS_TYPE == 2 // LANCZOS_TYPE_LUT
+#elif FFX_FSR2_OPTION_UPSAMPLE_USE_LANCZOS_TYPE == 1 // LANCZOS_TYPE_LUT
FfxFloat32 fSampleWeight = Lanczos2_UseLUT(length(fSrcSampleOffsetBiased));
+#elif FFX_FSR2_OPTION_UPSAMPLE_USE_LANCZOS_TYPE == 2 // LANCZOS_TYPE_APPROXIMATE
+ FfxFloat32 fSampleWeight = Lanczos2ApproxSq(dot(fSrcSampleOffsetBiased, fSrcSampleOffsetBiased));
#else
#error "Invalid Lanczos type"
#endif
diff --git a/src/ffx-fsr2-api/vk/ffx_fsr2_vk.cpp b/src/ffx-fsr2-api/vk/ffx_fsr2_vk.cpp
index be766b5..683ad9a 100644
--- a/src/ffx-fsr2-api/vk/ffx_fsr2_vk.cpp
+++ b/src/ffx-fsr2-api/vk/ffx_fsr2_vk.cpp
@@ -19,6 +19,8 @@
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
+#define fsr2GetPermutationBlobByIndex fsr2GetPermutationBlobByIndexVK
+
#include "../ffx_fsr2.h"
#include "ffx_fsr2_vk.h"
#include "shaders/ffx_fsr2_shaders_vk.h" // include all the precompiled VK shaders for the FSR2 passes
@@ -282,8 +284,6 @@ VkFormat getVKFormatFromSurfaceFormat(FfxSurfaceFormat fmt)
return VK_FORMAT_R32G32B32A32_SFLOAT;
case(FFX_SURFACE_FORMAT_R16G16B16A16_FLOAT):
return VK_FORMAT_R16G16B16A16_SFLOAT;
- case(FFX_SURFACE_FORMAT_R16G16B16A16_UNORM):
- return VK_FORMAT_R16G16B16A16_UNORM;
case(FFX_SURFACE_FORMAT_R32G32_FLOAT):
return VK_FORMAT_R32G32_SFLOAT;
case(FFX_SURFACE_FORMAT_R32_UINT):
@@ -409,8 +409,6 @@ FfxSurfaceFormat ffxGetSurfaceFormatVK(VkFormat fmt)
return FFX_SURFACE_FORMAT_R32G32B32A32_FLOAT;
case(VK_FORMAT_R16G16B16A16_SFLOAT):
return FFX_SURFACE_FORMAT_R16G16B16A16_FLOAT;
- case(VK_FORMAT_R16G16B16A16_UNORM):
- return FFX_SURFACE_FORMAT_R16G16B16A16_UNORM;
case(VK_FORMAT_R32G32_SFLOAT):
return FFX_SURFACE_FORMAT_R32G32_FLOAT;
case(VK_FORMAT_R32_UINT):
diff --git a/src/ffx-fsr2-api/vk/shaders/ffx_fsr2_shaders_vk.cpp b/src/ffx-fsr2-api/vk/shaders/ffx_fsr2_shaders_vk.cpp
index 230ae9b..5842275 100644
--- a/src/ffx-fsr2-api/vk/shaders/ffx_fsr2_shaders_vk.cpp
+++ b/src/ffx-fsr2-api/vk/shaders/ffx_fsr2_shaders_vk.cpp
@@ -19,6 +19,8 @@
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
+#define fsr2GetPermutationBlobByIndex fsr2GetPermutationBlobByIndexVK
+
#include "ffx_fsr2_shaders_vk.h"
#include "ffx_fsr2_autogen_reactive_pass_permutations.h"