Autoexposure example restoration by devshgraphicsprogramming · Pull Request #137 · Devsh-Graphics-Programming/Nabla-Examples-and-Tests

devshgraphicsprogramming · 2024-08-13T22:08:36Z

No description provided.

…Examples-and-Tests into autoexposure_ex

…update luma DSes

devshgraphicsprogramming · 2024-08-13T22:10:43Z

26_Autoexposure/app_resources/common.hlsl

+    float meteringWindowScaleX, meteringWindowScaleY;
+    float meteringWindowOffsetX, meteringWindowOffsetY;
+    float lumaMin, lumaMax;
+    uint32_t sampleCountX, sampleCountY;
+    uint32_t viewportSizeX, viewportSizeY;
+    uint64_t lumaMeterBDA;


you should pack these up intro structs in nbl::hlsl::luma_meter and friends

P.S. I see you already have, why not use things like nbl::hlsl::luma_meter::LumaMeteringWindow directly here?

AnastaZIuk · 2026-02-03T12:05:01Z

26_Autoexposure/main.cpp

+			const uint32_t2 dispatchSize = {
+				1 + ((gpuImgExtent.width / 2) - 1) / SubgroupSize,
+				1 + ((gpuImgExtent.height / 2) - 1) / SubgroupSize
+			};


Half dispatch with full viewport UV means you sample only the top left quarter, EV is computed from that quarter and then applied to the whole image so exposure drifts when luminance is not uniform

Nabla-Examples-and-Tests/26_Autoexposure/app_resources/avg_luma_meter.comp.hlsl

Lines 68 to 70 in 233db45

uint32_t texWidth, texHeight;

texture.GetDimensions(texWidth, texHeight);

meter.sampleLuma(pushData.window, val_accessor, tex, sdata, (float32_t2)(glsl::gl_WorkGroupID() * glsl::gl_WorkGroupSize()), float32_t2(texWidth, texHeight));

https://github.com/Devsh-Graphics-Programming/Nabla/blob/8466a9db6ee2d060abccb5710461f3f2d97ecf5a/include/nbl/builtin/hlsl/luma_meter/geom_mean.hlsl#L105

counterexample https://godbolt.org/z/de41xd9ro

AnastaZIuk · 2026-02-03T12:15:03Z

26_Autoexposure/main.cpp

+	inline void workLoopBody() override
+	{
+		memset(m_gatherMemory, 0, m_gatherBuffer->getSize());
+		memset(m_histoMemory, 0, m_gatherBuffer->getSize());


This typo keeps most bins dirty and EV is biased, use m_histoBuffer->getSize()

counterexample https://godbolt.org/z/Mozo84eYY

AnastaZIuk · 2026-02-03T18:18:12Z

26_Autoexposure/app_resources/avg_luma_tonemap.comp.hlsl

+    tonemapper::Reinhard<float32_t> reinhard = tonemapper::Reinhard<float32_t>::create(EV, 0.18f, 0.85f);
+    float32_t3 tonemappedColor = mul(colorspace::decode::XYZtoscRGB, reinhard(CIEColor));
+
+    textureOut[pos] = float32_t4(colorspace::oetf::sRGB(tonemappedColor), 1.0f);


No bounds guard and we have OOB write. We go 1280x720 and dispatch is ceil(Dimensions/SubgroupSize). With SubgroupSize 32 it is 1280x736 coverage

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 46 in 233db45

constexpr static inline uint32_t2 Dimensions = { 1280, 720 };

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Lines 841 to 844 in 233db45

const uint32_t2 dispatchSize = {

1 + ((Dimensions.x) - 1) / SubgroupSize,

1 + ((Dimensions.y) - 1) / SubgroupSize

};

I'm not sure about it, could you please confirm the expected metering region is $uv\in[0.1,0.9]^2$? If so then with current inputs

Dimensions=(1280,720)

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 47 in 59117a8

constexpr static inline uint32_t2 Dimensions = { 1280, 720 };

MeteringMinUV=(0.1,0.1)

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 48 in 59117a8

constexpr static inline float32_t2 MeteringMinUV = { 0.1f, 0.1f };

MeteringMaxUV=(0.9,0.9)

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 49 in 59117a8

constexpr static inline float32_t2 MeteringMaxUV = { 0.9f, 0.9f };

SamplingFactor=2

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 50 in 59117a8

constexpr static inline float32_t SamplingFactor = 2.f;

SubgroupSize=m_physicalDevice->getLimits().subgroupSize

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 791 in 59117a8

const uint32_t SubgroupSize = m_physicalDevice->getLimits().subgroupSize;

and flow like

meteringUVRange = MeteringMaxUV - MeteringMinUV

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 821 in 59117a8

const float32_t2 meteringUVRange = MeteringMaxUV - MeteringMinUV;

dispatchSize = ceil(dim * meteringUVRange / (SubgroupSize * SamplingFactor))

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 827 in 59117a8

const uint32_t2 dispatchSize = uint32_t2(hlsl::ceil(float32_t2(gpuImgExtent.width, gpuImgExtent.height) * meteringUVRange / (SubgroupSize * SamplingFactor)));

window = create(meteringUVRange / (dispatchSize * SubgroupSize), MeteringMinUV)

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 829 in 59117a8

pc.window = luma_meter::MeteringWindow::create(meteringUVRange / (float32_t2(dispatchSize) * float(SubgroupSize)), MeteringMinUV);

sampleLuma(..., tileOffset=WorkGroupID*WorkGroupSize, viewportSize=(texWidth,texHeight))

Nabla-Examples-and-Tests/26_Autoexposure/app_resources/avg_luma_meter.comp.hlsl

Line 71 in 59117a8

meter.sampleLuma(pushData.window, val_accessor, tex, sdata, (float32_t2)(glsl::gl_WorkGroupID() * glsl::gl_WorkGroupSize()), float32_t2(texWidth, texHeight));

shiftedCoord = (tileOffset + coord) / viewportSize
https://github.com/Devsh-Graphics-Programming/Nabla/blob/79fad2f7d16ba75347c94a8b4327042b4d786547/include/nbl/builtin/hlsl/luma_meter/geom_mean.hlsl#L126

uvPos = shiftedCoord * window.meteringWindowScale + window.meteringWindowOffset
https://github.com/Devsh-Graphics-Programming/Nabla/blob/79fad2f7d16ba75347c94a8b4327042b4d786547/include/nbl/builtin/hlsl/luma_meter/geom_mean.hlsl#L77

we have

dispatchSize=(16,9)

windowScale=(0.8/(16*32),0.8/(9*32))=(0.0015625,0.002777...)

(tileOffset+coord) spans [0,511]x[0,287]

and then sampled UV is approximately

$uv_x\in[0.100000,0.100624]$

$uv_y\in[0.100000,0.101107]$

this is a tiny rectangle near (0.1,0.1) not the intended (I guess) metering window because the coordinates are effectively normalized twice, first by viewportSize in shiftedCoord = (tileOffset + coord) / viewportSize and then scaled again by windowScale = meteringUVRange / (dispatchSize * SubgroupSize).

https://github.com/Devsh-Graphics-Programming/Nabla/blob/79fad2f7d16ba75347c94a8b4327042b4d786547/include/nbl/builtin/hlsl/luma_meter/geom_mean.hlsl#L126

Nabla-Examples-and-Tests/26_Autoexposure/main.cpp

Line 829 in 59117a8

pc.window = luma_meter::MeteringWindow::create(meteringUVRange / (float32_t2(dispatchSize) * float(SubgroupSize)), MeteringMinUV);

AnastaZIuk · 2026-02-03T18:19:43Z

26_Autoexposure/app_resources/median_luma_tonemap.comp.hlsl

+    tonemapper::Reinhard<float32_t> reinhard = tonemapper::Reinhard<float32_t>::create(EV, 0.18, 0.85f);
+    float32_t3 tonemappedColor = mul(colorspace::decode::XYZtoscRGB, reinhard(CIEColor));
+
+    textureOut[pos] = float32_t4(colorspace::oetf::sRGB(tonemappedColor), 1.0f);


the same OOB issue as in #137 (comment)

… in main

…ntiles

…se vulkan should do that

nipunG314 and others added 28 commits July 19, 2024 22:07

Add 26_Autoexposure

345ac7b

Change 26_Autoexposure to SimpleWindowedApplication

87d4794

Build a staging buffer and upload exr image

7a5ea7c

Merge branch 'master' of github.com:Devsh-Graphics-Programming/Nabla-…

5026a64

…Examples-and-Tests into autoexposure_ex

Init surface and create the swapchain

5d63d04

Load shaders and create the pipeline for full screen triagnle

640e6a3

Merge branch 'master' of github.com:Devsh-Graphics-Programming/Nabla-…

0148f60

…Examples-and-Tests into autoexposure_ex

Set window size according to loaded image

d69a111

Stop running if window is closed

54bf38f

Acquire swapchain image and present uploaded image to it

461efd3

Set window size directly and use that for swapchain rendering

734fea9

Merge branch 'master' of github.com:Devsh-Graphics-Programming/Nabla-…

fc9b0bb

…Examples-and-Tests into autoexposure_ex

m_computeSubgroupSize

4a11724

Merge branch 'master' of github.com:Devsh-Graphics-Programming/Nabla-…

734b887

…Examples-and-Tests into autoexposure_ex

Allocate buffer for gathered luma values

7d4895a

Create gpu resources for all passes

0e3e125

Create shaders and pipelines

cef80b3

Allocate and create texture for tonemapping

15e489f

Create separate ds for luma and present

c646c7d

Record luma meter commands

36d7097

Merge branch 'master' of github.com:Devsh-Graphics-Programming/Nabla-…

3a70977

…Examples-and-Tests into autoexposure_ex

fix layout issues with compute pipeline in 26_Autoexposure example + …

6addbf1

…update luma DSes

Create two sets from common lumaPresentLayout correctly

8434f20

Create compute and graphics resources separately and finish luma meter

bf08caa

Fix descriptor binding for luma_meter

b342c6c

Create separate pipeline layouts for luma and present

817c4a7

Setup luma_meter.comp.hlsl

7f89542

Pass push constants

defd45e

devshgraphicsprogramming commented Aug 13, 2024

View reviewed changes

Record draw pass correctly

f6f8154

keptsecret added 5 commits February 2, 2026 10:03

Merge branch 'master' into autoexposure_ex

7680a95

fixes to get it to compile after merge from master + old bugs in median

42ed25a

fixes histogram autoexposure

a15508e

clean up push constant usage, more values in push constant

9c330d3

rename defines to all caps, refactor luma meter template names

233db45

AnastaZIuk reviewed Feb 3, 2026

View reviewed changes

keptsecret added 21 commits February 4, 2026 10:26

Merge branch 'master' into autoexposure_ex

80fd063

fix change from master that broke

de38929

added temporal adaptation

ea585f5

minor fixes

e1d1cb1

refactor to use workgroup2 stuff in cpp and average metering

9f86c09

merge master, fix conflicts

5e27920

refactor to use workgroup2 with histogram metering, some more defines…

7c5d3b4

… in main

fix calculating sampling area uv for luma metering

d46e463

check range on tonemap output

77ec3d5

small fix to oob bug

59117a8

refactor uv fix, minor bug fixes

d104945

removed semantically incorrect naming of gl_WorkgroupSize()

986492e

added in temp dithering factor, fix usage of histogram metering perce…

07ad5db

…ntiles

ping pong buffer for last frame EV, removed tonemap output oetf becau…

1e012c0

…se vulkan should do that

removed sampleCount param (unused), made subgroup size member

df7f7ed

implement frames in flight

8b93ee5

add eotf, removed fake dither factor

0804333

use median + aces

f6b1b7d

made tonemapped image same size as input img

27ff451

moved push constant out to nabla

9482f15

refactor minor changes to luma meter

ca0e50e

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Comments

Autoexposure example restoration#137

Autoexposure example restoration#137
devshgraphicsprogramming wants to merge 97 commits intomasterfrom
autoexposure_ex

devshgraphicsprogramming commented Aug 13, 2024

Uh oh!

devshgraphicsprogramming Aug 13, 2024

Uh oh!

AnastaZIuk Feb 3, 2026

Uh oh!

keptsecret Feb 6, 2026

Uh oh!

AnastaZIuk Feb 3, 2026

Uh oh!

keptsecret Feb 6, 2026

Uh oh!

AnastaZIuk Feb 3, 2026

Uh oh!

keptsecret Feb 6, 2026

Uh oh!

AnastaZIuk Feb 6, 2026

Uh oh!

keptsecret Feb 9, 2026

Uh oh!

AnastaZIuk Feb 3, 2026

Uh oh!

keptsecret Feb 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

	uint32_t texWidth, texHeight;
	texture.GetDimensions(texWidth, texHeight);
	meter.sampleLuma(pushData.window, val_accessor, tex, sdata, (float32_t2)(glsl::gl_WorkGroupID() * glsl::gl_WorkGroupSize()), float32_t2(texWidth, texHeight));

	const uint32_t2 dispatchSize = {
	1 + ((Dimensions.x) - 1) / SubgroupSize,
	1 + ((Dimensions.y) - 1) / SubgroupSize
	};

Comments

Conversation

devshgraphicsprogramming commented Aug 13, 2024

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants