[WIP] Add VK_EXT_image_2d_view_of_3d support #2332

ncesario-lunarg · 2024-09-10T00:53:17Z

Given a 3D texture that is backed by a placement heap in MoltenVK, the
approach taken here is to create a 2D texture that is backed by memory
pointing into a 3D texture's memory.

While ideal compared to alternative implementation solutions for this
extension, this approach is sensitive to how Apple lays out the memory
for 3D textures. The solution here uses
heapTextureSizeAndAlignWithDescriptor to determine the overall size of a
given 3D texture and index into the beginning of each "slice" of the 3D
texture. So far this is good enough for storage images in CTS.

FYI @kocdemir.

billhollings

Thanks for getting this started!

In addition to the inline change requests, I recognize this is just to solve one immediate extension issue, but if we can do what we can to simplify a future move to using MTLHeaps, we should try as best we can.

billhollings · 2024-09-12T18:47:43Z

MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.mm

@@ -188,10 +188,6 @@
 // Can't create MTLHeaps of zero size.
 if (_allocationSize == 0) { return true; }

-#if MVK_MACOS
- // MTLHeaps on macOS must use private storage for now.
- if (_mtlStorageMode != MTLStorageModePrivate) { return true; }


Instead of being removed, does this need to be modified to include a test for non-Apple-Silicon devices?

See comment below.

billhollings · 2024-09-12T18:50:35Z

MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.mm

-#if MVK_MACOS
- // MTLHeaps on macOS must use private storage for now.
- if (_mtlStorageMode != MTLStorageModePrivate) { return true; }
-#endif
 #if MVK_IOS


Should both this section and the one above, be replaced with a single test for MVKMTLDeviceCapabilities.isAppleGPU?

TBH I'm not sure based on the docs, but I'll add that to be on the safe side. It looked like the usage of heaps in MoltenVK was mainly dependent on placement heaps, but my knowledge is still pretty sparse on the details here.

billhollings · 2024-09-12T18:52:53Z

MoltenVK/MoltenVK/GPUObjects/MVKImage.h

@@ -48,6 +48,16 @@ typedef struct {
 class MVKImagePlane : public MVKBaseObject {

 public:
+ struct HeapAllocation {


Should this declaration be moved to MVKDeviceMemory or MVKResource, since heaps can be used by buffers too?

Yep, I'll move it there.

@aitor-lunarg as this is something that may have a more straightforward approach with WIP: Proposal to base VkDeviceMemory on MTLHeap always #2309.

billhollings · 2024-09-12T19:37:43Z

MoltenVK/MoltenVK/GPUObjects/MVKImage.mm

@@ -1066,6 +1079,11 @@ static MTLRegion getMTLRegion(const ImgRgn& imgRgn) {
 return _memoryBindings[0]->_deviceMemory ? _memoryBindings[0]->_deviceMemory->getMTLCPUCacheMode() : MTLCPUCacheModeDefaultCache;
 }

+MVKImagePlane::HeapAllocation* MVKImage::getHeapAllocation(uint32_t planeIndex) {
+ auto& heapAllocation = _planes[planeIndex]->_heapAllocation;
+ return (heapAllocation) ? &heapAllocation : nullptr;


It surprised me that this was compiling, until I noticed the operator bool() on the struct above.

While clever, we don't use this design pattern anywhere else, so I'm concerned others will balk the same way I did.

Maybe rename operator bool() to bool isValid(), for clarity.

Changed to isValid().

billhollings · 2024-09-12T19:40:44Z

MoltenVK/MoltenVK/GPUObjects/MVKImage.mm

@@ -1251,6 +1269,8 @@ static MTLRegion getMTLRegion(const ImgRgn& imgRgn) {
 if (pExportInfo && pExportInfo->exportObjectType == VK_EXPORT_METAL_OBJECT_TYPE_METAL_IOSURFACE_BIT_EXT && !_ioSurface) {
 setConfigurationResult(useIOSurface(nil));
 }
+
+ _is2DViewCompatible = pCreateInfo->flags & VK_IMAGE_CREATE_2D_VIEW_COMPATIBLE_BIT_EXT;


Use mvkIsAnyFlagEnabled() for consistency.

billhollings · 2024-09-12T19:43:33Z

MoltenVK/MoltenVK/GPUObjects/MVKImage.h

@@ -395,6 +410,7 @@ class MVKImage : public MVKVulkanAPIDeviceObject {
 bool _hasMutableFormat;
 bool _shouldSupportAtomics;
 bool _isLinearForAtomics;
+ bool _is2DViewCompatible = false;


I assume this will apply only to 2D on 3D, and not for all 2D views?

Maybe rename to _is2DViewOn3DImageCompatible?

Yep, makes sense. Done.

billhollings · 2024-09-12T19:58:26Z

MoltenVK/MoltenVK/GPUObjects/MVKImage.h

+ size_t size = 0; // Total size of this allocation
+ size_t align = 0; // Allocation alignment requirement
+
+ operator bool() const {


See my comment below where this is used. Probably best to rename this to something like bool isValid().

billhollings · 2024-09-12T20:10:16Z

MoltenVK/MoltenVK/GPUObjects/MVKImage.h

+
+ // For now, this is only used for VK_EXT_image_2d_view_of_3d. Specifically, when the backing image of this view is 3D and the view is 2D,
+ // this texture will be allocated using a placement heap "on top of" the 3D textures backing memory.
+ id<MTLTexture> _mtlShadowTexture;


Since _mtlTexture end up holding a reference to the same object, does this need to be a separate iVar than _mtlTexture?

If it is needed for some reason, move it up just below the _mtlTexture declaration. We try to organize iVars to minimize gaps in the underlying class structure (ie- keep all 8-byte iVars together).

It also feels dangerous that the MTLTexture is not retained in both _mtlTexture and _mtlShadowTexture, and released from both in the destructor. If in the future, we might possibly have different objects in the two iVars, then we'll end up with a hard to discover memory leak.

Finally, maybe it's me, but when I see _mtlShadowTexture, I keep thinking "shadow map texture". Maybe rename it to _mtlAliasTexture or something.

Nope, there is no need for the extra MTLTexture in MVKImageViewPlane; I will remove it. I added this from a debug session and never reverted it (and also forgot to clean it up in the destructor when I was using it :/).

ncesario-lunarg · 2024-09-13T17:22:06Z

Thanks for the review @billhollings! I've made the changes you've suggested locally, but I'm hitting a crash on dEQP-VK.api.get_memory_commitment.memory_commitment at https://github.com/ncesario-lunarg/MoltenVK/blob/d01e7ecf13ef9bd4d9af3dc61b13ce683299d5ba/MoltenVK/MoltenVK/GPUObjects/MVKDeviceMemory.mm#L205. I'm going to resolve that before updating this PR.

Given a 3D texture that is backed by a placement heap in MoltenVK, the approach taken here is to create a 2D texture that is backed by memory pointing into a 3D texture's memory. While ideal compared to alternative implementation solutions for this extension, this approach is sensitive to how Apple lays out the memory for 3D textures. The solution here uses heapTextureSizeAndAlignWithDescriptor to determine the overall size of a given 3D texture and index into the beginning of each "slice" of the 3D texture. So far this is good enough for storage images in CTS.

Sets VkPhysicalDeviceImage2DViewOf3DFeaturesEXT::sampler2DViewOf3D to false, as CTS tests involving 2D views of sampled 3D textures are not currently working.

billhollings requested changes Sep 12, 2024

View reviewed changes

ncesario-lunarg added 2 commits September 13, 2024 18:02

Disable sampler2DViewOf3D

ba83997

Sets VkPhysicalDeviceImage2DViewOf3DFeaturesEXT::sampler2DViewOf3D to false, as CTS tests involving 2D views of sampled 3D textures are not currently working.

ncesario-lunarg force-pushed the 2dviewof3d-2 branch from d01e7ec to ba83997 Compare September 14, 2024 00:09

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP] Add VK_EXT_image_2d_view_of_3d support #2332

[WIP] Add VK_EXT_image_2d_view_of_3d support #2332

ncesario-lunarg commented Sep 10, 2024

billhollings left a comment •

edited

Loading

billhollings Sep 12, 2024

ncesario-lunarg Sep 12, 2024

billhollings Sep 12, 2024

ncesario-lunarg Sep 12, 2024

billhollings Sep 12, 2024

ncesario-lunarg Sep 12, 2024

billhollings Sep 12, 2024

ncesario-lunarg Sep 12, 2024

billhollings Sep 12, 2024

ncesario-lunarg Sep 12, 2024

billhollings Sep 12, 2024

ncesario-lunarg Sep 12, 2024

billhollings Sep 12, 2024

ncesario-lunarg Sep 12, 2024

billhollings Sep 12, 2024

ncesario-lunarg Sep 12, 2024

ncesario-lunarg commented Sep 13, 2024

[WIP] Add VK_EXT_image_2d_view_of_3d support #2332

Are you sure you want to change the base?

[WIP] Add VK_EXT_image_2d_view_of_3d support #2332

Conversation

ncesario-lunarg commented Sep 10, 2024

billhollings left a comment • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ncesario-lunarg commented Sep 13, 2024

billhollings left a comment •

edited

Loading