add support for image1d_buffer #609

rjodinchr · 2023-09-12T13:24:14Z

This implementation has been tested using the OpenCL-CTS patch with KhronosGroup/OpenCL-CTS#1806

Change-Id: I4eef91af393b5487bbc208bdc96d43daa4c5c462

remove 1DBUFFER mention in init_vulkan_image limit CL_DEVICE_IMAGE_MAX_BUFFER_SIZE WITH CL_DEVICE_MAX_MEM_ALLOC_SIZE

rjodinchr · 2023-10-13T06:25:25Z

This PR depends on an update of clspv

kpet

This change does not introduce CTS regressions at my end. As discussed I'm not a big fan of handling the buffer view special case in the API layer as this breaks encapsulation. In the interest of getting forward progress, I'm with this approach as a first step (and I think it likely makes the code a bit shorter).

src/memory.cpp

src/api.cpp

kpet · 2023-10-15T14:27:56Z

src/api.cpp

+            if ((components_sampled != components_storage) &&
+                (image_type == CL_MEM_OBJECT_IMAGE1D_BUFFER)) {
+                continue;
+            }


I'm a bit confused by this condition. Those swizzles are unused for 1D buffer images. Why is this needed?

I am not sure, this is coming from your initial commit I believe. Thinking about it, I don't see a reason to have it, maybe we should get rid of it?

yeah, if we pass the CTS without it, then we should remove it. If we somehow need this to pass the CTS, we should find a clearer expression of the conditions.

It seems that this test is needed. Otherwise the CTS thinks that we support CL_INTENSITY and CL_LUMINANCE with 1d buffer, which does not work with the actual implementation. As they are not mandatory, it feels ok to remove them with this test. What do you think?

If they're not supported we should remove them but the condition used to do that is a hack (you said it was part of my draft patch, I'm assuming I'd done that as a quick way of excluding those formats). I've added an entry to #625 so we don't forget to revisit this.

src/api.cpp

kpet · 2023-10-19T09:59:47Z

@rjodinchr In case you hadn't noticed: it looks like the 1DBufferImageFromSubBuffer test is hitting an assert.

rjodinchr · 2023-10-19T10:08:09Z

The tests are expected to fail one way or another as this PR is not updating clspv.

kpet · 2023-10-19T16:05:20Z

Ok, makes sense. Can you include the clspv update here (just the one commit we need please) so the CI passes? Thanks!

kpet

Almost there. I think we just need to resolve the one open thread about component mapping and update clspv (just the one commit this PR needs) so the CI can pass. Everything else looks nice!

kpet · 2023-10-19T16:14:47Z

src/api.cpp

+            if ((components_sampled != components_storage) &&
+                (image_type == CL_MEM_OBJECT_IMAGE1D_BUFFER)) {
+                continue;
+            }


yeah, if we pass the CTS without it, then we should remove it. If we somehow need this to pass the CTS, we should find a clearer expression of the conditions.

kpet · 2023-10-19T16:16:54Z

src/queue.hpp

@@ -957,6 +967,7 @@ struct cvk_command_buffer_image_copy final : public cvk_command_batchable {
    size_t m_offset;
    std::array<size_t, 3> m_origin;
    std::array<size_t, 3> m_region;
+    cl_command_type m_copy_type;


Reusing cl_command_type is a bit confusing but it'll do for a first implementation.

kpet · 2023-10-23T14:00:10Z

@rjodinchr I've had a look at the CI failures. The newly added test creates a sub-buffer at an offset that does not satisfy Vulkan constraints. clvk should have rejected the creation of the sub-buffer and texel buffer alignment constraints should be taken into account when calculating CL_DEVICE_MEM_BASE_ADDR_ALIGN. I suggest we don't fix sub-buffer error handling in this PR. In the interest of getting forward progress, you could just modify the offsets used in the test so that it works in the CI or query CL_DEVICE_MEM_BASE_ADDR_ALIGN, use it to pick the sub-buffer offset and change its implementation to take storageTexelBufferOffsetAlignmentBytes, storageTexelBufferOffsetSingleTexelAlignment, uniformTexelBufferOffsetAlignmentBytes, and uniformTexelBufferOffsetSingleTexelAlignment into account.

kpet

It seems swiftshader does not support VK_FORMAT_FEATURE_2_STORAGE_WRITE_WITHOUT_FORMAT_BIT_KHR on any format for texel buffers so you're going to have to disable the test for swiftshader.

kpet · 2023-10-23T15:31:41Z

src/memory.cpp

    VkFormatFeatureFlags format_feature_flags_WO;
-    format_feature_flags_WO = VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT;
+    if (type == CL_MEM_OBJECT_IMAGE1D_BUFFER) {
+        format_feature_flags_WO = VK_FORMAT_FEATURE_STORAGE_TEXEL_BUFFER_BIT;


Looking at the latest CI failures, it seems we also need VK_FORMAT_FEATURE_2_STORAGE_WRITE_WITHOUT_FORMAT_BIT_KHR which apparently is only available as a "2" format feature so we're dragging #497 as a dependency.

kpet · 2023-10-29T09:27:54Z

Merging this as is to unblock clspv updates. I will fix the CI with a separate commit on top.

kpet and others added 2 commits October 13, 2023 08:21

WIP texel buffers

c849f79

Change-Id: I4eef91af393b5487bbc208bdc96d43daa4c5c462

add support for image1d_buffer

1744028

remove 1DBUFFER mention in init_vulkan_image limit CL_DEVICE_IMAGE_MAX_BUFFER_SIZE WITH CL_DEVICE_MAX_MEM_ALLOC_SIZE

rjodinchr force-pushed the dev/1dbuffer branch from f5af614 to 1744028 Compare October 13, 2023 06:23

kpet requested changes Oct 15, 2023

View reviewed changes

kpet mentioned this pull request Oct 15, 2023

fix support for read of 3D images with unnormalised sampler #614

Merged

rjodinchr added 2 commits October 16, 2023 16:37

kpet feedback

14766e4

fix formatting

923bd44

rjodinchr requested a review from kpet October 16, 2023 14:45

update clspv

7feed5f

kpet reviewed Oct 19, 2023

View reviewed changes

use bigger padding in test

48be295

kpet reviewed Oct 23, 2023

View reviewed changes

kpet mentioned this pull request Oct 24, 2023

Tidy up support for CL_MEM_OBJECT_IMAGE1D_BUFFER #625

Open

5 tasks

kpet merged commit a25eb53 into kpet:main Oct 29, 2023
3 of 9 checks passed

rjodinchr deleted the dev/1dbuffer branch October 29, 2023 12:51

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

add support for image1d_buffer #609

add support for image1d_buffer #609

rjodinchr commented Sep 12, 2023

rjodinchr commented Oct 13, 2023

kpet left a comment

kpet Oct 15, 2023

rjodinchr Oct 16, 2023

kpet Oct 19, 2023

rjodinchr Oct 24, 2023

kpet Oct 29, 2023

kpet commented Oct 19, 2023

rjodinchr commented Oct 19, 2023

kpet commented Oct 19, 2023

kpet left a comment

kpet Oct 19, 2023

kpet Oct 19, 2023

kpet commented Oct 23, 2023

kpet left a comment

kpet Oct 23, 2023

kpet commented Oct 29, 2023

add support for image1d_buffer #609

add support for image1d_buffer #609

Conversation

rjodinchr commented Sep 12, 2023

rjodinchr commented Oct 13, 2023

kpet left a 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

kpet commented Oct 19, 2023

rjodinchr commented Oct 19, 2023

kpet commented Oct 19, 2023

kpet left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kpet commented Oct 23, 2023

kpet left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kpet commented Oct 29, 2023