Command buffer submitted to wrong queue (Intel A770, Chapter 5)

[dorian-apanel-intel]

When running on Intel A770, following core validation error is triggered in Chapter5:

VUID-vkQueueSubmit2-commandBuffer-03878(ERROR / SPEC): msgNum: -628003580 - Validation Error: [ VUID-vkQueueSubmit2-commandBuffer-03878 ] Object 0: handle = 0x22e857a8750, type = VK_OBJECT_TYPE_COMMAND_BUFFER; Object 1: handle = 0x22e853d52c0, type = VK_OBJECT_TYPE_QUEUE; | MessageID = 0xda916d04 | vkQueueSubmit2(): pSubmits[0].pCommandBufferInfos[0].commandBuffer Primary VkCommandBuffer 0x22e857a8750[] created in queue family 0 is being submitted on VkQueue 0x22e853d52c0[] from queue family 1. The Vulkan spec states: The commandBuffer member of any element of the pCommandBufferInfos member of any element of pSubmits must have been allocated from a VkCommandPool that was created for the same queue family queue belongs to (https://vulkan.lunarg.com/doc/view/1.3.239.0/windows/1.3-extensions/vkspec.html#VUID-vkQueueSubmit2-commandBuffer-03878)
    Objects: 2
       [0]  0x22e857a8750, type: 6, name: NULL
       [1]  0x22e853d52c0, type: 4, name: NULL
A breakpoint instruction (__debugbreak() statement or a similar call) was executed in RaptorEngine_Chapter5_Debug.exe.

Command buffer created for universal queue is submitted to compute only queue.

Log regarding queues:

GPU Used: Intel(R) Arc(TM) A770M Graphics
Family 0, flags 15 queue count 1
Family 1, flags 2 queue count 1
Family 2, flags 4 queue count 1

[dorian-apanel-intel]

Issue can be reproduced on NVidia by commenting those two lines (pretending that single universal queue is advertised):

        // Search for main queue that should be able to do all work (graphics, compute and transfer)
        if ( (queue_family.queueFlags & ( VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT )) == ( VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT ) ) {
            main_queue_family_index = fi;

            if ( queue_family.queueCount > 1 ) {
                // compute_queue_family_index = fi;
                // compute_queue_index = 1;
            }

            continue;
        }

[dorian-apanel-intel]

After grepping the whole solution, I do not see vulkan_compute_queue_family being used anywhere, Pool of vkCommandPools are created either for universal, or for transfer queues.

[theWatchmen]

Hi Dorian, many thanks for reporting this and a few other issues. Thank you also for the PRs. We'll take a look soon, we are in the middle of preparing for a conference :)

[theWatchmen]

I "merged" your changes. If you're happy with them, could you close this issue?

[dorian-apanel-intel]

Thanks for merging. Verified it with 47e8c21 on top.

Side note: On GPU's with no compute queue, there is crash when creating device. The same is when you have two GPUs, and one without compute queue is picked. (Family index passed is 0xFFFFFFFF)

[theWatchmen]

Thanks, I'll have a look at this as well. I thought we had checks in place for that scenario, but clearly they don't cover all cases.