Bug 108894

(In reply to Józef Kucia from comment #0)

> 
> I have a test case in the form of a vkd3d test:
> https://source.winehq.org/git/vkd3d.git/blob/HEAD:/tests/d3d12.c#l18279
> The test passes on RADV and Nvidia. Note that the current version of vkd3d
> doesn't issue a required pipeline barrier between vkCmdCopyBuffer() and
> vkCmdCopyQueryPoolResults().

Are you sure this is the test_query_timestamp() test?
It's passing for me and tracing the commands of that test, I can't see any vkCmdCopyBuffer() call.

It is test_resolve_non_issued_query_data().

Got it.

Yes, I think you need to put barriers because in Anv we're using the 3d pipeline to copy buffers, whereas timestamps are copied using the command streamer.
Those being 2 different pieces of hardware, there is no guarantee about which one is going to write to memory first/last.

Yes, but even with barriers I am not able to get the test passing on Anv.

(In reply to Józef Kucia from comment #4)
> Yes, but even with barriers I am not able to get the test passing on Anv.

I'm unfamiliar with d3d, what should I add to insert a barrier?

The issue is most likely because we accumulate barriers and delay actually applying the them until the next draw or copy command.  Should be an easy fix.

(In reply to Lionel Landwerlin from comment #5)
> I'm unfamiliar with d3d, what should I add to insert a barrier?

transition_resource_state(command_list, readback_buffer,                    
   D3D12_RESOURCE_STATE_COPY_DEST, D3D12_RESOURCE_STATE_COPY_DEST);    

produces the following Vulkan barrier:

vkCmdPipelineBarrier(commandBuffer, srcStageMask, dstStageMask, dependencyFlags, memoryBarrierCount, pMemoryBarriers, bufferMemoryBarrierCount, pBufferMemoryBarriers, imageMemoryBarrierCount, pImageMemoryBarriers) returns void:
    commandBuffer:                  VkCommandBuffer = 0x55a1a81b7430
    srcStageMask:                   VkPipelineStageFlags = 4096 (VK_PIPELINE_STAGE_TRANSFER_BIT)
    dstStageMask:                   VkPipelineStageFlags = 4096 (VK_PIPELINE_STAGE_TRANSFER_BIT)
    dependencyFlags:                VkDependencyFlags = 0
    memoryBarrierCount:             uint32_t = 0
    pMemoryBarriers:                const VkMemoryBarrier* = NULL
    bufferMemoryBarrierCount:       uint32_t = 1
    pBufferMemoryBarriers:          const VkBufferMemoryBarrier* = 0x7fffeabcdd00
        pBufferMemoryBarriers[0]:       const VkBufferMemoryBarrier = 0x7fffeabcdd00:
            sType:                          VkStructureType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER (44)
            pNext:                          const void* = NULL
            srcAccessMask:                  VkAccessFlags = 4096 (VK_ACCESS_TRANSFER_WRITE_BIT)
            dstAccessMask:                  VkAccessFlags = 4096 (VK_ACCESS_TRANSFER_WRITE_BIT)
            srcQueueFamilyIndex:            uint32_t = 4294967295
            dstQueueFamilyIndex:            uint32_t = 4294967295
            buffer:                         VkBuffer = 0x55a1a80f6fc0
            offset:                         VkDeviceSize = 0
            size:                           VkDeviceSize = 18446744073709551615
    imageMemoryBarrierCount:        uint32_t = 0
    pImageMemoryBarriers:           const VkImageMemoryBarrier* = NULL


with this change the test fails intermittently on Anv.

(In reply to Jason Ekstrand from comment #6)
> The issue is most likely because we accumulate barriers and delay actually
> applying the them until the next draw or copy command.  Should be an easy
> fix.

A call to genX(cmd_buffer_apply_pipe_flushes) seems to fix it for me.

> A call to genX(cmd_buffer_apply_pipe_flushes) seems to fix it for me.

I think we need to do a bit more than just that because MI writes aren't pipelined.  Probably something like:

/* Because MI writes aren't pipelined, we need a CS stall */
if (flushes & ANV_PIPE_FLUSH_BITS)
   flushes |= ANV_PIPE_CS_STALL_BIT;
genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);

Should be fixed in :

commit 37f9788e9a8e443772b5ad6f339567e6ae6a8320
Author: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Date:   Thu Nov 29 13:02:03 2018 +0000

    anv: flush pipeline before query result copies

*** Bug 108909 has been marked as a duplicate of this bug. ***