Testing: Dump DX12 resources for drawcall #1356

locke-lunarg · 2023-11-15T22:46:25Z

Test sample: D3D12HelloTexture. It's for renderpass test. I added renderpass, constant buffer, more triangles, and test code about copy resource and separated renderpass.
https://github.com/locke-lunarg/DirectX-Graphics-Samples/tree/test-dump

For not renderpass case, now, I tested HelloWorld project in DirectX-Graphics-Samples. I'm doing more tests.

Add option: --dump-resources drawcall-n . n is the index of drawcall.
To output readable data. Set TEST_READABLE = true in dx12_dump_resources.cpp.
Float for vertex, index, constant buffer.
Image for shader resource, render target.

The basic concept:
Not renderpass: Add copy resource command in before and after drawcall.
E.g.:
Original

cmd->Reset();
...
cmd->Draw();
...
cmd->Draw();  -> Target
...
cmd->End();

New

cmd->Reset();
...
cmd->Draw();
...
cmd->CopyResource(); -> Add
cmd->draw();  -> Target
cmd->CopyResource(); -> Add
...
cmd->End();

Renderpass: Add renderpasses in before and after drawcall.
E.g.:
Original

cmd->Reset();
...
cmd->BeginRenderPass();
...
cmd->Draw();
...
cmd->Draw();  -> Target
...
cmd->EndRenderPass();
...
cmd->End();

New

cmd->Reset();
...
cmd->BeginRenderPass();
...
cmd->Draw();
...
cmd->EndRenderPass();    -> Add
cmd->CopyResource();     -> Add
cmd->BeginRenderPass();  -> Add
cmd->Draw();  -> Target
cmd->EndRenderPass();    -> Add
cmd->CopyResource();     -> Add
cmd->BeginRenderPass();  -> Add
...
cmd->EndRenderPass();
...
cmd->End();

According to my test, it is ok to insert the renderpass in anyplace in the original renderpass. And then, the EndingAcccess of the renderpass and the BeginningAccess of following renderpasses have to be PRESERVE to keep the the rendering process for this following renderpass, or it will clear the render target.

ci-tester-lunarg · 2023-11-15T22:46:28Z

CI gfxreconstruct build queued with queue ID 84259.

ci-tester-lunarg · 2023-11-15T23:10:17Z

CI gfxreconstruct build queued with queue ID 84294.

ci-tester-lunarg · 2023-11-15T23:11:36Z

CI gfxreconstruct build queued with queue ID 84295.

ci-tester-lunarg · 2023-11-15T23:25:28Z

CI gfxreconstruct build queued with queue ID 84306.

ci-tester-lunarg · 2023-11-15T23:41:03Z

CI gfxreconstruct build queued with queue ID 84312.

ci-tester-lunarg · 2023-11-16T00:49:11Z

CI gfxreconstruct build queued with queue ID 84372.

ci-tester-lunarg · 2023-11-16T00:51:23Z

CI gfxreconstruct build queued with queue ID 84374.

ci-tester-lunarg · 2023-11-16T00:55:42Z

CI gfxreconstruct build queued with queue ID 84376.

ci-tester-lunarg · 2023-11-16T03:11:18Z

CI gfxreconstruct build queued with queue ID 84487.

ci-tester-lunarg · 2023-11-16T06:17:44Z

CI gfxreconstruct build queued with queue ID 84640.

locke-lunarg · 2023-11-16T06:40:55Z

framework/graphics/dx12_dump_resources.h

+
+// TODO: This class copys a lot of code to write json from VulkanExportJsonConsumerBase.
+//       We might need a class for writing json.
+class Dx12DumpResources


Dx12DumpResources output json and resources. A lot of json code are copied from VulkanExportJsonConsumerBase. We should have a common class for them.

locke-lunarg · 2023-11-16T06:42:39Z

framework/decode/dx12_browse_consumer.h

+    }
+};
+
+class Dx12BrowseConsumer : public Dx12Consumer


Dx12BrowseConsumer browses the file before replay to find the target drawcall code index and some key code indices, and track resource infos. That could help replay easier. But it doesn't track render target infos because the useful info is replay cpuDescriptor. It's only available in replay.

locke-lunarg · 2023-11-16T06:43:00Z

framework/decode/custom_dx12_replay_commands.h

+template <format::ApiCallId Id>
+struct CustomReplayPreCall
+{
+    template <typename... Args>
+    static void Dispatch(Dx12ReplayConsumerBase*, Args...)
+    {}
+};
+
+template <format::ApiCallId Id>
+struct CustomReplayPostCall
+{
+    template <typename... Args>
+    static void Dispatch(Dx12ReplayConsumerBase*, Args...)
+    {}
+};


Add CustomReplayPreCall and CustomReplayPostCall in every function call to track resources and add copy resource cmd.

locke-lunarg · 2023-11-16T06:44:25Z

framework/decode/dx12_object_info.h

@@ -303,7 +310,8 @@ struct D3D12ResourceInfo : DxObjectExtraInfo
    std::map<uint64_t, uint64_t>                       mapped_gpu_addresses;
    std::map<uint64_t, graphics::Dx12ShaderIdentifier> mapped_shader_ids;

-    D3D12_RESOURCE_DESC1 desc = {};
+    D3D12_RESOURCE_DESC1  desc = {};
+    D3D12_RESOURCE_STATES current_state{ D3D12_RESOURCE_STATE_COMMON };


It needs to track the current state of resources for change state for copy resources.

locke-lunarg · 2023-11-16T06:45:50Z

framework/decode/dx12_replay_consumer_base.cpp

+        (modified_initial_resource_state & D3D12_RESOURCE_STATE_ALL_SHADER_RESOURCE))
+    {
+        // shader resources can't change state after SetGraphicsRootDescriptorTable.
+        // But even if it changes state and copy resources before SetGraphicsRootDescriptorTable,
+        // it can't solve it, because it also has to copy resources after drawcall.
+        // In this case, adding a COPY_SOURCE state here, so it doesn't need to change state for copy resources.
+        modified_initial_resource_state |= D3D12_RESOURCE_STATE_COPY_SOURCE;


shader resources can't change state after SetGraphicsRootDescriptorTable. In this case, adding a COPY_SOURCE state for shader resource, so it doesn't need to change state for copy resources.

locke-lunarg · 2023-11-16T06:46:18Z

framework/decode/dx12_replay_consumer_base.cpp

+            if (barriers[i].Transition->decoded_value->StateAfter & D3D12_RESOURCE_STATE_ALL_SHADER_RESOURCE)
+            {
+                // shader resources can't change state after SetGraphicsRootDescriptorTable.
+                // Adding a COPY_SOURCE state here, so it doesn't need to change state for copy resources.
+                barriers[i].Transition->decoded_value->StateAfter |= D3D12_RESOURCE_STATE_COPY_SOURCE;


shader resources can't change state after SetGraphicsRootDescriptorTable. In this case, adding a COPY_SOURCE state for shader resource, so it doesn't need to change state for copy resources.

locke-lunarg · 2023-11-16T06:46:59Z

framework/decode/dx12_replay_consumer_base.cpp

+        hr = graphics::dx12::WaitForQueue(queue, fence, UINT64_MAX);
+        GFXRECON_ASSERT(SUCCEEDED(hr));
+
+        dump_resources_->WriteResources(track_dump_resources_);


After ExecuteCommandLists and WaitForQueue, output the resources.

locke-lunarg · 2023-11-16T06:47:38Z

framework/decode/dx12_replay_consumer_base.cpp

+        if (track_dump_resources_.target.begin_renderpass_code_index == 0)
+        {
+            // The program doesn't use renderpass.
+            AddCopyResourceCommandsForAfterDrawcall(commandlist);


Here add copy resources or renderpass command for after drawcall

After adding copy resource in after drawcall, it still has to run the original ExecuteCommandLists, not a new ExecuteCommandLists because some FIllMemoryCommand runs before the original ExecuteCommandLists. If it runs a new ExecuteCommandLists, it might miss FIllMemoryCommand.

locke-lunarg · 2023-11-16T06:50:42Z

framework/decode/dx12_replay_consumer_base.cpp

+        for (uint32_t i = 0; i < NumRenderTargets; ++i)
+        {
+            rt_descs[i].EndingAccess->decoded_value->Type = D3D12_RENDER_PASS_ENDING_ACCESS_TYPE_PRESERVE;
+        }
+
+        auto ds_desc = pDepthStencil->GetMetaStructPointer();
+        if (ds_desc)
+        {
+            ds_desc->DepthEndingAccess->decoded_value->Type   = D3D12_RENDER_PASS_ENDING_ACCESS_TYPE_PRESERVE;
+            ds_desc->StencilEndingAccess->decoded_value->Type = D3D12_RENDER_PASS_ENDING_ACCESS_TYPE_PRESERVE;


In BeginRednerPass, the EndingAcccess have to be PRESERVE to keep the the render process for this following inserted renderpass, or it will clear the render target.

ci-tester-lunarg · 2023-11-16T06:59:40Z

CI gfxreconstruct build queued with queue ID 84671.

ci-tester-lunarg · 2023-11-16T07:02:30Z

CI gfxreconstruct build queued with queue ID 84673.

locke-lunarg · 2023-11-16T07:04:41Z

framework/decode/dx12_replay_consumer_base.cpp

+            for (uint32_t i = 0; i < rt_descs_size; ++i)
+            {
+                rt_descs[i].cpuDescriptor        = track_dump_resources_.replay_render_target_handles[i];
+                rt_descs[i].BeginningAccess.Type = D3D12_RENDER_PASS_BEGINNING_ACCESS_TYPE_PRESERVE;
+                rt_descs[i].EndingAccess.Type    = D3D12_RENDER_PASS_ENDING_ACCESS_TYPE_PRESERVE;
+            }
+
+            D3D12_RENDER_PASS_DEPTH_STENCIL_DESC* p_ds_desc = nullptr;
+            D3D12_RENDER_PASS_DEPTH_STENCIL_DESC  ds_desc{};
+            if (track_dump_resources_.depth_stencil_heap_id != 0)
+            {
+                p_ds_desc                           = &ds_desc;
+                ds_desc.cpuDescriptor               = track_dump_resources_.replay_depth_stencil_handle;
+                ds_desc.DepthBeginningAccess.Type   = D3D12_RENDER_PASS_BEGINNING_ACCESS_TYPE_PRESERVE;
+                ds_desc.StencilBeginningAccess.Type = D3D12_RENDER_PASS_BEGINNING_ACCESS_TYPE_PRESERVE;
+                ds_desc.DepthEndingAccess.Type      = D3D12_RENDER_PASS_ENDING_ACCESS_TYPE_PRESERVE;
+                ds_desc.StencilEndingAccess.Type    = D3D12_RENDER_PASS_ENDING_ACCESS_TYPE_PRESERVE;


Here is to insert renderpass in before drawcall. The BeginningAccess an EndingAcccess of the renderpass have to be PRESERVE to keep the the render process from the previous renderpass and for this following inserted renderpass, or it will clear the render target.

ci-tester-lunarg · 2024-01-15T23:50:01Z

CI gfxreconstruct build # 3665 running.

framework/decode/dx12_replay_consumer_base.cpp

ci-tester-lunarg · 2024-01-16T01:33:32Z

CI gfxreconstruct build # 3665 failed.

ci-tester-lunarg · 2024-01-16T19:49:00Z

CI gfxreconstruct build queued with queue ID 117619.

ci-tester-lunarg · 2024-01-16T21:06:40Z

CI gfxreconstruct build # 3673 running.

It shouldn't change state in ResourceBarrier. It should track the AfterState in ResourceBarrier, and then change it in ExecuteCommandList. But the state could be wrong for CopyResource, if ResourceBarrier is in another commandlist, and the commandlist is executed after the target drawcall, we couldn't know the correct state.

1. To output readable data. Set "TEST_READABLE = true" in dx12_dump_resources.cpp. This feature is only for test because data type could be various. But here uses fixed type. float for vertex, index, constant buffer. image for shader resource, render target. 2. To skip dump shader resources, Set "TEST_SHADER_RES = false" in dx12_replay_consumer_base.cpp. Now, dump shader resources could cost too much memory to cause crash.

ci-tester-lunarg · 2024-01-16T21:38:05Z

CI gfxreconstruct build queued with queue ID 117698.

ci-tester-lunarg · 2024-01-16T22:50:41Z

CI gfxreconstruct build # 3674 running.

ci-tester-lunarg · 2024-01-16T23:22:37Z

CI gfxreconstruct build queued with queue ID 117731.

ci-tester-lunarg · 2024-01-17T00:34:31Z

CI gfxreconstruct build # 3675 running.

ci-tester-lunarg · 2024-01-17T02:18:02Z

CI gfxreconstruct build # 3675 passed.

ci-tester-lunarg · 2024-01-17T21:27:58Z

CI gfxreconstruct build queued with queue ID 118212.

ci-tester-lunarg · 2024-01-17T21:30:01Z

CI gfxreconstruct build # 3676 running.

ci-tester-lunarg · 2024-01-17T23:12:31Z

CI gfxreconstruct build # 3676 passed.

Keep track of pending resource state changes in command list info, then apply the pending state changes in execute command lists.

Use a map to track descriptors so that when a new descriptor is created in the same location as an existing descriptor, the existing descriptor is removed.

ci-tester-lunarg · 2024-01-19T20:57:13Z

CI gfxreconstruct build queued with queue ID 119197.

ci-tester-lunarg · 2024-01-19T21:00:01Z

CI gfxreconstruct build # 3682 running.

ci-tester-lunarg · 2024-01-19T22:42:52Z

CI gfxreconstruct build # 3682 passed.

ci-tester-lunarg · 2024-02-08T18:41:34Z

CI gfxreconstruct-extended build queued with queue ID 130708.

ci-tester-lunarg · 2024-02-09T01:06:14Z

CI gfxreconstruct-extended build # 315 running.

ci-tester-lunarg · 2024-02-09T16:57:36Z

CI gfxreconstruct-extended build # 315 passed.

locke-lunarg force-pushed the dx-dump branch from 6100b35 to bbf556e Compare November 15, 2023 23:10

locke-lunarg force-pushed the dx-dump branch from bbf556e to e532782 Compare November 15, 2023 23:11

locke-lunarg force-pushed the dx-dump branch from e532782 to 6482ca8 Compare November 15, 2023 23:25

locke-lunarg force-pushed the dx-dump branch from 6482ca8 to 4c76ce1 Compare November 15, 2023 23:41

locke-lunarg force-pushed the dx-dump branch from 4c76ce1 to 21e2836 Compare November 16, 2023 00:49

locke-lunarg force-pushed the dx-dump branch from 21e2836 to 1280f13 Compare November 16, 2023 00:51

locke-lunarg force-pushed the dx-dump branch from 1280f13 to 616acaa Compare November 16, 2023 00:55

locke-lunarg force-pushed the dx-dump branch from 616acaa to 8ef173a Compare November 16, 2023 03:11

locke-lunarg force-pushed the dx-dump branch from 8ef173a to eafb50a Compare November 16, 2023 06:17

locke-lunarg commented Nov 16, 2023

View reviewed changes

locke-lunarg force-pushed the dx-dump branch from eafb50a to c802431 Compare November 16, 2023 06:59

locke-lunarg force-pushed the dx-dump branch from c802431 to c61b4b3 Compare November 16, 2023 07:02

locke-lunarg commented Nov 16, 2023

View reviewed changes

locke-lunarg marked this pull request as ready for review November 16, 2023 07:13

locke-lunarg requested review from bradgrantham-lunarg, davidd-lunarg and davidlunarg November 16, 2023 07:13

davidd-lunarg reviewed Jan 16, 2024

View reviewed changes

framework/decode/dx12_replay_consumer_base.cpp Outdated Show resolved Hide resolved

locke-lunarg force-pushed the dx-dump branch from af91d9d to a41a388 Compare January 16, 2024 19:48

locke-lunarg added 2 commits January 16, 2024 14:37

locke-lunarg force-pushed the dx-dump branch from a41a388 to 70eaec4 Compare January 16, 2024 21:38

Support RTsSingleHandleToDescriptorRange

1db6462

Do not add D3D12_RESOURCE_STATE_COPY_SOURCE

304f743

davidd-lunarg added 3 commits January 19, 2024 10:45

Improve resource state tracking for dump resources

fe12f46

Keep track of pending resource state changes in command list info, then apply the pending state changes in execute command lists.

Improve descriptor tracking for dump resources

ad44077

Use a map to track descriptors so that when a new descriptor is created in the same location as an existing descriptor, the existing descriptor is removed.

Verify resource bytes per pixel before writing BMP

5835eec

locke-lunarg marked this pull request as draft February 16, 2024 20:57

locke-lunarg closed this Mar 4, 2024

Testing: Dump DX12 resources for drawcall #1356

Testing: Dump DX12 resources for drawcall #1356

Conversation

locke-lunarg commented Nov 15, 2023 • edited Loading

ci-tester-lunarg commented Nov 15, 2023

ci-tester-lunarg commented Nov 15, 2023

ci-tester-lunarg commented Nov 15, 2023

ci-tester-lunarg commented Nov 15, 2023

ci-tester-lunarg commented Nov 15, 2023

ci-tester-lunarg commented Nov 16, 2023

ci-tester-lunarg commented Nov 16, 2023

ci-tester-lunarg commented Nov 16, 2023

ci-tester-lunarg commented Nov 16, 2023

ci-tester-lunarg commented Nov 16, 2023

locke-lunarg Nov 16, 2023

Choose a reason for hiding this comment

locke-lunarg Nov 16, 2023

Choose a reason for hiding this comment

locke-lunarg Nov 16, 2023

Choose a reason for hiding this comment

locke-lunarg Nov 16, 2023

Choose a reason for hiding this comment

locke-lunarg Nov 16, 2023

Choose a reason for hiding this comment

locke-lunarg Nov 16, 2023

Choose a reason for hiding this comment

locke-lunarg Nov 16, 2023

Choose a reason for hiding this comment

locke-lunarg Nov 16, 2023 • edited Loading

Choose a reason for hiding this comment

locke-lunarg Nov 16, 2023 • edited Loading

Choose a reason for hiding this comment

ci-tester-lunarg commented Nov 16, 2023

ci-tester-lunarg commented Nov 16, 2023

locke-lunarg Nov 16, 2023 • edited Loading

Choose a reason for hiding this comment

ci-tester-lunarg commented Jan 15, 2024

ci-tester-lunarg commented Jan 16, 2024

ci-tester-lunarg commented Jan 16, 2024

ci-tester-lunarg commented Jan 16, 2024

ci-tester-lunarg commented Jan 16, 2024

ci-tester-lunarg commented Jan 16, 2024

ci-tester-lunarg commented Jan 16, 2024

ci-tester-lunarg commented Jan 17, 2024

ci-tester-lunarg commented Jan 17, 2024

ci-tester-lunarg commented Jan 17, 2024

ci-tester-lunarg commented Jan 17, 2024

ci-tester-lunarg commented Jan 17, 2024

ci-tester-lunarg commented Jan 19, 2024

ci-tester-lunarg commented Jan 19, 2024

ci-tester-lunarg commented Jan 19, 2024

ci-tester-lunarg commented Feb 8, 2024

ci-tester-lunarg commented Feb 9, 2024

ci-tester-lunarg commented Feb 9, 2024

locke-lunarg commented Nov 15, 2023 •

edited

Loading

locke-lunarg Nov 16, 2023 •

edited

Loading

locke-lunarg Nov 16, 2023 •

edited

Loading

locke-lunarg Nov 16, 2023 •

edited

Loading