Bug in input attachment validation

[ShabbyX]

A recent roll into ANGLE is exposing a VVL bug that says:

[ VUID-vkCmdDraw-None-02686 ] Validation Error: [ VUID-vkCmdDraw-None-02686 ] Object 0: handle = 0x555ebb092460, type = VK_OBJECT_TYPE_COMMAND_BUFFER; Object 1: handle = 0x6a000000006a, type = VK_OBJECT_TYPE_RENDER_PASS; | MessageID = 0x242cb6b8 | vkCmdDraw(): Fragment Shader's input attachment index #0 doesn't exist or VK_ATTACHMENT_UNUSED in VkRenderPass 0x6a000000006a[], subpass #0. The Vulkan spec states: Every input attachment used by the current subpass must be bound to the pipeline via a descriptor set (https://www.khronos.org/registry/vulkan/specs/1.2-extensions/html/vkspec.html#VUID-vkCmdDraw-None-02686)

To repro, apply this change to ANGLE, build and run:

$ ./angle_end2end_tests --gtest_filter=MultisampledRenderToTextureTest.2DColorAttachmentMultisampleDrawTest/ES3_Vulkan

In the above test, there's a single color attachment / resolve attachment pair. In the second render pass, there are two subpasses. The first of the two subpasses uses the resolve attachment as input attachment and renders into the color attachment. Subpass' pInputAttachments is not VK_ATTACHMENT_UNUSED, and the draw call binds the input attachment at set 0 binding 0 correctly. I verified these with RenderDoc.

[snowman791]

I have a similar problem:

Validation Error: [ VUID-vkCmdDraw-None-02686 ] Object 0: handle = 0x26956baba10, name = CommandBuffer: 0, type = VK_OBJECT_TYPE_COMMAND_BUFFER; Object 1: handle = 0x9c93ec00000001c2, name = Render Pass: result, type = VK_OBJECT_TYPE_RENDER_PASS; | MessageID = 0x242cb6b8 | vkCmdDraw(): VkRenderPass 0x9c93ec00000001c2[Render Pass: result], subpass #0, input attachment #1 is not used in fragment shader. Every input attachment used by the current subpass must be bound to the pipeline via a descriptor set

in the fragment shader the input is defined as array if subpassInputs, and not separated:

layout (input_attachment_index = 0, set = 0, binding = 0) uniform subpassInput inputColor[3];

[jonahryandavis]

This issue is still happening even with the above fix.
Examples documented here: https://bugs.chromium.org/p/angleproject/issues/detail?id=5228

[ShabbyX]

Wrong bug Jonah! This comment should have been on #2148.

[locke-lunarg]

@jonahryandavis @ShabbyX

I'm confused. Why is the error message is VkNonDispatchableHandle? It should have been VkPipelineLayout.

[ShabbyX]

@locke-lunarg it may be a false alarm. I had a change that fixes a pipeline-binding bug which was probably the cause of this validation error.

[ShabbyX]

@locke-lunarg Confirmed that the warning was actually due to an ANGLE bug.