khronos_validation.printf_to_stdout seems not work

[Inori]

khronos_validation.printf_to_stdout seems not work on my machine.

The validation messages shows that I'm using the correct settings file and other options works right (like khronos_validation.enables).

Only the printf_to_stdout seems doesn't make sense.
Whether I set it to true or false, I got debug messages from both debug callback and stdout.

Vulkan SDK version: 1.3.211.0

printf_to_stdout = true:

printf_to_stdout = false:

I've already commented out std::cerr line in my debug callback (and yes I've registered it and it is getting called), I should expect no output, right?

Originally posted by @Inori in #2464 (comment)

[Inori]

OK, I figured it out.
I should remove 'VK_DBG_LAYER_ACTION_LOG_MSG' from khronos_validation.debug_action and this way I can control the stdout.
But a new problem occurs, when I removed VK_DBG_LAYER_ACTION_LOG_MSG, debug printf seems only work when printf_to_stdout is true.
When I set it to false, I got message saying buffer size that was too small, even I set printf_buffer_size = 65535

In short, printf_to_stdout=false only works when VK_DBG_LAYER_ACTION_LOG_MSG is set.

[TonyBarbour]

I'm not quite sure what you are seeing.

With everything set to default, you should see debug printf messages, among other messages, going to the Debug Utils callback. This looks like your second screenshot above - notice that the extra debug information when the printfs go through the callback.

With everything set to default, except for khronos_validation.printf_to_stdout = true you should see all but the debug printf messages going to the callback, and debug printf messages going to stdout This looks like your first screenshot where only the debug printf text is sent and no debug info is attached.

You shouldn't need to change khronos_validation.debug_action from it's default to see debug printf messages on stdout or in the callback.

I do see that some of your shaders are failing validation. Debug printf will not instrument shaders that fail validation, and you won't get any debug printf messages from them when they run.

The value of printf_to_stdout should have no effect on the buffer size required to receive debug printf messages.

Aside from validation messages being printed from your callback with your cerr line commented out, which I can't explain, it looks like everything is working as expected.

[Inori]

Ok, I see, thanks for your reply.

[Inori]

@Tony-LunarG Hi Tony, sorry to disturb you again, I'm still suffering on similar issues.
I'm trying to print something from a compute shader, which local size is (10 x 10 x 1),
I got 2 problems:

1. the validation layer still complains buffer size not enough, it seems the expect size is extremely big for this shader, and it crashes randomly on the memset at function end:
   

2. I can't print more than one debug string, even if I emit 2 or 3 debug printf spir-v instructions, only the first one can be printed.

The spir-v bytecode is generated by my custom compiler, not compiled from glsl code, yet it can pass spirv-val check, for sure.
But I've compared the generated code, I didn't see any difference on debug printf instruction:

       OpExtension "SPV_KHR_non_semantic_info"
       %2249 = OpExtInstImport "NonSemantic.DebugPrintf"
       ...
       %2251 = OpString "exec %X
"
       %2253 = OpString "asuka %X
"
       ...
       %2250 = OpExtInst %void %2249 1 %2251 %2243
       %2252 = OpExtInst %void %2249 1 %2253 %2243

Can you give me some advice?

[TonyBarbour]

@Inori I'm on break this week but I should be able to more closely consider your first issue early next week.
One suggestion I have for now is that maybe you can limit the number of messages by conditionally using invocation ID to decide whether to print or not.
Also, you can increase memory available for debug printf using vkconfig or a layers settings file. The printf_buffer_size variable sets the size in bytes for each debug printf buffer (per draw / dispatch).

As for not receiving subsequent messages, there is only one debug printf buffer per draw, so if previous messages exhaust the available memory, subsequent messages will not appear - could that be what's happening?
Alternatively, are you sure that the subsequent messages are in a path executed by the shader invocation?

[Inori]

Limit message count using invocation ID together with a very big printf_buffer_size (over 65535) solves the problem.
Previously I though there must be something wrong, but it is really just not enough buffer size.
Thanks very much!