Why the evis verson of pre-process-rgb fails to register

[gdh1995]

I want to use the PRE_PROCESS node to convert nhwc-uint8-from-cv_imread into nchw-uint8-quant_per_tensor, but with the prebuilt SDK of 6.4.10.2, an A311D board reports such an error:

Kernel "com.vivantecorp.extension.evis.pre_process_rgb_copy_U8toU8" does not exist
E [TIM-VX/src/tim/vx/internal/src/kernel/vsi_nn_kernel.c:_kernel_init_obj:757]Add parameter 0 to kernel com.vivantecorp.extension.evis.pre_process_rgb_copy_U8toU8 fail. with -12.
E [TIM-VX/src/tim/vx/internal/src/kernel/vsi_nn_kernel.c:_kernel_init_obj:770]Finalize kernel com.vivantecorp.extension.evis.pre_process_rgb_copy_U8toU8 fail with -12.
E [TIM-VX/src/tim/vx/internal/src/kernel/vsi_nn_kernel.c:_kernel_init_obj:775]Remove kernel com.vivantecorp.extension.evis.pre_process_rgb_copy_U8toU8 fail with -10.
E [TIM-VX/src/tim/vx/internal/src/kernel/vsi_nn_kernel.c:vsi_nn_kernel_create_node:873]Register client kernel com.vivantecorp.extension.evis.pre_process_rgb_copy_U8toU8 fail with -10.

I've debugged the code, but it seems obj = vxAddKernelInProgram(...) has contains some errors - an inserted vxGetStatus((vx_reference)obj) returns -1, although obj is not NULL.

The CPU version named com.vivantecorp.extension.cpu.pre_process_rgb_sw can be created by vxAddUserKernel without any error.

So I wonder why the evis version breaks, and how I can work around it?

[sunshinemyson]

@gdh1995 ，

Thanks for report this. Can you share your patch with a PR? We'd like try it locally. Thanks

[gdh1995]

Um, I didn't make any meaningful change. All the logic keeps the same as commit f8741b4 .

My test code looks like:

// src/tim/vx/internal/src/kernel/vsi_nn_kernel.c:523
static vsi_status _gpu_register
{
    // ...
    obj = vxAddKernelInProgram(
        program,
        info->name,
        info->enumeration,
        info->numParams,
        info->validate,
        info->initialize,
        info->deinitialize
        );
    // region: I only added these 5 lines
    status = vxGetStatus( (vx_reference)obj );
    if (VSI_SUCCESS != status)
    {
        VSILOGE( "vxAddKernelInProgram %s fail: %d.", info->name, (int) status);
    }
    // endregion
    if( obj )
    {
        status = _kernel_init_obj( info, obj );
        // here status is -10.
        //vxReleaseKernel( &obj );
    }
}

[sherlock9991]

Um, I didn't make any meaningful change. All the logic keeps the same as commit f8741b4 .

My test code looks like:

// src/tim/vx/internal/src/kernel/vsi_nn_kernel.c:523
static vsi_status _gpu_register
{
    // ...
    obj = vxAddKernelInProgram(
        program,
        info->name,
        info->enumeration,
        info->numParams,
        info->validate,
        info->initialize,
        info->deinitialize
        );
    // region: I only added these 5 lines
    status = vxGetStatus( (vx_reference)obj );
    if (VSI_SUCCESS != status)
    {
        VSILOGE( "vxAddKernelInProgram %s fail: %d.", info->name, (int) status);
    }
    // endregion
    if( obj )
    {
        status = _kernel_init_obj( info, obj );
        // here status is -10.
        //vxReleaseKernel( &obj );
    }
}

Hi,

It looks like the kernel is invalid, please the kernel info

[gdh1995]

The original code without vxGetStatus( (vx_reference)obj ); still stops in the below _kernel_init_obj(info, obj), in which vxAddParameterToKernel returns -12, and then vxRemoveKernel returns -10.

[sherlock9991]

vxAddKernelInProgram returns failure, please check the parameters info and program.