Not available on win10

[863689877]

Hello, I can successfully compile and install lietorch on Windows, but I cannot pass the gpu related tests. When running the simplest multiplication, likes X1 = Ts * X0, it will get stuck and exit.

[madnananwer]

I'm having the same issue. Did you find the solution?

[yclicc]

I'm still not sure why it should be, but I managed to fix this on Windows by making all the CUDA kernels in lietorch_gpu.cu accept regular pointers where currently they accept const pointers. Hopefully someone who knows more about CUDA and Windows can make a proper pull request.

[carlosedubarreto]

Hello @yclicc , I have almost knowlegde of this type of programming. Can you tell how to change the const pointer to regular ones, like you said?

Or could you share your lietorch_gpu.cu?

thanks

[yclicc]

@carlosedubarreto well you just get rid of the word "const" in the function arguments. So for example

lietorch/lietorch/src/lietorch_gpu.cu

Line 21 in 0fa9ce8

__global__ void exp_forward_kernel(const scalar_t* a_ptr, scalar_t* X_ptr, int num_threads) {

becomes __global__ void exp_forward_kernel(scalar_t* a_ptr, scalar_t* X_ptr, int num_threads) {

In practice if you want to maintain the old behaviour (which presumably has some benefit) on non-windows platforms you can add before the first template the following:

#ifdef _WIN32
    #define NON_WINDOWS_CONST
#else
    #define NON_WINDOWS_CONST const
#endif

and then replace all the consts in function parameter lists with NON_WINDOWS_CONST. Then on Windows it will remove the const at compile time but keep it in on other platforms. So for our example above you would end up with __global__ void exp_forward_kernel(NON_WINDOWS_CONST scalar_t* a_ptr, scalar_t* X_ptr, int num_threads) {

I hope that helps!

[carlosedubarreto]

@yclicc thanks a lot for the detailed info.

I gave a shot just by replacing the "const" from nothing na it didnt work, maybe I did something wrong, I'll try again.

thanks a lot!!!!

[yclicc]

Did you replace all the consts in every function in lietorch_gpu.cu?

…

On Thu, 1 Feb 2024, 5:43 pm Carlos Barreto, ***@***.***> wrote: @yclicc <https://github.com/yclicc> thanks a lot for the detailed info. I gave a shot just by replacing the "const" from nothing na it didnt work, maybe I did something wrong, I'll try again. thanks a lot!!!! — Reply to this email directly, view it on GitHub <#22 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACK24JZUFC5ELV2CUCLT6H3YRPH25AVCNFSM6AAAAAAQT6JKX6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSMRRHA3DSMZTG4> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[carlosedubarreto]

@yclicc after reading the old comment you did, I replaced the text "const " to "" and compiled it again.
I had no success. but it was a quick test.

Now I'm gong to do it all from the start with more attention.

[carlosedubarreto]

I tested again doing a clean install, and it seems to be working, thanks A LOT @yclicc !!!!! I'm sharing here the final file I changed.
lietorch_gpu.zip

[yclicc]

Awesome, glad it worked for you as well as for me! Now we just need to ask someone with more knowledge of windows and cuda why it works!

[carlosedubarreto]

Im actually more curious to know how could you guess that removing the "const" would make it work 😊

[yclicc]

Lots of trial and error, and the Nsight debugger

[carlosedubarreto]

Oh Nsight debugger. Now I know that serves for something 😀, thanks a lot for the info 🦾

[gha890826]

I tested again doing a clean install, and it seems to be working, thanks A LOT @yclicc !!!!! I'm sharing here the final file I changed. lietorch_gpu.zip

Thanks for your sharing!
I solve same problem by this solution.