akaashrp opened a new pull request, #20410: URL: https://github.com/apache/tvm/pull/20410
#19900 moved the merge path axis of the GPU sort merge passes to `blockIdx.z` on every target. The WebGPU backend reserves `blockIdx.z` to extend `blockIdx.x` beyond 65535, so sort, argsort, and topk fail to compile for WebGPU whenever the input needs a merge pass. On WebGPU, fold the merge path and section axes into `blockIdx.x` and keep the batch on `blockIdx.y`; other targets keep the #19900 mapping. This also avoids the `blockIdx.y` overflow of the mapping before #19900, which produced wrong results for large batches (for example, a batch of 128 at size 151936). -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
