Error: zero size memory allocation when calling 'cuda_partition'

[hyeonjang]

Hello, I am a big fan of this lib.

During using CUDAArray, I encounter the error to allocate the memory with size zero.
With debugging, I found that It occurs from calling the function "cuda_partition"

enoki/src/cuda/horiz.cu

Lines 81 to 103 in 2a18afa

	size_t clamped_size = std::min(size, (size_t) 511);
	uint32_t *counts_h = (uint32_t *) cuda_host_malloc(sizeof(uint32_t) * (clamped_size + 1));
	void **ptrs_unique_h = (void **) cuda_host_malloc(sizeof(void *) * clamped_size);
	cuda_check(cudaMemcpyAsync(counts_h, counts, (clamped_size + 1) * sizeof(uint32_t), cudaMemcpyDeviceToHost));
	cuda_check(cudaMemcpyAsync(ptrs_unique_h, ptrs_unique, clamped_size * sizeof(void *), cudaMemcpyDeviceToHost));
	cuda_check(cudaDeviceSynchronize());
	size_t num_runs_h = (size_t) counts_h[0];
	if (num_runs_h > clamped_size) {
	cuda_host_free(counts_h);
	cuda_host_free(ptrs_unique_h);
	counts_h = (uint32_t *) cuda_host_malloc(sizeof(uint32_t) * (num_runs_h + 1));
	ptrs_unique_h = (void **) cuda_host_malloc(sizeof(void *) * num_runs_h);
	cuda_check(cudaMemcpyAsync(counts_h, counts, num_runs_h * sizeof(uint32_t), cudaMemcpyDeviceToHost));
	cuda_check(cudaMemcpyAsync(ptrs_unique_h, ptrs_unique, num_runs_h * sizeof(void *), cudaMemcpyDeviceToHost));
	cuda_check(cudaDeviceSynchronize());
	}

I think the line 99 should be changed as like the line 86.
cuda_check(cudaMemcpyAsync(counts_h, counts, (num_runs_h+1) * sizeof(uint32_t), cudaMemcpyDeviceToHost));

Or it will be fixed on the next version release?

[Speierers]

This will be fixed in the upcoming release and therefore we won't take the time to fix it in the current codebase.