Bulk transfer size too large

[niloo9876]

Came across an issue where the transfer size is larger than 16MB. So far this only happens when the DPU code is run on upmemcloud4.

To recreate:
Compile with 1 tasklet on the upmemcloud4 machine and run the decoder with -d option with any image.
./host-1 -d ./data/fox0.jpg

Error recieved:

Got 2546 dpus across 40 ranks (63 dpus per rank)
Warning: fewer input files than DPUs (1 < 2546)
api:W:2022-02-17|18:03:39: dpu_push_xfer_symbol: invalid symbol access (offset:0 + length:231996736 > size:16777216)
src/jpeg-host.c:85(scale_rank): DPU Error (invalid memory symbol access)

Reasoning:
In the following loop, the last dpu has a garbage value stored in its in_length variable. So the final length is set to this value which causes the transfer to fail. (Copied from )

	DPU_FOREACH(dpu_rank, dpu, dpu_id)
	{
		DPU_ASSERT(dpu_prepare_xfer(dpu, (void *) input[dpu_id].in_buffer));

		if (input[dpu_id].in_length > longest_length)
			longest_length = input[dpu_id].in_length;
	}
	DPU_ASSERT(dpu_push_xfer(dpu_rank, DPU_XFER_TO_DPU, "file_buffer", 0, ALIGN(longest_length, 8), DPU_XFER_DEFAULT));

This is because this for loop goes through DPUs [0,63] inclusive, whereas the value of in_length is only set for DPUs [0,62]. In other words, only 63 DPUs are allocated per rank, however, the DPU_FOREACH loop goes through all 64 DPUs in a rank. Here is the Code where length is set.

[niloo9876]

This might be an SDK error as I don't observe the same behavior on the upmemcloud5 even though not all 64 DPUs/rank are allocated. The way I tested this is that I printed the dpu_id in DPU_FOREACH(dpu_rank, dpu, dpu_id) loop. This value was between [0,63] and [0,60] on upmemcloud4 and upmemcloud5, respectively.

[niloo9876]

clang version 12.0.0 (https://github.com/upmem/llvm-project.git d36425841d9a4d1420b7aa155675f6ae8bcf9f08)
Target: dpu-upmem-dpurte
Thread model: posix