WIP: Cuda autotune

This PR introduces two one changes:

• rotate (32,1,1) depending on field strides to fastest dimension. So (1,1,32) for c-layout and (32,1,1) for fortran layout. So pystencils will be fast also for c-layout (this will always be performed)
• auto-tune the block dimensions to whatevers is fastest for a specific kernel on localhost. On first kernel call different layouts are tried and the kernel will be called henceforth with the fastest configuration (disk_cached). This could be intersting for OpenCL where we don't know which launch config is the fastest (on OpenCL the runtime can alternatively give a hint on that).

One drawback: the test calls are only correct if input and output fields do not overlap (so no in-place kernels).

added 2 commits

• 85dc6b1b - Fixup
• ebbb81c4 - Test also for 2d

Compare with previous version

marked as a Work In Progress

changed the description

added 1 commit

• e25c266c - Fixup

Compare with previous version

Ok, I forgot that this function already exists: https://i10git.cs.fau.de/seitz/pystencils/blob/e25c266c97247f38d1bd3d2146dd230073a72018/pystencils/gpucuda/indexing.py#L219-232

changed the description

added 1 commit

• 04ca021b - Use Martin's method to rotate block indices

Compare with previous version

added 1 commit

• 64e0dc69 - Also add AUTOTUNE_BLOCK_SIZES to LineIndexing

Compare with previous version

added 1 commit

• 15cbd9f0 - Add auto-tuning for CUDA call parameters

Compare with previous version

added 1 commit

• 4478d8b5 - Add auto-tuning for CUDA call parameters

Compare with previous version

added 1 commit

• 0800d84a - Add auto-tuning for CUDA call parameters

Compare with previous version

changed the description

added 1 commit

• d04b4789 - Add 'cuda' compiler config (with preferred_block_size and always_autotune)

Compare with previous version

added 1 commit

• ba7b20ac - Add 'cuda' compiler config (with preferred_block_size and always_autotune)

Compare with previous version

added 1 commit

• 98291ce5 - Change block_size -> block_and_thread_numbers

Compare with previous version

Weird: 0800d84a works as expected. Current version does not work for preferred_blocksize=(32,1,1) and always_autotune=True.

closed

changed the description