Implement pure OpenCL batch hashing.

[porcuquine]

This PR implements GPU batch hashing in pure OpenCL, implemented in the new proteus module. (Proteus and Triton are both moons of Neptune, hence the naming.)

This work is intended to introduce no change of behavior when the gpu feature flag is provided. If instead, the opencl feature flag is provided, a new BatcherType, BatcherType::OpenCL can be used instead of BatcherType::GPU.

This implementation provides the following benefits when compared to the extant neptune-triton GPU implementation:

• Better perfomance (almost 2x, see below).
• Fewer external dependencies (removes dependence on elaborate Futhark code-generation and toolchain).
• Much less total code.
• Much lower GPU memory usage.
• No known problem with multiple batch hashers being used at once. (vs. an outstanding bug in current neptune-triton code path).

Once the opencl feature has been tested and stabilized, it should be made the default, for all of these reasons.

Historical context: although replacing neptune-triton is an obvious next step now, its replacement benefits from the design which went into the Rust interface to neptune-triton, to the development of rust-gpu-tools and cl-ff-gen (neither of which existed at the time of the initial GPU implementation), and from the significant learning which went into neptune-triton's development. Although the current result is simpler, the path to it was not obvious from the outset.

Speedup is ~2x on column tree building. See gbench output using the same 2080Ti for both methods.

gbench with gpu feature:

RUST_LOG=info cargo run --release --features gpu,blst --no-default-features
    Finished release [optimized] target(s) in 0.07s
     Running `/home/porcuquine/dev/neptune/target/release/gbench`
[2021-01-16T00:07:05Z INFO  gbench] KiB: 4194304
[2021-01-16T00:07:05Z INFO  gbench] leaves: 134217728
[2021-01-16T00:07:05Z INFO  gbench] max column batch size: 400000
[2021-01-16T00:07:05Z INFO  gbench] max tree batch size: 700000
[2021-01-16T00:07:05Z INFO  gbench] GPU[Selector: BatcherType::GPU] --> Run 0
[2021-01-16T00:07:05Z INFO  gbench] GPU[Selector: BatcherType::GPU]: Creating ColumnTreeBuilder
[2021-01-16T00:07:05Z INFO  neptune::triton::cl] getting default futhark context
[2021-01-16T00:07:05Z INFO  neptune::triton::cl] getting context for ~Index(0)
[2021-01-16T00:07:06Z INFO  neptune::triton::cl] device: Device { brand: Nvidia, name: "GeForce RTX 2080 Ti", memory: 11551440896, bus_id: Some(33), platform: Platform(PlatformId(0x7f1f000b5590)), device: Device(DeviceId(0x7f1f000b5ae0)) }
[2021-01-16T00:07:11Z INFO  gbench] GPU[Selector: BatcherType::GPU]: ColumnTreeBuilder created
[2021-01-16T00:07:11Z INFO  gbench] GPU[Selector: BatcherType::GPU]: Using effective batch size 400000 to build columns
[2021-01-16T00:07:11Z INFO  gbench] GPU[Selector: BatcherType::GPU]: adding column batches
[2021-01-16T00:07:11Z INFO  gbench] GPU[Selector: BatcherType::GPU]: start commitment
...............................................................................................................................................................................................................................................................................................................................................
[2021-01-16T00:09:15Z INFO  gbench] GPU[Selector: BatcherType::GPU]: adding final column batch and building tree
[2021-01-16T00:09:31Z INFO  gbench] GPU[Selector: BatcherType::GPU]: end commitment
[2021-01-16T00:09:31Z INFO  gbench] GPU[Selector: BatcherType::GPU]: commitment time: 139.632183641s

gbench with opencl feature:

RUST_LOG=info cargo run --release --features opencl,blst --no-default-features
    Finished release [optimized] target(s) in 0.06s
     Running `/home/porcuquine/dev/neptune/target/release/gbench`
[2021-01-16T00:19:55Z INFO  gbench] KiB: 4194304
[2021-01-16T00:19:55Z INFO  gbench] leaves: 134217728
[2021-01-16T00:19:55Z INFO  gbench] max column batch size: 400000
[2021-01-16T00:19:55Z INFO  gbench] max tree batch size: 700000
[2021-01-16T00:19:55Z INFO  gbench] GPU[Selector: BatcherType::OpenCL] --> Run 0
[2021-01-16T00:19:55Z INFO  gbench] GPU[Selector: BatcherType::OpenCL]: Creating ColumnTreeBuilder
[2021-01-16T00:19:56Z INFO  neptune::proteus::gpu] device: Device { brand: Nvidia, name: "GeForce RTX 2080 Ti", memory: 11551440896, bus_id: Some(33), platform: Platform(PlatformId(0x7f7d240b5510)), device: Device(DeviceId(0x7f7d240b5a60)) }
[2021-01-16T00:19:58Z INFO  gbench] GPU[Selector: BatcherType::OpenCL]: ColumnTreeBuilder created
[2021-01-16T00:19:58Z INFO  gbench] GPU[Selector: BatcherType::OpenCL]: Using effective batch size 400000 to build columns
[2021-01-16T00:19:58Z INFO  gbench] GPU[Selector: BatcherType::OpenCL]: adding column batches
[2021-01-16T00:19:58Z INFO  gbench] GPU[Selector: BatcherType::OpenCL]: start commitment
...............................................................................................................................................................................................................................................................................................................................................
[2021-01-16T00:21:04Z INFO  gbench] GPU[Selector: BatcherType::OpenCL]: adding final column batch and building tree
[2021-01-16T00:21:14Z INFO  gbench] GPU[Selector: BatcherType::OpenCL]: end commitment
[2021-01-16T00:21:14Z INFO  gbench] GPU[Selector: BatcherType::OpenCL]: commitment time: 75.21889048s

[dignifiedquire]

Benchmarks on RTX3090

• GPU (current): 97.686103731s
• OpenCL (this PR): 52.741545444s

[dignifiedquire]

two small notes, nice work

[dignifiedquire]

shouldn’t this go under unreleased technically?

[porcuquine]

I'll change it. I was originally hoping to just release this version immediately after.

[dignifiedquire]

shouldn’t this be not any gpu or opencl?

[porcuquine]

Hmmm... I think that whole variant can be removed now. My latest does that and seems to build fine on macos now.