1. use `fence+relaxed` to replace `release` for fifo. `fence+relax` is
more efficient on A100
2. Update the deviceSyncer. Previous one cannot handle threadBlock
number change correctly. Use three counters to solve this issue. Reset
previous counter before sync on current counter.
3. Introduce relaxedWait which can be used with relaxedSignal for case
doesn't need guarantee the memory visibility
Remove __assert_fail for release build. This will reduce the number of
PTX instructions inside the loop. Also Trying to resolve this issue
reported in #497. Reduce the number of PTX instructions from 8 to 6.
8 ranks signal/wait will reduce from 3.2us->2.8us on NDv5
Also NDEBUG flag is confused here, sometime it will not be set. Use
customized flag for debug build.
Here is current PTX:
```
ld.u64 %rd12, [%rd2+-24];
mov.u64 %rd13, %rd12;
mov.u64 %rd11, %rd13;
ld.acquire.sys.b64 %rd10,[%rd11];
setp.lt.u64 %p1, %rd10, %rd3;
@%p1 bra $L__BB2_1;
```
If we change to `asm volatile("ld.global.acquire.sys.b64 %0, [%1];" :
"=l"(flag) : "l"(flag_addr));` will reduce to 4 instructions. We can get
2.1 us for 8 ranks signal/wait
```
ld.u64 %rd9, [%rd1+-24];
ld.global.acquire.sys.b64 %rd8, [%rd9];
setp.lt.u64 %p1, %rd8, %rd2;
@%p1 bra $L__BB2_1;
```
For mscclpp, to use nvls we require the buffer is allocated by
mscclpp::GpuBuffer. Due to cupy doesn't support bfloat16 yet, we export
the raw buffer to dlpack format.
User can use this feature to create buffer with type supported by
pytorch
```python
buffer = RawGpuBuffer(1024 * 2) # 2 for bfloat16
dl_pack = buffer.to_dlpack(str(torch.bfloat16))
tensor = torch.utils.dlpack.from_dlpack(dl_pack)
```
Use dlopen to load nccl/rccl Apis from shared library to
enable Allgather, Allreduce, Broadcast, ReduceScatter fallback to nccl/rccl operations.
Add three related environment variables
-x MSCCLPP_ENABLE_NCCL_FALLBACK=TRUE
-x MSCCLPP_NCCL_LIB_PATH=/path/libnccl.so/librccl.so
-x MSCCLPP_FORCE_NCCL_FALLBACK_OPERATION="allreduce,allgather,broadcast,reducescatter" or "all"
By default, if MSCCLPP_FORCE_NCCL_FALLBACK_OPERATION is not specified, all these operations will be fallback to nccl/rccl apis.
---------
Co-authored-by: Binyang Li <binyli@microsoft.com>
Enhancements to all-gather operation, a temporary solution to fix the
memory overhead when integrating msccl++ with pytorch.
This solution will not register input/output buffer to msccl++, so the
temp output buffer for allgather could be reused by torch automatically.
* Introduced a new `allgather8` kernel function in
`apps/nccl/src/allgather.hpp` to handle larger data sizes more
efficiently. This includes double buffering to hide synchronization
overhead and support for both in-place and out-of-place operations.
* Modified the `allgather` function to decide between `allgather6` and
`allgather8` based on data size and platform, improving performance for
large data sizes.
Configuration and environment improvements:
* Added a new environment variable `MSCCLPP_DISABLE_CHANNEL_CACHE` to
control whether the channel cache is disabled, enhancing
configurability. This variable is now part of the `Env` class and is
logged during environment initialization.
* Removed the redundant global variable `mscclppDisableChannelCache`
from `src/debug.cc` and updated its usage to refer to the new
environment variable.
* Renamed and moved mem alloc functions into the `mscclpp::detail::`
namespace (now `mscclpp::detail::gpuCalloc*<T>()`)
* Deprecated constructor-calling mem alloc functions
(`mscclpp::makeShared*<T>()` and `mscclpp::makeUnique*<T>()`)
* Added a new `mscclpp::GpuBuffer<T>()` class that should be used in
general for allocating communication buffers
* Added a new `mscclpp.utils.GpuBuffer` Python class that inherits
`cupy.ndarray` and allocates using `mscclpp::gpuMemAlloc`
* Renamed `mscclpp::memcpyCuda*<T>()` functions into
`mscclpp::gpuMemcpy*<T>()` for name consistency
* A few fixes in NVLS memory allocation
* Tackled minor compiler warnings
* Renamed `ProxyChannel` -> `BaseProxyChannel` and `SimpleProxyChannel`
-> `ProxyChannel`. It makes the interface more consistent by defining
channels to be associated with a certain src/dst memory region:
`ProxyChannel` as "sema + src/dst + fifo" and `SmChannel` as "sema +
src/dst". BaseProxyChannel is not associated with any memory regions, as
"sema + fifo".
* `ProxyChannelDeviceHandle` now inherits from
`BaseProxyChannelDeviceHandle`, instead of having one as a member.
- Support mote datatype for multicast operation
- Add new OP MULTI_LOAD_REDUCE_STORE to support NVLS
- Modify allocSharedPhysicalCuda, which return std::shared_ptr<T>
instead of std::shared_ptr<PhysicalCudaMemory>
- Add Python support for allocSharedPhysicalCuda
Test passed for `allreduce_nvls.json`
- Add C++ executor test
- Fix executor bugs for packet operation
- Enhance executor_test.py
---------
Co-authored-by: Binyang Li <binyli@microsoft.com>
