question15：add nvidia branch #903
Conversation
There was a problem hiding this comment.
Pull request overview
This pull request adds NVIDIA/CUDA implementations for 5 mathematical operators to support "question 15". The implementation includes complete test infrastructure, operator registration, CUDA kernels, CPU fallbacks, and Python bindings.
Key changes:
- Added 5 new operators:
atanh,addcmul,cdist,binary_cross_entropy_with_logits, andreciprocal - Implemented CUDA kernels for NVIDIA devices with CPU fallbacks
- Added comprehensive test suites for both infiniop and infinicore layers
- Updated Python bindings to expose new operators
Reviewed changes
Copilot reviewed 88 out of 88 changed files in this pull request and generated 8 comments.
Show a summary per file
| File | Description |
|---|---|
| test/infiniop/*.py | Test files for all 5 operators with test cases and tolerance maps |
| test/infiniop/libinfiniop/op_register.py | Registration of operator API bindings |
| src/infiniop/ops/*/operator.cc | Operator dispatchers for device-specific implementations |
| src/infiniop/ops//nvidia/.cu | CUDA kernel implementations |
| src/infiniop/ops//cpu/.cc | CPU implementations |
| src/infinicore/ops//.cc | InfiniCore operator implementations |
| src/infinicore/pybind11/ops/*.hpp | Python binding definitions |
| python/infinicore/ops/*.py | Python wrapper functions |
| include/infiniop/ops/*.h | C API header files |
| include/infinicore/ops/*.hpp | C++ API header files |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| @@ -0,0 +1,163 @@ | |||
| import torch | |||
| import ctypes | |||
There was a problem hiding this comment.
Module 'ctypes' is imported with both 'import' and 'import from'.
| @@ -0,0 +1,171 @@ | |||
| import torch | |||
| import ctypes | |||
There was a problem hiding this comment.
Module 'ctypes' is imported with both 'import' and 'import from'.
| @@ -0,0 +1,154 @@ | |||
| import torch | |||
| import ctypes | |||
There was a problem hiding this comment.
Module 'ctypes' is imported with both 'import' and 'import from'.
| @@ -0,0 +1,156 @@ | |||
| import torch | |||
| import ctypes | |||
There was a problem hiding this comment.
Module 'ctypes' is imported with both 'import' and 'import from'.
| @@ -0,0 +1,165 @@ | |||
| import torch | |||
| import ctypes | |||
There was a problem hiding this comment.
Module 'ctypes' is imported with both 'import' and 'import from'.
| @@ -0,0 +1,154 @@ | |||
| import torch | |||
| import ctypes | |||
| from ctypes import c_uint64, c_float, c_char_p | |||
There was a problem hiding this comment.
Import of 'c_float' is not used.
| @@ -0,0 +1,156 @@ | |||
| import torch | |||
| import ctypes | |||
| from ctypes import c_uint64, c_float, c_double | |||
There was a problem hiding this comment.
Import of 'c_float' is not used.
| constexpr int BCE_MAX_DIMS = 8; | ||
|
|
||
| struct BCETensorInfoDevice { | ||
| size_t ndim; | ||
| size_t shape[BCE_MAX_DIMS]; | ||
| ptrdiff_t strides[BCE_MAX_DIMS]; | ||
| }; | ||
|
|
||
| static inline BCETensorInfoDevice make_device_info(const BCETensorInfo &info) { | ||
| BCETensorInfoDevice dev{}; | ||
| dev.ndim = info.ndim; | ||
| for (size_t i = 0; i < info.ndim && i < static_cast<size_t>(BCE_MAX_DIMS); ++i) { | ||
| dev.shape[i] = info.dims[i]; | ||
| dev.strides[i] = info.stride[i]; | ||
| } |
There was a problem hiding this comment.
BCETensorInfoDevice stores shape and strides in fixed-size arrays of length BCE_MAX_DIMS, but make_device_info sets ndim directly from info.ndim and only truncates the arrays when copying. If info.ndim exceeds BCE_MAX_DIMS, later calls to indexToOffset(idx, logits_info.ndim, logits_info.shape, logits_info.strides) will index past the end of these arrays, producing out-of-bounds offsets and causing GPU memory corruption or unintended reads/writes on the logits/targets/outputs buffers. To harden this, either enforce info.ndim <= BCE_MAX_DIMS at descriptor creation or clamp and validate ndim before launching the CUDA kernel so that higher-rank tensors are rejected or handled safely.
1 participant
针对第15题，添加了NVIDIA实现