How many operations are still pending?
How does a thread ensure its activities are transferred before it ends? Because metrics are stored in thread local maps, it needs to get all activities and attribute them before it dies.
opencl-api.c
device_finalizer_register vs thread finalizer
pending_operations
| waka |
| #include <dlfcn.h> | |
| #include "tool.h" | |
| int main() { | |
| //void *handle = dlopen("./tool.so", RTLD_NOW); | |
| print_t func = (print_t)dlsym(RTLD_NEXT, "print"); | |
| func(); | |
| return 0; | |
| } |
| // | |
| // Generated by LLVM NVPTX Back-End | |
| // | |
| .version 8.0 | |
| .target sm_80 | |
| .address_size 64 | |
| // .globl triton__0d1d2d3d | |
| .visible .entry triton__0d1d2d3d( | |
| .param .u64 triton__0d1d2d3d_param_0, | |
| .param .u64 triton__0d1d2d3d_param_1, |
Install
git clone https://github.com/openai/triton.git;
cd triton/python;
pip install cmake; # build time dependency
pip install -e .
pip uninstall pytorch-triton -yExpected result (-0.1250)
| // | |
| // Generated by LLVM NVPTX Back-End | |
| // | |
| .version 8.0 | |
| .target sm_80 | |
| .address_size 64 | |
| // .globl triton__0d1d2d3d4d56d7d89d1011d1213d1415d1617d1819d2021d2223d2425d2627d2829d3031d3233d3435d3637d3839d4041d42d | |
| .extern .shared .align 1 .b8 global_smem[]; |
| import triton | |
| import pytest | |
| import torch | |
| import triton.language as tl | |
| import numpy as np | |
| from numpy.random import RandomState | |
| @pytest.mark.parametrize("M, N, K, num_warps, epilogue, allow_tf32, in_dtype, out_dtype, axis", | |
| [(*shape_nw, 'softmax', allow_tf32, in_dtype, out_dtype, axis) |
| import torch | |
| import sys | |
| device = torch.device('cpu') | |
| left = torch.zeros(100, device=device, requires_grad=True) | |
| right = torch.zeros(100, device=device, requires_grad=True) | |
| grad = torch.zeros(100, device=device) | |
| for _ in range(10): | |
| output = torch.add(left, right) |
| """ | |
| Matrix Multiplication | |
| ===================== | |
| In this tutorial, you will write a very short high-performance FP16 matrix multiplication kernel that achieves | |
| performance on par with cuBLAS or rocBLAS. | |
| You will specifically learn about: | |
| * Block-level matrix multiplications. |