Add Op dot #430
Add Op dot #430
Conversation
src/flag_gems/ops/dot.py
Outdated
|
|
||
| with torch_device_fn.device(x.device): | ||
| dot_kernel_1[grid_1](x, y, mid, N, block_size) | ||
| dot_kernel_2[grid_2](mid, out, mid_size, block_mid) |
There was a problem hiding this comment.
I think it's better to take tensor stride into consideration. but it's a good implementation!
src/flag_gems/ops/dot.py
Outdated
| dot_kernel_1[grid_1](x, y, mid, N, block_size) | ||
| dot_kernel_2[grid_2](mid, out, mid_size, block_mid) |
There was a problem hiding this comment.
Can we resort to a single persistent kernel when the input numel is small enough?
There was a problem hiding this comment.
I tried to implement dot in a single kernel using atomic_add, but even in very small input numel, the performance was not good, but I still kept the code in function dot_kernel.
this shows the performance of kernel1 and kernel2
this shows the performance of single kernel
There was a problem hiding this comment.
I probably didnt make myself clear. What I suggested is adding a one pass branch to handle small input. We don't have to use atomic_add on either branch. The two pass branch still exists.
|
@wlxjhyf, thanks for contributing to flaggems. Please resolve the conversions and complete this PR at your earliest convenience. |
I'm sorry I just saw it, I'll do it right now |
Don't be sorry. We are very grateful to you for your volunteering! |
|
When using a single kernel, N must be smaller than tl.TRITON_MAX_TENSOR_NUMEL (1048576).In my tests on A100, I found that when N is smaller than 4096, using a single operator can still maintain good performance.So, I currently design 4096 as the branching condition. 
And I found that the reason for the failure of the last submission was the failure of the test_index_put_acc_true test. Does this relate to my code? |
PR Category
Operator
Type of Change
Add new operator
Description
Implement dot operator, support Float32, Float16，BFloat16。
The operator implementation is to split the dot operator into two steps, the first step implementing elementwise level multiplication, and the second step implementing summation.
At present, in order to ensure accuracy requirements, the intermediate results are saved as float32 type in the first step.
Issue
#394
Progress
Performance
correctness
performance
