Same implementation for Sparse Multiplication for aligned and unaligned arrays #1274
Conversation
|
@tannergooding @eerhardt can you please take a look at this ? |
test/Microsoft.ML.CpuMath.PerformanceTests/SsePerformanceTests.cs
Outdated
Show resolved
Hide resolved
| while (ppos < pposEnd) | ||
| { | ||
| int col = *ppos; | ||
| Vector128<float> x1 = Sse.SetVector128(pm3[col], pm2[col], pm1[col], pm0[col]); |
There was a problem hiding this comment.
note to self: I want to check the codegen of this and ensure that it is being emitted "optimally" (two loads and three unpack with two folded loads; rather than as four loads and three unpack).
|
@tannergooding did u get the chance to look at the codegen for this ? |
|
No, not yet. |
| { | ||
| int col1 = *ppos; | ||
| int col2 = col1 + 4 * ccol; | ||
| Vector256<float> x1 = Avx.SetVector256(pm3[col2], pm2[col2], pm1[col2], pm0[col2], |
There was a problem hiding this comment.
Don't we have a helper method for this?
There was a problem hiding this comment.
no its different , the one we have the indexs are continous
return Avx.SetVector256(src[idx[7]], src[idx[6]], src[idx[5]], src[idx[4]], src[idx[3]], src[idx[2]], src[idx[1]], src[idx[0]]);
… place, span overloads and function for common code
Before
After
As the matrix becomes more dense , the new algorithm becomes faster cc @danmosemsft @eerhardt @tannergooding |
|
@tannergooding i have resolved the conflicts and addressed the feedback |
|
I have restarted the queue and new build passed successfully |
ghost
locked as resolved and limited conversation to collaborators
Working towards #1018