Adding vectorized implementations of Exp to Vector64/128/256/512 #97114
Conversation
ghost
assigned 
|
Note regarding the This serves as a reminder for when your PR is modifying a ref *.cs file and adding/modifying public APIs, please make sure the API implementation in the src *.cs file is documented with triple slash comments, so the PR reviewers can sign off that change. |
|
Tagging subscribers to this area: @dotnet/area-system-numerics Issue DetailsThis makes progress towards #93513
|
|
|
||
| private const float V_EXPF_MIN = -103.97208f; | ||
| private const float V_EXPF_MAX = 88.72284f; | ||
| private const float V_EXPF_MAX = +88.72284f; |
There was a problem hiding this comment.
Correct, just better aligning the numbers, which I personally find more visually appealing/easier to read.
| @@ -1426,6 +1426,50 @@ public static bool EqualsAny<T>(Vector128<T> left, Vector128<T> right) | |||
| || Vector64.EqualsAny(left._upper, right._upper); | |||
| } | |||
|
|
|||
| internal static Vector128<T> Exp<T>(Vector128<T> vector) | |||
There was a problem hiding this comment.
Is this used somewhere outside of this type, or could it be private?
There was a problem hiding this comment.
Copy/paste error. V128/256/512 should be calling the next smaller size on the upper/lower halves, as that gives the simplest implementation and tends to better performance on hardware that can't directly accelerate that size.
Log/Log2 are doing the right thing, just messed up here for Exp
| @@ -578,19 +814,19 @@ internal static class VectorMath | |||
|
|
|||
| if (typeof(TVectorInt64) == typeof(Vector64<long>)) | |||
| { | |||
| return (TVectorDouble)(object)Vector64.ConvertToDouble((Vector64<long>)(object)vector); | |||
| result = (TVectorDouble)(object)Vector64.ConvertToDouble((Vector64<long>)(object)vector); | |||
There was a problem hiding this comment.
What drove the changes here around how the result is returned?
There was a problem hiding this comment.
Consistency around the intent/way the code was written.
Most compilers tend to perform better and need to do less work with a single return path, the JIT has historically been no exception, and since we need to handle returning in the throw path anyways, I had intended for it to be setup like this originally.
tannergooding
force-pushed
the
fix-93513
branch
from
79eb2d0
to
398c78e
Compare
…net#97114) * Adding vectorized implementations of Exp to Vector64/128/256/512 * Accelerate TensorPrimitives.Exp for double * Ensure the right allowedVariance is used for the vectorized exp tests * Ensure V128/256/512 defers to the next smaller vector size by operating on the lower/upper halves * Ensure the right allowedVariance amounts are used for the vectorized Exp(float) tests * Ensure we call Exp and that the methods are properly inlined * Skip the Exp test for Vector128/256/512 on Mono due to dotnet#97176
This makes progress towards #93513