vectorize functional VS SIMD regular ops
Hi,
Let say we want to implent simple 2 vector multiplications.
the first implentation is using SIMD as data, we define 2 vectors using SIMD and do mul between them.
the second is using DTypePointer and then using vectorize functional.
What is the diffirence ? is it the same ?
thanks
5 Replies
If i understand your question correctly, the second process will be slower since you would be fetching data from the heap before performing the multiplication. Simply multiplying two SIMDs is faster than storing the data on the heap, reading it into two SIMD objects and then multiplying them.
Also, a word of caution for the future, the current plan as I understand it is to remove
DTypePointer in the future. You should use UnsafePointer instead.thanks Jake ! thanks for letting me know about the pointer things 👍 Regarding the question, for instance, if we examine the inner loop in the matrix multiplication example from Mojo, they implement it using vectorization. Therefore, if we convert the data to SIMD format and utilize SIMD operations, would the process would be faster ?
In Mojo, all values are SIMD values, so by loading data from the heap (via
SIMD.load) you are already creating a SIMD value that can use SIMD operations. Under the hood, all vectorize is doing is picking an optimal value of simd_width to use. There are no special conversions going on.Thanks a lot ... yes at the end i assumed that....
Congrats @Firas, you just advanced to level 4!