Modular•12mo ago

Are literals splatted in SIMD Operations?

Is there any performance difference if I write this

fn multiply_by_two[
    dtype: DType, simd_width: Int
](x: SIMD[dtype, simd_width]) -> SIMD[dtype, simd_width]:
    alias splatted_two = SIMD[dtype, simd_width](2.)
    return splatted_two * x

fn multiply_by_two[
    dtype: DType, simd_width: Int
](x: SIMD[dtype, simd_width]) -> SIMD[dtype, simd_width]:
    alias splatted_two = SIMD[dtype, simd_width](2.)
    return splatted_two * x

instead of this?

fn multiply_by_two[
    dtype: DType, simd_width: Int
](x: SIMD[dtype, simd_width]) -> SIMD[dtype, simd_width]:
    return 2.0 * x

fn multiply_by_two[
    dtype: DType, simd_width: Int
](x: SIMD[dtype, simd_width]) -> SIMD[dtype, simd_width]:
    return 2.0 * x

In other words, is the literal 2 transformed into a splatted SIMD vector at runtime every time multiply_by_two is called? Or is this something that the compiler is able to optimize?

3 Replies

guidorice•12mo ago

It seems like there must be some tiny overhead to ~~concert~~ convert from FloatLiteral to SIMD, but only by running the benchmark module could you say for sure.

sora•12mo ago

isn't that price prepaid at compile time?

Leandro CamposOP•11mo ago

This question was answered here: https://github.com/modularml/mojo/discussions/1416#discussioncomment-7986953

GitHub

Are literals splatted in SIMD Operations? · modularml mojo · Discus...

Is there any performance difference if I write this fn multiply_by_two[ dtype: DType, simd_width: Int ](x: SIMD[dtype, simd_width]) -> SIMD[dtype, simd_width]: alias splatted_two = SIMD[dtype, s...

Gaming

Programming

Are literals splatted in SIMD Operations?