vector dot product with SYCL esimd intrinsic functions

[mathbird]

Hi, is there a code sample for vector dot product, using SYCL esimd intrinsic functions?

Thanks,

D.

[alexbatashev]

@mathbird do you mean dp4a intrinsic? There's a test case for that: https://github.com/intel/llvm-test-suite/blob/intel/SYCL/ESIMD/dp4a.cpp

[mathbird]

I mean this:

let
A=(a_1,a_2,..., a_n)
B=(b_1,b_2,..., b_n)

do the calculation: sum =dot(A, B) = a_1 * b_2 + a_1 * b_2 + ... + a_n * b_n.

I would like to see how ESIMD uses shared local memory in the calculation.

In other words, how to rewrite the following program with SYCL ESIMD:

#include <stdlib.h> #include <math.h> #include <stdio.h> #include <iostream> #define N 1<<10 int main(int argc, char *argv[] ) { double *A = new double[N]; double *B = new double[N]; for(int i = 0; i < N; i++){ A[i] = 0.1; B[i] = 0.2; } double sum = 0.0; for(int i = 0; i < N; i++){ sum += A[i]*B[i]; } std::cout << " sum =" << sum << std::endl; delete []A; delete []B; return 0; } 

[AlexeySachkov]

Tagging @kbobrovs and @DenisBakhvalov here to help answering the question

[kbobrovs]

Looks like you are basically looking how to do a reduction in ESIMD using SLM. Good example is https://github.com/intel/llvm-test-suite/blob/intel/SYCL/ESIMD/histogram_256_slm.cpp, look for slm_init, slm_load, slm_atomic, slm_store.

[jinz2014]

Is SYCL ESMID portable (e.g. NVIDIA GPUs) ?

Is the dp4a example specific to Intel GPUs ?
https://github.com/intel/llvm-test-suite/blob/intel/SYCL/ESIMD/dp4a.cpp

[AlexeySachkov]

@jinz2014, it is not portable to other devices (even Intel CPUs).

It is documented as intel extension: https://github.com/intel/llvm/tree/sycl/sycl/doc/extensions/supported/sycl_ext_intel_esimd which generally means that it won't/can't be supported on HW from other vendors.

And it also explicitly states that this is an interface to Intel GPU HW capabilities:

The main motivation for introducing the "Explicit SIMD" SYCL extension (or simply "ESIMD") is enabling efficient low-level programming for Intel graphics architectures.

Essentially, it is not portable because it almost directly maps to Intel GPU ISA.