Intel® oneAPI DPC++/C++ Compiler Developer Guide and Reference

ID 767253
Date 11/07/2023
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

Efficiency with Structure of Arrays Example

This example demonstrates the efficiency of using a Structure of Arrays (SoA) approach by comparing the assembly generated from a simple SIMD loop using an Array of Structures (AoS) approach with the assembly generated using the SoA approach of SDLT.

Array of Structures: Non-unit stride access version

Source:

#include <stdio.h>

#define N 1024

typedef struct RGBs {
    float r;
    float g;
    float b;
} RGBTy;


void main()
{
    RGBTy a[N];  
    #pragma omp simd
    for (int k = 0; k<N; ++k) {
        a[k].r = k*1.5;  // non-unit stride access
        a[k].g = k*2.5;  // non-unit stride access
        a[k].b = k*3.5;  // non-unit stride access
    }
    std::cout << "k =" << 10 <<
        ", a[k].r =" << a[10].r <<
        ", a[k].g =" << a[10].g <<
        ", a[k].b =" << a[10].b << std::endl;
}

AVX2 assembly generated (69 instructions):

..TOP_OF_LOOP:
        vcvtdq2ps %ymm7, %ymm1
        lea       (%rax), %rcx
        vcvtdq2ps %ymm5, %ymm2
        vpaddd    %ymm3, %ymm7, %ymm7
        vpaddd    %ymm3, %ymm5, %ymm5
        vmulps    %ymm1, %ymm4, %ymm8
        vmulps    %ymm1, %ymm6, %ymm12
        vmulps    %ymm2, %ymm6, %ymm14
        vmulps    %ymm1, %ymm0, %ymm1
        vmulps    %ymm2, %ymm4, %ymm10
        addl      $16, %edx
        vextractf128 $1, %ymm8, %xmm9
        vmovss    %xmm8, (%rsp,%rcx)
        vmovss    %xmm9, 48(%rsp,%rcx)
        vextractps $1, %xmm8, 12(%rsp,%rcx)
        vextractps $2, %xmm8, 24(%rsp,%rcx)
        vextractps $3, %xmm8, 36(%rsp,%rcx)
        vmulps    %ymm2, %ymm0, %ymm8
        vextractps $1, %xmm9, 60(%rsp,%rcx)
        vextractps $2, %xmm9, 72(%rsp,%rcx)
        vextractps $3, %xmm9, 84(%rsp,%rcx)
        vextractf128 $1, %ymm12, %xmm13
        vextractf128 $1, %ymm14, %xmm15
        vextractf128 $1, %ymm1, %xmm2
        vextractf128 $1, %ymm8, %xmm9
        vmovss    %xmm12, 4(%rsp,%rax)
        vmovss    %xmm13, 52(%rsp,%rax)
        vextractps $1, %xmm12, 16(%rsp,%rax)
        vextractps $2, %xmm12, 28(%rsp,%rax)
        vextractps $3, %xmm12, 40(%rsp,%rax)
        vextractps $1, %xmm13, 64(%rsp,%rax)
        vextractps $2, %xmm13, 76(%rsp,%rax)
        vextractps $3, %xmm13, 88(%rsp,%rax)
        vmovss    %xmm14, 100(%rsp,%rax)
        vextractps $1, %xmm14, 112(%rsp,%rax)
        vextractps $2, %xmm14, 124(%rsp,%rax)
        vextractps $3, %xmm14, 136(%rsp,%rax)
        vmovss    %xmm15, 148(%rsp,%rax)
        vextractps $1, %xmm15, 160(%rsp,%rax)
        vextractps $2, %xmm15, 172(%rsp,%rax)
        vextractps $3, %xmm15, 184(%rsp,%rax)
        vmovss    %xmm1, 8(%rsp,%rax)
        vextractps $1, %xmm1, 20(%rsp,%rax)
        vextractps $2, %xmm1, 32(%rsp,%rax)
        vextractps $3, %xmm1, 44(%rsp,%rax)
        vmovss    %xmm2, 56(%rsp,%rax)
        vextractps $1, %xmm2, 68(%rsp,%rax)
        vextractps $2, %xmm2, 80(%rsp,%rax)
        vextractps $3, %xmm2, 92(%rsp,%rax)
        vmovss    %xmm8, 104(%rsp,%rax)
        vextractps $1, %xmm8, 116(%rsp,%rax)
        vextractps $2, %xmm8, 128(%rsp,%rax)
        vextractps $3, %xmm8, 140(%rsp,%rax)
        vmovss    %xmm9, 152(%rsp,%rax)
        vextractps $1, %xmm9, 164(%rsp,%rax)
        vextractps $2, %xmm9, 176(%rsp,%rax)
        vextractps $3, %xmm9, 188(%rsp,%rax)
        addq      $192, %rax
        vextractf128 $1, %ymm10, %xmm11
        vmovss    %xmm10, 96(%rsp,%rcx)
        vmovss    %xmm11, 144(%rsp,%rcx)
        vextractps $1, %xmm10, 108(%rsp,%rcx)
        vextractps $2, %xmm10, 120(%rsp,%rcx)
        vextractps $3, %xmm10, 132(%rsp,%rcx)
        vextractps $1, %xmm11, 156(%rsp,%rcx)
        vextractps $2, %xmm11, 168(%rsp,%rcx)
        vextractps $3, %xmm11, 180(%rsp,%rcx)
        cmpl      $1024, %edx
        jb        ..TOP_OF_LOOP

Structure of Arrays: Using SDLT for unit stride access

To introduce the use of SDLT, the code below will:

  • declare a primitive,

  • use an soa1d_container instead of an array

  • use an accessor inside a SIMD loop to generate efficient code

  • use a proxy object’s data member interface to access individual data members of an element inside the container

Source:

#include <stdio.h>
#include <sdlt/sdlt.h>

#define N 1024

typedef struct RGBs {
    float r;
    float g;
    float b;
} RGBTy;

SDLT_PRIMITIVE(RGBTy, r, g, b)

void main()
{
    // Use SDLT to get SOA data layout
    sdlt::soa1d_container<RGBTy> aContainer(N);
    auto a = aContainer.access();

    // use SDLT Data Member Interface to access struct members r, g, and b.
    // achieve unit-stride access after vectorization
    #pragma omp simd
    for (int k = 0; k<N; k++) {
        a[k].r() = k*1.5;
        a[k].g() = k*2.5;
        a[k].b() = k*3.5;
    }
    std::cout << "k =" << 10 <<
        ", a[k].r =" << a[10].r() <<
        ", a[k].g =" << a[10].g() <<
        ", a[k].b =" << a[10].b() << std::endl;
}

AVX2 assemply generated (19 instructions):

..TOP_OF_LOOP:
        vpaddd    %ymm4, %ymm3, %ymm12
        vcvtdq2ps %ymm3, %ymm7
        vcvtdq2ps %ymm12, %ymm10
        vmulps    %ymm7, %ymm2, %ymm5
        vmulps    %ymm7, %ymm1, %ymm6
        vmulps    %ymm7, %ymm0, %ymm8
        vmulps    %ymm10, %ymm2, %ymm3
        vmulps    %ymm10, %ymm1, %ymm9
        vmulps    %ymm10, %ymm0, %ymm11
        vmovups   %ymm5, (%r13,%rax,4)
        vmovups   %ymm6, (%r15,%rax,4)
        vmovups   %ymm8, (%rbx,%rax,4)
        vmovups   %ymm3, 32(%r13,%rax,4)
        vmovups   %ymm9, 32(%r15,%rax,4)
        vmovups   %ymm11, 32(%rbx,%rax,4)
        vpaddd    %ymm4, %ymm12, %ymm3
        addq      $16, %rax
        cmpq      $1024, %rax
        jb        ..TOP_OF_LOOP

Both versions appear to have unrolled the loop twice. When examining the assembly generated for AVX2 instruction set, we can see a measurable reduction in the number of instructions (19 vs. 69) when we are able to perform unit stride access using SDLT. Also, at runtime, the soa1d_container aligned its data allocation and will gain any of the architectural advantages that come with using aligned instead of unaligned SIMD stores.