vectorization / SIMD: How to shift a vector of float/double?

[mikeitexpert]

Hi Everyone,

I am trying to speed up a roll up / cumulative sum of an array using SIMD.  Below is the code ...

I need shift operator to achieve this. However, it gives casting error. I look up the Intel Intirinsics guide

But, all shift operators seems to working on integer values (ie. __m128i but not __m128).

Please let me know how to resolve this.

Much appreciate it.

#include <iostream>
#include <emmintrin.h>
#include <immintrin.h>


int main(){

	float *a = (float*) _mm_malloc((1 << 20)*sizeof(float), 512/8);
	float *as = (float*) _mm_malloc((1 << 20)*sizeof(float), 512/8);

	float seri_sum, vec_sum, vec_parallel_sum;

	for (int i = 1; i < (1 << 20); i++){
		as[i] += as[i-1];
	}
	seri_sum = as[(1 << 20)-1];
	std::cout << "seri_sum" << " = " << seri_sum << std::endl;

	float sum = 0;
	for(int i = 0; i < (1 << 20); i+=((128)/(sizeof(float)*8))){
		
		__m128 vsum; 
		vsum = _mm_load_ps1(&sum);
		
		__m128 vec;
		vec = _mm_load_ps(&a[i]);
		
		
		
		vsum = _mm_add_ps(vsum,vec);
		for(int j = 1; j < (1 << 20); i++){
			vec = (__m128) _mm_slli_epi32((__m128i) vec,1);
			vsum = _mm_add_ps(vsum,vec);
		}
		_mm_store_ps(&a[i],vsum);
		sum = a[i+((128)/(sizeof(float)*8))-1];
	}

	_mm_free(a);
	_mm_free(as);
	return 0;
}

Below is the error message:

C:\tmp\simdtests>make clean test_vec
rm -f *.exp *.lib *.supp *.exe *.obj *.optrpt
icl.exe -Qopt-report:5 -Qopt-report-phase:all /Ob0  -Qopenmp -Qsimd -Qopenmp-simd  -arch:avx  -Qdiag-error-limit:5   -c test_vec.cpp -Fo:test_vec.obj
Intel(R) C++ Intel(R) 64 Compiler for applications running on Intel(R) 64, Version 19.1.2.254 Build 20200623
Copyright (C) 1985-2020 Intel Corporation.  All rights reserved.

icl: remark #10397: optimization reports are generated in *.optrpt files in the output location
test_vec.cpp
test_vec.cpp(32): error: no suitable user-defined conversion from "__m128" to "__m128i" exists
                        vec = (__m128) _mm_slli_epi32((__m128i) vec,1);
                                                                ^

test_vec.cpp(32): error: no suitable user-defined conversion from "__m128i" to "__m128" exists
                        vec = (__m128) _mm_slli_epi32((__m128i) vec,1);
                                       ^

compilation aborted for test_vec.cpp (code 2)
make: *** [Makefile:32: test_vec.obj] Error 2

C:\tmp\simdtests>

[NoorjahanSk_Intel · Accepted solution]

Hi,

Thanks for reaching out to us.

We see that you are trying to do typecasting over here 

>>vec = (__m128) _mm_slli_epi32((__m128i) vec,1);

In Intrinsic guide we have type casting methods to do it.

For Example : To cast from __m128 to __m128i we can  use below method

 __m128i _mm_castps_si128 (__m128 a) 

Please refer:

https://software.intel.com/sites/landingpage/IntrinsicsGuide/#techs=SSE,SSE2,SSE3,SSSE3,SSE4_1,SSE4_2,AVX,AVX2&cats=Cast&expand=605,617

Please try this Intrinsic typecasting methods and let us know if you face any issues.

Thanks & Regards

Noorjahan. 

[mikeitexpert]

thanks for the hint ... I did like successfully below in case others are interested 

#include <iostream>
#include <emmintrin.h>
#include <immintrin.h>

int main(){

	float *a = (float*) _mm_malloc((1 << 4)*sizeof(float), 512/8);
	float *as = (float*) _mm_malloc((1 << 4)*sizeof(float), 512/8);

	for(int i = 0; i < (1 << 4); i++){
		a[i] = i;
		as[i] = i;
	}
	float seri_sum, vec_sum, vec_parallel_sum;

	for (int i = 1; i < (1 << 4); i++){
		as[i] += as[i-1];
	}
	seri_sum = as[(1 << 4)-1];
	std::cout << "seri_sum" << " = " << seri_sum << std::endl;

	float sum = 0;
	for(int i = 0; i < (1 << 4); i+=((128)/(sizeof(float)*8))){
		
		__m128 vsum; 
		vsum = _mm_load_ps1(&sum);
		
		__m128 vec;
		vec = _mm_load_ps(&a[i]);
		
		
		
		vsum = _mm_add_ps(vsum,vec);
		for(int j = 1; j < ((128)/(sizeof(float)*8)); j++){
			vec = _mm_castsi128_ps(_mm_bslli_si128(_mm_castps_si128(vec),sizeof(float)));
			vsum = _mm_add_ps(vsum,vec);
		}
		_mm_store_ps(&a[i],vsum);
		sum = a[i+((128)/(sizeof(float)*8))-1];
	}
	vec_sum = sum;
	std::cout << "vec_sum" << " = " << vec_sum << std::endl;

	_mm_free(a);
	_mm_free(as);
	return 0;
}

Issue resolved feel free to close  this thread.

Much appreciate it.

[NoorjahanSk_Intel]

Hi,

Thank you for accepting as a solution.

we will no longer respond to this thread.

If you require any additional assistance from Intel, please start a new thread.

Any further interaction in this thread will be considered community only.

Thanks & Regards

Noorjahan