A.nechi A.nechi - 1 year ago 63
C Question

Macro similar to SIMD instruction

I' trying to create some macros that are somehow similar to SIMD intrinsics. And the reason why I want to create these macros is that I'm working with a simulator called "Gem5" that does not support SIMD.

Well, first of all, i created a structure that defines a vector of 4 packed single precision floating point elements like so:

typedef struct
float vec1;
float vec2;
float vec3;
float vec4;
} __m128 __attribute__((aligned(16)));

then I've created the ADD macro:

#define __M128_MM_ADD_PS(dest, a,b) \
{ \
(dest)->vec1 = (a)->vec1 + (b)->vec1; \
(dest)->vec2 = (a)->vec2 + (b)->vec2; \
(dest)->vec3 = (a)->vec3 + (b)->vec3; \
(dest)->vec4 = (a)->vec4 + (b)->vec4; \

and another macro for storing the results in a float array:

#define __M128_MM_MOVA_PS(dest, a) \
{ \
dest[0] = a->vec1; \
dest[1] = a->vec2; \
dest[2] = a->vec3; \
dest[3] = a->vec4; \

and to them I've declared the variables as __m128 and a float array to store the results, this way(a little example):

void foo(){
__m128 bfly0_rv, x_n2_vec, x_N2_vec;
float *x;


and I'm getting these kind of error messages:

for the ADD macro
error: invalid type argument of ‘->’ (have ‘__m128’)
(dest)->vec4 = (a)->vec4 + (b)->vec4;

for the storing macro
error: invalid type argument of ‘->’ (have ‘__m128’)
dest[3] = a->vec4;

Anyone could have an idea about that?

Answer Source

All of your macros will work with pointers to __m128, but you pass just __m128 not __m128 *. Just replace -> with . inside macros or prepend macro arguments with &:

__M128_MM_ADD_PS(&bfly0_rv, &x_n2_vec, &x_N2_vec);
__M128_MM_MOVA_PS(&x[n2], &bfly0_rv);

Also don't forget to enclose dest and a in braces in __M128_MM_MOVA_PS.

P.S. It would be better to use do { ... } while(0) instead of { ... } in macro definitions.