SSE Instructions

Question

I have a question regarding SSE instruction.

I hope this is the right place to ask such a question if not pls let me know and I will remove this question.

My goal is to use SSE instructions to execute calculations on 3 chars in parallel.

I have a typedef struct which has the attribute that it is packed

typedef struct
{
        unsigned char x;
        unsigned char y;
        unsigned char z;
} __attribute__((packed)) Number;

For each char I have to go through a certain calculation.

As an example:

((Number[0].x * 20)  / 256);

I have to do a small calculation for every char and then add them together.

Since I have to write the code in assembly I have already done some research and stumble upon this instruction:

__m128i _mm_add_epi8 (__m128i a, __m128i b)

As far as I am concerned this should add two values (who have each the size of 8 bytes) together and save the result.

At least that's how I understand it: From this link

But since we only add two values together this defeats the whole purpose of executing multiple instructions at once.

Any help would be very apricated. Kind regards!

That's SSE not SEE. Yes, it does add two values, but it does 16 pairs at once. You could use 3 of those to perform calculation for 16 `Number` instances. — Jester, Dec 06 '18 at 17:11
And if you want to divide by 256, you're going to need a larger type than char... — Shawn, Dec 06 '18 at 17:12
you dont think the compiler will do that for you , modern compilers are wickedly good at optimization. — pm100, Dec 06 '18 at 17:20
@pm100, you'd think so, but they are not that hot when dealing with `char`s (unless there's an option i'm forgetting): https://gcc.godbolt.org/z/htGiqy — , Dec 06 '18 at 18:02

score 1 · Accepted Answer · answered Dec 06 '18 at 18:09

1

If you could provide more information about how you're actually using this it might be possible to better optimize it, but based on what you wrote I guess you'd want something like _mm_srli_epi32(_mm_mullo_epi32(_mm_set_epi32(n.x, n.y, n.z, 0), _mm_set1_epi32(20)), 8). It requires SSE 4.1, but if you want something which works for SSE 2 see SSE multiplication of 4 32-bit integers for a _mm_mullo_epi32 replacement.

You didn't specify what you want to do with the result, but you can use something like ((int*) &r_sse)[i] to access the results, where i is 1 for z, 2 for y, and 1 for x.

answered Dec 06 '18 at 18:09

nemequ

16,623
1
43
62

I am using it to convert an Image to gray. I have 3 chars red, blue and green and each of them has to go through a certain calculation and then need to be added together. The result of the calculation is then passed to each char(so every char gets the same value) and the result should be a gray pixel. – Mark Dec 06 '18 at 18:16
1

`_mm_mullo_epi32` is kind of slow though, can't we use 16bit arithmetic here? – harold Dec 07 '18 at 15:31

SSE Instructions

1 Answers1