SSE load/store memory transactions - c++

Question

There are two ways for memory-register interactions in use SSE intrinsics:
Intermediate pointers:
void f_sse(float *input, float *output, unsigned int n)
{
_m128 *input_sse = reinterpret_cast<__m128*>(input);//Input intermediate pointer
_m128 *output_sse = reinterpret_cast<__m128*>(output);//Output intermediate pointer
_m128 s = _mm_set1_ps(0.1f);
auto loop_size = n/4;
for(auto i=0; i<loop_size; ++i)
output_sse[i] = _mm_add_ps(input_sse[i], s);
}
Explicit fetch/store:
void f_sse(float *input, float *output, unsigned int n)
{
_m128 input_sse, output_sse, result;
_m128 s = _mm_set1_ps(0.1f);
for(auto i=0; i<n; i+=4)
{
input_sse = _mm_load_ps(input+i);
result = _mm_add_ps(input_sse, s);
_mm_store_ps(output+i, result);
}
}
What's the difference between mentioned approaches and which method is better in terms of perfomance? input and output pointers are aligned by _mm_malloc().

Answer 1

Compiled with g++ at optimization level O3 the assembly code of the inner loop (using objdump -d) are
20: 0f 28 04 07 movaps (%rdi,%rax,1),%xmm0
24: 0f 58 c1 addps %xmm1,%xmm0
27: 0f 29 04 06 movaps %xmm0,(%rsi,%rax,1)
2b: 48 83 c0 10 add $0x10,%rax
2f: 48 39 d0 cmp %rdx,%rax
32: 75 ec jne 20 <_Z5f_ssePfS_j+0x20>
and
10: 0f 28 04 07 movaps (%rdi,%rax,1),%xmm0
14: 83 c1 04 add $0x4,%ecx
17: 0f 58 c1 addps %xmm1,%xmm0
1a: 0f 29 04 06 movaps %xmm0,(%rsi,%rax,1)
1e: 48 83 c0 10 add $0x10,%rax
22: 39 ca cmp %ecx,%edx
24: 77 ea ja 10 <_Z5f_ssePfS_j+0x10>
They are pretty similar. In the first g++ manage to use only one counter (only one add instruction). So I guess its better.

Answer 2

I compiled both of your samples with g++ -O2, and the main difference I found was that the value in edx (n) is used differently, which leads to slightly different code.
First function:
0000000000000000 <_Z6f_sse2PfS_j>:
0: c1 ea 02 shr $0x2,%edx # loop_size = n / 4.
3: 85 d2 test %edx,%edx
5: 74 2d je 34 <_Z6f_sse2PfS_j+0x34>
7: 83 ea 01 sub $0x1,%edx
a: 0f 28 0d 00 00 00 00 movaps 0x0(%rip),%xmm1 # 11 <_Z6f_sse2PfS_j+0x11>
11: 48 83 c2 01 add $0x1,%rdx
15: 31 c0 xor %eax,%eax
17: 48 c1 e2 04 shl $0x4,%rdx // Adjust for loop size vs. index.
1b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
20: 0f 28 04 07 movaps (%rdi,%rax,1),%xmm0
24: 0f 58 c1 addps %xmm1,%xmm0
27: 0f 29 04 06 movaps %xmm0,(%rsi,%rax,1)
2b: 48 83 c0 10 add $0x10,%rax
2f: 48 39 d0 cmp %rdx,%rax
32: 75 ec jne 20 <_Z6f_sse2PfS_j+0x20>
34: f3 c3 repz retq
Second function:
0000000000000000 <_Z5f_ssePfS_j>:
0: 85 d2 test %edx,%edx
2: 74 22 je 26 <_Z5f_ssePfS_j+0x26>
4: 0f 28 0d 00 00 00 00 movaps 0x0(%rip),%xmm1 # b <_Z5f_ssePfS_j+0xb>
b: 31 c0 xor %eax,%eax
d: 31 c9 xor %ecx,%ecx
f: 90 nop
10: 0f 28 04 07 movaps (%rdi,%rax,1),%xmm0
14: 83 c1 04 add $0x4,%ecx
17: 0f 58 c1 addps %xmm1,%xmm0
1a: 0f 29 04 06 movaps %xmm0,(%rsi,%rax,1)
1e: 48 83 c0 10 add $0x10,%rax
22: 39 ca cmp %ecx,%edx
24: 77 ea ja 10 <_Z5f_ssePfS_j+0x10>
26: f3 c3 repz retq
I also looked at the code generated, and came up with this:
void f_sse2(float *input, float *output, unsigned int n)
{
__m128 *end = reinterpret_cast<__m128*>(&input[n]);
__m128 *input_sse = reinterpret_cast<__m128*>(input);//Input intermediate pointer
__m128 *output_sse = reinterpret_cast<__m128*>(output);//Output intermediate pointer
__m128 s = _mm_set1_ps(0.1f);
while(input_sse < end)
*output_sse++ = _mm_add_ps(*input_sse++, s);
}
which generates this code:
0000000000000000 <_Z6f_sse2PfS_j>:
0: 89 d2 mov %edx,%edx
2: 48 8d 04 97 lea (%rdi,%rdx,4),%rax
6: 48 39 c7 cmp %rax,%rdi
9: 73 23 jae 2e <_Z6f_sse2PfS_j+0x2e>
b: 0f 28 0d 00 00 00 00 movaps 0x0(%rip),%xmm1 # 12 <_Z6f_sse2PfS_j+0x12>
12: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
18: 0f 28 07 movaps (%rdi),%xmm0
1b: 48 83 c7 10 add $0x10,%rdi
1f: 0f 58 c1 addps %xmm1,%xmm0
22: 0f 29 06 movaps %xmm0,(%rsi)
25: 48 83 c6 10 add $0x10,%rsi
29: 48 39 f8 cmp %rdi,%rax
2c: 77 ea ja 18 <_Z6f_sse2PfS_j+0x18>
2e: f3 c3 repz retq
Which I think may be a tiny bit more efficient, but probably not worth changing it for. But it gave me something to do for 15 minutes.