问题
I have a multiply-add kernel inside my application and I want to increase its performance.
I use an Intel Core i7-960 (3.2 GHz clock) and have already manually implemented the kernel using SSE intrinsics as follows:
for(int i=0; i<iterations; i+=4) {
y1 = _mm_set_ss(output[i]);
y2 = _mm_set_ss(output[i+1]);
y3 = _mm_set_ss(output[i+2]);
y4 = _mm_set_ss(output[i+3]);
for(k=0; k<ksize; k++){
for(l=0; l<ksize; l++){
w = _mm_set_ss(weight[i+k+l]);
x1 = _mm_set_ss(input[i+k+l]);
y1 = _mm_add_ss(y1,_mm_mul_ss(w,x1));
…
x4 = _mm_set_ss(input[i+k+l+3]);
y4 = _mm_add_ss(y4,_mm_mul_ss(w,x4));
}
}
_mm_store_ss(&output[i],y1);
_mm_store_ss(&output[i+1],y2);
_mm_store_ss(&output[i+2],y3);
_mm_store_ss(&output[i+3],y4);
}
I know I can use packed fp vectors to increase the performance and I already did so succesfully, but I want to know why the single scalar code isn't able to meet the processor's peak performance.
The performance of this kernel on my machine is ~1.6 FP operations per cycle, while the maximum would be 2 FP operations per cycle (since FP add + FP mul can be executed in parallel).
If I'm correct from studying the generated assembly code, the ideal schedule would look like follows, where the mov instruction takes 3 cycles, the switch latency from the load domain to the FP domain for the dependent instructions takes 2 cycles, the FP multiply takes 4 cycles and the FP add takes 3 cycles. (Note that the dependence from the multiply -> add doesn't incur any switch latency because the operations belong to the same domain).
According to the measured performance (~80% of the maximum theoretical performance) there is an overhead of ~3 instructions per 8 cycles.
I am trying to either:
- get rid of this overhead, or
- explain where it comes from
Of course there is the problem with cache misses & data misalignment which can increase the latency of the move instructions, but are there any other factors that could play a role here? Like register read stalls or something?
I hope my problem is clear, thanks in advance for your responses!
Update: The assembly of the inner-loop looks as follows:
...
Block 21:
movssl (%rsi,%rdi,4), %xmm4
movssl (%rcx,%rdi,4), %xmm0
movssl 0x4(%rcx,%rdi,4), %xmm1
movssl 0x8(%rcx,%rdi,4), %xmm2
movssl 0xc(%rcx,%rdi,4), %xmm3
inc %rdi
mulss %xmm4, %xmm0
cmp $0x32, %rdi
mulss %xmm4, %xmm1
mulss %xmm4, %xmm2
mulss %xmm3, %xmm4
addss %xmm0, %xmm5
addss %xmm1, %xmm6
addss %xmm2, %xmm7
addss %xmm4, %xmm8
jl 0x401b52 <Block 21>
...
回答1:
I noticed in the comments that:
- The loop takes 5 cycles to execute.
- It's "supposed" to take 4 cycles. (since there's 4 adds and 4 mulitplies)
However, your assembly shows 5 SSE movssl instructions. According to Agner Fog's tables all floating-point SSE move instructions are at least 1 inst/cycle reciprocal throughput for Nehalem.
Since you have 5 of them, you can't do better than 5 cycles/iteration.
So in order to get to peak performance, you need to reduce the # of loads that you have. How you can do that I can't see immediately this particular case - but it might be possible.
One common approach is to use tiling. Where you add nesting levels to improve locality. Although it's used mostly for improving cache access, it can also be used in registers to reduce the # of load/stores that are needed.
Ultimately, your goal is to reduce the number of loads to be less than the numbers of add/muls. So this might be the way to go.
回答2:
Thanks a lot for your answers, this explained a lot. Continuing on my question, when i use packed instructions instead of scalar instructions the code using intrinsics would look very similar:
for(int i=0; i<size; i+=16) {
y1 = _mm_load_ps(output[i]);
…
y4 = _mm_load_ps(output[i+12]);
for(k=0; k<ksize; k++){
for(l=0; l<ksize; l++){
w = _mm_set_ps1(weight[i+k+l]);
x1 = _mm_load_ps(input[i+k+l]);
y1 = _mm_add_ps(y1,_mm_mul_ps(w,x1));
…
x4 = _mm_load_ps(input[i+k+l+12]);
y4 = _mm_add_ps(y4,_mm_mul_ps(w,x4));
}
}
_mm_store_ps(&output[i],y1);
…
_mm_store_ps(&output[i+12],y4);
}
The measured performance of this kernel is about 5.6 FP operations per cycle, although i would expect it to be exactly 4x the performance of the scalar version, i.e. 4.1,6=6,4 FP ops per cycle.
Taking the move of the weight factor into account (thanks for pointing that out), the schedule looks like:
It looks like the schedule doesn't change, although there is an extra instruction after the movss operation that moves the scalar weight value to the XMM register and then uses shufps to copy this scalar value in the entire vector. It seems like the weight vector is ready to be used for the mulps in time taking the switching latency from load to the floating point domain into account, so this shouldn't incur any extra latency.
The movaps (aligned, packed move),addps & mulps instructions that are used in this kernel (checked with assembly code) have the same latency & throughput as their scalar versions, so this shouldn't incur any extra latency either.
Does anybody have an idea where this extra cycle per 8 cycles is spent on, assuming the maximum performance this kernel can get is 6.4 FP ops per cycle and it is running at 5.6 FP ops per cycle?
Thanks again for all of your help!
回答3:
Making this an answer from my comment.
On a non-server Linux distro I believe the interrupt timer is usually set to 250Hz by default, though that varies by distro it's almost always over 150. That speed is necessary to provide a 30+fps interactive GUI. That interrupt timer is used to preempt code. That means 150+ times per second your code is interrupted and the scheduler code runs and decides what to give more time to. It sounds like you're doing great to simply get 80% of max speed, no problems there. If you need better install say, Ubuntu Server (100Hz default) and tweak the kernel (preemption off) a bit
EDIT: On a 2+ core system this has much less impact as your process will almost definitely be slapped onto one core and more-or-less left to do its own thing.
来源：https://stackoverflow.com/questions/9992054/c-code-loop-performance