Context
I'm writing some high-performance code for ARM64 using NEON SIMD instructions, which I am trying to further optimize. I only use integer operations, no floating-point. This code is fully CPU- or memory-bound: it does not perform system calls or I/O of any kind (filesystem, networking, or anything else). The code is single-threaded by design -- any parallelism should be handled by calling the code from different CPUs with different arguments. The data working set should be small enough to fit in my CPU's L1 D-cache, and if it overflows a little, it will definitely fit in L2 with lots of space to spare.
My development environment is an Apple laptop with the M1 processor, running macOS; as such, the prime choice for a performance investigation tool is Apple's Instruments. I know VTune has some more advanced features such as top-down microarchitecture analysis, but evidently this isn't available for ARM.
The problem
I had an idea that, at a high level, works like this: a certain function f(x, y) can be broken down into two functions g() and h(). I can calculate x2 = g(x), y2 = g(y) and then h(x2, y2), obtaining the same result as f(x, y). However, it turns out that I compute f() many times with different combinations of the same input arguments. By applying all these inputs to g() and caching their outputs, I can directly call the output of h()with these cached values and save some time recomputing theg()-part of f()`.
Benchmarks
I confirmed the basic idea is sound by microbenchmarking with Google Benchmark. If f() takes 100 X (where X is some arbitrary unit of time), then each call to g() takes 14 X, and a call to h() takes 78 X. While it's longer to call g() twice then h() rather than f(), suppose I need to compute f(x, y) and f(x, z), which would ordinarily take 200 X. I can instead compute x2 = g(x), y2 = g(y) and z2 = g(z), taking 3*14 = 42 X, and then h(x2, y2) and h(x2, z2), taking 2*78 = 156 X. In total, I spend 156 + 42 = 198 X, which is already less than 200 X, and the savings would add up for larger examples, up to maximum of 22%, since this is how much less h() costs compared to f() (assuming I compute h() much more often than g()). This would represent a significant speedup for my application.
I proceeded to test this idea on a more realistic example: I have some code which does a bunch of things, plus 3 calls to f() which, among themselves, use combinations of the same 2 arguments. So, I replace 3 calls to f() by 2 calls to g() and 3 calls to h(). The benchmarks above indicate this should reduce execution time by 3*100 - 2*14 - 3*78 = 38 X. However, benchmarking the modified code shows that execution time increases by ~700 X!
I tried replacing each call to f() individually with 2 calls to g() for its arguments and a call to h(). This should increase execution time by 2*14 + 78 - 100 = 6 X, but instead, execution time increases by 230 X (not coincidentally, approximately 1/3 of 700 X).
Performance counter results using Apple Instruments
To bring some data to the discussion, I ran both codes under Apple Instruments using the CPU counters template, monitoring some performance counters I thought might be relevant.
For reference, the original code executes in 7.6 seconds (considering only number of iterations times execution time per iteration, i.e. disregarding Google Benchmark overhead), whereas the new code executes in 9.4 seconds; i.e. a difference of 1.8 seconds. Both versions use the exact same number of iterations and work on the same input, producing the same output. The code runs on the M1's performance core, which I assume is running at its maximum 3.2 GHz clock speed.
Parameter
Original code
New code
Total cycles
22,199,155,777
27,510,276,704
MAP_DISPATCH_BUBBLE
78,611,658
6,438,255,204
L1D_CACHE_MISS_LD
892,442
1,808,341
L1D_CACHE_MISS_ST
2,163,402
4,830,661
L1I_CACHE_MISS_DEMAND
2,620,793
7,698,674
INST_SIMD_ALU
79,448,291,331
78,253,076,740
INST_SIMD_LD
17,254,640,147
16,867,679,279
INST_SIMD_ST
14,169,912,790
14,029,275,120
INST_INT_ALU
4,512,600,211
4,379,585,445
INST_INT_LD
550,965,752
546,134,341
INST_INT_ST
455,541,070
455,298,056
INST_ALL
119,683,934,968
118,972,558,207
MAP_STALL_DISPATCH
6,307,551,337
5,470,291,508
SCHEDULE_UOP
116,252,941,232
113,882,670,763
MAP_REWIND
16,293,616
11,787,119
FLUSH_RESTART_OTHER_NONSPEC
58,616
90,955
FETCH_RESTART
27,417,457
28,119,690
BRANCH_MISPRED_NONSPEC
432,761
465,697
L1I_TLB_MISS_DEMAND
754,161
1,492,705
L2_TLB_MISS_INSTRUCTION
485,702
1,217,474
MMU_TABLE_WALK_INSTRUCTION
486,812
1,219,082
BRANCH_MISPRED_NONSPEC
377,750
440,382
INST_BRANCH
1,194,614,553
1,151,040,641
Instruments won't let me add all these counters to the same run, so some results are from different runs. However, since the code is fully deterministic and runs the same number of iterations, any differences between runs should be just random noise.
EDIT: playing around with Instruments, I found one performance counter that has wildly differing values between the original code and the new code, which is MAP_DISPATCH_BUBBLE. Still doing research on what it means, whether it might explain the issues I'm seeing, and how to work around this.
EDIT 2: I decided to test this code on other ARM processors I have access to (Cortex-X2 and Cortex-A72). On the Cortex-X2, both versions perform identically, and on the Cortex-A72, there was a small (~1.5%) increase in performance with the new code. So I'm more inclined than ever to believe that I hit an M1 front-end bottleneck.
Hypotheses and data analysis
Having faced previous performance problems with this code base before, some ideas sprung to mind:
Memory alignment: SIMD code is sometimes sensitive to memory alignment, particularly for memory-bound code, which I suspect my code may be. However, adding or removing __attribute__((aligned(64))) made no difference, so I don't think that's it.
D-cache misses: the new code allocates some new arrays to cache the output of g(), so it might lead to more cache misses. And indeed there are 3.6 million more L1 D-cache misses (load + store) than the original code. However, as I've mentioned at the beginning, the working set easily fits into L2. Assuming a 10-cycle L2 cache miss cost, that's only 36 million cycles. At 3.2 GHz, that's just 1.1 ms, i.e. < 0.1% of the observed performance difference.
I-cache misses: a similar situation: there's an extra 5.1 million L1 I-cache misses, but at a 10-cycle cost, we're looking at 1.6 ms, again < 0.1% of the observed performance difference.
Inlining/unrolling: I employ aggressive inlining and loop unrolling on my code, as well as LTO and unity builds, since performance is the #1 priority and code size is irrelevant (unless it affects performance via e.g. I-cache misses). I considered the possibility that the new code might be inlining/unrolling less aggressively due to the compiler hitting some kind of heuristic for maximum code size. This might result in more instructions being executed, such as compares/branches for loops, and CALL/RET and function prologues/epilogues for function call. However, the table shows that the new code executes a bit fewer instructions of each kind (as I would expect), and of course, in total (INST_ALL).
Somehow, the original code simply achieves a higher IPC, and I have no idea why. Also, to be clear: both codes perform the same operation using the same algorithm. What I did was to basically the code for f() (a bunch of function calls to other subroutines) between g() and h().
The question
This brings me to my question: what could possibly be making the new code run slower than the old code? What other performance counters could I look at in Instruments to give me insight into this issue?
Beyond answers to this specific question, I'm looking for general advice on how to approach similar problems like this in the future. I've found some books about debugging performance problems, but they generally fall into two camps. The first just describes the profiling process I'm familiar with: find out which functions take the longest to execute and optimize them. The second is represented by books like Systems Performance: Enterprise and the Cloud and The Every Computer Performance Book, and is closer to what I'm looking for. However, they look at system-level issues like I/O, kernel calls, etc.; the kind of code I write is CPU- and maybe memory-bound, with many opportunities to convert to SIMD, and no interaction with the outside world. Basically, I'd like to know how to design meaningful experiments using a profiler and CPU performance counters (cycle counters, cache misses, instructions executed by type such as ALU, memory, etc.) to solve these kinds of performance issues with my code when they arise.
I got access to the AMD Zen4 server and tested AVX-512 packed double performance. I chose Harmonic Series Sum[1/n over positive integers] and compared the performance using standard doubles, AVX2 (4 packed doubles) and AVX-512 (8 packed doubles). The test code is here.
AVX-256 version runs four times faster than the standard double version. I was expecting the AVX-512 version to run two times faster than the AVX-256 version, but there was barely any improvement in runtimes:
Method Runtime (minutes:seconds)
HarmonicSeriesPlain 0:41.33
HarmonicSeriesAVX256 0:10.32
HarmonicSeriesAVX512 0:09.82
I was scratching my head over the results and tested individual operations. See full results. Here is runtime for the division:
Method Runtime (minutes:seconds)
div_plain 1:53.80
div_avx256f 0:28.47
div_avx512f 0:14.25
Interestingly, div_avx256f takes 28 seconds, while HarmonicSeriesAVX256 takes only 10 seconds to complete. HarmonicSeriesAVX256 is doing more operations than div_avx256f - summing up the results and increasing the denominator each time (the number of packed divisions is the same). The speed-up has to be due to the instructions pipelining.
However, I need help finding out more details.
The analysis with the llvm-mca (LLVM Machine Code Analyzer) fails because it does not support Zen4 yet:
gcc -O3 -mavx512f -mfma -S "$file" -o - | llvm-mca -iterations 10000 -timeline -bottleneck-analysis -retire-stats
error: found an unsupported instruction in the input assembly sequence.
note: instruction: vdivpd %zmm0, %zmm4, %zmm2
On the Intel platform, I would use
perf stat -M pipeline binary
to find more details, but this metricgroup is not available on Zen4. Any more suggestions on how to analyze the instructions pipelining on Zen4? I have tried these perf stat events:
cycles,stalled-cycles-frontend,stalled-cycles-backend,cache-misses,sse_avx_stalls,fp_ret_sse_avx_ops.all,fp_ret_sse_avx_ops.div_flops,fpu_pipe_assignment.total,fpu_pipe_assignment.total0,
fpu_pipe_assignment.total1,fpu_pipe_assignment.total2,fpu_pipe_assignment.total3
and got the results here.
From this I can see, that the workload is backed bound. AMD's performance event fp_ret_sse_avx_ops.all ( the number of retired SSE/AVX operations) helps, but I still want to get better insights into instructions pipelining on Zen4. Any tips?
Zen 4 execution units are mostly 256-bit wide; handling a 512-bit uop occupies it for 2 cycles. It's normal that 512-bit vectors don't have more raw throughput for any math instructions in general on Zen 4. Although using them on Zen4 does mean more work per uop so out-of-order exec has an easier time.
Or in the case of division, they're occupied for longer since division isn't fully pipelined, like on all modern CPUs. Division is hard to implement.
On Intel Ice Lake for example, divpd throughput is 2 doubles per 4 clocks whether you're using 128-bit, 256-bit, or 512-bit vectors. 512-bit takes extra uops, so we can infer that the actual divider execution unit is 256-bit wide in Ice Lake, but that divpd xmm can use the two halves of it independently. (Unlike AMD).
https://agner.org/optimize/ has instructing timing tables (and his microarch PDF has details on how CPUs work that are essential to making sense of them). https://uops.info/ also has good automated microbenchmark results, free from typos and other human error except sometimes in choosing what to benchmark. (But the actual instruction sequences tested are available, so you can check what they actually tested.) Unfortunately they don't yet have Zen 4 results up, only up to Zen 3.
Zen4 has execution units 256-bit wide for the most part, so 512-bit instructions are single uop but take 2 cycles on most execution units. (Unlike Zen1 where they took 2 uops and thus hurt OoO exec). And it has efficient 512-bit shuffles, and lets you use the power of new AVX-512 instructions for 256-bit vector width, which is where a lot of the real value is. (Better shuffles, masking, vpternlogd, vector popcount, etc.)
Division isn't fully pipelined on any modern x86 CPU. Even on Intel CPUs 512-bit vdivpd zmm has about the same doubles-per-clock throughput as vdivpd ymm (Floating point division vs floating point multiplication has some older data on the YMM vs. XMM situation which is similar, although Zen4 apparently can't send different XMM vectors through the halves of its 256-bit-wide divide unit; vdivpd xmm has the same instruction throughput as vdivpd ymm)
Fast-reciprocal + Newton iterations
For something that's almost entirely bottlenecked on division throughput (not front-end or other ports), you might consider approximate-reciprocal with a Newton-Raphson iteration or two to refine the accuracy to close to 1 ulp. (Not quite the 0.5 ulp you'd get from exact division).
AVX-512 has vrcp14pd approx-reciprocal for packed-double. So two rounds of Newton iterations should double the number of correct bits each time, to 28 then 56 (which is more than the 53-bit mantissa of a double). Fast vectorized rsqrt and reciprocal with SSE/AVX depending on precision mostly talks about rsqrt, but similar idea.
SSE/AVX1 only had single-precision versions of the fast-reciprocal and rsqrt instructions, with only 12-bit precision. e.g. rcpps.
AVX-512ER has 28-bit precision versions, but only Xeon Phi ever had those; mainstream CPUs haven't included them. (Xeon Phi had very vdivps / pd exact division, so it was much better to use the reciprocals.)
I got the answer for the question from title: How to analyze the instructions pipelining on Zen4? directly from AMD:
For determining if a workload is backend-bound, the recommended
method on Zen 4 is to use the pipeline utilization metrics. We are
the process of providing similar metrics and metric groups through
the perf JSON event files for Zen 4 and they will be out very soon.
Read more details in this email thread
AMD has already posted the patches.
Before the patches land in favorite Linux distribution, you can use the raw events on Zen4. Check this example
I stumbled upon a peculiar performance issue when running the following c++ code on some Intel Xeon processors:
// array_a contains permutation of [0, n - 1]
// array_b and inverse are initialized arrays
for (int i = 0; i < n; ++i) {
array_b[i] = array_a[i];
inverse[array_b[i]] = i;
}
The first line of the loop sequentially copies array_a into array_b (very few cache misses expected). The second line computes the inverse of array_b (many cache misses expected since array_b is a random permutation). We may also split the code into two separate loops:
for (int i = 0; i < n; ++i)
array_b[i] = array_a[i];
for (int i = 0; i < n; ++i)
inverse[array_b[i]] = i;
I would have expected the two versions (single vs. dual loop) to perform almost identically on relatively modern hardware. However, it appears that some Xeon processors are incredibly slow when executing the single loop version.
Below you can see the wall time in nano-seconds divided by n when running the snippet on a range of different processors. For the purpose of testing, the code was compiled using GCC 7.5.0 with flags -O3 -funroll-loops -march=native on a system with a Xeon E5-4620v4. Then, the same binary was used on all systems, using numactl -m 0 -N 0 on systems with multiple NUMA domains.
The used code is available on github. The interesting stuff is in the file runner.cpp.
[EDIT:] The assembly is provided here.
[EDIT:] New results including AMD EPYC.
On the various i7 models, the results are mostly as expected. Using the single loop is only slightly slower than the dual loops. This also holds for the Xeon E3-1271v3, which is basically the same hardware as an i7-4790. The AMC EPYC 7452 performed best by far, with almost no difference between the single and dual loop implementation. However, on the Xeon E5-2690v4 and E5-4620v4 systems using the single loop is incredibly slow.
In previous tests I also observed this strange performance issue on Xeon E5-2640 and E5-2640v4 systems. In contrast to that, there was no performance issue on several AMD EPYC and Opteron systems, and also no issues on Intel i5 and i7 mobile processors.
My question to the CPU experts thus is: Why does Intel's most high-end product-line perform so poorly compared to other CPUs? I am by far no expert in CPU architectures, so your knowledge and thoughts are much appreciated!
Maybe this is related to avx-512 frequendy throttling on Intel processors. These instructions generate a lot of heat, and if used on some circunstances the processor reduces the working frequency.
Here are some benchmarks for OpenSSL that show the effect. There is a rant from Linus Torvalds on this topic.
If avx-512 instructions are generated with "-march=native" you may be suffering this effect. Try disabling avx-512 with:
gcc -mno-avx512f
I have a C/C++ program which involves intensive 32-bit floating-point matrix math computations such as addition, subtraction, multiplication, division, etc.
Can I speed up my program by converting 32-bit floating-point numbers into 16-bit fixed-point numbers ? How much speed gain can I get ?
Currently I'm working on a Intel I5 CPU. I'm using Openblas to perform the matrix calculations. How should I re-implement Openblas functions such as cblas_dgemm to perform fixed-point calculations ?
I know that SSE(Simple SIMD Extensions) operates on 4x32=8x16=128 bit data at one time, i.e., 4 32-bit floating-point type or 8 16-bit fixed-point type. I guess that after conversion from 32-bit floating-point to 16-bit fixed-point, my program would be twice faster.
Summary: Modern FPU hardware is hard to beat with fixed-point, even if you have twice as many elements per vector.
Modern BLAS library are typically very well tuned for cache performance (with cache blocking / loop tiling) as well as for instruction throughput. That makes them very hard to beat. Especially DGEMM has lots of room for this kind of optimization, because it does O(N^3) work on O(N^2) data, so it's worth transposing just a cache-sized chunk of one input, and stuff like that.
What might help is reducing memory bottlenecks by storing your floats in 16-bit half-float format. There is no hardware support for doing math on them in that format, just a couple instructions to convert between that format and normal 32-bit element float vectors while loading/storing: VCVTPH2PS (__m256 _mm256_cvtph_ps(__m128i)) and VCVTPS2PH (__m128i _mm256_cvtps_ph(__m256 m1, const int imm8_rounding_control). These two instructions comprise the F16C extension, first supported by AMD Bulldozer and Intel IvyBridge.
IDK if any BLAS libraries support that format.
Fixed point:
SSE/AVX doesn't have any integer division instructions. If you're only dividing by constants, you might not need a real div instruction, though. So that's one major stumbling block for fixed point.
Another big downside of fixed point is the extra cost of shifting to correct the position of the decimal (binary?) point after multiplies. That will eat into any gain you could get from having twice as many elements per vector with 16-bit fixed point.
SSE/AVX actually has quite a good selection of packed 16-bit multiplies (better than for any other element size). There's packed multiply producing the low half, high half (signed or unsigned), and even one that takes 16 bits from 2 bits below the top, with rounding (PMULHRSW.html). Skylake runs those at two per clock, with 5 cycle latency. There are also integer multiply-add instructions, but they do a horizontal add between pairs of multiply results. (See Agner Fog's insn tables, and also the x86 tag wiki for performance links.) Haswell and previous don't have as many integer-vector add and multiply execution units. Often code bottlenecks on total uop throughput, not on a specific execution port anyway. (But a good BLAS library might even have hand-tuned asm.)
If your inputs and outputs are integer, it's often faster to work with integer vectors, instead of converting to floats. (e.g. see my answer on Scaling byte pixel values (y=ax+b) with SSE2 (as floats)?, where I used 16-bit fixed-point to deal with 8-bit integers).
But if you're really working with floats, and have a lot of multiplying and dividing to do, just use the hardware FPUs. They're amazingly powerful in modern CPUs, and have made fixed-point mostly obsolete for many tasks. As #Iwill points out, FMA instructions are another big boost for FP throughput (and sometimes latency).
Integer add/subtract/compare instructions (but not multiply) are also lower latency than their FP counterparts.