Cowmoogun Cowmoogun - 3 months ago 14
C++ Question

Deoptimizing a program for the pipeline in Intel Sandybridge-family CPUs

I've been racking my brain for a week trying to complete this assignment and I'm hoping someone here can lead me toward the right path. Let me start with the instructor's instructions:


Your assignment is the opposite of our first lab assignment, which was to optimize a prime number program. Your purpose in this assignment is to pessimize the program, i.e. make it run slower. Both of these are CPU-intensive programs. They take a few seconds to run on our lab PCs. You may not change the algorithm.

To deoptimize the program, use your knowledge of how the Intel i7 pipeline operates. Imagine ways to re-order instruction paths to introduce WAR, RAW, and other hazards. Think of ways to minimize the effectiveness of the cache. Be diabolically incompetent.


The assignment gave a choice of Whetstone or Monte-Carlo programs. The cache-effectiveness comments are mostly only applicable to Whetstone, but I chose the Monte-Carlo simulation program:

// Un-modified baseline for pessimization, as given in the assignment
#include <algorithm> // Needed for the "max" function
#include <cmath>
#include <iostream>

// A simple implementation of the Box-Muller algorithm, used to generate
// gaussian random numbers - necessary for the Monte Carlo method below
// Note that C++11 actually provides std::normal_distribution<> in
// the <random> library, which can be used instead of this function
double gaussian_box_muller() {
double x = 0.0;
double y = 0.0;
double euclid_sq = 0.0;

// Continue generating two uniform random variables
// until the square of their "euclidean distance"
// is less than unity
do {
x = 2.0 * rand() / static_cast<double>(RAND_MAX)-1;
y = 2.0 * rand() / static_cast<double>(RAND_MAX)-1;
euclid_sq = x*x + y*y;
} while (euclid_sq >= 1.0);

return x*sqrt(-2*log(euclid_sq)/euclid_sq);
}

// Pricing a European vanilla call option with a Monte Carlo method
double monte_carlo_call_price(const int& num_sims, const double& S, const double& K, const double& r, const double& v, const double& T) {
double S_adjust = S * exp(T*(r-0.5*v*v));
double S_cur = 0.0;
double payoff_sum = 0.0;

for (int i=0; i<num_sims; i++) {
double gauss_bm = gaussian_box_muller();
S_cur = S_adjust * exp(sqrt(v*v*T)*gauss_bm);
payoff_sum += std::max(S_cur - K, 0.0);
}

return (payoff_sum / static_cast<double>(num_sims)) * exp(-r*T);
}

// Pricing a European vanilla put option with a Monte Carlo method
double monte_carlo_put_price(const int& num_sims, const double& S, const double& K, const double& r, const double& v, const double& T) {
double S_adjust = S * exp(T*(r-0.5*v*v));
double S_cur = 0.0;
double payoff_sum = 0.0;

for (int i=0; i<num_sims; i++) {
double gauss_bm = gaussian_box_muller();
S_cur = S_adjust * exp(sqrt(v*v*T)*gauss_bm);
payoff_sum += std::max(K - S_cur, 0.0);
}

return (payoff_sum / static_cast<double>(num_sims)) * exp(-r*T);
}

int main(int argc, char **argv) {
// First we create the parameter list
int num_sims = 10000000; // Number of simulated asset paths
double S = 100.0; // Option price
double K = 100.0; // Strike price
double r = 0.05; // Risk-free rate (5%)
double v = 0.2; // Volatility of the underlying (20%)
double T = 1.0; // One year until expiry

// Then we calculate the call/put values via Monte Carlo
double call = monte_carlo_call_price(num_sims, S, K, r, v, T);
double put = monte_carlo_put_price(num_sims, S, K, r, v, T);

// Finally we output the parameters and prices
std::cout << "Number of Paths: " << num_sims << std::endl;
std::cout << "Underlying: " << S << std::endl;
std::cout << "Strike: " << K << std::endl;
std::cout << "Risk-Free Rate: " << r << std::endl;
std::cout << "Volatility: " << v << std::endl;
std::cout << "Maturity: " << T << std::endl;

std::cout << "Call Price: " << call << std::endl;
std::cout << "Put Price: " << put << std::endl;

return 0;
}


The changes I have made seemed to increase the code running time by a second but I'm not entirely sure what I can change to stall the pipeline without adding code. A point to the right direction would be awesome, I appreciate any responses.




Update: the professor who gave this assignment posted some details



The highlights are:


  • It's a second semester architecture class at a community college (using the Hennessy and Patterson textbook).

  • the lab computers have Haswell CPUs

  • The students have been exposed to the
    CPUID
    instruction and how to determine cache size, as well as intrinsics and the
    CLFLUSH
    instruction.

  • any compiler options are allowed, and so is inline asm.

  • Writing your own square root algorithm was announced as being outside the pale



Cowmoogun's comments on the meta thread indicate that it wasn't clear compiler optimizations could be part of this, and assumed
-O0
, and that a 17% increase in run-time was reasonable.

So it sounds like the goal of the assignment was to get students to re-order the existing work to reduce instruction-level parallelism or things like that, but it's not a bad thing that people have delved deeper and learned more.




Keep in mind that this is a computer-architecture question, not a question about how to make C++ slow in general.

Answer

Important background reading: Agner Fog's microarch pdf, and probably also Ulrich Drepper's What Every Programmer Should Know About Memory. See also the other links in the tag wiki, especially Intel's optimization manuals, and David Kanter's analysis of the Haswell microarchitecture, with diagrams.

Very cool assignment; much better than the ones I've seen where students were asked to optimize some code for gcc -O0, learning a bunch of tricks that don't matter in real code. In this case, you're being asked to learn about the CPU pipeline and use that to guide your de-optimization efforts, not just blind guessing. The most fun part of this one is justifying each pessimization with "diabolical incompetence", not intentional malice.


Problems with the assignment wording and code:

The uarch-specific options for this code are limited. It doesn't use any arrays, and much of the cost is calls to exp/log library functions. There isn't an obvious way to have more or less instruction-level parallelism, and the loop-carried dependency chain is very short.

I'd love to see an answer that attempted to get a slowdown from re-arranging the expressions to change the dependencies, to reduce ILP just from dependencies (hazards). I haven't attempted it.

Intel Sandybridge-family CPUs are aggressive out-of-order designs that spend lots of transistors and power to find parallelism and avoid hazards (dependencies) that would trouble a classic RISC in-order pipeline. Usually the only traditional hazards that slow it down are RAW "true" dependencies that cause throughput to be limited by latency.

WAR and WAW hazards for registers are pretty much not an issue, thanks to register renaming. (except for popcnt/lzcnt/tzcnt, which have a false dependency their destination on Intel CPUs, even though it's write-only. i.e. WAW being handled as a RAW hazard + a write). For memory ordering, modern CPUs use store queues to delay commit into cache until retirement, also avoiding WAR and WAW hazards.

The "i7" brand-name was introduced with Nehalem (successor to Core2), and some Intel manuals even say "Core i7" when they seem to mean Nehalem, but they kept the "i7" branding for Sandybridge and later microarchitectures. SnB is when the P6-family evolved into a new species, the SnB-family. In many ways, Nehalem has more in common with Pentium III than with Sandybridge (e.g. register read stalls and ROB-read stalls don't happen on SnB, because it changed to using a physical register file. Also a uop cache and a different internal uop format). The term "i7 architecture" is not useful, because it makes no sense to group the SnB-family with Nehalem but not Core2. (Nehalem did introduce the shared inclusive L3 cache architecture for connecting multiple cores together, though. And also integrated GPUs. So chip-level, the naming makes more sense.)


Summary of the good ideas that diabolical incompetence can justify

Even the diabolically incompetent are unlikely to add obviously useless work or an infinite loop, and making a mess with C++/Boost classes is beyond the scope of the assignment.

  • Multi-thread with a single shared std::atomic<uint64_t> loop counter, so the right total number of iterations happen. Atomic uint64_t is especially bad with -m32 -march=i586. For bonus points, arrange for it to be misaligned, and crossing a page boundary with an uneven split (not 4:4).
  • False sharing for some other non-atomic variable -> memory-order mis-speculation pipeline clears, as well as extra cache misses.
  • Instead of using - on FP variables, XOR the high byte with 0x80 to flip the sign bit, causing store-forwarding stalls.
  • Time each iteration independently, with something even heavier than RDTSC. e.g. CPUID / RDTSC or a time function that makes a system call. Serializing instructions are inherently pipeline-unfriendly.
  • Change multiplies by constants to divides by their reciprocal ("for ease of reading"). div is slow and not fully pipelined.
  • Vectorize the multiply/sqrt with AVX (SIMD), but fail to use vzeroupper before calls to scalar math-library exp() and log() functions, causing AVX<->SSE transition stalls.
  • Store the RNG output in a linked list, or in arrays which you traverse out of order. Same for the result of each iteration, and sum at the end.

Also covered in this answer but excluded from the summary: suggestions that would be just as slow on a non-pipelined CPU, or that don't seem to be justifiable even with diabolical incompetence. e.g. many gimp-the-compiler ideas that produce obviously different / worse asm.


Multi-thread badly

Maybe use OpenMP to multi-thread loops with very few iterations, with way more overhead than speed gain. Your monte-carlo code has enough parallelism to actually get a speedup, though, esp. if we succeed at making each iteration slow. (Each thread computes a partial payoff_sum, added at the end). #omp parallel on that loop would probably be an optimization, not a pessimization.

Multi-thread but force both threads to share the same loop counter (with atomic increments so the total number of iterations is correct). This seems diabolically logical. This means using a static variable as a loop counter. This justifies use of atomic for loop counters, and creates actual cache-line ping-ponging (as long as the threads don't run on the same physical core with hyperthreading; that might not be as slow). Anyway, this is much slower than the un-contended case for lock inc. And lock cmpxchg8b to atomically increment a contended uint64_t on a 32bit system will have to retry in a loop instead of having the hardware arbitrate an atomic inc.

Also create false sharing, where multiple threads keep their private data (e.g. RNG state) in different bytes of the same cache line. (Intel tutorial about it, including perf counters to look at). There's a microarchitecture-specific aspect to this: Intel CPUs speculate on memory mis-ordering not happening, and there's a memory-order machine-clear perf event to detect this, at least on P4. The penalty might not be as large on Haswell. As that link points out, a locked instruction assumes this will happen, avoiding mis-speculation. A normal load speculates that other cores won't invalidate a cache line between when the load executes and when it retires in program-order (unless you use pause). True sharing without locked instructions is usually a bug. It would be interesting to compare a non-atomic shared loop counter with the atomic case. To really pessimize, keep the shared atomic loop counter, and cause false sharing in the same or a different cache line for some other variable.


Random uarch-specific ideas:

If you can introduce any unpredictable branches, that will pessimize the code substantially. Modern x86 CPUs have quite long pipelines, so a mispredict costs ~15 cycles (when running from the uop cache).


Dependency chains:

I think this was one of the intended parts of the assignment.

Defeat the CPU's ability to exploit instruction-level parallelism by choosing an order of operations that has one long dependency chain instead of multiple short dependency chains. Compilers aren't allowed to change the order of operations for FP calculations unless you use -ffast-math, because that can change the results (as discussed below).

To really make this effective, increase the length of a loop-carried dependency chain. Nothing leaps out as obvious, though: The loops as written have very short loop-carried dependency chains: just an FP add. (3 cycles). Multiple iterations can have their calculations in-flight at once, because they can start well before the payoff_sum += at the end of the previous iteration. (log() and exp take many instructions, but not a lot more than Haswell's out-of-order window for finding parallelism: ROB size=192 fused-domain uops, and scheduler size=60 unfused-domain uops. As soon as execution of the current iteration progresses far enough to make room for instructions from the next iteration to issue, any parts of it that have their inputs ready (i.e. independent/separate dep chain) can start executing when older instructions leave the execution units free (e.g. because they're bottlenecked on latency, not throughput.).

The RNG state will almost certainly be a longer loop-carried dependency chain than the addps.


Use slower/more FP operations (esp. more division):

Divide by 2.0 instead of multiplying by 0.5, and so on. FP multiply is heavily pipelined in Intel designs, and has one per 0.5c throughput on Haswell and later. FP divsd/divpd is only partially pipelined. (Although Skylake has an impressive one per 4c throughput for divpd xmm, with 13-14c latency, vs not pipelined at all on Nehalem (7-22c)).

The do { ...; euclid_sq = x*x + y*y; } while (euclid_sq >= 1.0); is clearly testing for a distance, so clearly it would be proper to sqrt() it. :P (sqrt is even slower than div).

As @Paul Clayton suggests, rewriting expressions with associative/distributive equivalents can introduce more work (as long as you don't use -ffast-math to allow the compiler to re-optimize). (exp(T*(r-0.5*v*v)) could become exp(T*r - T*v*v/2.0). Note that while math on real numbers is associative, floating point math is not, even without considering overflow/NaN (which is why -ffast-math isn't on by default). See Paul's comment for a very hairy nested pow() suggestion.

If you can scale the calculations down to very small numbers, then FP math ops take ~120 extra cycles to trap to microcode when an operation on two normal numbers produces a denormal. See Agner Fog's microarch pdf for the exact numbers and details. This is unlikely since you have a lot of multiplies, so the scale factor would be squared and underflow all the way to 0.0. I don't see any way to justify the necessary scaling with incompetence (even diabolical), only intentional malice.


If you can use intrinsics (<immintrin.h>)

Use movnti to evict your data from cache. Diabolical: it's new and weakly-ordered, so that should let the CPU run it faster, right? Or see that linked question for a case where someone was in danger of doing exactly this (for scattered writes where only some of the locations were hot). clflush is probably impossible without malice.

Use integer shuffles between FP math operations to cause bypass delays.

Mixing SSE and AVX instructions without proper use of vzeroupper causes large stalls in pre-Skylake (and a different penalty in Skylake). Even without that, vectorizing badly can be worse than scalar (more cycles spent shuffling data into/out of vectors than saved by doing the add/sub/mul/div/sqrt operations for 4 Monte-Carlo iterations at once, with 256b vectors). add/sub/mul execution units are fully pipelined and full-width, but div and sqrt on 256b vectors aren't as fast as on 128b vectors (or scalars), so the speedup isn't dramatic for double.

exp() and log() don't have hardware support, so that part would require extracting vector elements back to scalar and calling the library function separately, then shuffling the results back into a vector. libm is typically compiled to only use SSE2, so will use the legacy-SSE encodings of scalar math instructions. If your code uses 256b vectors and calls exp without doing a vzeroupper first, then you stall. After returning, an AVX-128 instruction like vmovsd to set up the next vector element as an arg for exp will also stall. And then exp() will stall again when it runs an SSE instruction. This is exactly what happened in this question, causing a 10x slowdown. (Thanks @ZBoson).

See also Nathan Kurz's experiments with Intel's math lib vs. glibc for this code. Future glibc will come with vectorized implementations of exp() and so on.


If targeting pre-IvB, or esp. Nehalem, try to get gcc to cause partial-register stalls with 16bit or 8bit operations followed by 32bit or 64bit operations. In most cases, gcc will use movzx after an 8 or 16bit operation, but here's a case where gcc modifies ah and then reads ax


With (inline) asm:

With (inline) asm, you could break the uop cache: A 32B chunk of code that doesn't fit in three 6uop cache lines forces a switch from the uop cache to the decoders. An incompetent ALIGN using many single-byte nops instead of a couple long nops on a branch target inside the inner loop might do the trick. Or put the alignment padding after the label, instead of before. :P This only matters if the frontend is a bottleneck, which it won't be if we succeeded at pessimizing the rest of the code.

Use self-modifying code to trigger pipeline clears (aka machine-nukes).

LCP stalls from 16bit instructions with immediates too large to fit in 8 bits are unlikely to be useful. The uop cache on SnB and later means you only pay the decode penalty once. On Nehalem (the first i7), it might work for a loop that doesn't fit in the 28 uop loop buffer. gcc will sometimes generate such instructions, even with -mtune=intel and when it could have used a 32bit instruction.


A common idiom for timing is CPUID(to serialize) then RDTSC. Time every iteration separately with a CPUID/RDTSC to make sure the RDTSC isn't reordered with earlier instructions will be slow things down a lot. (In real life, the smart way to time is to time all the iterations together, instead of timing each separately and adding them up).


Cause lots of cache misses and other memory slowdowns

Use a union { double d; char a[8]; } for some of your variables. Cause a store-forwarding stall by doing a narrow store (or Read-Modify-Write) to just one of the bytes. (That wiki article also covers a lot of other microarchitectural stuff for load/store queues). e.g. flip the sign of a double using XOR 0x80 on just the high byte, instead of a - operator. The diabolically incompetent developer may have heard that FP is slower than integer, and thus try to do as much as possible using integer ops. (A very good compiler targeting FP math in SSE registers may possibly compile this to an xorps with a constant in another xmm register, but the only way this isn't terrible for x87 is if the compiler realizes that it's negating the value and replaces the next add with a subtract.)


Use volatile if you're compiling with -O3 and not using std::atomic, to force the compiler to actually store/reload all over the place. Global variables (instead of locals) will also force some stores/reloads, but the C++ memory model's weak ordering doesn't require the compiler to spill/reload to memory all the time.

Replace local vars with members of a big struct, so you can control the memory layout.

Use arrays in the struct for padding (and storing random numbers, to justify their existence).

Choose your memory layout so everything goes into a different line in the same "set" in the L1 cache. It's only 8-way associative, i.e. each set has 8 "ways". Cache lines are 64B.

Even better, put things exactly 4096B apart, since loads have a false dependency on stores to different pages but with the same offset within a page. Aggressive out-of-order CPUs use Memory Disambiguation to figure out when loads and stores can be reordered without changing the results, and Intel's implementation has false-positives that prevent loads from starting early. Probably they only check bits below the page offset, so the check can start before the TLB has translated the high bits from a virtual page to a physical page. As well as Agner's guide, see an answer from Stephen Canon, and also a section near the end of @Krazy Glew's answer on the same question. (Andy Glew was one of the architects of Intel's original P6 microarchitecture.)

Use __attribute__((packed)) to let you mis-align variables so they span cache-line or even page boundaries. (So a load of one double needs data from two cache-lines). Misaligned loads have no penalty in any Intel i7 uarch, except when crossing cache lines and page lines. Cache-line splits still take extra cycles. Skylake dramatically reduces the penalty for page split loads, from 100 to 5 cycles. (Section 2.1.3). Perhaps related to being able to do two page walks in parallel.

A page-split on an atomic<uint64_t> should be just about the worst case, esp. if it's 5 bytes in one page and 3 bytes in the other page, or anything other than 4:4. Even splits down the middle are more efficient for cache-line splits with 16B vectors on some uarches, IIRC. Put everything in a alignas(4096) struct __attribute((packed)) (to save space, of course), including an array for storage for the RNG results. Achieve the misalignment by using uint8_t or uint16_t for something before the counter.

If you can get the compiler to use indexed addressing modes, that will defeat uop micro-fusion. Maybe by using #defines to replace simple scalar variables with my_data[constant].

If you can introduce an extra level of indirection, so load/store addresses aren't known early, that can pessimize further.


Traverse arrays in non-contiguous order

I think we can come up with incompetent justification for introducing an array in the first place: It lets us separate the random number generation from the random number use. Results of each iteration could also be stored in an array, to be summed later (with more diabolical incompetence).

For "maximum randomness", we could have a thread looping over the random array writing new random numbers into it. The thread consuming the random numbers could generate a random index to load a random number from. (There's some make-work here, but microarchitecturally it helps for load-addresses to be known early so any possible load latency can be resolved before the loaded data is needed.) Having a reader and writer on different cores will cause memory-ordering mis-speculation pipeline clears (as discussed earlier for the false-sharing case).

For maximum pessimization, loop over your array with a stride of 4096 bytes (i.e. 512 doubles). e.g.

for (int i=0 ; i<512; i++)
    for (int j=i ; j<UPPER_BOUND ; j+=512)
        monte_carlo_step(rng_array[j]);

So the access pattern is 0, 4096, 8192, ...,
8, 4104, 8200, ...
16, 4112, 8208, ...

This is what you'd get for accessing a 2D array like double rng_array[MAX_ROWS][512] in the wrong order (looping over rows, instead of columns within a row in the inner loop, as suggested by @JesperJuhl). If diabolical incompetence can justify a 2D array with dimensions like that, garden variety real-world incompetence easily justifies looping with the wrong access pattern. This happens in real code in real life.

Adjust the loop bounds if necessary to use many different pages instead of reusing the same few pages, if the array isn't that big. Hardware prefetching doesn't work (as well/at all) across pages. The prefetcher can track one forward and one backward stream within each page (which is what happens here), but will only act on it if the memory bandwidth isn't already saturated with non-prefetch.

This will also generate lots of TLB misses, unless the pages get merged into a hugepage (Linux does this opportunistically for anonymous (not file-backed) allocations like malloc/new that use mmap(MAP_ANONYMOUS)).

Instead of an array to store the list of results, you could use a linked list. Then every iteration would require a pointer-chasing load (a RAW true dependency hazard for the load-address of the next load). With a bad allocator, you might manage to scatter the list nodes around in memory, defeating cache. With a diabolically incompetent allocator, it could put every node at the beginning of its own page. (e.g. allocate with mmap(MAP_ANONYMOUS) directly, without breaking up pages or tracking object sizes to properly support free).


These aren't really microarchitecture-specific, and have little to do with the pipeline (most of these would also be a slowdown on a non-pipelined CPU).

Somewhat off-topic: make the compiler generate worse code / do more work:

Use C++11 std::atomic<int> and std::atomic<double> for the most pessimal code. The MFENCEs and locked instructions are quite slow even without contention from another thread.

-m32 will make slower code, because x87 code will be worse than SSE2 code. The stack-based 32bit calling convention takes more instructions, and passes even FP args on the stack to functions like exp(). atomic<uint64_t>::operator++ on -m32 requires a lock cmpxchg8B loop (i586). (So use that for loop counters! [Evil laugh]).

-march=i386 will also pessimize (thanks @Jesper). FP compares with fcom are slower than 686 fcomi. Pre-586 doesn't provide an atomic 64bit store, (let alone a cmpxchg), so all 64bit atomic ops compile to libgcc function calls (which is probably compiled for i686, rather than actually using a lock). Try it on the Godbolt Compiler Explorer link in the last paragraph.

Use long double / sqrtl / expl for extra precision and extra slowness in ABIs where sizeof(long double) is 10 or 16 (with padding for alignment). (IIRC, 64bit Windows uses 8byte long double equivalent to double. (Anyway, load/store of 10byte (80bit) FP operands is 4 / 7 uops, vs. float or double only taking 1 uop each for fld m64/m32/fst). Forcing x87 with long double defeats auto-vectorization even for gcc -m64 -march=haswell -O3.

If not using atomic<uint64_t> loop counters, use long double for everything, including loop counters.

atomic<double> compiles, but read-modify-write operations like += aren't supported for it (even on 64bit). atomic<long double> has to call a library function just for atomic loads/stores. It's probably really inefficient, because the x86 ISA doesn't naturally support atomic 10byte loads/stores, and the only way I can think of without locking (cmpxchg16b) requires 64bit mode.


At -O0, breaking up a big expression by assigning parts to temporary vars will cause more store/reloads. Without volatile or something, this won't matter with optimization settings that a real build of real code would use.

C aliasing rules allow a char to alias anything, so storing through a char* forces the compiler to store/reload everything before/after the byte-store, even at -O3. (This is a problem for auto-vectorizing code that operates on an array of uint8_t, for example.)

Try uint16_t loop counters, to force truncation to 16bit, probably by using 16bit operand-size (potential stalls) and/or extra movzx instructions (safe). Signed overflow is undefined behaviour, so unless you use -fwrapv or at least -fno-strict-overflow, signed loop counters don't have to be re-sign-extended every iteration, even if used as offsets to 64bit pointers.


Force conversion from integer to float and back again. And/or double<=>float conversions. The instructions have greater-than-one latency, and scalar int->float (cvtsi2ss) is badly designed to not zero the rest of the xmm register. (gcc inserts an extra pxor to break dependencies, for this reason.)


Frequently set your CPU affinity to a different CPU (suggested by @Egwor). diabolical reasoning: You don't want one core to get overheated from running your thread for a long time, do you? Maybe swapping to another core will let that core turbo to a higher clock speed. (In reality: they're so thermally close to each other that this is highly unlikely except in a multi-socket system). Now just get the tuning wrong and do it way too often. Besides the time spent in the OS saving/restoring thread state, the new core has cold L2/L1 caches, uop cache, and branch predictors.

Introducing frequent unnecessary system calls can slow you down no matter what they are. Although some important but simple ones like gettimeofday may be implemented in user-space with, with no transition to kernel mode. (glibc on Linux does this with the kernel's help, since the kernel exports code in the vdso).