faster alternative to memcpy?

Question

I have a function that is doing memcpy, but it\'s taking up an enormous amount of cycles. Is there a faster alternative/approach than using memcpy to move a piece of memory?

Answer 1

This is an answer for x86_64 with AVX2 instruction set present. Though something similar may apply for ARM/AArch64 with SIMD.

On Ryzen 1800X with single memory channel filled completely (2 slots, 16 GB DDR4 in each), the following code is 1.56 times faster than memcpy() on MSVC++2017 compiler. If you fill both memory channels with 2 DDR4 modules, i.e. you have all 4 DDR4 slots busy, you may get further 2 times faster memory copying. For triple-(quad-)channel memory systems, you can get further 1.5(2.0) times faster memory copying if the code is extended to analogous AVX512 code. With AVX2-only triple/quad channel systems with all slots busy are not expected to be faster because to load them fully you need to load/store more than 32 bytes at once (48 bytes for triple- and 64-bytes for quad-channel systems), while AVX2 can load/store no more than 32 bytes at once. Though multithreading on some systems can alleviate this without AVX512 or even AVX2.

So here is the copy code that assumes you are copying a large block of memory whose size is a multiple of 32 and the block is 32-byte aligned.

For non-multiple size and non-aligned blocks, prologue/epilogue code can be written reducing the width to 16 (SSE4.1), 8, 4, 2 and finally 1 byte at once for the block head and tail. Also in the middle a local array of 2-3 __m256i values can be used as a proxy between aligned reads from the source and aligned writes to the destination.

#include <immintrin.h>
#include <cstdint>
/* ... */
void fastMemcpy(void *pvDest, void *pvSrc, size_t nBytes) {
  assert(nBytes % 32 == 0);
  assert((intptr_t(pvDest) & 31) == 0);
  assert((intptr_t(pvSrc) & 31) == 0);
  const __m256i *pSrc = reinterpret_cast<const __m256i*>(pvSrc);
  __m256i *pDest = reinterpret_cast<__m256i*>(pvDest);
  int64_t nVects = nBytes / sizeof(*pSrc);
  for (; nVects > 0; nVects--, pSrc++, pDest++) {
    const __m256i loaded = _mm256_stream_load_si256(pSrc);
    _mm256_stream_si256(pDest, loaded);
  }
  _mm_sfence();
}

A key feature of this code is that it skips CPU cache when copying: when CPU cache is involved (i.e. AVX instructions without _stream_ are used), the copy speed drops several times on my system.

My DDR4 memory is 2.6GHz CL13 . So when copying 8GB of data from one array to another I got the following speeds:

memcpy(): 17,208,004,271 bytes/sec.
Stream copy: 26,842,874,528 bytes/sec.

Note that in these measurements the total size of both input and output buffers is divided by the number of seconds elapsed. Because for each byte of the array there are 2 memory accesses: one to read the byte from the input array, another to write the byte to the output array. In the other words, when copying 8GB from one array to another, you do 16GB worth of memory access operations.

Moderate multithreading can further improve performance about 1.44 times, so total increase over memcpy() reaches 2.55 times on my machine. Here's how stream copy performance depends on the number of threads used on my machine:

Stream copy 1 threads: 27114820909.821 bytes/sec
Stream copy 2 threads: 37093291383.193 bytes/sec
Stream copy 3 threads: 39133652655.437 bytes/sec
Stream copy 4 threads: 39087442742.603 bytes/sec
Stream copy 5 threads: 39184708231.360 bytes/sec
Stream copy 6 threads: 38294071248.022 bytes/sec
Stream copy 7 threads: 38015877356.925 bytes/sec
Stream copy 8 threads: 38049387471.070 bytes/sec
Stream copy 9 threads: 38044753158.979 bytes/sec
Stream copy 10 threads: 37261031309.915 bytes/sec
Stream copy 11 threads: 35868511432.914 bytes/sec
Stream copy 12 threads: 36124795895.452 bytes/sec
Stream copy 13 threads: 36321153287.851 bytes/sec
Stream copy 14 threads: 36211294266.431 bytes/sec
Stream copy 15 threads: 35032645421.251 bytes/sec
Stream copy 16 threads: 33590712593.876 bytes/sec

The code is:

void AsyncStreamCopy(__m256i *pDest, const __m256i *pSrc, int64_t nVects) {
  for (; nVects > 0; nVects--, pSrc++, pDest++) {
    const __m256i loaded = _mm256_stream_load_si256(pSrc);
    _mm256_stream_si256(pDest, loaded);
  }
}

void BenchmarkMultithreadStreamCopy(double *gpdOutput, const double *gpdInput, const int64_t cnDoubles) {
  assert((cnDoubles * sizeof(double)) % sizeof(__m256i) == 0);
  const uint32_t maxThreads = std::thread::hardware_concurrency();
  std::vector<std::thread> thrs;
  thrs.reserve(maxThreads + 1);

  const __m256i *pSrc = reinterpret_cast<const __m256i*>(gpdInput);
  __m256i *pDest = reinterpret_cast<__m256i*>(gpdOutput);
  const int64_t nVects = cnDoubles * sizeof(*gpdInput) / sizeof(*pSrc);

  for (uint32_t nThreads = 1; nThreads <= maxThreads; nThreads++) {
    auto start = std::chrono::high_resolution_clock::now();
    lldiv_t perWorker = div((long long)nVects, (long long)nThreads);
    int64_t nextStart = 0;
    for (uint32_t i = 0; i < nThreads; i++) {
      const int64_t curStart = nextStart;
      nextStart += perWorker.quot;
      if ((long long)i < perWorker.rem) {
        nextStart++;
      }
      thrs.emplace_back(AsyncStreamCopy, pDest + curStart, pSrc+curStart, nextStart-curStart);
    }
    for (uint32_t i = 0; i < nThreads; i++) {
      thrs[i].join();
    }
    _mm_sfence();
    auto elapsed = std::chrono::high_resolution_clock::now() - start;
    double nSec = 1e-6 * std::chrono::duration_cast<std::chrono::microseconds>(elapsed).count();
    printf("Stream copy %d threads: %.3lf bytes/sec\n", (int)nThreads, cnDoubles * 2 * sizeof(double) / nSec);

    thrs.clear();
  }
}

Answer 2

Sometimes functions like memcpy, memset, ... are implemented in two different ways:

• once as a real function
• once as some assembly that's immediately inlined

Not all compilers take the inlined-assembly version by default, your compiler may use the function variant by default, causing some overhead because of the function call. Check your compiler to see how to take the intrinsic variant of the function (command line option, pragma's, ...).

Edit: See http://msdn.microsoft.com/en-us/library/tzkfha43%28VS.80%29.aspx for an explanation of intrinsics on the Microsoft C compiler.

Answer 3

Here is an alternative C version of memcpy that is inlineable and I find it outperforms memcpy for GCC for Arm64 by about 50% in the application I used it for. It is 64-bit platform independent. The tail processing can be removed if the usage instance does not need it for a bit more speed. Copies uint32_t arrays, smaller datatypes not tested but might work. Might be able to adapt for other datatypes. 64-bit copy (two indexes are copied simultaneously). 32-bit should also work but slower. Credits to Neoscrypt project.

    static inline void newmemcpy(void *__restrict__ dstp, 
                  void *__restrict__ srcp, uint len)
        {
            ulong *dst = (ulong *) dstp;
            ulong *src = (ulong *) srcp;
            uint i, tail;

            for(i = 0; i < (len / sizeof(ulong)); i++)
                *dst++ = *src++;
            /*
              Remove below if your application does not need it.
              If console application, you can uncomment the printf to test
              whether tail processing is being used.
            */
            tail = len & (sizeof(ulong) - 1);
            if(tail) {
                //printf("tailused\n");
                uchar *dstb = (uchar *) dstp;
                uchar *srcb = (uchar *) srcp;

                for(i = len - tail; i < len; i++)
                    dstb[i] = srcb[i];
            }
        }

Answer 4

Please offer us more details. On i386 architecture it is very possible that memcpy is the fastest way of copying. But on different architecture for which the compiler doesn't have an optimized version it is best that you rewrite your memcpy function. I did this on a custom ARM architecture using assembly language. If you transfer BIG chunks of memory then DMA is probably the answer you are looking for.

Please offer more details - architecture, operating system (if relevant).

Answer 5

Usually the standard library shipped with the compiler will implement memcpy() the fastest way possible for the target platform already.

Answer 6

It's generally faster not to make a copy at all. Whether you can adapt your function to not copy I don't know but it's worth looking in to.