Threaded wisdom failure when avx512 is enabled and plan rigor > FFTW_ESTIMATE

[dfarns]

When planning for multiple threads with AVX512 enabled (e.g., skylake-avx512 or knl), fftw_export_wisdom_to_string returns only wisdom header and footer. This occurs when FFTW_MEASURE or FFTW_PATIENT is used; FFTW_ESTIMATE returns wisdom as expected.

This behavior is not observed for single thread planning with AVX512, nor for threaded planning on skylake-avx512 cpus when AVX512 is omitted from the build (e.g., configured using --enable-avx2 and --enable-openmp only).

The library build was configured with minimal options (e.g., --enable-avx512 --enable-openmp) and built with gcc 6.1.0 and 8.1.0. Adding the recommended --enable-avx2 does not help. This occurs for each of the libfftw3_omp and libfftw3_threads (built with --enable-threads) libs. Built and tested on centos7 and sles12 with same behavior.

** NOTE: The plans returned with FFTW_MEASURE and FFTW_PATIENT are the same as that for FFTW_ESTIMATE, suggesting that the planner is finding no applicable (or none at all) wisdom and reverting to the FFTW_ESTIMATE behavior.

Sample output for a N=1024 C2C in-place transform (using FFTW_MEASURE):

./a.omp.out -nthreads 2
Planning with 2 threads.

plan =
(dft-thr-ct-dit-x2/32
  (dftw-direct-32/32 "t3fv_32_avx512")
  (dftw-direct-32/32 "t3fv_32_avx512")
  (dft-buffered-32-x32/32-6
    (dft-thr-vrank>=1-x2/1
      (dft-direct-32-x16 "n2fv_32_avx512")
      (dft-direct-32-x16 "n2fv_32_avx512"))
    (dft-r2hc-1
      (rdft-thr-vrank>=1-x2/1
        (rdft-rank0-iter-ci/64-x16)
        (rdft-rank0-iter-ci/64-x16)))
    (dft-nop)))

wisdom =
(fftw-3.3.8 fftw_wisdom #x563d20a7 #x21166e89 #x51240a27 #x4128fc72
)

Build script:

#!/bin/bash

ARCH="x86_skylake"
enable_optimizations='--enable-avx512'
GCC_VERSION="8.1.0"

BINPATH=/opt/gcc/${GCC_VERSION}/bin
CONFIG_CC=${BINPATH}/gcc
CONFIG_CPP=${BINPATH}/cpp
CONFIG_CXX=${BINPATH}/g++
CONFIG_F77=${BINPATH}/gfortran

PREFIX=$(pwd)/fftw_install/${ARCH}

make distclean 2>&1 | tee log.make-distclean.txt

./configure \
    --prefix="$PREFIX"              \
    $enable_optimizations           \
    --enable-openmp                 \
    --enable-threads                \
    CC=$CONFIG_CC                   \
    CPP=$CONFIG_CPP                 \
    CXX=$CONFIG_CXX                 \
    F77=$CONFIG_F77                 \
    2>&1 | tee log.configure.txt

make -j 16 2>&1 | tee log.make.txt
make install 2>&1 | tee log.make-install.txt

Reproducer:

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <fftw3.h>
#ifdef _USEOMP
#  include <omp.h>
#endif

int main(int argc, char *argv[]) {

  char *wisstring;
  const int N = 1024;
  fftw_plan plan;
  fftw_complex *z1=NULL;
  int NthreadsMax, Nthreads = 1;
  int ti, iArg;

  for ( iArg = 1; iArg < argc; iArg++ ) {
   if ( strcmp(argv[iArg],"-nthreads") == 0 && iArg < argc - 1) {
      Nthreads = (int) atoi(argv[++iArg]);
    }
  }

# ifdef _USEOMP
  NthreadsMax = omp_get_max_threads();
  if ( Nthreads > NthreadsMax ) {
    printf("%d threads requested but only %d threads available.  Setting Nthreads = %d\n", Nthreads, NthreadsMax, NthreadsMax);
    Nthreads = NthreadsMax;
  }
# endif

  if ( Nthreads > 1 ) {
    ti = fftw_init_threads();
    fftw_plan_with_nthreads(Nthreads);
  }

  // Create a simple plan, then print it out
  printf("Planning with %d threads.\n", Nthreads);
  z1 = (fftw_complex*) fftw_malloc(N*sizeof(fftw_complex));
  plan = fftw_plan_dft_1d(N, z1, z1, FFTW_FORWARD, FFTW_MEASURE);
  printf("\n");
  printf("plan = \n");
  fftw_print_plan(plan);
  printf("\n");

  // Print out the wisdom, including the wisdom header
  wisstring = fftw_export_wisdom_to_string();
  printf("\n");
  if ( wisstring != NULL ) {
    printf("wisdom = \n%s\n", wisstring);
  } else {
    printf("wisstring is NULL.");
  }

  // Cleanup
  if ( plan != NULL ) {
    fftw_destroy_plan(plan);
  }
  if ( wisstring != NULL ) {
    free(wisstring);
  }
  if ( z1 != NULL ) {
    fftw_free(z1);
  }

  if ( Nthreads > 1 ) {
    fftw_cleanup_threads();
  } else {
    fftw_cleanup();
  }

  return 0;
}

Here is the same test on a skylake-avx512 cpu with the library configured with avx2 only and run with FFTW_MEASURE (demonstrating that it's unlikely to be a cpu issue):

./a.omp.out -nthreads 2
Planning with 2 threads.

plan =
(dft-thr-ct-dit-x2/16
  (dftw-direct-16/16 "t3fv_16_avx2")
  (dftw-direct-16/16 "t3fv_16_avx2")
  (dft-directbuf/66-64-x16 "n2fv_64_avx2"))

wisdom =
(fftw-3.3.8 fftw_wisdom #x44b8b2bb #xf6415fbb #x528d8d0a #x01cf6e58
  (fftw_codelet_n2fv_64_avx2 1 #x11bdd #x11bdd #x0 #x1922ec23 #xa3fc4eb9 #xce8a9d6e #xac6cc7f0)
  (fftw_codelet_t3fv_16_avx2 1 #x11bdd #x11bdd #x0 #xa737e14d #xd2ed5d58 #x701e154a #x4351c9d2)
)

Ideas?

[dfarns]

Some printf statements in mkapiplan, mkplan, etc are suggesting that the wisdom generated by the planner is considered "bogus" in mkplan() [i.e., (plnr->wisdom_state == WISDOM_IS_BOGUS) is true] and thus rejected. This would explain the empty wisdom string returned by fftw_export_wisdom as well as the planner reversion to FFTW_ESTIMATE rigor.

Any reason why this might happen only when avx512 codelets are enabled and nthreads>1? Possible race condition / wisdom table corruption?

[matteo-frigo]

Probably fixed in ebde7c4e4607afb6bbba7e6609fae56ff0fda01b, can you verify?

[dfarns]

I can verify that the simple reproducer I posted no longer exhibits the deviant behavior. Full validation will take some time.