Check that the `which` command exists

check_dependencies.sh depends on `which`, but some distributions, such as those intended for use with Docker, do not include it. Check to see if is installed.

Revert "introduce a new splice configuration for tdnn+xent on swbd as default…"

introduce a new splice configuration for tdnn+xent on swbd as default…

fixed the usage of lmrescore_const_arpa.sh

Speed up LogSoftmaxComponent::Backprop

bench result:
New: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 16, speed was 0.0152883 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 16, speed was 0.00217375 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 32, speed was 0.0577221 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 32, speed was 0.00867094 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 64, speed was 0.267811 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 64, speed was 0.035306 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 128, speed was 0.878541 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 128, speed was 0.134737 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 256, speed was 2.8799 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 256, speed was 0.491975 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 512, speed was 6.20522 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 512, speed was 1.34159 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 1024, speed was 10.4197 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 1024, speed was 2.4438 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 2048, speed was 10.5138 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 2048, speed was 2.97796 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 4096, speed was 10.3679 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<float>, for dim = 4096, speed was 3.25972 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 16, speed was 0.0139596 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 16, speed was 0.00193458 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 32, speed was 0.0573372 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 32, speed was 0.0073193 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 64, speed was 0.197072 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 64, speed was 0.0282332 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 128, speed was 0.751801 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 128, speed was 0.111315 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 256, speed was 2.43203 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 256, speed was 0.394491 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 512, speed was 4.53031 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 512, speed was 0.930698 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 1024, speed was 5.43358 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 1024, speed was 1.52317 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 2048, speed was 5.47013 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 2048, speed was 1.84648 gigaflops.
New: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 4096, speed was 5.23873 gigaflops.
Old: For CuMatrix::DiffLogSoftmaxPerRow<double>, for dim = 4096, speed was 1.87967 gigaflops.

Conflicts:
	src/cudamatrix/cu-kernels-ansi.h
	src/cudamatrix/cu-kernels.h

naming of diff log softmax

Replace implementation of atomic addition.

Old version was based on atomicExch(), while this version uses CUDA's
built-in atomicAdd(), added in SM 2.0. When tested in isolation (test
code not provided in this commit), on a K10 (Kepler), the built-in
atomicAdd() is two times faster than the old version of atomic_add()
here, and on a 950M (Maxwell), 3 times faster.

Speed up to forward backward, however, is marginal for an
nnet3-chain-train call on the TEDLIUM version 1 dataset:

Times reported on a K10. Note speedup in BetaDashGeneralFrame(),
which is the only code calling the atomic add function.

New code:

[cudevice profile]
AddRows	0.468516s
AddVecVec	0.553152s
MulRowsVec	0.614542s
CuMatrix::SetZero	0.649105s
CopyRows	0.748831s
TraceMatMat	0.777907s
AddVecToRows	0.780592s
CuMatrix::Resize	0.850884s
AddMat	1.23867s
CuMatrixBase::CopyFromMat(from other CuMatrixBase)	2.04559s
AddDiagMatMat	2.18652s
AddMatVec	3.67839s
AlphaGeneralFrame	6.42574s
BetaDashGeneralFrame	8.69981s
AddMatMat	29.9714s
Total GPU time:	63.8273s (may involve some double-counting)
-----

Old code:

[cudevice profile]
AddRows	0.469031s
AddVecVec	0.553298s
MulRowsVec	0.615624s
CuMatrix::SetZero	0.658105s
CopyRows	0.750856s
AddVecToRows	0.782937s
TraceMatMat	0.786361s
CuMatrix::Resize	0.91639s
AddMat	1.23964s
CuMatrixBase::CopyFromMat(from other CuMatrixBase)	2.05253s
AddDiagMatMat	2.18863s
AddMatVec	3.68707s
AlphaGeneralFrame	6.42885s
BetaDashGeneralFrame	9.03617s
AddMatMat	29.9942s
Total GPU time:	64.3928s (may involve some double-counting)
-----

Updates to SRE08 example

nnet1: adding <ParametricRelu> component,

nnet3/report : Added plotting capability for parameter differences.

nnet1: minor cosmetic change,

WIP: Multi-database English LVCSR recipe

Removing little-used feature: time-reversed, and fwd-bkwd, decoding.

Various unrelated fixes: add --iter options to TIMIT sclite scoring; improve how syncfiles are removed in queue.pl; minor cosmetic and efficiency improvements in nnet3 code.

Extract wav - perturb_data_dir_speed.sh implementation

In scripts such as perturb-speed and perturb-volume scp lines are
tranformed into piped command with the appropropriate sox command. The
case that the scp file has file offsets was not handled. This commit
both generalizes the wav-copy command to work also on xfilenames and
fixes the two perturb scripts to use this command in case of file
offsets.

Speed up log softmax