Any ETA on the Intel Torch release date?

Will be great if you can also run following two benchmarks and share performance of Torch on Phi 72xx series, so we can compare it against NVidia cuDNN performance.

1. https://github.com/jcjohnson/cnn-benchmarks
2. https://github.com/soumith/convnet-benchmarks

Thanks!

When installing the MKL branch intel/torch implementation, I am subsequently unable to install nn lua rock:

luarock install nn

Here's the error that I get:

Cloning into 'nn'...
/tmp/luarocks_nn-scm-1-7871/nn/lib/THNN/generic/Abs.c: In function 'THNN_FloatAbs_updateGradInput':
/tmp/luarocks_nn-scm-1-7871/nn/lib/THNN/init.c:31:5: error: unknown type name 'ptrdiff_t'
     ptrdiff_t n1 = THTensor_(nElement)(I1);     \
     ^
/tmp/luarocks_nn-scm-1-7871/nn/lib/THNN/generic/Abs.c:20:3: note: in expansion of macro 'THNN_CHECK_NELEMENT'
   THNN_CHECK_NELEMENT(input, gradOutput);
   ^
/tmp/luarocks_nn-scm-1-7871/nn/lib/THNN/init.c:32:5: error: unknown type name 'ptrdiff_t'
     ptrdiff_t n2 = THTensor_(nElement)(I2);                                 \
     ^

The solution, according to multiple sources, and presented around the same time as code here was first pushed, is to upgrade to the latest torch, with explanation being a broken dependency between torch and this lua rock.

When is the next drop of intel/torch expected?

I've found a bug in the index function which corrupts the data whenever it is not indexing a Tensor across the 1st dimension.

-- Create random data and random slice indices
x = torch.rand(2000,20000)+3
ind = torch.randperm(x:size(2))[{{1,10}}]:long()

-- Slice x along dimension 2 using the supplied indices in ind
sx = x:index(2,ind)

print("Original Range: " .. x:min() .. "," .. x:max())
print("Slice Range: " .. sx:min() .. "," .. sx:max())

The original range should report 3,4 while the sliced range reports an invalid range of 0,~4 (the upper and lower can vary due to randomness, but it should never be lower than 3).

This indicates that the internal data is invalid when using the index function. It certainly has something to do with the size of the matrix itself since you get perfectly fine copies of Tensor sizes of [200,10]:

x = torch.rand(200,10)+3
ind = torch.LongTensor{1,2,3,4,5,6,7,8,9,10} -- Select all indices in order
sx = x:index(2,ind)
print("Matching: ".. x:eq(sx):sum() .. "/" .. x:nElement())

This produces Matching: 2000/2000, however the critical point seems to be when the first dimension is OVER 720.

A matrix of size [720,10] with the above test produces Matching: 7200/7200, however with a matrix of size [721,10] it produces Matching: 129/7210 which is absolutely not correct.

I'd really appreciate any insights into this error.

Hi,

icc (ICC) 17.0.4 20170411 fails to compile the code.

It looks like there are some cuda components (eg, host_config.h and cuda_runtime.h) that only allow ICC15 and 16.

Is this expected?

Thanks,

Pedro N.

Hi, thank you for great work. It runs faster than original torch on Xeon.

However, I'm having one trouble while trying to load trained .t7 files. These .t7 files are trained on GPGPU torch but converted for CPU. Original torch on CPU (without GPGPU) works well with these .t7 files. However, Intel torch shows following error message.

$ th ...
| loading model file...
/home/.../torch/inteltorch/install/bin/lua: .../torch/inteltorch/install/share/lua/5.2/torch/
File.lua:301: Failed to load function from bytecode: binary string: not a precompiled chunk
stack traceback:
        [C]: in function 'error'
        .../torch/inteltorch/install/share/lua/5.2/torch/File.lua:301: in function 'readObject'
        .../torch/inteltorch/install/share/lua/5.2/torch/File.lua:369: in function 'readObject'

Any idea? Thanks.

Following up on this instruction:

Be sure you have installed the lastest MKL version: parallel_studio_xe_2017.

I can't find a free download link - I do see that Parallel Studio is a commercial product.