winograd.lua

require 'cutorch'
require 'cunn'
require 'cudnn'
require 'sys'

-- WINOGRAD benchmark
-- required: cuDNN v5, cudnn.torch R5 branch

function create_model(ch) -- simple 3x3 conv model
   local model = nn.Sequential()
   model:add(cudnn.SpatialConvolution(ch, 64, 3, 3, 1, 1, 1, 1))
   model:add(cudnn.ReLU())
   model:add(cudnn.SpatialConvolution(64, 64, 3, 3, 1, 1, 1, 1))
   model:add(cudnn.ReLU())
   return model
end
function setMode(model, modes)
   local modules = model:findModules("cudnn.SpatialConvolution")
   for i = 1, #modules do
      modules[i]:setMode(table.unpack(modes))
      modules[i].workspace_limit = (modules[i].nInputPlane * modules[i].nOutputPlane) * 64
      --modules[i].fastest = true
   end
end

cudnn.benchmark = false -- don't use auto tuning
cudnn.fastest = true
cudnn.verbose = true

model1 = create_model(64):cuda()
model2 = model1:clone()
setMode(model1, {'CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM'})
setMode(model2, {'CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD'}) -- use WINOGRAD

input = torch.Tensor(1, 64, 512, 512):cuda()
-- warm :)
sum = 0
for i = 1, 10 do
   sum = sum + model1:forward(input:fill(i)):sum()
end
-- test model1
t = sys.clock()
sum = 0
for i = 1, 10 do
   sum = sum + model1:forward(input:fill(i)):sum()
end
t1 = sys.clock() - t
print(t1, sum)

-- test model2
t = sys.clock()
sum = 0
for i = 1, 10 do
   sum = sum + model2:forward(input:fill(i)):sum()
end
t2 = sys.clock() - t
print(t2, sum)

print(("%.2f x faster!"):format(t1 / t2))

0.30708813667297 112713675.625
0.11358094215393 112713691.875
2.70 x faster!

Can I get Winograd code in C++?