Merge pull request #1303 from nouiz/gpusoftmax

doc/dev_start_guide.txt

@@ -505,5 +505,6 @@ Other tools that can help you
  * `line_profiler <http://pypi.python.org/pypi/line_profiler/>`_: Line-by-line profiler.
  * `memory_profiler <http://fseoane.net/blog/2012/line-by-line-report-of-memory-usage/>`_: memory profiler
  * `runsnake <http://www.vrplumber.com/programming/runsnakerun/>`_: Gui for cProfile(time profiler) and Meliae(memory profiler)
+ * `Guppy <https://pypi.python.org/pypi/guppy/>`_: Supports object and heap memory sizing, profiling and debugging.
  * `hub <https://github.com/defunkt/hub>`_: A tool that adds github commands to the git command line.
  * `git pull-requests <http://www.splitbrain.org/blog/2011-06/19-automate_github_pull_requests>`_: Another tool for git/github command line.

theano/misc/check_blas.py

@@ -198,6 +198,8 @@ if __name__ == "__main__":
 
         cuda version      5.0    4.2    4.1    4.0    3.2    3.0   # note
         gpu
+        K20m/ECC          0.07s
+        K20/NOECC         0.07s
         M2070             0.25s         0.27s         0.32s
         M2050(Amazon)     0.25s
         C2075                    0.25s
@@ -215,7 +217,7 @@ if __name__ == "__main__":
         GTX 285                  0.452s        0.452s        0.40s # cuda 3.0 seems faster? driver version?
         GTX 550 Ti                             0.57s
         GT 520                   2.68s                3.06s
-        520M                                          3.19s        # with bumblebee on Ubuntu 12.04
+        520M              2.44s                       3.19s        # with bumblebee on Ubuntu 12.04
         GT 220                                        3.80s
         GT 210                                 6.35s
         8500 GT                                              10.68s

theano/sandbox/cuda/tests/test_nnet.py

@@ -172,8 +172,8 @@ def test_softmax_with_bias():
     x = T.fmatrix('x')
     # We can't use zeros_like(x[0,::]) as this don't allow to test with
     # 0 shape.
-    z = T.nnet.softmax_with_bias(x, T.alloc(numpy.asarray(0, dtype='float32'),
-                                            x.shape[1]))
+    z = T.nnet.softmax_with_bias(x, T.arange(x.shape[1] * 2,
+                                             dtype='float32')[::2])
 
     f = theano.function([x], z, mode=mode_without_gpu)
     f_gpu = theano.function([x], z, mode=mode_with_gpu)
@@ -181,24 +181,12 @@ def test_softmax_with_bias():
     assert isinstance(f_gpu.maker.fgraph.toposort()[-2].op,
                       cuda.nnet.GpuSoftmaxWithBias)
 
-    def cmp(n, m, catch=False):
-        """Some old card won't accet the configuration arguments of
-        this implementation. For those cases set catch=True to skip
-        those errors.
-        """
-        try:
+    def cmp(n, m):
         #print "test_softmax",n,m
         data = numpy.arange(n * m, dtype='float32').reshape(n, m)
         out = f(data)
         gout = f_gpu(data)
         assert numpy.allclose(out, gout), numpy.absolute(out - gout)
-        except RuntimeError, e:
-            if not catch:
-                raise
-            # Different CUDA driver have different error message
-            assert (e.args[0].startswith(
-              'Cuda error: kSoftmaxWithBias_node_0: invalid configuration argument.\n') or
-            e.args[0].startswith('Cuda error: kSoftmaxWithBias_node_0: invalid argument.\n'))
 
     cmp(2, 5)
     #we need to test n>32*1024 to check that we make the block loop.
@@ -211,7 +199,11 @@ def test_softmax_with_bias():
     cmp(4, 2000)
     cmp(4, 2024)
     #GTX285 don't have enough shared mem for this case.
-    cmp(4, 4074, True)
+    cmp(4, 4074)
+    # The GTX580, 680 and kepler don't have enough shared memory.
+    cmp(2, 10000)
+    cmp(128, 16 * 1024)
+    cmp(128, 64 * 1024)
 
 
 def test_softmax():
@@ -219,9 +211,7 @@ def test_softmax():
     This is basic test for GpuSoftmax
 
     We check that we loop when their is too much block
-
-    TODO: check that we loop when their is too much thread.(THIS IS
-    NOT IMPLEMENTED)
+    We use slower code when there isn't enough shared memory
     """
     x = T.fmatrix('x')
 
@@ -232,25 +222,12 @@ def test_softmax():
     assert isinstance(f_gpu.maker.fgraph.toposort()[-2].op,
                       cuda.nnet.GpuSoftmax)
 
-    def cmp(n, m, catch=False):
-        """Some old card won't accept the configuration arguments of
-        this implementation. For those cases set catch=True to skip
-        those errors.
-
-        """
-        try:
+    def cmp(n, m):
         #print "test_softmax",n,m
         data = numpy.arange(n * m, dtype='float32').reshape(n, m)
         out = f(data)
         gout = f_gpu(data)
         assert numpy.allclose(out, gout), numpy.absolute(out - gout)
-        except RuntimeError, e:
-            if not catch:
-                raise
-            # Different CUDA driver have different error message
-            assert (e.args[0].startswith(
-              'Cuda error: kSoftmax_node_0: invalid configuration argument.\n') or
-            e.args[0].startswith('Cuda error: kSoftmax_node_0: invalid argument.\n'))
 
     #we need to test n>32*1024 to check that we make the block loop.
     cmp(2, 5)
@@ -262,5 +239,9 @@ def test_softmax():
     cmp(4, 1024)
     cmp(4, 2000)
     cmp(4, 2024)
-    #GTX285 don't have enough shared mem for this case.
-    cmp(4, 4074, True)
+    # The GTX285 don't have enough shared memory.
+    cmp(4, 4074)
+    # The GTX580, 680 and kepler don't have enough shared memory.
+    cmp(2, 10000)
+    cmp(128, 16 * 1024)
+    cmp(128, 64 * 1024)