Merge pull request #244 from facebookresearch/mul-red-test

unit test for mul reduction, use pytest + gen=2/pop=1 to improve test time

Author: prigoyal

.gitignore

@@ -8,6 +8,7 @@ third-party/llvm_sources*
 third-party-install/*
 buck-out
 .buckd
+.pytest_cache
 conda
 .nfs*
 */.nfs*

.jenkins/build.sh

@@ -72,6 +72,7 @@ if [[ "$DISTRIB_RELEASE" == 14.04 ]]; then
     source activate tc-env
     conda install -y -c pytorch pytorch
     conda install -y pyyaml
+    conda install -yc conda-forge pytest
     WITH_PYTHON_C2=OFF CORES=$(nproc) CLANG_PREFIX=/usr/local/clang+llvm-tapir5.0 BUILD_TYPE=Release ./build.sh --all
   else
     echo "Building TC in non-conda env"
@@ -86,6 +87,7 @@ if [[ "$DISTRIB_RELEASE" == 16.04 ]]; then
     source activate tc-env
     conda install -y pytorch cuda90 -c pytorch
     conda install -y pyyaml
+    conda install -yc conda-forge pytest
     WITH_PYTHON_C2=OFF CORES=$(nproc) CLANG_PREFIX=/usr/local/clang+llvm-tapir5.0 BUILD_TYPE=Release ./build.sh --all
   else
     echo "Building TC in non-conda env"

tensor_comprehensions/tc_unit.py

@@ -31,13 +31,13 @@
 
 # these are quick options for finishing autotuning
 autotuner_settings = {
-    "threads": 32, "generations": 2, "pop_size": 5,
+    "threads": 32, "generations": 1, "pop_size": 2,
 }
 
 # TC prunes autotuning for kernels which require < 256 threads. So to tune small
 # size kernels, we set the min kernel threads to 1
 small_sizes_autotuner_settings = {
-    "threads": 32, "generations": 2, "pop_size": 5, "tuner_min_launch_total_threads": 1,
+    "threads": 32, "generations": 1, "pop_size": 2, "tuner_min_launch_total_threads": 1,
 }
 
 ###############################################################################

test_python/layers/test_matmul_train.py

@@ -20,7 +20,7 @@ def matmul_grad(float(M,N) A, float(N,K) B, float(M,K) O_grad) -> (A_grad, B_gra
         }
         """
 
-        matmul = tc.define(MATMUL_LANG, name="matmul", training=True, backward="matmul_grad")
+        matmul = tc.define(LANG, name="matmul", training=True, backward="matmul_grad")
         mat1 = Parameter(torch.randn(3, 4).cuda())
         mat2 = Variable(torch.randn(4, 5).cuda(), requires_grad=True)
         out = matmul(mat1, mat2, options=[tc.Options("mlp"), tc.Options("mlp")])

test_python/layers/test_multiply_reduction.py

@@ -0,0 +1,39 @@
+# Copyright (c) 2017-present, Facebook, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+##############################################################################
+
+import tensor_comprehensions as tc
+
+import torch
+import unittest
+
+
+class TestMultReduction(unittest.TestCase):
+
+   def test_mult_reduction(self):
+       LANG = """
+       def conv_mult(float(N,C,H,W) I, float(M,C,KH,KW) W1) -> (O) {
+           O(n, m, h, w) *=! I(n, c, h + kh, w + kw) * W1(m, c, kh, kw)
+       }
+       """
+       N, C, H, W, O, kH, kW = 64, 10, 24, 24, 10, 7, 7
+       conv_mult = tc.define(LANG, name="conv_mult")
+       I, W1 = torch.ones(N, C, H, W).cuda(), torch.ones(O, C, kH, kW).cuda()
+       # Note: There is no bias here
+       out = conv_mult(I, W1)
+       assert out.data.min() > 0
+
+
+if __name__ == '__main__':
+   unittest.main()

test_python/run_test.sh

@@ -30,43 +30,7 @@ $PYTHON test_tc_torch.py -v
 # PyTorch layer tests
 ###############################################################################
 echo "Running all PyTorch layers tests"
-$PYTHON layers/test_absolute.py -v
-$PYTHON layers/test_autotuner.py -v
-$PYTHON layers/test_avgpool.py -v
-$PYTHON layers/test_avgpool_autotune.py -v
-$PYTHON layers/test_batchmatmul.py -v
-$PYTHON layers/test_batchnorm.py -v
-$PYTHON layers/test_layernorm.py -v
-$PYTHON layers/test_cast.py -v
-$PYTHON layers/test_concat.py -v
-$PYTHON layers/test_convolution.py -v
-$PYTHON layers/test_convolution_strided.py -v
-$PYTHON layers/test_convolution_reorder.py -v
-$PYTHON layers/test_convolution_strided_autotune.py -v
-$PYTHON layers/test_convolution_train.py -v
-$PYTHON layers/test_copy.py -v
-$PYTHON layers/test_cos.py -v
-$PYTHON layers/test_cosine_similarity.py -v
-$PYTHON layers/test_dump_cuda.py -v
-$PYTHON layers/test_external_cuda_injection.py -v
-$PYTHON layers/test_fc.py -v
-$PYTHON layers/test_fusion_fcrelu.py -v
-$PYTHON layers/test_broadcast_fcrelu.py -v
-$PYTHON layers/test_group_convolution.py -v
-$PYTHON layers/test_group_convolution_strided.py -v
-$PYTHON layers/test_indexing.py -v
-$PYTHON layers/test_lookup_table.py -v
-$PYTHON layers/test_matmul.py -v
-$PYTHON layers/test_matmul_reuse_outputs.py -v
-$PYTHON layers/test_maxpool.py -v
-$PYTHON layers/test_relu.py -v
-$PYTHON layers/test_scale.py -v
-$PYTHON layers/test_sigmoid.py -v
-$PYTHON layers/test_small_mobilenet.py -v
-$PYTHON layers/test_softmax.py -v
-$PYTHON layers/test_tanh.py -v
-$PYTHON layers/test_tensordot.py -v
-$PYTHON layers/test_transpose.py -v
+$PYTHON -m pytest -v --full-trace --junit-xml="/tmp/tensorcomp/python/result.xml" layers/
 
 echo "All PyTorch layer tests have finished"