Add Large Dim Checks for linalg Operators (apache#18816)

* initial * test * gemm and gemm2 * type fix * syrk trmm trsm * gelqf * move tests from test_large_array.py to test_large_vector.py * fix white space issue Co-authored-by: Ubuntu <[email protected]>
bgawrych · Jul 31, 2020 · f4e62df · f4e62df
1 parent 84c9e0d
commit f4e62df
Show file tree

Hide file tree

Showing 2 changed files with 73 additions and 1 deletion.
diff --git a/src/operator/tensor/la_op.h b/src/operator/tensor/la_op.h
@@ -157,6 +157,14 @@ struct LaTrianParam : public dmlc::Parameter<LaTrianParam> {
   }
 };
 
+// check if any dim will overflow 32-bit int
+inline void check_large_dim(std::vector<dim_t> dims) {
+  for (dim_t dim : dims) {
+    CHECK_LE(dim, INT_MAX)
+      << "Large matrix dimensions (>= 2^31) are not supported";
+  }
+}
+
 // Common function for shape inference for matrix mult and matrix mac.
 inline bool LaMatrixMultMacOpShape(const nnvm::NodeAttrs& attrs,
                                    mxnet::ShapeVector* in_attrs,
@@ -181,6 +189,11 @@ inline bool LaMatrixMultMacOpShape(const nnvm::NodeAttrs& attrs,
     const int ndim((*in_attrs)[0].ndim()), axis(axis_param < 0 ? ndim + axis_param : axis_param);
     CHECK(axis >= 0 && axis < ndim-1)
       << "Invalid row axis (" << axis_param << ")";
+    // Check if input matrix dims are too large
+    check_large_dim({(*in_attrs)[0][axis],
+                     (*in_attrs)[0][ndim-1],
+                     (*in_attrs)[1][axis],
+                     (*in_attrs)[1][ndim-1]});
     std::vector<int> oshape(ndim);
     for ( int i = 0; i < ndim-1; ++i ) {
       if (i != axis) {
@@ -225,6 +238,10 @@ inline bool LaTriangMatrixMultOpShape(const nnvm::NodeAttrs& attrs,
         << "Shapes of inputs 0, 1 must be the same, except on last two dimensions";
       oshape[i] = (*in_attrs)[0][i];
     }
+    // Check if the input matrix dims are too large; it suffices to check the second
+    // input only because the first is square whose size is bounded by memory
+    check_large_dim({(*in_attrs)[1][ndim-1],
+                     (*in_attrs)[1][ndim-2]});
     if ( param.rightside ) {
       // We compute B * A where A is the first and B the second input.
       CHECK_EQ((*in_attrs)[0][ndim-2], (*in_attrs)[1][ndim-1])
@@ -341,6 +358,9 @@ inline bool LaSyrkShape(const nnvm::NodeAttrs& attrs,
   bool transpose = nnvm::get<LaSyrkParam>(attrs.parsed).transpose;
   const int ndim = in_attr.ndim();
   if ( ndim >= 2 ) {
+    // Check if input matrix dims are too large
+    check_large_dim({in_attr[ndim-1],
+                     in_attr[ndim-2]});
     // Forward shape inference.
     std::vector<int> oshape(ndim);
     for ( int i = 0; i < ndim-2; ++i ) {
@@ -371,6 +391,9 @@ inline bool LaLQFactShape(const nnvm::NodeAttrs& attrs,
     const int ndim(in_a.ndim());
     CHECK_LE(in_a[ndim-2], in_a[ndim-1])
       << "Input A shape wrong: Last dimension must be >= than second to last";
+    // Check if the last dimension is too large; it suffices to check the last dim
+    // only since the second to last dim <= last dim
+    check_large_dim({in_a[ndim-1]});
     // Q must have same shape as A
     SHAPE_ASSIGN_CHECK(*out_attrs, 0, in_a);
     std::vector<int> oshape_l(ndim);

diff --git a/tests/nightly/test_large_vector.py b/tests/nightly/test_large_vector.py
@@ -27,13 +27,15 @@
 
 from mxnet.test_utils import rand_ndarray, assert_almost_equal, rand_coord_2d, create_vector
 from mxnet import gluon, nd
-from tests.python.unittest.common import with_seed
+from tests.python.unittest.common import with_seed, assertRaises
+from mxnet.base import MXNetError
 from nose.tools import with_setup
 import unittest
 
 # dimension constants
 LARGE_X = 4300000000
 MEDIUM_X = 1000000000
+INT32_MAX = 2**31-1
 
 
 def test_nn():
@@ -1064,6 +1066,53 @@ def check_minimum():
     check_minimum()
 
 
+# openblas and cublas are known to not work well with large
+# matrix dims under current configuration. checks are added
+# to exit from such use cases
+def test_linalg_large_dim():
+    def check_gemm():
+        A = nd.ones(shape=(1, INT32_MAX + 1, 1))
+        B = nd.ones(shape=(1, INT32_MAX + 1, 1))
+        C = nd.ones(shape=(1, 1, 1))
+        assertRaises(MXNetError, nd.linalg.gemm, \
+            A, B, C, transpose_b=True)
+
+    def check_gemm2():
+        A = nd.ones(shape=(1, 1, INT32_MAX + 1))
+        B = nd.ones(shape=(1, 1, INT32_MAX + 1))
+        assertRaises(MXNetError, nd.linalg.gemm2, \
+            A, B, transpose_b=True)
+
+    def check_trmm():
+        A = nd.ones(shape=(1, 1, 1))
+        B = nd.ones(shape=(1, INT32_MAX + 1, 1))
+        assertRaises(MXNetError, nd.linalg.trmm, \
+            A, B, rightside=True)
+
+    def check_trsm():
+        A = nd.ones(shape=(1, 1, 1))
+        B = nd.ones(shape=(1, 1, INT32_MAX + 1))
+        assertRaises(MXNetError, nd.linalg.trsm, \
+            A, B, rightside=False)
+
+    def check_syrk():
+        A = nd.ones(shape=(1, INT32_MAX + 1, 1))
+        assertRaises(MXNetError, nd.linalg.syrk, A)
+        assertRaises(MXNetError, nd.linalg.syrk, A, transpose=True)
+
+    def check_gelqf():
+        A = nd.ones(shape=(1, 1, INT32_MAX + 1))
+        assertRaises(MXNetError, nd.linalg.gelqf, A)
+
+    # batch input
+    check_gemm()
+    check_gemm2()
+    check_trmm()
+    check_trsm()
+    check_syrk()
+    check_gelqf()
+
+
 if __name__ == '__main__':
     import nose
     nose.runmodule()