Merge pull request #5525 from toslunar/use-cupy

Use `cupy.linalg.det` in `F.det`

chainer/functions/math/det.py

@@ -1,49 +1,11 @@
-import numpy
-
 import chainer
 from chainer.backends import cuda
 from chainer import function_node
 import chainer.functions
-from chainer.functions.math import matmul
-from chainer import utils
 from chainer.utils import precision
 from chainer.utils import type_check
 
 
-def _det_gpu(b):
-    # We do a batched LU decomposition on the GPU to compute
-    # and compute the determinant by multiplying the diagonal.
-    # Change the shape of the array to be size=1 minibatch if necessary.
-    # Also copy the matrix as the elments will be modified in-place.
-    a = matmul._as_batch_mat(b).copy()
-    n = a.shape[1]
-    n_matrices = len(a)
-    # Pivot array
-    p = cuda.cupy.zeros((n_matrices, n), dtype='int32')
-    # Output array
-    # These arrays hold information on the execution success
-    # or if the matrix was singular.
-    info = cuda.cupy.zeros(n_matrices, dtype=numpy.intp)
-    ap = matmul._mat_ptrs(a)
-    _, lda = matmul._get_ld(a)
-    if b.dtype == numpy.float32:
-        cuda.cublas.sgetrfBatched(cuda.Device().cublas_handle, n, ap.data.ptr,
-                                  lda, p.data.ptr, info.data.ptr, n_matrices)
-    elif b.dtype == numpy.float64:
-        cuda.cublas.dgetrfBatched(cuda.Device().cublas_handle, n, ap.data.ptr,
-                                  lda, p.data.ptr, info.data.ptr, n_matrices)
-    else:
-        assert False
-    det = cuda.cupy.prod(a.diagonal(axis1=1, axis2=2), axis=1)
-    # The determinant is equal to the product of the diagonal entries
-    # of `a` where the sign of `a` is flipped depending on whether
-    # the pivot array is equal to its index.
-    rng = cuda.cupy.arange(1, n + 1, dtype='int32')
-    parity = cuda.cupy.sum(p != rng, axis=1) % 2
-    sign = 1. - 2. * parity.astype(b.dtype, copy=False)
-    return det * sign, info
-
-
 class BatchDet(function_node.FunctionNode):
 
     @property
@@ -61,17 +23,12 @@ def check_type_forward(self, in_types):
         type_check.expect(a_type.shape[-1] == a_type.shape[-2])
 
     @precision._fp16_mixed_precision_helper
-    def forward_cpu(self, x):
-        self.retain_inputs((0,))
-        self.retain_outputs((0,))
-        detx = utils.force_array(numpy.linalg.det(x[0]))
-        return detx,
-
-    @precision._fp16_mixed_precision_helper
-    def forward_gpu(self, x):
+    def forward(self, inputs):
         self.retain_inputs((0,))
         self.retain_outputs((0,))
-        detx, _ = _det_gpu(x[0])
+        x, = inputs
+        xp = cuda.get_array_module(x)
+        detx = xp.linalg.det(x)
         return detx,
 
     def backward(self, indexes, gy):