gpu opt, don't move float16 ops or cast to/from float32 in the graph for some…

theano/gpuarray/opt.py

@@ -1327,6 +1327,8 @@ theano.tensor.nnet.conv2d()
 @op_lifter([SparseBlockGemv])
 @register_opt2([SparseBlockGemv], 'fast_compile')
 def local_gpua_sparseblockgemv(op, context_name, inputs, outputs):
+    if inputs[0].dtype == 'float16':
+        return
     if op.inplace:
         return gpu_sparse_block_gemv_inplace
     else:
@@ -1337,6 +1339,8 @@ def local_gpua_sparseblockgemv(op, context_name, inputs, outputs):
 @op_lifter([SparseBlockOuter])
 @register_opt2([SparseBlockOuter], 'fast_compile')
 def local_gpua_sparseblockouter(op, context_name, inputs, outputs):
+    if inputs[0].dtype == 'float16':
+        return
     if op.inplace:
         return gpu_sparse_block_outer_inplace
     else:
@@ -1990,9 +1994,15 @@ def local_gpu_maxandargmax(op, context_name, inputs, outputs):
 def local_gpu_solve(op, context_name, inputs, outputs):
     if not cusolver_available:
         return
+    if inputs[0].dtype not in ['float16', 'float32']:
+        return
     if op.A_structure not in MATRIX_STRUCTURES_SOLVE:
         return
-    return GpuCusolverSolve(A_structure=op.A_structure)
+    op = GpuCusolverSolve(A_structure=op.A_structure)
+    if inputs[0].dtype == 'float16':
+        return op(inputs[0].astype('float32'),
+                  inputs[1].astype('float32')).astype('float16')
+    return op
 
 
 @register_inplace()
@@ -2010,7 +2020,13 @@ def local_inplace_gpu_solve(node):
 def local_gpu_cholesky(op, context_name, inputs, outputs):
     if not cusolver_available:
         return
-    return GpuCholesky(lower=op.lower, inplace=op.destructive)
+    if inputs[0].dtype not in ['float16', 'float32']:
+        return
+    op = GpuCholesky(lower=op.lower, inplace=op.destructive)
+    if inputs[0].dtype == 'float16':
+        return op(inputs[0].astype('float32')).astype('float16')
+
+    return op
 
 
 @register_inplace()
@@ -2026,7 +2042,12 @@ def local_inplace_cholesky(node):
 def local_gpu_matrix_inverse(op, context_name, inputs, outputs):
     if not config.magma.enabled:
         return
-    return GpuMagmaMatrixInverse()
+    if inputs[0].dtype not in ['float16', 'float32']:
+        return
+    op = GpuMagmaMatrixInverse()
+    if inputs[0].dtype == 'float16':
+        return op(inputs[0].astype('float32')).astype('float16')
+    return op
 
 
 @register_inplace()
@@ -2043,9 +2064,13 @@ def local_inplace_matrix_inverse_inplace(node):
 def local_gpu_svd(op, context_name, inputs, outputs):
     if not config.magma.enabled:
         return
-    return GpuMagmaSVD(full_matrices=op.full_matrices,
-                       compute_uv=op.compute_uv)
-
+    if inputs[0].dtype not in ['float16', 'float32']:
+        return
+    op = GpuMagmaSVD(full_matrices=op.full_matrices,
+                     compute_uv=op.compute_uv)
+    if inputs[0].dtype == 'float16':
+        return op(inputs[0].astype('float32')).astype('float16')
+    return op
 
 # Do not register in fast_run or fast_compile.
 # It will be added to fast_run if the GPU is enabled.