“8d3c8d674c446d66f5539814a17d5aabc1ea72b0”上不存在“mobile/src/operators/sequence_ops/sequence_pool_op.h”
提交 acf0ed94 编写于 作者: 李寅

Merge branch 'fused-bn' into 'master'

Memory opt: use int to calculate memory size.

See merge request !201
...@@ -53,7 +53,7 @@ class MemoryOptimizer(object): ...@@ -53,7 +53,7 @@ class MemoryOptimizer(object):
self.mem_block[mem_id] = [0, 0] self.mem_block[mem_id] = [0, 0]
mem_size = self.mem_block[mem_id] mem_size = self.mem_block[mem_id]
mem_size[1] = max(mem_size[1], op.output_shape[0].dims[0] * op.output_shape[0].dims[1]) mem_size[1] = max(mem_size[1], op.output_shape[0].dims[0] * op.output_shape[0].dims[1])
mem_size[0] = max(mem_size[0], op.output_shape[0].dims[2] * (op.output_shape[0].dims[3]+3)/4) mem_size[0] = max(mem_size[0], op.output_shape[0].dims[2] * int((op.output_shape[0].dims[3]+3)/4))
# de-ref input tensor mem # de-ref input tensor mem
for ipt in op.input: for ipt in op.input:
...@@ -79,6 +79,7 @@ class MemoryOptimizer(object): ...@@ -79,6 +79,7 @@ class MemoryOptimizer(object):
continue continue
origin_mem_size += reduce(operator.mul, op.output_shape[0].dims, 1) origin_mem_size += reduce(operator.mul, op.output_shape[0].dims, 1)
for mem in self.mem_block: for mem in self.mem_block:
print mem, self.mem_block[mem]
optimized_mem_size += reduce(operator.mul, self.mem_block[mem], 4) optimized_mem_size += reduce(operator.mul, self.mem_block[mem], 4)
print('origin mem: %d, optimized mem: %d', origin_mem_size, optimized_mem_size) print('origin mem: %d, optimized mem: %d', origin_mem_size, optimized_mem_size)
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册