Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Update passes in quant2_int8_mkldnn_pass #38912

Merged
merged 3 commits into from
Jan 27, 2022
Merged

Update passes in quant2_int8_mkldnn_pass #38912

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d3e1d67
Upadate pass in quant2_int8_mkldnn_pass
wozna Jan 11, 2022
4f2e240
Back to the previous scale_matmul order
wozna Jan 17, 2022
95d68aa
Change place of cpu_quantize_placement_pass
wozna Jan 19, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 22 additions & 4 deletions python/paddle/fluid/contrib/slim/quantization/quant2_int8_mkldnn_pass.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -397,6 +397,7 @@ def _remove_ctrl_vars(self, graph):
def _optimize_fp32_graph(self, graph):
graph = self._update_activations(graph)
graph = self._remove_ctrl_vars(graph)
graph = self._apply_pass(graph, 'layer_norm_fuse_pass')
graph = self._apply_pass(graph, 'attention_lstm_fuse_pass')
graph = self._apply_pass(graph, 'seqconv_eltadd_relu_fuse_pass')
# graph = self._apply_pass(graph, 'seqpool_concat_fuse_pass')
Expand All @@ -409,24 +410,39 @@ def _optimize_fp32_graph(self, graph):
graph = self._apply_pass(graph, 'multi_gru_fuse_pass')
graph = self._apply_pass(graph, 'multi_gru_seq_fuse_pass')
graph = self._apply_pass(graph, 'seq_concat_fc_fuse_pass')
graph = self._apply_pass(graph, 'squeeze2_matmul_fuse_pass')
graph = self._apply_pass(graph, 'reshape2_matmul_fuse_pass')
graph = self._apply_pass(graph, 'flatten2_matmul_fuse_pass')
graph = self._apply_pass(graph, 'matmul_v2_scale_fuse_pass')
graph = self._apply_pass(graph, 'squared_mat_sub_fuse_pass')
graph = self._apply_pass(graph, 'is_test_pass')
graph = self._apply_pass(graph, 'map_matmul_v2_to_mul_pass')
graph = self._apply_pass(graph, 'map_matmul_v2_to_matmul_pass')
graph = self._apply_pass(graph, 'matmul_scale_fuse_pass')
graph = self._apply_pass(graph, 'map_matmul_to_mul_pass')
graph = self._apply_pass(graph, 'repeated_fc_relu_fuse_pass')
graph = self._apply_pass(graph, 'mkldnn_placement_pass',
['mkldnn_enabled_op_types'], [set()])
graph = self._apply_pass(graph, 'depthwise_conv_mkldnn_pass')
graph = self._apply_pass(graph, 'conv_bn_fuse_pass')
graph = self._apply_pass(graph, 'conv_eltwiseadd_bn_fuse_pass')
graph = self._apply_pass(graph, 'conv_affine_channel_fuse_pass')
graph = self._apply_pass(graph,
'conv_eltwiseadd_affine_channel_fuse_pass')
graph = self._apply_pass(graph, 'conv_transpose_bn_fuse_pass')
graph = self._apply_pass(graph,
'conv_transpose_eltwiseadd_bn_fuse_pass')
graph = self._apply_pass(graph, 'conv_bias_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_transpose_bias_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_elementwise_add_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_concat_relu_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_relu_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_leaky_relu_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_relu6_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_swish_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_hard_swish_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_hard_sigmoid_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'conv_gelu_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'fc_fuse_pass',
['use_gpu', 'use_fc_padding'], [False, False])
graph = self._apply_pass(graph, 'repeated_fc_relu_fuse_pass')
Expand All @@ -436,6 +452,8 @@ def _optimize_fp32_graph(self, graph):
graph = self._apply_pass(graph, 'fc_act_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'matmul_transpose_reshape_fuse_pass')
graph = self._apply_pass(graph, 'matmul_v2_transpose_reshape_fuse_pass')
graph = self._apply_pass(graph, 'batch_norm_act_fuse_pass')
graph = self._apply_pass(graph, 'softplus_activation_mkldnn_fuse_pass')
# the following pass should be the last one since it will work on all fused ops.
graph = self._apply_pass(graph, 'runtime_context_cache_pass')
return graph
Expand Down Expand Up @@ -638,15 +656,15 @@ def _get_data_layout(self, graph):
return 'NHWC' if self._is_conv_quantized(graph) else 'NCHW'

def _quantize_fp32_graph(self, graph):
graph = self._apply_pass(
graph, 'cpu_quantize_placement_pass',
['quantize_enabled_op_types', 'quantize_excluded_op_ids'],
[self._ops_to_quantize, self._find_avg_pooling_ids(graph)])
graph = self._apply_pass(graph, 'scale_matmul_fuse_pass')
graph = self._apply_pass(graph,
'reshape_transpose_matmul_mkldnn_fuse_pass')
graph = self._apply_pass(graph,
'reshape_transpose_matmul_v2_mkldnn_fuse_pass')
graph = self._apply_pass(
graph, 'cpu_quantize_placement_pass',
['quantize_enabled_op_types', 'quantize_excluded_op_ids'],
[self._ops_to_quantize, self._find_avg_pooling_ids(graph)])
graph = self._apply_pass(
graph, 'cpu_quantize_pass', ['quant_var_scales', 'data_layout'],
[self._var_quant_scales, self._get_data_layout(graph)])
Expand Down