From 6724be2b0df66c28c64b0a470080755f30b8f94c Mon Sep 17 00:00:00 2001
From: xiaolil1 <39753926+xiaolil1@users.noreply.github.com>
Date: Wed, 27 Feb 2019 10:09:53 +0800
Subject: [PATCH] INT8 Pool kernel Key Creation Optimization. (#15883)

* Optimize key creation of INT8 pool kernel to improve the peformance of ResNet-50 and MobileNet, especially for latency.
test=develop

* Optimize key creation of pool fp32 grad.
test=develop
---
 .../fluid/operators/mkldnn/pool_mkldnn_op.cc  | 43 ++++++++++---------
 paddle/fluid/platform/mkldnn_reuse.h          |  5 ++-
 2 files changed, 26 insertions(+), 22 deletions(-)

diff --git a/paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc
index 38a65b50bd2..5d8e8192115 100644
--- a/paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc
@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/data_layout_transform.h"
 #include "paddle/fluid/operators/pool_op.h"
 #include "paddle/fluid/platform/mkldnn_helper.h"
+#include "paddle/fluid/platform/mkldnn_reuse.h"
 
 namespace paddle {
 namespace operators {
@@ -29,23 +30,23 @@ using mkldnn::stream;
 using platform::to_void_cast;
 
 // Generate keys for storing/retriving primitives for this operator
-// TODO(jczaja): Make hashing function more optimial
-static std::string gethash(const memory::dims& input_dims,
-                           const std::string& pooling_type,
-                           const std::vector<int>& ksize,
-                           const std::vector<int>& strides,
-                           const std::vector<int>& paddings,
-                           const memory::data_type& dt,
-                           const std::string& suffix) {
-  auto dims2str = [](const memory::dims& operand_dims) {
-    std::string dstr = "";
-    for (size_t i = 0; i < operand_dims.size(); ++i) {
-      dstr += std::to_string(operand_dims[i]) + "-";
-    }
-    return dstr;
-  };
-  return dims2str(input_dims) + dims2str(ksize) + dims2str(strides) +
-         dims2str(paddings) + std::to_string(dt) + pooling_type + suffix;
+std::string CreateKey(const paddle::framework::ExecutionContext& ctx,
+                      const memory::dims& input_dims,
+                      const std::string& pooling_type,
+                      const std::vector<int>& ksize,
+                      const std::vector<int>& strides,
+                      const std::vector<int>& paddings,
+                      const memory::data_type& dt, const std::string& suffix) {
+  std::string key;
+  key.reserve(platform::MKLDNNHandler::MaxKeyLength);
+  platform::MKLDNNHandler::AppendKeyDims(&key, input_dims);
+  platform::MKLDNNHandler::AppendKey(&key, pooling_type);
+  platform::MKLDNNHandler::AppendKeyVec(&key, ksize);
+  platform::MKLDNNHandler::AppendKeyVec(&key, strides);
+  platform::MKLDNNHandler::AppendKeyVec(&key, paddings);
+  platform::MKLDNNHandler::AppendKey(&key, std::to_string(dt));
+  platform::MKLDNNHandler::AppendKey(&key, suffix);
+  return key;
 }
 
 static inline int ComputeCeiledOutput(int input_size, int kernel_size,
@@ -114,8 +115,8 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 
     mkldnn::memory::data_type dt =
         paddle::framework::ToMKLDNNDataType(input->type());
-    const std::string key = gethash(src_tz, pooling_type, ksize, strides,
-                                    paddings, dt, ctx.op().Output("Out"));
+    const std::string key = CreateKey(ctx, src_tz, pooling_type, ksize, strides,
+                                      paddings, dt, ctx.op().Output("Out"));
     const std::string key_pool_p = key + "@pool_p";
     const std::string key_pool_pd = key + "@pool_pd";
     const std::string key_pool_src_mem_p = key + "@pool_src_mem_p";
@@ -294,8 +295,8 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     // Get an unique name from "argument" name of "Out" variable
     // This name will be used as key when referring info from device context
     const std::string key =
-        gethash(diff_src_tz, pooling_type, ksize, strides, paddings,
-                memory::data_type::f32, ctx.op().Input("Out"));
+        CreateKey(ctx, diff_src_tz, pooling_type, ksize, strides, paddings,
+                  memory::data_type::f32, ctx.op().Input("Out"));
     const std::string key_pool_bwd_p = key + "@pool_bwd_p";
     const std::string key_pool_diff_src_mem_p = key + "@pool_diff_src_mem_p";
     const std::string key_pool_diff_dst_mem_p = key + "@pool_diff_dst_mem_p";
diff --git a/paddle/fluid/platform/mkldnn_reuse.h b/paddle/fluid/platform/mkldnn_reuse.h
index 4a674ca526f..4fa6774f028 100644
--- a/paddle/fluid/platform/mkldnn_reuse.h
+++ b/paddle/fluid/platform/mkldnn_reuse.h
@@ -271,7 +271,6 @@ class MKLDNNHandler {
     AppendKey(key, suffix);
   }
 
- protected:
   static void AppendKeyDims(std::string* key,
                             const mkldnn::memory::dims& dims) {
     for (unsigned int i = 0; i < dims.size(); i++) {
@@ -289,6 +288,7 @@ class MKLDNNHandler {
     key->append(s);
   }
 
+ protected:
   static std::string dims2str(const mkldnn::memory::dims& operand_dims) {
     std::string dstr = "";
     for (size_t i = 0; i < operand_dims.size(); ++i) {
@@ -302,6 +302,9 @@ class MKLDNNHandler {
   mkldnn::engine engine_;
   std::string key_;
   bool is_reusing_;
+
+ public:
+  static constexpr int MaxKeyLength = 256;
 };
 
 class TransposeMKLDNNHandler : public MKLDNNHandler {
-- 
GitLab