Change backward_guard to optimize_guard to maximize the allreduce overlap. (#19506)

Change backward_guard to optimize_guard to maximize the allreduce overlap

paddle/fluid/framework/operator_kernel_configs.h

@@ -81,6 +81,8 @@ TAlgorithm framework::AlgorithmsCache<TAlgorithm>::GetAlgorithm(
   seed ^= hashFn(static_cast<int64_t>(algorithmFlags)) + 0x9e3779b9 +
           (seed << 6) + (seed >> 2) + 5;
 
+  VLOG(10) << "seed:" << seed << ", hash_.size:" << hash_.size();
+
   if (seed == 0) return gen_func();
 
   if (hash_.find(seed) == hash_.end()) {

paddle/fluid/operators/conv_cudnn_helper.h

@@ -22,6 +22,14 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+template <typename T>
+std::ostream& operator<<(std::ostream& out, const std::vector<T>& v) {
+  out << "[";
+  for (auto const& tmp : v) out << tmp << ",";
+  out << "]";
+  return out;
+}
+
 using framework::AlgorithmsCache;
 
 struct ConvArgs {
@@ -119,6 +127,11 @@ struct SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t> {
       auto x_dims = framework::vectorize(args.x->dims());
       auto w_dims = framework::vectorize(args.w->dims());
 
+      VLOG(10) << "cudnnConvolutionFwdAlgoPerf_t algo_cache_id:"
+               << algo_cache_id << ", x_dims:" << x_dims
+               << ", w_dims:" << w_dims << ", args.s" << args.s << ", args.p"
+               << args.p << ", args.d" << args.d;
+
       algo = algo_cache.GetAlgorithm(
           x_dims, w_dims, args.s, args.p, args.d, 0, [&]() {
             int returned_algo_count;
@@ -247,6 +260,11 @@ struct SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t> {
       auto x_dims = framework::vectorize(args.x->dims());
       auto w_dims = framework::vectorize(args.w->dims());
 
+      VLOG(10) << "cudnnConvolutionFwdAlgoPerf_t algo_cache_id:"
+               << algo_cache_id << ", x_dims:" << x_dims
+               << ", w_dims:" << w_dims << ", args.s" << args.s << ", args.p"
+               << args.p << ", args.d" << args.d;
+
       algo = algo_cache.GetAlgorithm(
           x_dims, w_dims, args.s, args.p, args.d, 0, [&]() {
             int returned_algo_count;
@@ -368,6 +386,11 @@ struct SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t> {
       auto x_dims = framework::vectorize(args.x->dims());
       auto w_dims = framework::vectorize(args.w->dims());
 
+      VLOG(10) << "cudnnConvolutionFwdAlgoPerf_t algo_cache_id:"
+               << algo_cache_id << ", x_dims:" << x_dims
+               << ", w_dims:" << w_dims << ", args.s" << args.s << ", args.p"
+               << args.p << ", args.d" << args.d;
+
       algo = algo_cache.GetAlgorithm(
           x_dims, w_dims, args.s, args.p, args.d, 0, [&]() {
             int returned_algo_count;

python/paddle/fluid/contrib/mixed_precision/fp16_utils.py

@@ -80,8 +80,9 @@ def create_master_params_grads(params_grads, main_prog, startup_prog,
         A list of master parameters and gradients. 
     """
     master_params_grads = []
-    with main_prog._backward_role_guard():
-        for p, g in params_grads:
+    for p, g in params_grads:
+        # create master parameters
+        with main_prog._optimized_guard([p, g]):
             # create master parameters
             master_param = copy_to_master_param(p, main_prog.global_block())
             startup_master_param = startup_prog.global_block()._clone_variable(

python/paddle/fluid/data_feeder.py

@@ -278,8 +278,8 @@ class DataFeeder(object):
 
         for each_sample in iterable:
             assert len(each_sample) == len(converter), (
-                "The number of fields in data (%s) does not match " +
-                "len(feed_list) (%s)") % (len(each_sample), len(converter))
+                "The number of fields in data (%d) does not match " +
+                "len(feed_list) (%d)") % (len(each_sample), len(converter))
             for each_converter, each_slot in six.moves.zip(converter,
                                                            each_sample):
                 each_converter.feed(each_slot)