[bf16] Refine BF16 amp-o1 logic (#39815)

* refine bf16 amp-o1 logic * refine amp GLOG * refine unittest * refine unittest
PaddlePaddle · Feb 28, 2022 · 18ee051 · 18ee051
1 parent d1595c2
commit 18ee051
Show file tree

Hide file tree

Showing 4 changed files with 45 additions and 12 deletions.
diff --git a/paddle/fluid/imperative/amp_auto_cast.cc b/paddle/fluid/imperative/amp_auto_cast.cc
@@ -273,8 +273,9 @@ static inline std::shared_ptr<VarType> CastToBF16(
 
 template <typename VarType>
 static inline framework::proto::VarType::Type GetPromoteType(
-    const std::string& op_type, const NameVarMap<VarType>& ins) {
-  auto dst_type = framework::proto::VarType::FP16;
+    const std::string& op_type, const NameVarMap<VarType>& ins,
+    const framework::proto::VarType::Type amp_dtype) {
+  auto dst_type = amp_dtype;
   for (const auto& pair : ins) {
     for (const auto& var : pair.second) {
       if (GetDataType<VarType>(var) == framework::proto::VarType::FP32) {
@@ -337,7 +338,8 @@ NameVarMap<VarType> AutoCastInputs(const std::string& op_type,
     }
     return new_ins;
   } else {
-    auto dst_type = GetPromoteType<VarType>(op_type, ins);
+    auto dst_type =
+        GetPromoteType<VarType>(op_type, ins, framework::proto::VarType::FP16);
 
     // NOTE(zhiqiu): if the op has op fp16 kernel, fall back to fp32.
     if (dst_type == framework::proto::VarType::FP16 &&
@@ -435,7 +437,7 @@ NameVarMap<VarType> AutoCastBF16Inputs(const std::string& op_type,
       }
     }
     return new_ins;
-  } else {
+  } else if (AmpOperators::Instance().GetMutableBlockOps()->count(op_type)) {
     for (auto& pair : new_ins) {
       VLOG(5) << "Op(" << op_type << "): Cast " << pair.first << " from "
               << GetDtypeStr(*pair.second.cbegin()) << " to float";
@@ -444,6 +446,26 @@ NameVarMap<VarType> AutoCastBF16Inputs(const std::string& op_type,
       }
     }
     return new_ins;
+  } else {
+    auto dst_type =
+        GetPromoteType<VarType>(op_type, ins, framework::proto::VarType::BF16);
+    // NOTE(zhangbo): if the op has op fp16 kernel, fall back to fp32.
+    if (dst_type == framework::proto::VarType::BF16 &&
+        AmpOperators::Instance().GetMutableUnsupportedBf16Ops()->count(
+            op_type)) {
+      dst_type = framework::proto::VarType::FP32;
+    }
+    for (auto& pair : new_ins) {
+      VLOG(5) << "Op(" << op_type << "): Cast " << pair.first << " from "
+              << GetDtypeStr(*pair.second.cbegin()) << " to "
+              << framework::DataTypeToString(dst_type);
+      for (auto& var : pair.second) {
+        var = (dst_type == framework::proto::VarType::FP32
+                   ? CastToFP32<VarType>(var)
+                   : CastToBF16<VarType>(var));
+      }
+    }
+    return new_ins;
   }
   return new_ins;
 }

diff --git a/paddle/fluid/imperative/tracer.cc b/paddle/fluid/imperative/tracer.cc
@@ -205,17 +205,19 @@ void Tracer::TraceOp(const std::string& type, const NameVarMap<VarType>& ins,
 
   NameVarMap<VarType> new_ins = ins;
   if (amp_level_ == AmpLevel::O1) {
-    VLOG(5) << "Auto mixed precision run operator: " << type;
     if (amp_dtype_ == phi::DataType::FLOAT16) {
+      VLOG(5) << "Float16 Auto Mixed Precision O1 run operator: " << type;
       new_ins = AutoCastInputs<VarType>(type, ins);
     } else if (amp_dtype_ == phi::DataType::BFLOAT16) {
+      VLOG(5) << "BFloat16 Auto Mixed Precision O1 run operator: " << type;
       new_ins = AutoCastBF16Inputs<VarType>(type, ins);
     }
   } else if (amp_level_ == AmpLevel::O2) {
-    VLOG(5) << "Pure fp16 run operator: " << type;
     if (amp_dtype_ == phi::DataType::FLOAT16) {
+      VLOG(5) << "Float16 Auto Mixed Precision O2 run operator: " << type;
       new_ins = CastPureFp16Inputs<VarType>(type, ins);
     } else if (amp_dtype_ == phi::DataType::BFLOAT16) {
+      VLOG(5) << "BFloat16 Auto Mixed Precision O2 run operator: " << type;
       new_ins = CastPureBf16Inputs<VarType>(type, ins);
     }
   }

diff --git a/python/paddle/fluid/dygraph/amp/auto_cast.py b/python/paddle/fluid/dygraph/amp/auto_cast.py
@@ -75,7 +75,7 @@
     'lookup_table', 'lookup_table_v2', 'scatter', 'scatter_grad'
 }
 
-BF16_WHITE_LIST = {'conv2d'}
+BF16_WHITE_LIST = {'conv2d', 'matmul_v2'}
 BF16_BLACK_LIST = {' '}
 
 _g_amp_state_ = None

diff --git a/python/paddle/fluid/tests/unittests/test_imperative_auto_mixed_precision.py b/python/paddle/fluid/tests/unittests/test_imperative_auto_mixed_precision.py
@@ -1131,20 +1131,29 @@ class TestBf16(unittest.TestCase):
     test amp for BF16 
     '''
 
-    def train(self, enable_amp=True):
+    def train(self, enable_amp=True, amp_level='O1'):
         paddle.seed(100)
         input = paddle.uniform((2, 4, 8, 8), dtype='float32', min=-1., max=1.)
         conv = paddle.nn.Conv2D(4, 6, (3, 3))
         with paddle.amp.auto_cast(
-                enable=enable_amp, level='O2', dtype='bfloat16'):
+                enable=enable_amp, level=amp_level, dtype='bfloat16'):
             output = conv(input)
         output = output.cast('float32')
         return output.numpy()
 
     def test_bf16(self):
-        out_fp32 = self.train(enable_amp=False)
-        out_bf16 = self.train(enable_amp=True)
-        self.assertTrue(np.allclose(out_fp32, out_bf16, rtol=1.e-3, atol=1.e-1))
+        if fluid.core.is_compiled_with_cuda():
+            cudnn_version = paddle.device.get_cudnn_version()
+            if cudnn_version is not None and cudnn_version >= 8100:
+                out_fp32 = self.train(enable_amp=False)
+                out_bf16_O1 = self.train(enable_amp=True, amp_level='O1')
+                out_bf16_O2 = self.train(enable_amp=True, amp_level='O2')
+                self.assertTrue(
+                    np.allclose(
+                        out_fp32, out_bf16_O1, rtol=1.e-3, atol=1.e-1))
+                self.assertTrue(
+                    np.allclose(
+                        out_fp32, out_bf16_O2, rtol=1.e-3, atol=1.e-1))
 
 
 class TestPyLayerWithAmp(unittest.TestCase):