[Ascend]zq/update batch_norm by aclnn (#1181)

DeepLink-org · Jun 5, 2024 · ec68771 · ec68771
1 parent 160952d
commit ec68771
Show file tree

Hide file tree

Showing 3 changed files with 31 additions and 128 deletions.
diff --git a/impl/ascend/aclnn/adaptor.hpp b/impl/ascend/aclnn/adaptor.hpp
@@ -76,6 +76,7 @@ inline aclTensor* createAclTensorFromDiopiTensor(diopiConstTensorHandle_t tensor
     if (tensor == nullptr) {
         return nullptr;
     }
+
     diopiSize_t shape{};
     diopiGetTensorShape(tensor, &shape);
     diopiSize_t stride{};

diff --git a/impl/ascend/functions/batch_norm.cpp b/impl/ascend/functions/batch_norm.cpp
@@ -4,146 +4,48 @@
  * @copyright  (c) 2023, DeepLink.
  */
 
-#include "../common/acloprunner.hpp"
+#include "../aclnn/acl_scalar.hpp"
+#include "../aclnn/adaptor.hpp"
 
 namespace impl {
 namespace ascend {
 
-void updateInputAscendTensorDim(AscendTensor& inputAt, bool training) {
-    int64_t dim = inputAt.dim();
-    if (2 == dim) {
-        inputAt.unsqueeze(2);
-        inputAt.unsqueeze(3);
-    } else if (3 == dim) {
-        inputAt.unsqueeze(3);
-    } else if (5 == dim && !training) {
-        std::vector<int64_t> shape4d{inputAt.shape(0), inputAt.shape(1), inputAt.shape(2), inputAt.shape(3) * inputAt.shape(4)};
-        inputAt.view(shape4d);
-    }
-}
-
-void batchNormBackwardTrainingUpdate(diopiContextHandle_t ctx, diopiTensorHandle_t gradWeight, diopiTensorHandle_t gradBias, AscendTensor gradOutputAt,
-                                     AscendTensor inputAt, diopiConstTensorHandle_t saveMean, diopiConstTensorHandle_t saveInvstd, double eps) {
-    std::string name = (inputAt.dim() == 5) ? "BN3DTrainingUpdateGrad" : "BNTrainingUpdateGrad";
-    AclOpRunner<4, 2>(name, ctx)
-        .addInput(gradOutputAt)
-        .addInput(inputAt)
-        .addInput(saveMean)
-        .addInput(saveInvstd)
-        .addOutput(gradWeight)
-        .addOutput(gradBias)
-        .setAttr<float>("epsilon", static_cast<float>(eps))
-        .run();
-}
-
-void batchNormBackwardTrainingReduceNocheck(diopiContextHandle_t ctx, AscendTensor gradInputAt, diopiConstTensorHandle_t gradWeight,
-                                            diopiConstTensorHandle_t gradBias, AscendTensor gradOutputAt, AscendTensor inputAt, diopiConstTensorHandle_t weight,
-                                            diopiConstTensorHandle_t saveMean, diopiConstTensorHandle_t saveInvstd, double eps) {
-    std::string name = (inputAt.dim() == 5) ? "BN3DTrainingReduceGrad" : "BNTrainingReduceGrad";
-    AclOpRunner<7, 1>(name, ctx)
-        .addInput(gradOutputAt)
-        .addInput(inputAt)
-        .addInput(gradWeight)
-        .addInput(gradBias)
-        .addInput(weight)
-        .addInput(saveMean)
-        .addInput(saveInvstd)
-        .addOutput(gradInputAt)
-        .setAttr<float>("epsilon", static_cast<float>(eps))
-        .run();
-}
-
 diopiError_t diopiBatchNorm(diopiContextHandle_t ctx, diopiTensorHandle_t out, diopiTensorHandle_t saveMean, diopiTensorHandle_t saveInvstd,
                             diopiConstTensorHandle_t input, diopiConstTensorHandle_t weight, diopiConstTensorHandle_t bias, diopiTensorHandle_t runningMean,
                             diopiTensorHandle_t runningVar, bool training, double momentum, double eps) {
-    AscendTensor inputAt(input), outputAt(out);
-    updateInputAscendTensorDim(inputAt, training);
-    outputAt.view(inputAt.getAclMemShape());
-
-    std::vector<int64_t> batchShapeV{inputAt.shape(1)};
-    diopiSize_t batchShapeSizeT{batchShapeV.data(), static_cast<int64_t>(batchShapeV.size())};
-    diopiTensorHandle_t weightTemp = createTensorIfNullptrOrConstCast(ctx, weight, batchShapeSizeT, inputAt.dtype(), true, 1);
-    diopiTensorHandle_t biasTemp = createTensorIfNullptrOrConstCast(ctx, bias, batchShapeSizeT, inputAt.dtype(), true, 0);
-    diopiTensorHandle_t runningMeanTemp = createTensorIfNullptrOrConstCast(ctx, runningMean, batchShapeSizeT, inputAt.dtype(), true, 0);
-    diopiTensorHandle_t runningVarTemp = createTensorIfNullptrOrConstCast(ctx, runningVar, batchShapeSizeT, inputAt.dtype(), true, 1);
-
-    if (!training) {
-        AclOpRunner<5, 1>("BNInfer", ctx)
-            .addInput(inputAt)
-            .addInput(weightTemp)
-            .addInput(biasTemp)
-            .addInput(runningMeanTemp)
-            .addInput(runningVarTemp)
-            .addOutput(outputAt)
-            .setAttr("epsilon", static_cast<float>(eps))
-            .run();
-
-        diopiTensorHandle_t runningVarBroadcasted;
-        makeTensorLike(ctx, &runningVarBroadcasted, input);
-        AscendTensor runningVarAt(runningVar);
-        runningVarAt.unsqueeze(0);
-        runningVarAt.unsqueeze(2);
-        runningVarAt.unsqueeze(3);
-        AclOpRunner<2, 1>("BroadcastTo", ctx).addInput(runningVarAt).addConstInput(inputAt.shape()).addOutput(runningVarBroadcasted).run();
-    } else {
-        diopiTensorHandle_t sum = nullptr, squareSum = nullptr;
-        diopiSize_t shape, stride;
-        diopiGetTensorShape(runningMeanTemp, &shape);
-        diopiGetTensorStride(runningMeanTemp, &stride);
-        diopiRequireTensor(ctx, &sum, &shape, &stride, diopiDtype_t::diopi_dtype_float32, diopi_device);
-        diopiRequireTensor(ctx, &squareSum, &shape, &stride, diopiDtype_t::diopi_dtype_float32, diopi_device);
-        AclOpRunner<1, 2>("BNTrainingReduce", ctx).addInput(inputAt).addOutput(sum).setAttr("epsilon", static_cast<float>(eps)).addOutput(squareSum).run();
-        AclOpRunner<7, 5>("BNTrainingUpdate", ctx)
-            .addInput(inputAt)
-            .addInput(sum)
-            .addInput(squareSum)
-            .addInput(weightTemp)
-            .addInput(biasTemp)
-            .addInput(runningMeanTemp)
-            .addInput(runningVarTemp)
-            .setAttr("epsilon", static_cast<float>(eps))
-            .setAttr("factor", static_cast<float>(momentum))
-            .addOutput(outputAt)
-            .addOutput(runningMeanTemp)
-            .addOutput(runningVarTemp)
-            .addOutput(saveMean)
-            .addOutput(saveInvstd)
-            .run();
-    }
+    DIOPI_ASCEND_CALL_ACLNN(aclnnBatchNorm, ctx, input, weight, bias, runningMean, runningVar, training, momentum, eps, out, saveMean, saveInvstd);
     return diopiSuccess;
 }
 
 diopiError_t diopiBatchNormBackward(diopiContextHandle_t ctx, diopiTensorHandle_t gradInput, diopiTensorHandle_t gradWeight, diopiTensorHandle_t gradBias,
                                     diopiConstTensorHandle_t gradOutput, diopiConstTensorHandle_t input, diopiConstTensorHandle_t weight,
-                                    diopiConstTensorHandle_t runninMean, diopiConstTensorHandle_t runningVar, diopiConstTensorHandle_t saveMean,
+                                    diopiConstTensorHandle_t runningMean, diopiConstTensorHandle_t runningVar, diopiConstTensorHandle_t saveMean,
                                     diopiConstTensorHandle_t saveInvstd, bool training, double eps) {
-    AscendTensor inputAt(input), gradOutputAt(gradOutput), gradInputAt(gradInput);
-    updateInputAscendTensorDim(inputAt, training);
-    gradOutputAt.view(inputAt.getAclMemShape());
-    gradInputAt.view(inputAt.getAclMemShape());
-
-    if (!training) {
-        batchNormBackwardTrainingUpdate(ctx, gradWeight, gradBias, gradOutputAt, inputAt, runninMean, runningVar, eps);
-
-        AclOpRunner<3, 1>("BNInferGrad", ctx)
-            .addInput(gradOutputAt)
-            .addInput(weight)
-            .addInput(runningVar)
-            .addOutput(gradInputAt)
-            .setAttr<float>("epsilon", static_cast<float>(eps))
-            .run();
-
-        diopiTensorHandle_t runningVarBroadcasted;
-        makeTensorLike(ctx, &runningVarBroadcasted, input);
-        AscendTensor runningVarAt(runningVar);
-        runningVarAt.unsqueeze(0);
-        runningVarAt.unsqueeze(2);
-        runningVarAt.unsqueeze(3);
-        AclOpRunner<2, 1>("BroadcastTo", ctx).addInput(runningVarAt).addConstInput(inputAt.shape()).addOutput(runningVarBroadcasted).run();
-    } else {
-        batchNormBackwardTrainingUpdate(ctx, gradWeight, gradBias, gradOutputAt, inputAt, saveMean, saveInvstd, eps);
-        batchNormBackwardTrainingReduceNocheck(ctx, gradInputAt, gradWeight, gradBias, gradOutputAt, inputAt, weight, saveMean, saveInvstd, eps);
+    std::array<bool, 3> gradMask = {true, true, true};
+    if (nullptr == gradInput) {
+        gradMask[0] = false;
+    }
+    if (nullptr == gradWeight) {
+        gradMask[1] = false;
+    }
+    if (nullptr == gradBias) {
+        gradMask[2] = false;
     }
+    DIOPI_ASCEND_CALL_ACLNN(aclnnBatchNormBackward,
+                            ctx,
+                            gradOutput,
+                            input,
+                            weight,
+                            runningMean,
+                            runningVar,
+                            saveMean,
+                            saveInvstd,
+                            training,
+                            eps,
+                            gradMask,
+                            gradInput,
+                            gradWeight,
+                            gradBias);
     return diopiSuccess;
 }
 

diff --git a/impl/ascend_npu/ascend_config.yaml b/impl/ascend_npu/ascend_config.yaml
@@ -20,6 +20,8 @@ ascend:
 - diopiAtanInp
 - diopiBaddbmm
 - diopiBaddbmmInp
+- diopiBatchNorm
+- diopiBatchNormBackward
 - diopiBitwiseNot
 - diopiBitwiseNotInp
 - diopiBitwiseAnd
@@ -219,8 +221,6 @@ ascend:
 - diopiZeros
 ascend_npu:
 - diopiAdamW
-- diopiBatchNorm
-- diopiBatchNormBackward
 - diopiNonzero
 - diopiCat
 - diopiCopyInp