[NPU] Support fused_linear_param_grad_add (PaddlePaddle#976)

zlzzl · web-flow · commit a923e01583dc · 2024-03-06T12:02:36.000+08:00
diff --git a/backends/npu/kernels/fused_linear_param_grad_add_kernel.cc b/backends/npu/kernels/fused_linear_param_grad_add_kernel.cc
@@ -0,0 +1,133 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "kernels/funcs/npu_funcs.h"
+#include "kernels/funcs/npu_op_runner.h"
+
+namespace custom_kernel {
+
+template <typename T, typename Context>
+void FusedLinearParamGradAdd(const Context &dev_ctx,
+                             const phi::DenseTensor &x,
+                             const phi::DenseTensor &dout,
+                             const paddle::optional<phi::DenseTensor> &dweight,
+                             const paddle::optional<phi::DenseTensor> &dbias,
+                             bool multi_precision,
+                             bool has_bias,
+                             phi::DenseTensor *dweight_out,
+                             phi::DenseTensor *dbias_out) {
+  using MT = typename phi::dtype::MPTypeTrait<T>::Type;
+
+  int64_t K = x.dims()[x.dims().size() - 1];
+  int64_t N = dout.dims()[dout.dims().size() - 1];
+
+  phi::DenseTensor reshape_x(x);
+  reshape_x.Resize({x.numel() / K, K});
+  phi::DenseTensor reshape_dout(dout);
+  reshape_dout.Resize({dout.numel() / N, N});
+
+  if (dweight_out && dweight) {
+    *dweight_out = dweight.get();
+    if (multi_precision) {
+      PADDLE_ENFORCE_EQ(
+          dweight_out->dtype(),
+          phi::CppTypeToDataType<MT>::Type(),
+          phi::errors::InvalidArgument("Invaid data type error."));
+    } else {
+      PADDLE_ENFORCE_EQ(
+          dweight_out->dtype(),
+          phi::CppTypeToDataType<T>::Type(),
+          phi::errors::InvalidArgument("Invaid data type error."));
+    }
+  } else {
+    dweight_out->Resize(phi::make_ddim({K, N}));
+    if (multi_precision) {
+      dev_ctx.template Alloc<MT>(dweight_out);
+    } else {
+      dev_ctx.template Alloc<T>(dweight_out);
+    }
+  }
+
+  if (has_bias && dbias_out) {
+    dev_ctx.template Alloc<T>(dbias_out);
+  }
+
+  float alpha = 1.0;
+  float beta = 1.0;
+
+  phi::DenseTensor new_dweight;
+  if (dweight) {
+    new_dweight = dweight.get();
+  } else {
+    phi::DenseTensorMeta dweight_meta = {x.dtype(), {K, N}};
+    new_dweight.set_meta(dweight_meta);
+    FillNpuTensorWithConstant<T>(&new_dweight, dev_ctx, static_cast<T>(0));
+    new_dweight.Resize({K, N});
+  }
+
+  int64_t trans_a = 1;
+  int64_t trans_b = 0;
+  int8_t cube_math_type = 0;
+  bool keep_dim = false;
+  EXEC_NPU_CMD(aclnnGemm,
+               dev_ctx,
+               reshape_x,
+               reshape_dout,
+               new_dweight,
+               alpha,
+               beta,
+               trans_a,
+               trans_b,
+               *dweight_out,
+               cube_math_type);
+  if (has_bias) {
+    phi::IntArray axis = {0};
+
+    phi::DenseTensor new_dbias;
+    if (dbias) {
+      new_dbias = dbias.get();
+    } else {
+      phi::DenseTensorMeta new_dbias_meta = {x.dtype(), {N}};
+      new_dbias.set_meta(new_dbias_meta);
+      FillNpuTensorWithConstant<T>(&new_dbias, dev_ctx, static_cast<T>(0));
+    }
+
+    auto dst_dtype = ConvertToNpuDtype(reshape_dout.dtype());
+
+    phi::DenseTensor bias_sum;
+    phi::DenseTensorMeta bias_sum_meta = {x.dtype(), {N}};
+    bias_sum.set_meta(bias_sum_meta);
+    dev_ctx.template Alloc<T>(&bias_sum);
+    EXEC_NPU_CMD(aclnnReduceSum,
+                 dev_ctx,
+                 reshape_dout,
+                 axis,
+                 keep_dim,
+                 dst_dtype,
+                 bias_sum);
+    phi::Scalar add_alpha = 1.0;
+    EXEC_NPU_CMD(aclnnAdd, dev_ctx, bias_sum, new_dbias, add_alpha, *dbias_out);
+  }
+}
+
+}  // namespace custom_kernel
+
+PD_REGISTER_PLUGIN_KERNEL(fused_linear_param_grad_add,
+                          npu,
+                          ALL_LAYOUT,
+                          custom_kernel::FusedLinearParamGradAdd,
+                          float,
+                          double,
+                          phi::dtype::float16,
+                          phi::dtype::bfloat16) {}
diff --git a/backends/npu/tests/unittests/test_fused_linear_param_grad_add_op_npu.py b/backends/npu/tests/unittests/test_fused_linear_param_grad_add_op_npu.py
@@ -0,0 +1,165 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import numpy as np
+
+import paddle
+from paddle import _C_ops
+from npu_utils import check_soc_version
+
+
+def promote_dtype(x):
+    if x.dtype in [paddle.float16, paddle.bfloat16]:
+        return x.astype(paddle.float32)
+    else:
+        return x
+
+
+def recreate(x, multi_precision):
+    if isinstance(x, (list, tuple)):
+        return [recreate(item, multi_precision) for item in x]
+
+    if x is None:
+        return None
+
+    if multi_precision:
+        x = promote_dtype(x)
+
+    return paddle.to_tensor(x.numpy())
+
+
+def run_ground_truth(x, dy, dweight, dbias, multi_precision, has_bias):
+    x, dy, dweight, dbias = recreate([x, dy, dweight, dbias], multi_precision)
+
+    dweight_tmp = paddle.matmul(
+        x.reshape([-1, x.shape[-1]]),
+        dy.reshape([-1, dy.shape[-1]]),
+        transpose_x=True,
+    )
+    if dweight is None:
+        dweight = dweight_tmp
+    else:
+        assert dweight.shape == dweight_tmp.shape
+        assert dweight.dtype == dweight.dtype
+        dweight += dweight_tmp
+
+    if has_bias:
+        dbias_tmp = dy.reshape([-1, dy.shape[-1]]).sum(axis=0)
+        if dbias is None:
+            dbias = dbias_tmp
+        else:
+            assert dbias.shape == dbias_tmp.shape
+            assert dbias.dtype == dbias_tmp.dtype
+            dbias += dbias_tmp
+
+        return promote_dtype(dweight).numpy(), promote_dtype(dbias).numpy()
+    else:
+        return promote_dtype(dweight).numpy()
+
+
+def run_fused_linear_param_grad_add(x, dy, dweight, dbias, multi_precision, has_bias):
+    dweight_new, dbias_new = _C_ops.fused_linear_param_grad_add(
+        x, dy, dweight, dbias, multi_precision, has_bias
+    )
+    if dweight is not None:
+        assert dweight_new.data_ptr() == dweight.data_ptr()
+    if has_bias:
+        return (
+            promote_dtype(dweight_new).numpy(),
+            promote_dtype(dbias_new).numpy(),
+        )
+    else:
+        return promote_dtype(dweight_new).numpy()
+
+
+class TestMainClassBase(unittest.TestCase):
+    def setUp(self):
+        self.shape = [3, 4, 32]
+        self.output_size = 128
+        self.dtype = paddle.float16
+
+    def config(self):
+        pass
+
+    def rand(self, shape, dtype=None):
+        x = np.random.randint(low=-5, high=5, size=shape)
+        x = paddle.to_tensor(x)
+        return x.astype(dtype or self.dtype)
+
+    def generate_rand_inputs(self, has_dweight, has_dbias, multi_precision, has_bias):
+        x_shape = self.shape
+        dy_shape = self.shape[:-1] + [self.output_size]
+        dweight_shape = [self.shape[-1], self.output_size]
+        dbias_shape = [self.output_size]
+
+        x = self.rand(x_shape)
+        dy = self.rand(dy_shape)
+        if has_dweight:
+            dweight = self.rand(dweight_shape)
+            if multi_precision:
+                dweight = promote_dtype(dweight)
+        else:
+            dweight = None
+
+        if has_bias and has_dbias:
+            dbias = self.rand(dbias_shape)
+            if multi_precision:
+                dbias = promote_dtype(dbias)
+        else:
+            dbias = None
+        return x, dy, dweight, dbias
+
+    def check_main(self, has_dweight, has_dbias, multi_precision, has_bias):
+        x, dy, dweight, dbias = self.generate_rand_inputs(
+            has_dweight, has_dbias, multi_precision, has_bias
+        )
+        res1 = run_ground_truth(x, dy, dweight, dbias, multi_precision, has_bias)
+        res2 = run_fused_linear_param_grad_add(
+            x, dy, dweight, dbias, multi_precision, has_bias
+        )
+        self.assertEqual(len(res1), len(res2))
+        for r1, r2 in zip(res1, res2):
+            max_diff = np.max(np.abs(r1 - r2))
+            self.assertLess(max_diff, 1e-10)
+
+    @check_soc_version
+    def test_main(self):
+        for has_dweight in [False, True]:
+            for has_bias in [False, True]:
+                for has_dbias in [False, True]:
+                    for multi_precision in [False, True]:
+                        self.check_main(
+                            has_dweight, has_dbias, multi_precision, has_bias
+                        )
+
+
+class TestMainClassBF16(TestMainClassBase):
+    def config(self):
+        self.dtype = paddle.bfloat16
+
+
+class TestMainClassFP32(TestMainClassBase):
+    def config(self):
+        self.dtype = paddle.float32
+
+
+class TestMainClassFP64(TestMainClassBase):
+    def config(self):
+        self.dtype = paddle.float64
+
+
+if __name__ == "__main__":
+    unittest.main()
