|
| 1 | +/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. |
| 2 | +Licensed under the Apache License, Version 2.0 (the "License"); |
| 3 | +you may not use this file except in compliance with the License. |
| 4 | +You may obtain a copy of the License at |
| 5 | + http://www.apache.org/licenses/LICENSE-2.0 |
| 6 | +Unless required by applicable law or agreed to in writing, software |
| 7 | +distributed under the License is distributed on an "AS IS" BASIS, |
| 8 | +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 9 | +See the License for the specific language governing permissions and |
| 10 | +limitations under the License. */ |
| 11 | + |
| 12 | +#include "paddle/phi/backends/onednn/onednn_reuse.h" |
| 13 | +#include "paddle/phi/core/kernel_registry.h" |
| 14 | + |
| 15 | +namespace phi { |
| 16 | + |
| 17 | +static DDim ValidateShape(const std::vector<int64_t>& shape, |
| 18 | + const DDim& in_dims) { |
| 19 | + const int64_t in_size = product(in_dims); |
| 20 | + auto in_dims_vec = vectorize(in_dims); |
| 21 | + bool all_positive = std::all_of(in_dims_vec.cbegin(), |
| 22 | + in_dims_vec.cend(), |
| 23 | + [](int64_t i) { return i > 0; }); |
| 24 | + // only one dimension can be set to -1, whose size will be automatically |
| 25 | + // infered |
| 26 | + const int64_t unk_dim_val = -1; |
| 27 | + const int64_t copy_dim_val = 0; |
| 28 | + |
| 29 | + std::vector<int64_t> output_shape(shape.size(), 0); |
| 30 | + int64_t capacity = 1; |
| 31 | + int unk_dim_idx = -1; |
| 32 | + for (size_t i = 0; i < shape.size(); ++i) { |
| 33 | + if (shape[i] == unk_dim_val) { |
| 34 | + PADDLE_ENFORCE_EQ( |
| 35 | + unk_dim_idx, |
| 36 | + -1, |
| 37 | + errors::InvalidArgument( |
| 38 | + "Only one dimension value of 'shape' in ReshapeOp can " |
| 39 | + "be -1. But received shape = [%s], shape[%d] is also -1.", |
| 40 | + make_ddim(shape), |
| 41 | + i)); |
| 42 | + unk_dim_idx = i; |
| 43 | + } else if (shape[i] == copy_dim_val) { |
| 44 | + PADDLE_ENFORCE_LT( |
| 45 | + static_cast<int>(i), |
| 46 | + in_dims.size(), |
| 47 | + errors::InvalidArgument( |
| 48 | + "The index of 0 in `shape` must be less than " |
| 49 | + "the input tensor X's dimensions. " |
| 50 | + "But received shape = [%s], shape[%d] = 0, X's shape = [%s], " |
| 51 | + "X's dimensions = %d.", |
| 52 | + make_ddim(shape), |
| 53 | + i, |
| 54 | + in_dims, |
| 55 | + in_dims.size())); |
| 56 | + } else { |
| 57 | + PADDLE_ENFORCE_GT( |
| 58 | + shape[i], |
| 59 | + 0, |
| 60 | + errors::InvalidArgument( |
| 61 | + "Each dimension value of 'shape' in ReshapeOp must not " |
| 62 | + "be negative except one unknown dimension. " |
| 63 | + "But received shape = [%s], shape[%d] = %d.", |
| 64 | + make_ddim(shape), |
| 65 | + i, |
| 66 | + shape[i])); |
| 67 | + } |
| 68 | + |
| 69 | + capacity *= (shape[i] ? shape[i] : in_dims[i]); |
| 70 | + output_shape[i] = (shape[i] ? static_cast<int64_t>(shape[i]) : in_dims[i]); |
| 71 | + } |
| 72 | + |
| 73 | + if (unk_dim_idx != -1) { |
| 74 | + if (all_positive) { |
| 75 | + // in_size < 0 and is un-determinate in compile time, skip the check, |
| 76 | + // for example, in_dims = [-1, 8, 1, 1], shape = [-1, 3, 8], |
| 77 | + // capacity = -24, in_size = -8, output_shape[0] = 0 |
| 78 | + // the following check will fail. |
| 79 | + output_shape[unk_dim_idx] = -in_size / capacity; |
| 80 | + PADDLE_ENFORCE_EQ( |
| 81 | + output_shape[unk_dim_idx] * capacity, |
| 82 | + -in_size, |
| 83 | + errors::InvalidArgument( |
| 84 | + "The 'shape' attribute in ReshapeOp is invalid. " |
| 85 | + "The input tensor X'size must be divisible by known " |
| 86 | + "capacity of 'shape'. " |
| 87 | + "But received X's shape = [%s], X's size = %d, " |
| 88 | + "'shape' is [%s], known capacity of 'shape' is %d.", |
| 89 | + in_dims, |
| 90 | + in_size, |
| 91 | + make_ddim(shape), |
| 92 | + capacity)); |
| 93 | + } else { |
| 94 | + output_shape[unk_dim_idx] = -1; |
| 95 | + } |
| 96 | + } else { |
| 97 | + if (all_positive) { |
| 98 | + PADDLE_ENFORCE_EQ( |
| 99 | + capacity, |
| 100 | + in_size, |
| 101 | + errors::InvalidArgument( |
| 102 | + "The 'shape' in ReshapeOp is invalid. " |
| 103 | + "The input tensor X'size must be equal to the capacity of " |
| 104 | + "'shape'. " |
| 105 | + "But received X's shape = [%s], X's size = %d, 'shape' is " |
| 106 | + "[%s], the capacity of 'shape' is %d.", |
| 107 | + in_dims, |
| 108 | + in_size, |
| 109 | + make_ddim(shape), |
| 110 | + capacity)); |
| 111 | + } |
| 112 | + } |
| 113 | + return make_ddim(output_shape); |
| 114 | +} |
| 115 | + |
| 116 | +template <typename T, typename Context> |
| 117 | +void ExecuteReshape(const Context& dev_ctx, |
| 118 | + const DenseTensor& x, |
| 119 | + const IntArray& shape, |
| 120 | + const DDim& x_dims, |
| 121 | + DenseTensor* out) { |
| 122 | + auto out_dims = ValidateShape(shape.GetData(), x_dims); |
| 123 | + auto x_vec_dims = vectorize(x_dims); |
| 124 | + |
| 125 | + funcs::ReorderOneDNNHandler reorder_handler( |
| 126 | + x_vec_dims, |
| 127 | + x.dtype(), |
| 128 | + funcs::ToOneDNNDataType(x.dtype()), |
| 129 | + dev_ctx.GetEngine()); |
| 130 | + |
| 131 | + auto reorder_src_memory_p = reorder_handler.AcquireSrcMemory( |
| 132 | + x.mem_desc(), funcs::to_void_cast(x.data<T>())); |
| 133 | + out->Resize(x_dims); // to match x numel, format is changed later |
| 134 | + // reorder is done into a plain tag to allow usage with blocked formats |
| 135 | + auto reorder_dst_memory_p = reorder_handler.AcquireDstMemory( |
| 136 | + out, funcs::GetPlainOneDNNFormat(x_dims.size()), dev_ctx.GetPlace()); |
| 137 | + auto reorder_p = reorder_handler.AcquireReorder(reorder_dst_memory_p, |
| 138 | + reorder_src_memory_p); |
| 139 | + |
| 140 | + auto& astream = OneDNNContext::tls().get_stream(); |
| 141 | + reorder_p->execute(astream, *reorder_src_memory_p, *reorder_dst_memory_p); |
| 142 | + |
| 143 | + astream.wait(); |
| 144 | + |
| 145 | + out->Resize(out_dims); |
| 146 | + out->set_mem_desc( |
| 147 | + reorder_dst_memory_p->get_desc().reshape(vectorize(out_dims))); |
| 148 | +} |
| 149 | + |
| 150 | +template <typename T, typename Context> |
| 151 | +void ReshapeKernel(const Context& dev_ctx, |
| 152 | + const DenseTensor& x, |
| 153 | + const IntArray& shape, |
| 154 | + DenseTensor* out) { |
| 155 | + auto x_dims = x.dims(); |
| 156 | + ExecuteReshape<T, Context>(dev_ctx, x, shape, x_dims, out); |
| 157 | +} |
| 158 | + |
| 159 | +template <typename T, typename Context> |
| 160 | +void ReshapeWithXShape(const Context& dev_ctx, |
| 161 | + const DenseTensor& x, |
| 162 | + const IntArray& shape, |
| 163 | + DenseTensor* out, |
| 164 | + DenseTensor* xshape) { |
| 165 | + auto x_dims = slice_ddim(xshape->dims(), 1, xshape->dims().size()); |
| 166 | + ExecuteReshape<T, Context>(dev_ctx, x, shape, x_dims, out); |
| 167 | +} |
| 168 | + |
| 169 | +} // namespace phi |
| 170 | + |
| 171 | +PD_REGISTER_KERNEL( |
| 172 | + reshape, OneDNN, ONEDNN, phi::ReshapeKernel, float, phi::dtype::bfloat16) {} |
| 173 | + |
| 174 | +PD_REGISTER_KERNEL(reshape_with_xshape, |
| 175 | + OneDNN, |
| 176 | + ONEDNN, |
| 177 | + phi::ReshapeWithXShape, |
| 178 | + float, |
| 179 | + phi::dtype::bfloat16) {} |
0 commit comments