|
| 1 | +/* |
| 2 | + * Copyright (c) Meta Platforms, Inc. and affiliates. |
| 3 | + * All rights reserved. |
| 4 | + * |
| 5 | + * This source code is licensed under the BSD-style license found in the |
| 6 | + * LICENSE file in the root directory of this source tree. |
| 7 | + */ |
| 8 | + |
| 9 | +#include <executorch/kernels/optimized/cpu/binary_ops.h> |
| 10 | +#include <executorch/kernels/optimized/vec/functional.h> |
| 11 | +#include <executorch/kernels/optimized/vec/vec.h> |
| 12 | +#include <executorch/kernels/portable/cpu/scalar_utils.h> |
| 13 | +#include <executorch/kernels/portable/cpu/util/broadcast_util.h> |
| 14 | +#include <executorch/runtime/kernel/kernel_includes.h> |
| 15 | +#include <executorch/runtime/platform/assert.h> |
| 16 | + |
| 17 | +namespace torch { |
| 18 | +namespace executor { |
| 19 | +namespace kernels { |
| 20 | +namespace impl { |
| 21 | + |
| 22 | +namespace { |
| 23 | +template < |
| 24 | + bool can_cast, |
| 25 | + typename CTYPE_A, |
| 26 | + typename CTYPE_B, |
| 27 | + typename CTYPE_IN, |
| 28 | + typename CTYPE_OUT> |
| 29 | +struct AddInner; |
| 30 | + |
| 31 | +template < |
| 32 | + typename CTYPE_A, |
| 33 | + typename CTYPE_B, |
| 34 | + typename CTYPE_IN, |
| 35 | + typename CTYPE_OUT> |
| 36 | +struct AddInner<true, CTYPE_A, CTYPE_B, CTYPE_IN, CTYPE_OUT> { |
| 37 | + static void |
| 38 | + run(const Tensor& a, const Tensor& b, CTYPE_IN alpha_val, Tensor& out) { |
| 39 | + apply_binary_elementwise_fn<CTYPE_A, CTYPE_B, CTYPE_OUT>( |
| 40 | + // NOLINTNEXTLINE(facebook-hte-ConstantArgumentPassByValue) |
| 41 | + [alpha_val](const CTYPE_A val_a, const CTYPE_B val_b) { |
| 42 | + CTYPE_IN a_casted = static_cast<CTYPE_IN>(val_a); |
| 43 | + CTYPE_IN b_casted = static_cast<CTYPE_IN>(val_b); |
| 44 | + CTYPE_IN value = a_casted + alpha_val * b_casted; |
| 45 | + |
| 46 | + return static_cast<CTYPE_OUT>(value); |
| 47 | + }, |
| 48 | + a, |
| 49 | + b, |
| 50 | + out); |
| 51 | + } |
| 52 | +}; |
| 53 | + |
| 54 | +template <typename CTYPE_IN> |
| 55 | +struct ReportCanCastBug { |
| 56 | + static void run(const Tensor&, const Tensor&, CTYPE_IN, Tensor&) { |
| 57 | + ET_DCHECK_MSG(false, "BUG: canCast should have been checked above"); |
| 58 | + } |
| 59 | +}; |
| 60 | + |
| 61 | +template < |
| 62 | + typename CTYPE_A, |
| 63 | + typename CTYPE_B, |
| 64 | + typename CTYPE_IN, |
| 65 | + typename CTYPE_OUT> |
| 66 | +struct AddInner<false, CTYPE_A, CTYPE_B, CTYPE_IN, CTYPE_OUT> |
| 67 | + : public ReportCanCastBug<CTYPE_IN> {}; |
| 68 | + |
| 69 | +} // namespace |
| 70 | + |
| 71 | +using Tensor = executorch::aten::Tensor; |
| 72 | +using ScalarType = executorch::aten::ScalarType; |
| 73 | + |
| 74 | +Tensor& opt_add_sub_out_impl( |
| 75 | + KernelRuntimeContext& ctx, |
| 76 | + const Tensor& a, |
| 77 | + const Tensor& b, |
| 78 | + const Scalar& alpha, |
| 79 | + Tensor& out, |
| 80 | + const bool is_sub) { |
| 81 | + (void)ctx; |
| 82 | + |
| 83 | + ScalarType a_type = a.scalar_type(); |
| 84 | + ScalarType b_type = b.scalar_type(); |
| 85 | + ScalarType out_type = out.scalar_type(); |
| 86 | + |
| 87 | + auto selected_optimized_path = select_optimized_path(a, b, out); |
| 88 | + if (selected_optimized_path == ElementwiseOptimizedPath::kTreatAs1d) { |
| 89 | + // Resize for dynamic shape |
| 90 | + auto error = resize_tensor(out, a.sizes()); |
| 91 | + ET_KERNEL_CHECK_MSG( |
| 92 | + ctx, |
| 93 | + error == Error::Ok, |
| 94 | + InvalidArgument, |
| 95 | + out, |
| 96 | + "Failed to resize output tensor."); |
| 97 | + |
| 98 | + ET_SWITCH_REALB_TYPES(a_type, ctx, "add.out", CTYPE, [&]() { |
| 99 | + CTYPE alpha_val; |
| 100 | + ET_KERNEL_CHECK( |
| 101 | + ctx, |
| 102 | + torch::executor::native::utils::extract_scalar(alpha, &alpha_val), |
| 103 | + InvalidArgument, ); |
| 104 | + if (is_sub) { |
| 105 | + alpha_val = -alpha_val; |
| 106 | + } |
| 107 | + using Vec = executorch::vec::Vectorized<CTYPE>; |
| 108 | + executorch::vec::map2<CTYPE>( |
| 109 | + [alpha_val](Vec x, Vec y) { return x + Vec(alpha_val) * y; }, |
| 110 | + out.mutable_data_ptr<CTYPE>(), |
| 111 | + a.const_data_ptr<CTYPE>(), |
| 112 | + b.const_data_ptr<CTYPE>(), |
| 113 | + out.numel()); |
| 114 | + }); |
| 115 | + } else if (selected_optimized_path != ElementwiseOptimizedPath::kNone) { |
| 116 | + // Cannot apply the trick of -alpha here because alpha is Scalar without |
| 117 | + // support for - operator. At least not right now. |
| 118 | + if (is_sub) { |
| 119 | + if (selected_optimized_path == |
| 120 | + ElementwiseOptimizedPath::kBroadcast2dBy1dReverseArguments || |
| 121 | + selected_optimized_path == |
| 122 | + ElementwiseOptimizedPath::kBroadcastLastDimReverseArguments || |
| 123 | + selected_optimized_path == |
| 124 | + ElementwiseOptimizedPath::kBroadcastNdByNdReverseArguments) { |
| 125 | + auto add_lambda = [](auto x, auto y, auto alpha_val) { |
| 126 | + return y - alpha_val * x; |
| 127 | + }; |
| 128 | + return torch::executor::handle_broadcast_elementwise<BinaryOpType::kSub>( |
| 129 | + ctx, add_lambda, a, b, out, selected_optimized_path, alpha); |
| 130 | + } else { |
| 131 | + auto add_lambda = [](auto x, auto y, auto alpha_val) { |
| 132 | + return x - alpha_val * y; |
| 133 | + }; |
| 134 | + return torch::executor::handle_broadcast_elementwise<BinaryOpType::kSub>( |
| 135 | + ctx, add_lambda, a, b, out, selected_optimized_path, alpha); |
| 136 | + } |
| 137 | + } else { |
| 138 | + if (selected_optimized_path == |
| 139 | + ElementwiseOptimizedPath::kBroadcast2dBy1dReverseArguments || |
| 140 | + selected_optimized_path == |
| 141 | + ElementwiseOptimizedPath::kBroadcastLastDimReverseArguments || |
| 142 | + selected_optimized_path == |
| 143 | + ElementwiseOptimizedPath::kBroadcastNdByNdReverseArguments) { |
| 144 | + auto add_lambda = [](auto x, auto y, auto alpha_val) { |
| 145 | + return y + alpha_val * x; |
| 146 | + }; |
| 147 | + return torch::executor::handle_broadcast_elementwise<BinaryOpType::kAdd>( |
| 148 | + ctx, add_lambda, a, b, out, selected_optimized_path, alpha); |
| 149 | + } else { |
| 150 | + auto add_lambda = [](auto x, auto y, auto alpha_val) { |
| 151 | + return x + alpha_val * y; |
| 152 | + }; |
| 153 | + return torch::executor::handle_broadcast_elementwise<BinaryOpType::kAdd>( |
| 154 | + ctx, add_lambda, a, b, out, selected_optimized_path, alpha); |
| 155 | + } |
| 156 | + } |
| 157 | + } else { |
| 158 | + ScalarType common_type = |
| 159 | + promoteTypes(a_type, b_type, /*half_to_float*/ true); |
| 160 | + ET_KERNEL_CHECK(ctx, canCast(common_type, out_type), InvalidArgument, out); |
| 161 | + |
| 162 | + ET_KERNEL_CHECK( |
| 163 | + ctx, |
| 164 | + resize_to_broadcast_target_size(a, b, out) == Error::Ok, |
| 165 | + InvalidArgument, |
| 166 | + out); |
| 167 | + |
| 168 | + ET_SWITCH_REALHBBF16_TYPES(a_type, ctx, "add.out", CTYPE_A, [&]() { |
| 169 | + ET_SWITCH_REALHBBF16_TYPES(b_type, ctx, "add.out", CTYPE_B, [&]() { |
| 170 | + using CTYPE_IN = typename torch::executor:: |
| 171 | + promote_types<CTYPE_A, CTYPE_B, /*half_to_float*/ true>::type; |
| 172 | + ET_DCHECK(CppTypeToScalarType<CTYPE_IN>::value == common_type); |
| 173 | + ET_SWITCH_REALHBBF16_TYPES(out_type, ctx, "add.out", CTYPE_OUT, [&]() { |
| 174 | + CTYPE_IN alpha_val; |
| 175 | + ET_KERNEL_CHECK( |
| 176 | + ctx, |
| 177 | + torch::executor::native::utils::extract_scalar(alpha, &alpha_val), |
| 178 | + InvalidArgument, ); |
| 179 | + if (is_sub) { |
| 180 | + alpha_val = -alpha_val; |
| 181 | + } |
| 182 | + |
| 183 | + AddInner< |
| 184 | + can_cast<CTYPE_IN, CTYPE_OUT>::value, |
| 185 | + CTYPE_A, |
| 186 | + CTYPE_B, |
| 187 | + CTYPE_IN, |
| 188 | + CTYPE_OUT>::run(a, b, alpha_val, out); |
| 189 | + }); |
| 190 | + }); |
| 191 | + }); |
| 192 | + } |
| 193 | + |
| 194 | + return out; |
| 195 | +} |
| 196 | + |
| 197 | +} // namespace impl |
| 198 | +} // namespace kernels |
| 199 | +} // namespace executor |
| 200 | +} // namespace torch |
0 commit comments