support load_state_dict after ipex.optimize (#1326)

* support load_state_dict after ipex.optimize

* remove un-used import

* fix inference test for NotEqual

* add inplace test && remove unneccessary check

* improve comments

Author: zhuhaozhe

csrc/jit/cpu/kernels/ContextConvTranspose.h

@@ -14,7 +14,7 @@ struct ContextConvTranspose final {
   // at_weight will share same memory with weight_packed_
   // at_weight is used for autograd and optimizer update
   at::Tensor at_weight_;
-  c10::optional<at::Tensor> bias_;
+  c10::optional<at::Tensor> at_bias_;
   // paddings_, strided_, dilation_, output_padding_ here are expanded and
   // might different with those stored on ConvTransposeOpContext.
   // For example, aten deconv2d can accept padding = 2, but onednn deconv2d need
@@ -48,7 +48,7 @@ struct ContextConvTranspose final {
       : original_desc_(std::move(original_desc)),
         weight_packed_(std::move(weight_packed)),
         at_weight_(std::move(at_weight)),
-        bias_(std::move(bias)),
+        at_bias_(std::move(bias)),
         padding_(padding),
         output_padding_(output_padding),
         stride_(stride),

csrc/jit/cpu/kernels/ContextLinear.h

@@ -13,7 +13,7 @@ struct ContextLinear final {
   // at_weight will share same memory with weight_packed_
   // at_weight is used for autograd and optimizer update
   at::Tensor at_weight_;
-  c10::optional<at::Tensor> bias_;
+  c10::optional<at::Tensor> at_bias_;
 
   ContextLinear() = delete;
 
@@ -25,7 +25,7 @@ struct ContextLinear final {
       : original_desc_(std::move(original_desc)),
         weight_packed_(std::move(weight_packed)),
         at_weight_(std::move(at_weight)),
-        bias_(std::move(bias)) {}
+        at_bias_(std::move(bias)) {}
 
   ContextLinear(ContextLinear&&) = default;
   ContextLinear& operator=(ContextLinear&&) = default;

csrc/jit/cpu/kernels/ContextLinearMKL.h

@@ -9,9 +9,9 @@ namespace cpu {
 namespace detail {
 struct ContextLinearMKL final {
   std::vector<int64_t> sgemm_sizes_ = {0, 0, 0};
-  at::Tensor mkl_weight_;
-  at::Tensor ori_weight_;
-  c10::optional<at::Tensor> bias_;
+  at::Tensor at_weight_; // packed at weight
+  at::Tensor ori_weight_; // non-packed at weight
+  c10::optional<at::Tensor> at_bias_;
 
   ContextLinearMKL() = delete;
 
@@ -21,9 +21,9 @@ struct ContextLinearMKL final {
       at::Tensor&& ori_weight,
       c10::optional<at::Tensor>&& bias)
       : sgemm_sizes_(std::move(sgemm_sizes)),
-        mkl_weight_(std::move(mkl_weight)),
+        at_weight_(std::move(mkl_weight)),
         ori_weight_(std::move(ori_weight)),
-        bias_(std::move(bias)) {}
+        at_bias_(std::move(bias)) {}
 
   ContextLinearMKL(ContextLinearMKL&&) = default;
   ContextLinearMKL& operator=(ContextLinearMKL&&) = default;

csrc/jit/cpu/kernels/ConvTransposePacked.cpp

@@ -352,7 +352,7 @@ at::Tensor run(
   check_shape_forward(
       input_.sizes(),
       context.origin_weight_dims_,
-      context.bias_,
+      context.at_bias_,
       context.padding_,
       context.stride_,
       context.dilation_,
@@ -361,7 +361,7 @@ at::Tensor run(
   return conv_transpose_kernel_impl(
       input_,
       context.weight_packed_,
-      context.bias_,
+      context.at_bias_,
       context.stride_,
       context.padding_,
       context.output_padding_,
@@ -397,7 +397,7 @@ at::Tensor& run(
   check_shape_forward(
       input_.sizes(),
       context.origin_weight_dims_,
-      context.bias_,
+      context.at_bias_,
       context.padding_,
       context.stride_,
       context.dilation_,
@@ -406,7 +406,7 @@ at::Tensor& run(
   conv_transpose_out_kernel_impl(
       input_,
       context.weight_packed_,
-      context.bias_,
+      context.at_bias_,
       accumu,
       context.stride_,
       context.padding_,

csrc/jit/cpu/kernels/LinearMKLPacked.cpp

@@ -60,7 +60,7 @@ at::Tensor run(ContextLinearMKL& context, const at::Tensor& input) {
       "Check the shapes of mat1 and mat2, they cannot be multiplied!");
   auto input_ = input.contiguous();
   c10::MaybeOwned<at::Tensor> bias_maybe_owned =
-      at::borrow_from_optional_tensor(context.bias_);
+      at::borrow_from_optional_tensor(context.at_bias_);
   const at::Tensor& bias = *bias_maybe_owned;
   int64_t input_batch = (int64_t)(input_.numel() / K);
 
@@ -71,7 +71,7 @@ at::Tensor run(ContextLinearMKL& context, const at::Tensor& input) {
   if (input_batch != context.sgemm_sizes_[0])
     return mkl_sgemm_kernel(input_, context.ori_weight_, bias);
   return mkl_prepack_sgemm_kernel(
-      input_, context.mkl_weight_, bias, context.sgemm_sizes_[2]);
+      input_, context.at_weight_, bias, context.sgemm_sizes_[2]);
 }
 
 at::Tensor& run(
@@ -84,14 +84,14 @@ at::Tensor& run(
       "Check the shapes of mat1 and mat2, they cannot be multiplied!");
   auto input_ = input.contiguous();
   c10::MaybeOwned<at::Tensor> bias_maybe_owned =
-      at::borrow_from_optional_tensor(context.bias_);
+      at::borrow_from_optional_tensor(context.at_bias_);
   const at::Tensor& bias = *bias_maybe_owned;
   int64_t input_batch = (int64_t)(input_.numel() / K);
   if (input_batch != context.sgemm_sizes_[0]) {
     mkl_sgemm_kernel_output(input_, context.ori_weight_, bias, accumu);
   } else {
     mkl_prepack_sgemm_kernel_output(
-        input_, context.mkl_weight_, bias, context.sgemm_sizes_[2], accumu);
+        input_, context.at_weight_, bias, context.sgemm_sizes_[2], accumu);
   }
   return accumu;
 }

csrc/jit/cpu/kernels/LinearPacked.cpp

@@ -218,7 +218,7 @@ at::Tensor run(
       "Check the shapes of mat1 and mat2, they cannot be multiplied!");
   auto input_ = input.contiguous();
   c10::MaybeOwned<at::Tensor> bias_maybe_owned =
-      at::borrow_from_optional_tensor(context.bias_);
+      at::borrow_from_optional_tensor(context.at_bias_);
   const at::Tensor& bias = *bias_maybe_owned;
   return linear_kernel(input_, context.weight_packed_, bias, attr);
 }
@@ -233,7 +233,7 @@ at::Tensor& run(
       "Check the shapes of mat1 and mat2, they cannot be multiplied!");
   auto input_ = input.contiguous();
   c10::MaybeOwned<at::Tensor> bias_maybe_owned =
-      at::borrow_from_optional_tensor(context.bias_);
+      at::borrow_from_optional_tensor(context.at_bias_);
   const at::Tensor& bias = *bias_maybe_owned;
   linear_kernel_output(input_, context.weight_packed_, bias, accumu, attr);
   return accumu;
@@ -250,8 +250,8 @@ void run_core(
   TORCH_CHECK(
       input.size(input.dim() - 1) == context.weight_packed_.get_dims()[1],
       "Check the shapes of mat1 and mat2, they cannot be multiplied!");
-  if (context.bias_) {
-    auto mkl_bias = itensor_view_from_dense(*context.bias_);
+  if (context.at_bias_) {
+    auto mkl_bias = itensor_view_from_dense(*context.at_bias_);
     ideep::inner_product_forward::prepare(
         param,
         mkldnn_input,
@@ -280,7 +280,7 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor> run_backward(
       context.at_weight_,
       output_mask,
       context.weight_packed_,
-      context.bias_);
+      context.at_bias_);
 }
 
 at::Tensor pack(ContextLinear& context, const at::Tensor& tensor) {

csrc/jit/cpu/kernels/OpContext.cpp

@@ -8,6 +8,32 @@
 namespace torch_ipex {
 namespace cpu {
 
+template <typename T1, typename T2>
+void load_from_ctx_template(T1* self, c10::intrusive_ptr<T2> other) {
+  auto& other_ctx_ = other->get_context();
+  auto loaded_weight = other_ctx_.at_weight_;
+  auto loaded_bias = other_ctx_.at_bias_;
+  self->get_context().at_weight_.copy_(loaded_weight);
+  if (loaded_bias.has_value()) {
+    self->get_context().at_bias_.value().copy_(loaded_bias.value());
+  }
+  return;
+}
+
+template <>
+void load_from_ctx_template<IpexLinearMKLOpContext, MKLOpContext>(
+    IpexLinearMKLOpContext* self,
+    c10::intrusive_ptr<MKLOpContext> other) {
+  auto& other_ctx_ = other->get_context();
+  auto loaded_weight = other_ctx_.at_weight_;
+  auto loaded_bias = other_ctx_.at_bias_;
+  self->get_context().at_weight_.copy_(loaded_weight);
+  if (loaded_bias.has_value()) {
+    self->get_context().at_bias_.value().copy_(loaded_bias.value());
+  }
+  self->get_context().ori_weight_.copy_(other->get_context().ori_weight_);
+  return;
+}
 c10::intrusive_ptr<ConvolutionOpContext> IpexConvolutionOpContext::
     create_context(
         at::Tensor&& weight,
@@ -99,6 +125,11 @@ at::Tensor IpexConvolutionOpContext::get_data_handle() {
   return ptr;
 }
 
+void IpexConvolutionOpContext::load_from_ctx(
+    c10::intrusive_ptr<ConvolutionOpContext> other) {
+  load_from_ctx_template(this, other);
+}
+
 c10::intrusive_ptr<LinearOpContext> IpexLinearOpContext::create_context(
     at::Tensor&& weight,
     c10::optional<at::Tensor>&& bias,
@@ -153,6 +184,11 @@ at::Tensor IpexLinearOpContext::to_public(const at::Tensor& tensor) {
   return torch_ipex::cpu::detail::linear::unpack(op_context_, tensor);
 }
 
+void IpexLinearOpContext::load_from_ctx(
+    c10::intrusive_ptr<LinearOpContext> other) {
+  load_from_ctx_template(this, other);
+}
+
 c10::intrusive_ptr<ConvTransposeOpContext> IpexConvTransposeOpContext::
     create_context(
         at::Tensor&& weight,
@@ -194,7 +230,7 @@ c10::intrusive_ptr<MKLOpContext> IpexLinearMKLOpContext::create_context(
 }
 
 at::Tensor IpexLinearMKLOpContext::get_at_packed_weight() {
-  return op_context_.mkl_weight_;
+  return op_context_.at_weight_;
 }
 
 at::Tensor IpexLinearMKLOpContext::get_data_handle() {
@@ -221,7 +257,7 @@ at::Tensor IpexLinearMKLOpContext::to_public(const at::Tensor& tensor) {
   return op_context_.ori_weight_.clone();
 }
 
-detail::ContextLinearMKL& IpexLinearMKLOpContext::get_mkl_context() {
+detail::ContextLinearMKL& IpexLinearMKLOpContext::get_context() {
   return op_context_;
 }
 
@@ -233,6 +269,11 @@ int64_t IpexLinearMKLOpContext::get_in_features() {
   return op_context_.sgemm_sizes_[1];
 }
 
+void IpexLinearMKLOpContext::load_from_ctx(
+    c10::intrusive_ptr<MKLOpContext> other) {
+  load_from_ctx_template(this, other);
+}
+
 at::Tensor IpexConvTransposeOpContext::run(
     const at::Tensor& input,
     const ideep::attr_t& attr) {
@@ -288,5 +329,10 @@ detail::ContextConvTranspose& IpexConvTransposeOpContext::get_context() {
   return op_context_;
 }
 
+void IpexConvTransposeOpContext::load_from_ctx(
+    c10::intrusive_ptr<ConvTransposeOpContext> other) {
+  load_from_ctx_template(this, other);
+}
+
 } // namespace cpu
 } // namespace torch_ipex