jit: weight prepack for convolution and linear in graph (intel#212)

* jit: weight prepack for convolution

* jit: weight prepack for n-D weight convolution

* jit: weight prepack for linear

* fix code format issue

* add record function

* remove some dead code

* check add's inputs shape for fusion add op

* fix rebase confilct

* refact code

* simplify get_conv_expected_weights_desc function

* fix code format

* simpilfy get_conv_packed_weight api

* change prepack to pack for weight pack function

Author: XiaobingSuper

tests/cpu/test_jit.py

@@ -8,6 +8,27 @@
 namespace torch_ipex {
 namespace cpu {
 
+void convolution_kernel_output(
+    const at::Tensor& input,
+    const ideep::tensor& mkldnn_weight,
+    const c10::optional<at::Tensor>& bias_opt,
+    at::Tensor& output,
+    at::IntArrayRef stride,
+    at::IntArrayRef padding,
+    at::IntArrayRef dilation,
+    int64_t groups,
+    const ideep::attr_t& attr);
+
+at::Tensor convolution_kernel(
+    const at::Tensor& input,
+    const ideep::tensor& mkldnn_weight,
+    const c10::optional<at::Tensor>& bias_opt,
+    at::IntArrayRef stride,
+    at::IntArrayRef padding,
+    at::IntArrayRef dilation,
+    int64_t groups,
+    const ideep::attr_t& attr);
+
 std::vector<int64_t> calc_conv_output_size(
     at::IntArrayRef input_size,
     at::IntArrayRef kernel_size,

torch_ipex/csrc/cpu/Conv.cpp

@@ -35,50 +35,6 @@ at::Tensor AtenIpexJITDev::dil_convolution_base(
   return convolution_impl(input, weight, bias, stride, padding, dilation, groups, ideep::attr_t());
 }
 
-at::Tensor AtenIpexJITDev::dil_convolution_swish(
-    const at::Tensor& input,
-    const at::Tensor& weight,
-    const at::Tensor& bias,
-    at::IntArrayRef stride,
-    at::IntArrayRef padding,
-    at::IntArrayRef dilation,
-    int64_t groups) {
-#if defined(IPEX_PROFILE_OP)
-  RECORD_FUNCTION("AtenIpexJITDev::dil_convolution_swish", std::vector<c10::IValue>({}));
-#endif
-  return convolution_impl(
-    input,
-    weight,
-    bias,
-    stride,
-    padding,
-    dilation,
-    groups,
-    ideep::attr_t::fuse_swish());
-}
-
-at::Tensor AtenIpexJITDev::dil_convolution_sigmoid(
-    const at::Tensor& input,
-    const at::Tensor& weight,
-    const at::Tensor& bias,
-    at::IntArrayRef stride,
-    at::IntArrayRef padding,
-    at::IntArrayRef dilation,
-    int64_t groups) {
-#if defined(IPEX_PROFILE_OP)
-  RECORD_FUNCTION("AtenIpexJITDev::dil_convolution_sigmoid", std::vector<c10::IValue>({}));
-#endif
-  return convolution_impl(
-    input,
-    weight,
-    bias,
-    stride,
-    padding,
-    dilation,
-    groups,
-    ideep::attr_t::fuse_sigmoid());
-}
-
 /**
  * Dispatch at::matmul + at::div pattern to ipex for jit inference, but only
  * one-element tensor and channel dim boadcast is enabled in oneDNN 2.2.0 now.
@@ -139,30 +95,6 @@ at::Tensor  AtenIpexJITDev::dil_matmul_div(
   }
 }
 
-at::Tensor AtenIpexJITDev::dil_convolution_clamp(
-    const at::Tensor& input,
-    const at::Tensor& weight,
-    const at::Tensor& bias,
-    at::IntArrayRef stride,
-    at::IntArrayRef padding,
-    at::IntArrayRef dilation,
-    int64_t groups,
-    float lower_bound,
-    float upper_bound) {
-#if defined(IPEX_PROFILE_OP)
-  RECORD_FUNCTION("AtenIpexJITDev::dil_convolution_clamp", std::vector<c10::IValue>({}));
-#endif
-  return convolution_impl(
-    input,
-    weight,
-    bias,
-    stride,
-    padding,
-    dilation,
-    groups,
-    ideep::attr_t::fuse_clamp(lower_bound, upper_bound));
-}
-
 at::Tensor AtenIpexJITDev::dil_convolution_relu(
     const at::Tensor& input,
     const at::Tensor& weight,
@@ -185,33 +117,6 @@ at::Tensor AtenIpexJITDev::dil_convolution_relu(
     ideep::attr_t::fuse_relu());
 }
 
-at::Tensor AtenIpexJITDev::dil_convolution_elu(
-    const at::Tensor& input,
-    const at::Tensor& weight,
-    const at::Tensor& bias,
-    at::IntArrayRef stride,
-    at::IntArrayRef padding,
-    at::IntArrayRef dilation,
-    int64_t groups,
-    float alpha,
-    at::Scalar scale,
-    at::Scalar input_scale) {
-#if defined(IPEX_PROFILE_OP)
-  RECORD_FUNCTION("AtenIpexJITDev::dil_convolution_elu", std::vector<c10::IValue>({}));
-#endif
-  auto scale_value = scale.to<float>();
-  auto input_scale_value = input_scale.to<float>();
-  return convolution_impl(
-    input,
-    weight,
-    bias,
-    stride,
-    padding,
-    dilation,
-    groups,
-    ideep::attr_t::fuse_elu(scale_value, alpha, input_scale_value));
-}
-
 at::Tensor AtenIpexJITDev::dil_conv_transpose2d(
     const at::Tensor& input,
     const at::Tensor& weight,
@@ -305,52 +210,6 @@ at::Tensor AtenIpexJITDev::dil_max_pool2d(
       ideep::algorithm::pooling_max);
 }
 
-at::Tensor AtenIpexJITDev::dil_linear(
-    const at::Tensor& self,
-    const at::Tensor& weight,
-    const at::Tensor& bias) {
-#if defined(IPEX_PROFILE_OP)
-  RECORD_FUNCTION("AtenIpexJITDev::dil_linear", std::vector<c10::IValue>({}));
-#endif
-  return linear_impl(self, weight, bias, ideep::attr_t());
-}
-
-at::Tensor AtenIpexJITDev::dil_linear_fuse_eltwise(
-    const at::Tensor& self,
-    const at::Tensor& weight,
-    const at::Tensor& bias,
-    const ideep::attr_t& attr) {
-#if defined(IPEX_PROFILE_OP)
-  RECORD_FUNCTION("AtenIpexJITDev::dil_linear_fuse_eltwise", std::vector<c10::IValue>({}));
-#endif
-  return linear_impl(self, weight, bias, attr);
-}
-
-/**
- *Dispatch Linear + Add fusion pattern to ipex oneDNN kernel for inference mode.
- *This feature might improve performance for cases like residual learning blocks
- *Pattern: accum = accum * alpha + Linear(self, weight, bias)
- *
- *@param self Activatin input for Linear
- *@param weight Weight for Linear
- *@param bias Bias for Linear
- *@param accum One input for add operation, another is the output of Linear
- *@param alpha Scale for accum when doing add operation.
- *
- *@return Value for the fusion pattern output.
- */
-at::Tensor AtenIpexJITDev::dil_linear_add(const at::Tensor &self,
-                                          const at::Tensor &weight,
-                                          const at::Tensor &bias,
-                                          at::Tensor &accumu,
-                                          at::Scalar alpha) {
-#if defined(IPEX_PROFILE_OP)
-  RECORD_FUNCTION("AtenIpexJITDev::dil_linear_add", std::vector<c10::IValue>({}));
-#endif
-  auto scale = alpha.to<float>();
-  return linear_inplace_impl(self, weight, bias, accumu, ideep::attr_t::fuse_sum(scale));
-}
-
 //Dispatch softmax to oneDNN path for jit inference
 at::Tensor AtenIpexJITDev::dil_softmax(
     const at::Tensor& input,