Fix static initialization issue for static build (pytorch#90133)

Fixes pytorch#83255

Code comes from pytorch#83258 after fixing merge conflicts.

Pull Request resolved: pytorch#90133
Approved by: https://github.com/soumith, https://github.com/malfet

Author: ppwwyyxx

aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear.cpp

@@ -247,6 +247,7 @@ class QLinearInt8 final {
 };
 
 TORCH_LIBRARY_IMPL(sparse, QuantizedCPU, m) {
+  register_linear_params();
   m.impl(
       TORCH_SELECTIVE_NAME("sparse::qlinear"),
       TORCH_FN(QLinearInt8<false>::run));

aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear_prepack.cpp

@@ -240,6 +240,7 @@ class QLinearPackWeightInt8 final {
 };
 
 TORCH_LIBRARY_IMPL(sparse, QuantizedCPU, m) {
+  register_linear_params();
   m.impl(
       TORCH_SELECTIVE_NAME("sparse::qlinear_prepack"),
       TORCH_FN(QLinearPackWeightInt8::run));

aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear_unpack.cpp

@@ -133,6 +133,7 @@ class QLinearUnpackWeightInt8 final {
 };
 
 TORCH_LIBRARY_IMPL(sparse, CatchAll, m) {
+  register_linear_params();
   m.impl(
       TORCH_SELECTIVE_NAME("sparse::qlinear_unpack"),
       TORCH_FN(QLinearUnpackWeightInt8::run));

aten/src/ATen/native/quantized/cpu/qlinear.cpp

@@ -955,12 +955,14 @@ class QLinearInt8FusedQDQ final {
 };
 
 TORCH_LIBRARY_IMPL(quantized, QuantizedCPU, m) {
+  register_linear_params();
   m.impl(TORCH_SELECTIVE_NAME("quantized::linear"), TORCH_FN(QLinearInt8<false>::run));
   m.impl(TORCH_SELECTIVE_NAME("quantized::linear_relu"), TORCH_FN(QLinearInt8<true>::run));
   m.impl(TORCH_SELECTIVE_NAME("quantized::linear_leaky_relu"), TORCH_FN(QLinearLeakyReluInt8::run));
 }
 
 TORCH_LIBRARY_IMPL(_quantized, QuantizedCPU, m) {
+  register_linear_params();
   m.impl(TORCH_SELECTIVE_NAME("_quantized::linear"), TORCH_FN(QLinearInt8<false>::run));
 }
 

aten/src/ATen/native/quantized/cpu/qlinear_dynamic.cpp

@@ -662,6 +662,7 @@ class QLinearDynamicFp16 final {
 };
 
 TORCH_LIBRARY_IMPL(quantized, CPU, m) {
+  register_linear_params();
   m.impl(
       TORCH_SELECTIVE_NAME("quantized::linear_dynamic"),
       TORCH_FN(QLinearDynamicInt8<false>::run));
@@ -677,6 +678,7 @@ TORCH_LIBRARY_IMPL(quantized, CPU, m) {
 }
 
 TORCH_LIBRARY_IMPL(_quantized, CPU, m) {
+  register_linear_params();
   m.impl(
       TORCH_SELECTIVE_NAME("_quantized::linear_dynamic"),
       TORCH_FN(QLinearDynamicInt8<false>::run));

aten/src/ATen/native/quantized/cpu/qlinear_prepack.cpp

@@ -381,20 +381,24 @@ class QLinearPackWeightFp16Legacy final {
 };
 
 TORCH_LIBRARY_IMPL(quantized, QuantizedCPU, m) {
+  register_linear_params();
   m.impl(TORCH_SELECTIVE_NAME("quantized::linear_prepack"), TORCH_FN(QLinearPackWeightInt8::run));
   m.impl(TORCH_SELECTIVE_NAME("quantized::linear_prepack_legacy"), TORCH_FN(QLinearPackWeightInt8Legacy::run));
 }
 
 TORCH_LIBRARY_IMPL(quantized, CPU, m) {
+  register_linear_params();
   m.impl(TORCH_SELECTIVE_NAME("quantized::linear_prepack_fp16"), TORCH_FN(QLinearPackWeightFp16::run));
   m.impl(TORCH_SELECTIVE_NAME("quantized::linear_prepack_fp16_legacy"), TORCH_FN(QLinearPackWeightFp16Legacy::run));
 }
 
 TORCH_LIBRARY_IMPL(_quantized, QuantizedCPU, m) {
+  register_linear_params();
   m.impl(TORCH_SELECTIVE_NAME("_quantized::linear_prepack"), TORCH_FN(QLinearPackWeightInt8::run));
 }
 
 TORCH_LIBRARY_IMPL(_quantized, CPU, m) {
+  register_linear_params();
   m.impl(TORCH_SELECTIVE_NAME("_quantized::linear_prepack_fp16"), TORCH_FN(QLinearPackWeightFp16::run));
   m.impl(TORCH_SELECTIVE_NAME("_quantized::linear_prepack_fp16_legacy"), TORCH_FN(QLinearPackWeightFp16Legacy::run));
 }

aten/src/ATen/native/quantized/qconv_unpack.cpp

@@ -28,6 +28,12 @@ and /cudnn/ConvUnpackImpl.cpp, for cudnn.
 #include <ATen/ops/from_blob.h>
 #endif
 
+template <int kSpatialDim = 2>
+int register_conv_params();
+
+extern template int register_conv_params<2>();
+extern template int register_conv_params<3>();
+
 
 namespace at {
 namespace native {
@@ -192,6 +198,8 @@ unpack_quantized_prepacked_sizes_conv2d(const IValue& ivalue) {
 }
 
 TORCH_LIBRARY_IMPL(quantized, CatchAll, m) {
+  register_conv_params<2>();
+  register_conv_params<3>();
   // conv_unpack is deprecated, please use conv2d_unpack for 2D conv.
   m.impl(TORCH_SELECTIVE_NAME("quantized::conv_unpack"), TORCH_FN(QConvUnpackWeightsInt8<2>::run));
   // We use  conv2d_unpack to be consistent with conv3d_unpack

aten/src/ATen/native/quantized/qlinear_unpack.cpp

@@ -13,6 +13,8 @@ and /cudnn/linear_unpack_impl.cpp, for cudnn.
 #include <torch/custom_class.h>
 #include <torch/library.h>
 
+int register_linear_params();
+
 namespace at {
 namespace native {
 namespace {
@@ -68,6 +70,7 @@ TORCH_LIBRARY_IMPL(quantized, CPU, m) {
 }
 
 TORCH_LIBRARY_IMPL(quantized, CatchAll, m) {
+  register_linear_params();
   m.impl(TORCH_SELECTIVE_NAME("quantized::linear_unpack"), TORCH_FN(QLinearUnpackWeightInt8::run));
   m.impl(TORCH_SELECTIVE_NAME("quantized::linear_unpack_fp16"), TORCH_FN(QLinearUnpackWeightFp16::run));
 }