Remove set_quantizer_ from native_functions.yaml (#49463)

Summary: Pull Request resolved: #49463 set_quantizer_ takes a ConstQuantizerPtr argument, which is neither supported by JIT nor by c10. Also, it doesn't get dispatched (CPU and CUDA have the same implementation) and it is excluded from python bindings generation. So there is no real reason why this needs to be in native_functions.yaml Removing it unblocks the migration to c10-fullness since this is an op that would have been hard to migrate. See https://fb.quip.com/QRtJAin66lPN ghstack-source-id: 118710663 Test Plan: waitforsandcastle Reviewed By: ezyang Differential Revision: D25587763 fbshipit-source-id: 8fab921f4c256c128d48d82dac731f04ec9bad92
pytorch · Dec 17, 2020 · 26974e6 · 26974e6
1 parent f5b68e7
commit 26974e6
Show file tree

Hide file tree

Showing 9 changed files with 19 additions and 24 deletions.
diff --git a/aten/src/ATen/native/Copy.cpp b/aten/src/ATen/native/Copy.cpp
@@ -171,7 +171,7 @@ static Tensor & copy_impl(Tensor & self, const Tensor & src, bool non_blocking)
     TORCH_CHECK(self.qscheme() == src.qscheme(),
                 "Quantized Copy only works with same qscheme");
     TORCH_CHECK(self.scalar_type() == src.scalar_type());
-    self.set_quantizer_(src.quantizer());
+    set_quantizer_(self, src.quantizer());
   }
 
   if (!self.is_quantized() && src.is_quantized()) {

diff --git a/aten/src/ATen/native/native_functions.yaml b/aten/src/ATen/native/native_functions.yaml
@@ -5455,11 +5455,6 @@
     CPU: set_cpu_
     CUDA: set_cuda_
 
-- func: set_quantizer_(Tensor(a!) self, ConstQuantizerPtr quantizer) -> Tensor(a!)
-  variants: method
-  dispatch:
-    QuantizedCPU, QuantizedCUDA: set_quantizer_
-
 - func: is_set_to(Tensor self, Tensor tensor) -> bool
   use_c10_dispatcher: full
   variants: method

diff --git a/aten/src/ATen/native/quantized/QTensor.cpp b/aten/src/ATen/native/quantized/QTensor.cpp
@@ -128,11 +128,6 @@ QScheme qscheme_quant(const Tensor& self) {
   return quantizer->qscheme();
 }
 
-Tensor& set_quantizer_(Tensor& self, ConstQuantizerPtr quantizer) {
-  get_qtensorimpl(self)->set_quantizer_(quantizer);
-  return self;
-}
-
 Tensor quantized_clone(
     const Tensor& self,
     c10::optional<c10::MemoryFormat> optional_memory_format) {

diff --git a/aten/src/ATen/native/quantized/cpu/fbgemm_utils.cpp b/aten/src/ATen/native/quantized/cpu/fbgemm_utils.cpp
@@ -237,8 +237,8 @@ template <>
 Tensor TransposeConvTensorUnpackConversion<2>(const Tensor& src, int groups) {
   // OC IC/G HW -> IC OC/G HW logically
   auto oc_g_ic_g_hw_tensors = src.chunk(groups);
-  auto fused_tensor =
-      at::cat(oc_g_ic_g_hw_tensors, 1).set_quantizer_(src.quantizer());
+  auto fused_tensor = at::cat(oc_g_ic_g_hw_tensors, 1);
+  set_quantizer_(fused_tensor, src.quantizer());
   return fused_tensor.permute({1, 0, 2, 3});
 }
 
@@ -284,8 +284,8 @@ template <>
 Tensor TransposeConvTensorUnpackConversion<3>(const Tensor& src, int groups) {
   // OC IC/G DHW -> IC OC/G DHW logically
   auto oc_g_ic_g_hw_tensors = src.chunk(groups);
-  auto fused_tensor =
-      at::cat(oc_g_ic_g_hw_tensors, 1).set_quantizer_(src.quantizer());
+  auto fused_tensor = at::cat(oc_g_ic_g_hw_tensors, 1);
+  set_quantizer_(fused_tensor, src.quantizer());
   return fused_tensor.permute({1, 0, 2, 3, 4});
 }
 
@@ -302,8 +302,8 @@ Tensor ConvertConvWeightsToChannelLastTensor<2>(
         for (auto& tensor : ic_g_oc_g_hw_tensors) {
           tensor = tensor.unsqueeze(0);
         }
-        auto fused_tensor =
-            at::cat(ic_g_oc_g_hw_tensors).set_quantizer_(src.quantizer());
+        auto fused_tensor = at::cat(ic_g_oc_g_hw_tensors);
+        set_quantizer_(fused_tensor, src.quantizer());
         return fused_tensor.permute({0, 2, 3, 4, 1})
             .contiguous(c10::MemoryFormat::Contiguous);
       }()

diff --git a/aten/src/ATen/native/quantized/cpu/qadd.cpp b/aten/src/ATen/native/quantized/cpu/qadd.cpp
@@ -88,7 +88,7 @@ Tensor _add_scalar_out(Tensor& out, const Tensor& self, Scalar other) {
     if (q_min > z - c_q) {
       s_prime = (((double)q_max - (z - c_q))) / ((double)q_max - q_min) * s;
       z_prime = q_min;
-      out.set_quantizer_(make_per_tensor_affine_quantizer(
+      set_quantizer_(out, make_per_tensor_affine_quantizer(
           s_prime, z_prime, self.scalar_type()));
       if (ReLUFused) {
         qadd_scalar_relu_stub(self.device().type(), out, self, c_q);
@@ -98,7 +98,7 @@ Tensor _add_scalar_out(Tensor& out, const Tensor& self, Scalar other) {
     } else if (q_max < z - c_q) {
       s_prime = ((double)(z - c_q) - q_min) / ((double)q_max - q_min) * s;
       z_prime = q_max;
-      out.set_quantizer_(make_per_tensor_affine_quantizer(
+      set_quantizer_(out, make_per_tensor_affine_quantizer(
           s_prime, z_prime, self.scalar_type()));
       if (ReLUFused) {
         qadd_scalar_relu_stub(self.device().type(), out, self, c_q);
@@ -109,7 +109,7 @@ Tensor _add_scalar_out(Tensor& out, const Tensor& self, Scalar other) {
       s_prime = s;
       z_prime = z - c_q;
       out.copy_(self);
-      out.set_quantizer_(make_per_tensor_affine_quantizer(
+      set_quantizer_(out, make_per_tensor_affine_quantizer(
           s_prime, z_prime, self.scalar_type()));
       if (ReLUFused) {
         at::native::relu_quantized_cpu_(out);

diff --git a/aten/src/ATen/native/quantized/cpu/qmul.cpp b/aten/src/ATen/native/quantized/cpu/qmul.cpp
@@ -59,7 +59,7 @@ Tensor _mul_scalar_out(Tensor& out, const Tensor& self, Scalar other) {
       } else {
         out.copy_(self);
       }
-      out.set_quantizer_(make_per_tensor_affine_quantizer(
+      set_quantizer_(out, make_per_tensor_affine_quantizer(
           scale_prime, zero_point_prime, self.scalar_type()));
     } else if (other_val == 0.0) {
       scale_prime = 1.0;
@@ -74,7 +74,7 @@ Tensor _mul_scalar_out(Tensor& out, const Tensor& self, Scalar other) {
           [&](Vec256<scalar_t> vec) -> Vec256<scalar_t> {
             return Vec256<scalar_t>(scalar_t(0));
           });
-      out.set_quantizer_(make_per_tensor_affine_quantizer(
+      set_quantizer_(out, make_per_tensor_affine_quantizer(
           scale_prime, zero_point_prime, self.scalar_type()));
     } else /* other_val < 0.0 */ {
       scale_prime = std::abs(other_val) * self_scale;
@@ -91,7 +91,7 @@ Tensor _mul_scalar_out(Tensor& out, const Tensor& self, Scalar other) {
             }
             return a;
           });
-      out.set_quantizer_(make_per_tensor_affine_quantizer(
+      set_quantizer_(out, make_per_tensor_affine_quantizer(
           scale_prime, zero_point_prime, self.scalar_type()));
     }
   });

diff --git a/aten/src/ATen/quantized/Quantizer.cpp b/aten/src/ATen/quantized/Quantizer.cpp
@@ -207,4 +207,8 @@ Tensor PerChannelAffineFloatQParamsQuantizer::dequantize(Tensor qtensor) {
 
 Quantizer::~Quantizer() {}
 
+C10_EXPORT void set_quantizer_(const Tensor& self, ConstQuantizerPtr quantizer) {
+  get_qtensorimpl(self)->set_quantizer_(quantizer);
+}
+
 } // namespace at
diff --git a/aten/src/ATen/quantized/Quantizer.h b/aten/src/ATen/quantized/Quantizer.h
@@ -225,4 +225,6 @@ CAFFE2_API Tensor new_qtensor(
     const TensorOptions& options,
     QuantizerPtr quantizer);
 
+CAFFE2_API void set_quantizer_(const Tensor& self, ConstQuantizerPtr quantizer);
+
 } // namespace at
diff --git a/tools/autograd/gen_python_functions.py b/tools/autograd/gen_python_functions.py
@@ -78,7 +78,6 @@
     'copy_sparse_to_sparse_', 'copy_',
     'numpy_T',  # this needs to be an attribute in Python, not a function
     'nonzero(_(out|numpy))?',
-    'set_quantizer_',  # return types not supported yet
     'set_data',
     '.*_overrideable',  # overrideable functions for backend extension
     'data', 'is_leaf', 'output_nr', '_version', 'requires_grad_', 'retain_grad', 'set_'