Minor fixes in per channel support for qconv kernel (pytorch#25182)

Summary: Pull Request resolved: pytorch#25182 Removing empty line, an unused variable and adding a check for supported quantized data types. ghstack-source-id: 8929318 Test Plan: buck test mode/dev caffe2/test:quantized -- --print-passing-details Reviewed By: jianyuh Differential Revision: D17052234 fbshipit-source-id: dbe470f0cd73fa4fca44bd15424adbaf7ceca469
atalman · Aug 30, 2019 · fcab254 · fcab254
1 parent 0f928dc
commit fcab254
Show file tree

Hide file tree

Showing 4 changed files with 5 additions and 5 deletions.
diff --git a/aten/src/ATen/native/quantized/TensorFactories.cpp b/aten/src/ATen/native/quantized/TensorFactories.cpp
@@ -37,6 +37,9 @@ Tensor empty_per_channel_affine_quantized_cpu(
   TORCH_CHECK(
       options.has_dtype(),
       "Must provide data type for Tensor creation functions.");
+  TORCH_CHECK(
+      options.dtype() == kQInt8 || options.dtype() == kQUInt8,
+      "Supported data type for tensor creation is int8 or uint8");
   TORCH_CHECK(scales.dim() == 1, "scale tensor must have dimension 1");
   TORCH_CHECK(
       zero_points.dim() == 1, "zero_points tensor must have dimension 1")

diff --git a/aten/src/ATen/native/quantized/cpu/fbgemm_utils.h b/aten/src/ATen/native/quantized/cpu/fbgemm_utils.h
@@ -1,7 +1,6 @@
 #pragma once
 
 #include <c10/core/QScheme.h>
-
 #ifdef USE_FBGEMM
 #include "fbgemm/Fbgemm.h"
 #include "fbgemm/QuantUtils.h"

diff --git a/aten/src/ATen/native/quantized/cpu/qconv.cpp b/aten/src/ATen/native/quantized/cpu/qconv.cpp
@@ -139,7 +139,6 @@ class QConv2dInt8 final : public c10::OperatorKernel {
     int32_t act_zero_point = act.q_zero_point();
 
     std::vector<float> output_multiplier_float(1, 0.0);
-    auto qtype = kPerTensorAffine;
     TORCH_CHECK(
         pack_ptr.w_scale.size() == pack_ptr.w_zp.size(),
         "Weight scales and zero points vectors should have the same size.");
@@ -185,7 +184,7 @@ class QConv2dInt8 final : public c10::OperatorKernel {
           act_ptr,
           *packB,
           reinterpret_cast<uint8_t*>(output.data_ptr<c10::quint8>()),
-          buffer.data<int32_t>(),
+          buffer.data_ptr<int32_t>(),
           outputProcObj,
           0 /* thread_id*/,
           1 /* num_threads */);

diff --git a/aten/src/ATen/native/quantized/cpu/qconv_unpack.cpp b/aten/src/ATen/native/quantized/cpu/qconv_unpack.cpp
@@ -52,12 +52,11 @@ class QConvUnpackWeightsInt8 final : public c10::OperatorKernel {
         pack_ptr.w_zp.size(),
         device(kCPU).dtype(kInt));
 
-    // The output channel axis is 0
     unpacked_weights = _empty_per_channel_affine_quantized_like(
         scales.toType(kDouble),
         zero_points.toType(kLong),
         {output_channels, kernel_h, kernel_w, C_per_G},
-        {0},
+        {0}, /* The output channel axis is 0 */
         device(kCPU).dtype(kQInt8));
    }
     int8_t* unpacked_weights_p =