conv code refactoring (pytorch#856)

Summary:
Pull Request resolved: pytorch#856

To simplify D33523940, separate out code refactoring changes like removing deprecated interface, reducing code duplication

Reviewed By: jiyuanzFB

Differential Revision: D33533769

fbshipit-source-id: 00e26cd87102364c12c4abbae78cae149f6a3b30

include/fbgemm/Fbgemm.h

@@ -1362,27 +1362,6 @@ FBGEMM_API void fbgemmPacked(
  *        Note: Currently threading is not supported. This function does
  *              nothing for thread_ids > 0, i.e., returns early.
  *
- * @param rowOffsetBuf nullptr if B uses symmetric quantization
- */
-
-template <
-    typename packed_W,
-    typename outType,
-    typename processOutputType,
-    int SPATIAL_DIM = 2>
-FBGEMM_API void fbgemmGroupwiseConv(
-    const conv_param_t<SPATIAL_DIM>& conv_param,
-    const std::uint8_t* activations,
-    std::int32_t a_zero_point,
-    std::int32_t* rowOffsetBuf,
-    packed_W& packed_weights,
-    outType* out,
-    std::int32_t* outBuffer,
-    const processOutputType& outProcess,
-    int thread_id,
-    int num_threads);
-
-/**
  * @param rowOffsetBuf nullptr if B uses symmetric quantization
  *        Note: Currently threading is not supported. This function does
  *              nothing for thread_ids > 0, i.e., returns early.
@@ -1414,22 +1393,6 @@ template <int SPATIAL_DIM = 2>
 FBGEMM_API int rowOffsetBufferSizeGConv(
     const conv_param_t<SPATIAL_DIM>& conv_param);
 
-/**
- * @brief Perform depthwise separable convolution
- */
-template <
-    typename packingAMatrix,
-    typename packingBMatrix,
-    typename outT,
-    typename processOutputType>
-void convDepthwiseSeparable(
-    const conv_param_t<>& conv_param_dw,
-    const conv_param_t<>& conv_param_1x1,
-    packingAMatrix& packdw,
-    packingBMatrix& packed_1x1,
-    outT* out,
-    const processOutputType& output);
-
 /**
  * @brief Is this depthwise convolution optimized?
  */

include/fbgemm/FbgemmI8DepthwiseAvx2.h

@@ -83,59 +83,6 @@ FBGEMM_API void depthwise_2d_same_pad(
     int thread_id = 0,
     int num_threads = 1);
 
-template <typename BIAS_TYPE = std::int32_t>
-FBGEMM_API void depthwise_2d_same_pad(
-    int N,
-    int H,
-    int W,
-    int IC_OC,
-    int stride_h,
-    int stride_w,
-    std::int32_t A_zero_point,
-    const std::uint8_t* A,
-    std::int32_t B_zero_point,
-    const PackedDepthWiseConvMatrix& Bp,
-    float C_multiplier,
-    std::int32_t C_zero_point,
-    std::uint8_t* C,
-    const std::int32_t* col_offsets,
-    const BIAS_TYPE* bias,
-    bool fuse_relu = false,
-    float act_times_w_scale = 1.0f,
-    int thread_id = 0,
-    int num_threads = 1);
-
-/**
- * Depth-wise convolution that results in the same output feature size as the
- * input feature. That is PAD_T = PAD_B = (R - 1) / 2 and PAD_L = PAD_R =
- * (S - 1) / 2. This function also does requantization and uses per-channel
- * quantization.
- * @param col_offsets nullptr if col_offsets are folded into bias
- * @param act_times_w_scale Only used if BIAS_TYPE is float, i.e., bias is
- *                          unquantized.
- */
-template <typename BIAS_TYPE = std::int32_t>
-FBGEMM_API void depthwise_2d_per_channel_quantization_same_pad(
-    int N,
-    int H,
-    int W,
-    int IC_OC,
-    int stride_h,
-    int stride_w,
-    std::int32_t A_zero_point,
-    const std::uint8_t* A,
-    const std::int32_t* B_zero_point,
-    const PackedDepthWiseConvMatrix& Bp,
-    const float* C_multiplier,
-    std::int32_t C_zero_point,
-    std::uint8_t* C,
-    const std::int32_t* col_offsets,
-    const BIAS_TYPE* bias,
-    bool fuse_relu = false,
-    const float* act_times_w_scale = nullptr,
-    int thread_id = 0,
-    int num_threads = 1);
-
 /**
  * @param col_offsets nullptr if col_offsets are folded into bias
  */

src/FbgemmI8Depthwise2DAvx2-inl.h

@@ -119,8 +119,8 @@ template <
     bool HAS_BIAS,
     bool A_SYMMETRIC,
     bool B_SYMMETRIC,
-    typename BIAS_TYPE,
-    QuantizationGranularity Q_GRAN>
+    QuantizationGranularity Q_GRAN,
+    typename BIAS_TYPE>
 static ALWAYS_INLINE void depthwise_2d_(
     int N,
     int H,
@@ -185,8 +185,7 @@ static ALWAYS_INLINE void depthwise_2d_(
             HAS_BIAS,
             A_SYMMETRIC,
             B_SYMMETRIC,
-            Q_GRAN,
-            BIAS_TYPE>(
+            Q_GRAN>(
             H,
             W,
             IC,
@@ -216,8 +215,7 @@ static ALWAYS_INLINE void depthwise_2d_(
             HAS_BIAS,
             A_SYMMETRIC,
             B_SYMMETRIC,
-            Q_GRAN,
-            BIAS_TYPE>(
+            Q_GRAN>(
             H,
             W,
             IC,
@@ -247,8 +245,7 @@ static ALWAYS_INLINE void depthwise_2d_(
             HAS_BIAS,
             A_SYMMETRIC,
             B_SYMMETRIC,
-            Q_GRAN,
-            BIAS_TYPE>(
+            Q_GRAN>(
             H,
             W,
             IC,
@@ -292,8 +289,7 @@ static ALWAYS_INLINE void depthwise_2d_(
             HAS_BIAS,
             A_SYMMETRIC,
             B_SYMMETRIC,
-            Q_GRAN,
-            BIAS_TYPE>(
+            Q_GRAN>(
             H,
             W,
             IC,
@@ -341,8 +337,7 @@ static ALWAYS_INLINE void depthwise_2d_(
             HAS_BIAS,
             A_SYMMETRIC,
             B_SYMMETRIC,
-            Q_GRAN,
-            BIAS_TYPE>(
+            Q_GRAN>(
             H,
             W,
             IC,
@@ -373,8 +368,7 @@ static ALWAYS_INLINE void depthwise_2d_(
             HAS_BIAS,
             A_SYMMETRIC,
             B_SYMMETRIC,
-            Q_GRAN,
-            BIAS_TYPE>(
+            Q_GRAN>(
             H,
             W,
             IC,
@@ -406,8 +400,7 @@ static ALWAYS_INLINE void depthwise_2d_(
             HAS_BIAS,
             A_SYMMETRIC,
             B_SYMMETRIC,
-            Q_GRAN,
-            BIAS_TYPE>(
+            Q_GRAN>(
             H,
             W,
             IC,
@@ -437,8 +430,7 @@ static ALWAYS_INLINE void depthwise_2d_(
             HAS_BIAS,
             A_SYMMETRIC,
             B_SYMMETRIC,
-            Q_GRAN,
-            BIAS_TYPE>(
+            Q_GRAN>(
             H,
             W,
             IC,
@@ -468,8 +460,7 @@ static ALWAYS_INLINE void depthwise_2d_(
             HAS_BIAS,
             A_SYMMETRIC,
             B_SYMMETRIC,
-            Q_GRAN,
-            BIAS_TYPE>(
+            Q_GRAN>(
             H,
             W,
             IC,
@@ -502,8 +493,8 @@ template <
     int S,
     bool FUSE_RELU,
     bool HAS_BIAS,
-    typename BIAS_TYPE,
-    QuantizationGranularity Q_GRAN>
+    QuantizationGranularity Q_GRAN,
+    typename BIAS_TYPE>
 static void depthwise_2d_(
     int N,
     int H,
@@ -534,7 +525,6 @@ static void depthwise_2d_(
           HAS_BIAS,
           true /*A_symmetric*/,
           true /*B_symmetric*/,
-          BIAS_TYPE,
           Q_GRAN>(
           N,
           H,
@@ -563,7 +553,6 @@ static void depthwise_2d_(
           HAS_BIAS,
           true /*A_symmetric*/,
           false /*B_symmetric*/,
-          BIAS_TYPE,
           Q_GRAN>(
           N,
           H,
@@ -594,7 +583,6 @@ static void depthwise_2d_(
           HAS_BIAS,
           false /*A_symmetric*/,
           true /*B_symmetric*/,
-          BIAS_TYPE,
           Q_GRAN>(
           N,
           H,
@@ -623,7 +611,6 @@ static void depthwise_2d_(
           HAS_BIAS,
           false /*A_symmetric*/,
           false /*B_symmetric*/,
-          BIAS_TYPE,
           Q_GRAN>(
           N,
           H,
@@ -654,8 +641,8 @@ static void depthwise_2d_(
 template <
     int S,
     bool FUSE_RELU,
-    typename BIAS_TYPE,
-    QuantizationGranularity Q_GRAN>
+    QuantizationGranularity Q_GRAN,
+    typename BIAS_TYPE>
 static void depthwise_2d_(
     int N,
     int H,
@@ -677,7 +664,7 @@ static void depthwise_2d_(
     int thread_id,
     int num_threads) {
   if (bias) {
-    depthwise_2d_<S, FUSE_RELU, true /*HAS_BIAS*/, BIAS_TYPE, Q_GRAN>(
+    depthwise_2d_<S, FUSE_RELU, true /*HAS_BIAS*/, Q_GRAN>(
         N,
         H,
         W,
@@ -698,7 +685,7 @@ static void depthwise_2d_(
         thread_id,
         num_threads);
   } else {
-    depthwise_2d_<S, FUSE_RELU, false /*HAS_BIAS*/, BIAS_TYPE, Q_GRAN>(
+    depthwise_2d_<S, FUSE_RELU, false /*HAS_BIAS*/, Q_GRAN>(
         N,
         H,
         W,