From ebbe4f4fca119e9787f47b769591643ddfc8c4a7 Mon Sep 17 00:00:00 2001
From: Jongsoo Park <jongsoo@fb.com>
Date: Mon, 10 Dec 2018 22:59:28 -0800
Subject: instantiate more kernels for PackAmatrix (#47)

Summary:
Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/47

PackAMatrix (compared to PackAWithRowOffset) can be a faster alternative when B_zero_point = 0

Reviewed By: jianyuh

Differential Revision: D13413605

fbshipit-source-id: 2cac4560e8f166d19c58c65ae25400d1b0795b19
---
 src/ExecuteKernelU8S8.cc | 39 +++++++++++++++++++--------------------
 1 file changed, 19 insertions(+), 20 deletions(-)

(limited to 'src/ExecuteKernelU8S8.cc')

diff --git a/src/ExecuteKernelU8S8.cc b/src/ExecuteKernelU8S8.cc
index 152d7f1..3b6e059 100644
--- a/src/ExecuteKernelU8S8.cc
+++ b/src/ExecuteKernelU8S8.cc
@@ -243,25 +243,30 @@ void ExecuteKernel<
 
 ////////////////////////////////////////////////////////////////////////////////
 // ReQuantizeOutput
-#define INSTANTIATE_BASE(ACC_T, RELU, Q_GRAN) \
-  template class ExecuteKernel<               \
-      PackAWithRowOffset<uint8_t, ACC_T>,     \
-      PackBMatrix<int8_t, ACC_T>,             \
-      uint8_t,                                \
+#define INSTANTIATE_BASE(PACK_A, ACC_T, RELU, Q_GRAN) \
+  template class ExecuteKernel<                       \
+      PACK_A<uint8_t, ACC_T>,                         \
+      PackBMatrix<int8_t, ACC_T>,                     \
+      uint8_t,                                        \
       ReQuantizeOutput<RELU, Q_GRAN>>;
 
-#define INSTANTIATE_Q_GRANS(ACC_T, RELU)                          \
-  INSTANTIATE_BASE(ACC_T, RELU, QuantizationGranularity::TENSOR); \
-  INSTANTIATE_BASE(ACC_T, RELU, QuantizationGranularity::GROUP);  \
-  INSTANTIATE_BASE(ACC_T, RELU, QuantizationGranularity::OUT_CHANNEL);
+#define INSTANTIATE_Q_GRANS(PACK_A, ACC_T, RELU)                          \
+  INSTANTIATE_BASE(PACK_A, ACC_T, RELU, QuantizationGranularity::TENSOR); \
+  INSTANTIATE_BASE(PACK_A, ACC_T, RELU, QuantizationGranularity::GROUP);  \
+  INSTANTIATE_BASE(PACK_A, ACC_T, RELU, QuantizationGranularity::OUT_CHANNEL);
 
-#define INSTANTIATE_RELU(ACC_T) \
-  INSTANTIATE_Q_GRANS(ACC_T, false);   \
-  INSTANTIATE_Q_GRANS(ACC_T, true);
+#define INSTANTIATE_RELU(PACK_A, ACC_T)      \
+  INSTANTIATE_Q_GRANS(PACK_A, ACC_T, false); \
+  INSTANTIATE_Q_GRANS(PACK_A, ACC_T, true);
 
-INSTANTIATE_RELU(int32_t);
-INSTANTIATE_RELU(int16_t);
+#define INSTANTIATE_ACC_T(PACK_A)    \
+  INSTANTIATE_RELU(PACK_A, int32_t); \
+  INSTANTIATE_RELU(PACK_A, int16_t);
+
+INSTANTIATE_ACC_T(PackAMatrix);
+INSTANTIATE_ACC_T(PackAWithRowOffset);
 
+#undef INSTANTIATE_ACC_T
 #undef INSTANTIATE_RELU
 #undef INSTANTIATE_Q_GRANS
 #undef INSTANTIATE_BASE
@@ -295,12 +300,6 @@ INSTANTIATE_RELU(int16_t);
 #undef INSTANTIATE_Q_GRANS
 #undef INSTANTIATE_BASE
 
-template class ExecuteKernel<
-    PackAMatrix<uint8_t, int16_t>,
-    PackBMatrix<int8_t, int16_t>,
-    uint8_t,
-    ReQuantizeOutput<false>>;
-
 ////////////////////////////////////////////////////////////////////////////////
 // ReQuantizeForFloat
 #define INSTANTIATE_BASE(PACK_A, RELU, Q_GRAN) \
-- 
cgit v1.2.3