From ff0bccfb4697c591d569db9c2dc223f2e311a7d3 Mon Sep 17 00:00:00 2001
From: Anthony Barbier <anthony.barbier@arm.com>
Date: Fri, 30 Nov 2018 10:42:40 +0000
Subject: COMPMID-1497: Add support for interleaved B reshaping in
 gemm_interleaved

Change-Id: I2171e1bf707bdcfa221c18d7a8904979e110020d
Reviewed-on: https://review.mlplatform.org/326
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
---
 .../NEGEMMInterleavedMatrixMultiplyWrapper.cpp     | 12 +++++++++-
 .../NEGEMMInterleavedPrepareBWrapperKernel.cpp     | 27 ++++++++++++++++++----
 2 files changed, 33 insertions(+), 6 deletions(-)

(limited to 'src/core/NEON/kernels/assembly')

diff --git a/src/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.cpp b/src/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.cpp
index 2c9cd320f0..3b2975dd80 100644
--- a/src/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.cpp
+++ b/src/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.cpp
@@ -101,6 +101,14 @@ void NEGEMMInterleavedMatrixMultiplyWrapperTemplate<To, Tr, use_dot>::create_wor
     using strategy = typename Kernel<To, use_dot>::strategy;
 
     unsigned int offset_transformed_b = 0;
+    unsigned int wl_index             = 0;
+    unsigned int num_buffers = 0, reshaped_block_size = 0;
+
+    if(!_b_is_pretransposed)
+    {
+        num_buffers         = _transformed_b->info()->tensor_shape()[1];
+        reshaped_block_size = _transformed_b->info()->tensor_shape()[0];
+    }
     execute_window_loop(_block_walker, [&](const Coordinates & id)
     {
         const unsigned int x0    = id.x();
@@ -122,7 +130,9 @@ void NEGEMMInterleavedMatrixMultiplyWrapperTemplate<To, Tr, use_dot>::create_wor
         }
         else
         {
-            ARM_COMPUTE_ERROR("Not supported");
+            // Rotate through the BufferManager's buffers:
+            wl_index++;
+            offset_transformed_b = (wl_index % num_buffers) * reshaped_block_size;
         }
     });
 }
diff --git a/src/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.cpp b/src/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.cpp
index 41a031c1c7..7fc57f3c02 100644
--- a/src/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.cpp
+++ b/src/core/NEON/kernels/assembly/NEGEMMInterleavedPrepareBWrapperKernel.cpp
@@ -35,10 +35,18 @@ namespace arm_compute
 namespace
 {
 // Call the lambda function for each workload generated by the passed window.
-template <typename To, bool use_dot, typename Lambda>
+template <typename To, bool use_dot, bool use_buffer_manager, typename Lambda>
 void for_each_element_in_window(const Window &window, const ITensor *b, ITensor *transformed_b, unsigned int N, unsigned int K, Lambda &&lambda)
 {
-    using strategy = typename Kernel<To, use_dot>::strategy;
+    using strategy           = typename Kernel<To, use_dot>::strategy;
+    unsigned int wl_index    = 0;
+    unsigned int num_buffers = 0, reshaped_block_size = 0;
+
+    if(use_buffer_manager)
+    {
+        num_buffers         = transformed_b->info()->tensor_shape()[1];
+        reshaped_block_size = transformed_b->info()->strides_in_bytes().y();
+    }
 
     unsigned int offset_transformed_b = transformed_b->info()->offset_first_element_in_bytes();
     execute_window_loop(window, [&](const Coordinates & coordinates)
@@ -62,7 +70,16 @@ void for_each_element_in_window(const Window &window, const ITensor *b, ITensor
         lambda(PrepareBWorkload(offset_b, offset_transformed_b, x0, xmax, k0, kmax));
 
         //Each workload represents one block:
-        offset_transformed_b += (x_size * k_size * sizeof(To));
+        if(use_buffer_manager)
+        {
+            // Rotate through the BufferManager's buffers:
+            wl_index++;
+            offset_transformed_b = (wl_index % num_buffers) * reshaped_block_size;
+        }
+        else
+        {
+            offset_transformed_b += (x_size * k_size * sizeof(To));
+        }
     });
 }
 
@@ -142,7 +159,7 @@ void NEGEMMInterleavedPrepareBWrapperKernelTemplate<To, use_dot>::transform(cons
 template <typename To, bool use_dot>
 void NEGEMMInterleavedPrepareBWrapperKernelTemplate<To, use_dot>::create_workloads(std::vector<PrepareBWorkload> &workloads)
 {
-    for_each_element_in_window<To, use_dot>(window(), _b, _transformed_b, _Nsize, _Ksize, [&workloads](PrepareBWorkload && wl)
+    for_each_element_in_window<To, use_dot, true>(window(), _b, _transformed_b, _Nsize, _Ksize, [&workloads](PrepareBWorkload && wl)
     {
         workloads.push_back(std::move(wl));
     });
@@ -152,7 +169,7 @@ template <typename To, bool use_dot>
 void NEGEMMInterleavedPrepareBWrapperKernelTemplate<To, use_dot>::run(const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(window, INEKernel::window());
-    for_each_element_in_window<To, use_dot>(window, _b, _transformed_b, _Nsize, _Ksize, [&](PrepareBWorkload && wl)
+    for_each_element_in_window<To, use_dot, false>(window, _b, _transformed_b, _Nsize, _Ksize, [&](PrepareBWorkload && wl)
     {
         this->transform(wl, info);
     });
-- 
cgit v1.2.1