COMPMID-797 Integrate Mobilenet QASYMM8 with new graph.

Change-Id: I4df63ec2f4eb27a8a6eec2bea27741bf8dec6910
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/126966
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

4 files changed, 31 insertions, 20 deletions

diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index 5f747bc477..711b006ede 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -102,7 +102,10 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
         matrix_b = &_tmp_b;
 
         _memory_group.manage(&_tmp_a);
-        _memory_group.manage(&_tmp_b);
+        if(!_reshape_b_only_on_first_run)
+        {
+            _memory_group.manage(&_tmp_b);
+        }
 
         // Configure interleave kernel
         _mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height);
@@ -119,7 +122,10 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
     {
         TensorInfo info_vector_sum_col(compute_reductionA_shape(*b->info()), 1, DataType::S32);
         _vector_sum_col.allocator()->init(info_vector_sum_col);
-        _memory_group.manage(&_vector_sum_col);
+        if(!_reshape_b_only_on_first_run)
+        {
+            _memory_group.manage(&_vector_sum_col);
+        }
 
         // Configure Matrix B reduction kernel
         _mtx_b_reduction_kernel.configure(b, &_vector_sum_col);
diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp
index 18e6e919c3..e0859be93e 100644
--- a/src/runtime/NEON/functions/NEGEMM.cpp
+++ b/src/runtime/NEON/functions/NEGEMM.cpp
@@ -107,7 +107,10 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe
 
             // Manage intermediate buffers
             _memory_group.manage(&_tmp_a);
-            _memory_group.manage(&_tmp_b);
+            if(!_reshape_b_only_on_first_run)
+            {
+                _memory_group.manage(&_tmp_b);
+            }
 
             int m = a->info()->dimension(1);
             int n = b->info()->dimension(0);
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index 7f25c2e717..3c48d691ed 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -109,14 +109,6 @@ Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, co
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
     }
 
-    // Checks performed when biases are present
-    if(append_bias)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
-    }
-
     if(transpose1xW)
     {
         TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, append_bias));
@@ -344,7 +336,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
     _memory_group.manage(&_input_im2col_reshaped);
 
     // Create tensor (interleave) to prepare input tensor for GEMM
-    if(!_is_fully_connected_convolution && !run_optimised)
+    if(!_is_fully_connected_convolution && !run_optimised && _is_interleaved)
     {
         TensorShape shape_interleaved(shape_im2col);
         shape_interleaved.set(0, shape_interleaved.x() * 4);
@@ -362,7 +354,9 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
     TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position());
     info_gemm.set_quantization_info(output->info()->quantization_info());
     _gemm_output.allocator()->init(info_gemm);
-    _memory_group.manage(&_gemm_output);
+
+    // FIXME: enabling memory manager for _gemm_output gives incorrect results (maybe bound to the assembly kernel in GEMMLowp?)
+    //  _memory_group.manage(&_gemm_output);
 
     // Configure kernels
     // Configure im2col
@@ -491,7 +485,7 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
         reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution));
         ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */));
     }
-    else
+    else if(!is_quantized)
     {
         TensorShape reshaped_weights_shape;
 
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
index 7372c6ca57..cbec73fc31 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
@@ -43,19 +43,19 @@ using namespace arm_compute::misc::shape_calculator;
 NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _asm_glue_unsigned(), _asm_glue_signed(), _mm_kernel(nullptr), _mtx_a_reshape_kernel(nullptr), _mtx_b_reshape_kernel(nullptr), _mtx_a_reduction_kernel(),
       _mtx_b_reduction_kernel(), _offset_contribution_kernel(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _workspace(), _a_offset(0), _b_offset(0), _run_vector_matrix_multiplication(false),
-      _dot_product_path(false)
+      _dot_product_path(false), _is_first_run(true), _reshape_b_only_on_first_run(false)
 {
 }
 
 void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, ITensor *output, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
-    ARM_COMPUTE_UNUSED(gemm_info);
     ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), output->info(), gemm_info));
 
     _a_offset                         = a->info()->quantization_info().offset;
     _b_offset                         = b->info()->quantization_info().offset;
     _run_vector_matrix_multiplication = a->info()->dimension(1) < 2;
+    _reshape_b_only_on_first_run      = gemm_info.reshape_b_only_on_first_run();
 
 #ifdef __aarch64__
     switch(a->info()->data_type())
@@ -98,7 +98,10 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
             _tmp_a.allocator()->init(info_a);
             _tmp_b.allocator()->init(info_b);
             _memory_group.manage(&_tmp_a);
-            _memory_group.manage(&_tmp_b);
+            if(!_reshape_b_only_on_first_run)
+            {
+                _memory_group.manage(&_tmp_b);
+            }
 
             // Configure interleave kernel
             {
@@ -129,7 +132,10 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
         TensorInfo info_vector_sum_col(compute_reductionA_shape(*b->info()), 1, DataType::S32);
 
         _vector_sum_col.allocator()->init(info_vector_sum_col);
-        _memory_group.manage(&_vector_sum_col);
+        if(!_reshape_b_only_on_first_run)
+        {
+            _memory_group.manage(&_vector_sum_col);
+        }
 
         // Configure Matrix B reduction kernel
         _mtx_b_reduction_kernel.configure(b, &_vector_sum_col, a->info()->dimension(0), false);
@@ -252,7 +258,7 @@ void NEGEMMLowpMatrixMultiplyCore::run()
             NEScheduler::get().schedule(_mtx_a_reshape_kernel.get(), Window::DimY);
         }
 
-        if(_mtx_b_reshape_kernel)
+        if(_mtx_b_reshape_kernel && (_is_first_run || !_reshape_b_only_on_first_run))
         {
             NEScheduler::get().schedule(_mtx_b_reshape_kernel.get(), Window::DimY);
         }
@@ -278,7 +284,7 @@ void NEGEMMLowpMatrixMultiplyCore::run()
     }
 
     // Run matrix B reduction kernel only if _a_offset is not equal to 0
-    if(_a_offset != 0)
+    if(_a_offset != 0 && (_is_first_run || !_reshape_b_only_on_first_run))
     {
         NEScheduler::get().schedule(&_mtx_b_reduction_kernel, Window::DimX);
     }
@@ -287,4 +293,6 @@ void NEGEMMLowpMatrixMultiplyCore::run()
     NEScheduler::get().schedule(&_offset_contribution_kernel, Window::DimY);
 
     _memory_group.release();
+
+    _is_first_run = false;
 }