commit 11f099921570422e8ee1c02ecfda44201ceec092
author Georgios Pinitas <georgios.pinitas@arm.com> Mon Nov 27 11:18:34 2017 +0000
committer Anthony Barbier <anthony.barbier@arm.com> Fri Nov 02 16:41:17 2018 +0000
tree 334e0cf7079f68f399a22047869788b49f9c81ba
parent 27066c2bed8fb88843308a70f375fd49835edd55

COMPMID-556: Fix lws_hint CLSoftmaxLayer

Change-Id: I15347bcfc68d76d47efa06b6036e1ed81548d227
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110678
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

src/core/CL/ICLSimple3DKernel.cpp

diff --git a/src/core/CL/ICLSimple3DKernel.cpp b/src/core/CL/ICLSimple3DKernel.cpp
index 7b0d011..0bd9d15 100644
--- a/src/core/CL/ICLSimple3DKernel.cpp
+++ b/src/core/CL/ICLSimple3DKernel.cpp
@@ -41,7 +41,7 @@
         unsigned int idx = 0;
         add_3D_tensor_argument(idx, _input, slice);
         add_3D_tensor_argument(idx, _output, slice);
-        enqueue(queue, *this, slice);
+        enqueue(queue, *this, slice, _lws_hint);
     }
     while(window.slide_window_slice_3D(slice));
 }

src/core/CL/kernels/CLSoftmaxLayerKernel.cpp

diff --git a/src/core/CL/kernels/CLSoftmaxLayerKernel.cpp b/src/core/CL/kernels/CLSoftmaxLayerKernel.cpp
index 53a78f7..5d71424 100644
--- a/src/core/CL/kernels/CLSoftmaxLayerKernel.cpp
+++ b/src/core/CL/kernels/CLSoftmaxLayerKernel.cpp
@@ -288,7 +288,7 @@
         add_3D_tensor_argument(idx, _max, slice);
         add_3D_tensor_argument(idx, _output, slice);
         add_3D_tensor_argument(idx, _sum, slice);
-        enqueue(queue, *this, slice);
+        enqueue(queue, *this, slice, _lws_hint);
     }
     while(window_collapsed.slide_window_slice_3D(slice));
 }
@@ -334,8 +334,6 @@
     build_opts.add_option_if(is_data_type_fixed_point(dt) && (beta != 1.0f), "-DBETA=" + support::cpp11::to_string(beta_int));
     build_opts.add_option_if(is_data_type_float(dt) && (beta != 1.0f), "-DBETA=" + float_to_string_with_full_precision(beta));
 
-    // Setting _lws_hint in this way can also communicate grid_size to CLLogits1DMaxShiftExpSumKernel::run().
-    // A single workgroup performs reduction in dimension 0 in the parallel case, hence lws[0]==gws[0].
     _lws_hint                                     = cl::NullRange;
     std::string           kernel_name             = std::string("softmax_layer_max_shift_exp_sum_serial");
     ParallelReductionInfo parallel_reduction_info = is_parallel_reduction(reduction_dim_size);
@@ -355,6 +353,9 @@
         // Handle boundary conditions.
         const unsigned int multiple_grid_size = (reduction_dim_size / vector_size) % _grid_size;
         build_opts.add_option_if((multiple_grid_size != 0) || ((reduction_dim_size % vector_size) != 0), "-DNON_MULTIPLE_OF_GRID_SIZE");
+        // Setting _lws_hint in this way can also communicate grid_size to CLLogits1DMaxShiftExpSumKernel::run().
+        // A single workgroup performs reduction in dimension 0 in the parallel case, hence lws[0]==gws[0].
+        _lws_hint = cl::NDRange(_grid_size);
     }
 
     // Create kernel.
@@ -548,7 +549,7 @@
         add_3D_tensor_argument(idx, _input, slice);
         add_3D_tensor_argument(idx, _sum, sum_slice);
         add_3D_tensor_argument(idx, _output, slice);
-        enqueue(queue, *this, slice);
+        enqueue(queue, *this, slice, _lws_hint);
     }
     while(window_collapsed.slide_window_slice_3D(slice));
 }