[GPU] Fix detection output stage-0 kernel (openvinotoolkit#10262)

- Change the constant value to the maximum work group size - Add CLK_GLOBAL_MEM_FENCE barrier to synchronize storing result in intermediate buffer - Add condition to prevent access local array out of range Signed-off-by: Andrew Kwangwoong Park <[email protected]>
kwende · Feb 10, 2022 · 51c89df · 51c89df
1 parent 89c3a18
commit 51c89df
Show file tree

Hide file tree

Showing 2 changed files with 16 additions and 11 deletions.
diff --git a/.../src/kernel_selector/core/actual_kernels/detection_output/detection_output_kernel_ref.cpp b/.../src/kernel_selector/core/actual_kernels/detection_output/detection_output_kernel_ref.cpp
@@ -210,6 +210,7 @@ KernelsData DetectionOutputKernelRef::GetKernelsData(const Params& params, const
     auto num_classes = detectOutParams.detectOutParams.num_classes;
     auto num_loc_classes = (detectOutParams.detectOutParams.share_location) ? 1 : num_classes;
     auto num_prior_boxes = (loc_feature_num / (num_loc_classes * prior_box_size));
+    auto max_wg = detectOutParams.engineInfo.maxWorkGroupSize;
 
     constexpr size_t buffer_bytes = 10;  // The size of struct Scores in detection_output_gpu_ref.cl
     size_t buffer_stride = num_prior_boxes * buffer_bytes;
@@ -238,7 +239,7 @@ KernelsData DetectionOutputKernelRef::GetKernelsData(const Params& params, const
                     cldnnJit.AddConstants({MakeJitConstant("DO_STAGE_" + std::to_string(i) + "_CAFFE_OPT", "true")});
                 }
                 size_t num_bit_mask = CeilDiv(num_prior_boxes, 8);
-                size_t num_score_per_item = RoundUp(CeilDiv(num_prior_boxes, 256), 8);
+                size_t num_score_per_item = RoundUp(CeilDiv(num_prior_boxes, max_wg), 8);
                 size_t num_score_block = CeilDiv(num_prior_boxes, num_score_per_item);
                 cldnnJit.AddConstants({MakeJitConstant("NUM_BIT_MASK", num_bit_mask),
                                        MakeJitConstant("NUM_PRIORS_PER_ITEM", num_score_per_item),

diff --git a/src/plugins/intel_gpu/src/kernel_selector/core/cl_kernels/detection_output_gpu_ref.cl b/src/plugins/intel_gpu/src/kernel_selector/core/cl_kernels/detection_output_gpu_ref.cl
@@ -241,9 +241,10 @@ KERNEL (detection_output_stage_0_scores_caffe)(__global INPUT1_TYPE* input_confi
     __local char4 bit_mask[NUM_BIT_MASK];
     __local int4 block_num[NUM_PRIOR_BLOCKS];
 
-    block_num[box_gid] = (int4)(0, 0, 0, 0);
-
     {
+        // to prevent access array out of range
+        if (start_bid < end_bid)
+            block_num[box_gid] = (int4)(0, 0, 0, 0);
         int mask_id = start_bid / 8;
         for (int i = start_bid; i < end_bid; i += 8) {
             bit_mask[mask_id] = (char4)(0, 0, 0, 0);
@@ -277,10 +278,12 @@ KERNEL (detection_output_stage_0_scores_caffe)(__global INPUT1_TYPE* input_confi
             }
         }
     }
-    barrier(CLK_LOCAL_MEM_FENCE);
+    barrier(CLK_GLOBAL_MEM_FENCE | CLK_LOCAL_MEM_FENCE);
 
     {
-        int4 write_offsets = block_num[box_gid];
+        int4 write_offsets = (int4)(0, 0, 0, 0);
+        if (start_bid < end_bid)
+            write_offsets = block_num[box_gid];
         int mask_id = start_bid >> 3;
         for (int i = start_bid; i < end_bid; i += 8) {
             for (int bi = 0; bi < 8; bi++) {
@@ -319,9 +322,10 @@ KERNEL (detection_output_stage_0_scores_caffe)(__global INPUT1_TYPE* input_confi
     __local char bit_mask[NUM_BIT_MASK];
     __local int block_num[NUM_PRIOR_BLOCKS];
 
-    block_num[box_gid] = 0;
-
     {
+        // to prevent access array out of range
+        if (start_bid < end_bid)
+            block_num[box_gid] = 0;
         int mask_id = start_bid / 8;
         for (int i = start_bid; i < end_bid; i += 8) {
             bit_mask[mask_id] = 0;
@@ -336,7 +340,6 @@ KERNEL (detection_output_stage_0_scores_caffe)(__global INPUT1_TYPE* input_confi
             mask_id++;
         }
     }
-
     barrier(CLK_LOCAL_MEM_FENCE);
 
     {
@@ -350,11 +353,12 @@ KERNEL (detection_output_stage_0_scores_caffe)(__global INPUT1_TYPE* input_confi
             buffer1[batchId * NUM_CLASSES_ACC + classId] = acc_num;
         }
     }
-
-    barrier(CLK_LOCAL_MEM_FENCE);
+    barrier(CLK_GLOBAL_MEM_FENCE | CLK_LOCAL_MEM_FENCE);
 
     {
-        int write_offset = block_num[box_gid];
+        int write_offset = 0;
+        if (start_bid < end_bid)
+            write_offset = block_num[box_gid];
         int mask_id = start_bid >> 3;
         for (int i = start_bid; i < end_bid; i += 8) {
             for (int bi = 0; bi < 8; bi++) {