Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[GPU] Fix clBuildProgram failure with ssd_mobilnet_v1_coco and batch=256 #7121

Merged
merged 1 commit into from
Aug 23, 2021
Merged

[GPU] Fix clBuildProgram failure with ssd_mobilnet_v1_coco and batch=256 #7121

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion ...lDNN/kernel_selector/core/actual_kernels/detection_output/detection_output_kernel_ref.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,7 +123,7 @@ DetectionOutputKernelRef::DispatchData SetDefault(const detection_output_params&
dispatchData.lws = {1, 1, 1};
} else {
dispatchData.gws = {input.Batch().v, 1, 1};
dispatchData.lws = {input.Batch().v, 1, 1};
dispatchData.lws = {1, 1, 1};
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can LWS downgrade has potential impact on performance for cases with big batches?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

As a result of checking performance impact, The latency for final stage improved by changing to work-groups as many as batches rather than work-items as many as batches in single work-group.

}
} else {
dispatchData.gws = {1, 1, 1};
Expand Down
11 changes: 5 additions & 6 deletions ...rence-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/detection_output_gpu_ref.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -658,7 +658,7 @@ KERNEL (detection_output_stage_final_caffe)(__global INPUT0_TYPE* input_location
__global int *buffer1) {
const int batchId = get_global_id(0);

__local int class_offset[LOCAL_BATCHES_NUM * NUM_CLASSES_ACC];
__local int class_offset[NUM_CLASSES_ACC];

const int total_det = FUNC_CALL(get_accumulated_detections)(buffer1, batchId);
buffer1[batchId * NUM_CLASSES_ACC + NUM_CLASSES] = total_det;
Expand Down Expand Up @@ -689,9 +689,9 @@ KERNEL (detection_output_stage_final_caffe)(__global INPUT0_TYPE* input_location
}

// calculate starting point of each class
class_offset[scores_size_offset] = 0;
class_offset[0] = 0;
for (int i = 1; i < NUM_CLASSES_ACC; ++i) {
class_offset[scores_size_offset + i] = class_offset[scores_size_offset + i - 1] + buffer1[scores_size_offset + i - 1];
class_offset[i] = class_offset[i - 1] + buffer1[scores_size_offset + i - 1];
}

barrier(CLK_LOCAL_MEM_FENCE);
Expand All @@ -700,7 +700,7 @@ KERNEL (detection_output_stage_final_caffe)(__global INPUT0_TYPE* input_location
for (uint idx_num_det = 0; idx_num_det < KEEP_TOP_K; idx_num_det++) {
SCORES_INFO score_info;
score_info = scoresList[scores_offset + idx_num_det];
const int idx = startIdx + class_offset[batchId * NUM_CLASSES_ACC + score_info.classId];
const int idx = startIdx + class_offset[score_info.classId];
output[idx * OUTPUT_ROW_SIZE] = TO_OUTPUT_TYPE(batchId);
output[idx * OUTPUT_ROW_SIZE + 1] = TO_OUTPUT_TYPE((DECREASE_LABEL_ID) ? score_info.classId - 1 : score_info.classId);
output[idx * OUTPUT_ROW_SIZE + 2] = TO_OUTPUT_TYPE(score_info.score);
Expand All @@ -719,7 +719,7 @@ KERNEL (detection_output_stage_final_caffe)(__global INPUT0_TYPE* input_location
ymax = max(TO_INPUT0_TYPE(0.0), min(TO_INPUT0_TYPE(1.0), ymax));
}
vstore4((OUTPUT_TYPE4)(xmin, ymin, xmax, ymax), 0, output + idx * OUTPUT_ROW_SIZE + 3);
class_offset[batchId * NUM_CLASSES_ACC + score_info.classId]++;
class_offset[score_info.classId]++;
}
} else {
const int startIdx = FUNC_CALL(get_start_idx)(buffer1, batchId);
Expand Down Expand Up @@ -753,7 +753,6 @@ KERNEL (detection_output_stage_final_caffe)(__global INPUT0_TYPE* input_location
}
}

barrier(CLK_GLOBAL_MEM_FENCE);
if(batchId == 0) {
const int final_detections = FUNC_CALL(get_final_detections)(buffer1);
unroll_for (uint i = final_detections; i < NUM_OF_IMAGES * KEEP_TOP_K; i++) {
Expand Down