Improve ANSCV
This commit is contained in:
@@ -91,9 +91,14 @@ namespace ANSCENTER {
|
||||
}
|
||||
|
||||
if (!m_trtEngine) {
|
||||
// Enable batch support
|
||||
m_options.optBatchSize = 8;
|
||||
m_options.maxBatchSize = 32;
|
||||
// Enable batch support. maxBatchSize controls the TRT workspace
|
||||
// allocation (~linear in batch); opt is the kernel-selection sweet
|
||||
// spot. Max=4 was picked to fit 4 concurrent face crops per frame
|
||||
// comfortably on 8 GB GPUs while freeing ~1.5 GB VRAM vs max=32
|
||||
// — most scenes have ≤4 faces visible, so throughput cost is
|
||||
// near-zero (amortized per-face latency drops too at lower batch).
|
||||
m_options.optBatchSize = 4;
|
||||
m_options.maxBatchSize = 4;
|
||||
|
||||
m_options.maxInputHeight = GPU_FACE_HEIGHT;
|
||||
m_options.minInputHeight = GPU_FACE_HEIGHT;
|
||||
|
||||
Reference in New Issue
Block a user