Xenova's picture
Xenova HF Staff
sync 6fdf6301e2bb
863d9ca verified
Raw History Blame
45.4 kB
{
"cases": [
{
"name": "small_64_boxes",
"inputs": {
"boxes": { "dtype": "float32", "shape": [1, 64, 4], "data": { "kind": "constant", "value": 0.0 } },
"scores": { "dtype": "float32", "shape": [1, 1, 64], "data": { "kind": "constant", "value": 1.0 } },
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 10 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.5 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.1 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [10, 3] } },
"bench": {
"metrics": [
{
"type": "gflops",
"value": "7 * dim(shapes.boxes, 0) * dim(shapes.scores, 1) * dim(shapes.boxes, 1) * dim(shapes.boxes, 1)"
}
]
}
},
{
"name": "nms-1class-2048-boxes",
"preset": "smoke",
"inputs": {
"boxes": { "dtype": "float32", "shape": [1, 2048, 4], "data": { "kind": "constant", "value": 0.0 } },
"scores": { "dtype": "float32", "shape": [1, 1, 2048], "data": { "kind": "constant", "value": 1.0 } },
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 64 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.5 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.5 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [64, 3] } },
"bench": {
"primary": true,
"metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }]
}
},
{
"name": "nms-8class-1024-boxes",
"preset": "smoke",
"inputs": {
"boxes": { "dtype": "float32", "shape": [1, 1024, 4], "data": { "kind": "constant", "value": 0.0 } },
"scores": { "dtype": "float32", "shape": [1, 8, 1024], "data": { "kind": "constant", "value": 1.0 } },
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 32 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.5 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.5 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [256, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_single_class_4096_boxes_max_keep_wg_parallel",
"preset": "stress",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": { "dtype": "float32", "shape": [1, 4096, 4], "data": { "kind": "constant", "value": 0.0 } },
"scores": { "dtype": "float32", "shape": [1, 1, 4096], "data": { "kind": "constant", "value": 1.0 } },
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 4096 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.01 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [4096, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "yolo-80class-8400-boxes-top100",
"preset": "stress",
"provenance": {
"notes": "Worst-case exact-output workload: zero-area boxes never overlap and unit scores pass the threshold, so every class reaches max_output_boxes_per_class."
},
"attrs": { "center_point_box": 1 },
"inputs": {
"boxes": { "dtype": "float32", "shape": [1, 8400, 4], "data": { "kind": "constant", "value": 0.0 } },
"scores": { "dtype": "float32", "shape": [1, 80, 8400], "data": { "kind": "constant", "value": 1.0 } },
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.25 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [8000, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "private_flags_n255_wg256",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 255, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 1019.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 255],
"data": { "kind": "cycle", "values": [0.9, 0.25, -0.1, 0.1] }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.45] } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.25] } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [15, 3] } },
"provenance": {
"notes": "Non-overlapping positive-area boxes and tied scores above, equal to, and below the threshold. Five independent classes check exact score filtering and class-local output limits."
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "private_flags_n256_wg256",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 256, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 1023.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 256],
"data": { "kind": "cycle", "values": [0.9, 0.25, -0.1, 0.1] }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.45] } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.25] } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [15, 3] } },
"provenance": {
"notes": "Non-overlapping positive-area boxes and tied scores above, equal to, and below the threshold. Five independent classes check exact score filtering and class-local output limits."
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "private_flags_n257_wg256",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 257, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 1027.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 257],
"data": { "kind": "cycle", "values": [0.9, 0.25, -0.1, 0.1] }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.45] } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.25] } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [15, 3] } },
"provenance": {
"notes": "Non-overlapping positive-area boxes and tied scores above, equal to, and below the threshold. Five independent classes check exact score filtering and class-local output limits."
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "private_flags_n8192_wg256",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8192, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 32767.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 8192],
"data": { "kind": "cycle", "values": [0.9, 0.25, -0.1, 0.1] }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.45] } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.25] } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [15, 3] } },
"provenance": {
"notes": "Non-overlapping positive-area boxes and tied scores above, equal to, and below the threshold. Five independent classes check exact score filtering and class-local output limits."
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "private_flags_n8193_wg256",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8193, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 32771.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 8193],
"data": { "kind": "cycle", "values": [0.9, 0.25, -0.1, 0.1] }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.45] } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.25] } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [15, 3] } },
"provenance": {
"notes": "Non-overlapping positive-area boxes and tied scores above, equal to, and below the threshold. Five independent classes check exact score filtering and class-local output limits."
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "private_flags_n32768_wg256",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 32768, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 131071.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 32768],
"data": { "kind": "cycle", "values": [0.9, 0.25, -0.1, 0.1] }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.45] } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.25] } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [15, 3] } },
"provenance": {
"notes": "Non-overlapping positive-area boxes and tied scores above, equal to, and below the threshold. Five independent classes check exact score filtering and class-local output limits."
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "private_flags_n32769_wg256",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 32769, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 131075.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 32769],
"data": { "kind": "cycle", "values": [0.9, 0.25, -0.1, 0.1] }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.45] } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.25] } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [15, 3] } },
"provenance": {
"notes": "Non-overlapping positive-area boxes and tied scores above, equal to, and below the threshold. Five independent classes check exact score filtering and class-local output limits."
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "private_flags_n1025_wg128",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 1025, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 4099.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 1025],
"data": { "kind": "cycle", "values": [0.9, 0.25, -0.1, 0.1] }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.45] } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.25] } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [15, 3] } },
"provenance": {
"notes": "Non-overlapping positive-area boxes and tied scores above, equal to, and below the threshold. Five independent classes check exact score filtering and class-local output limits."
},
"tunables": { "WORKGROUP_SIZE": 128 },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms-multiclass-sparse-scores",
"preset": "stress",
"provenance": {
"notes": "Separated positive-area boxes never overlap. One score per sixteen candidates exceeds the threshold; every class still supplies at least one hundred candidates. The selected indices have an independent exact arithmetic reference."
},
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8400, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 33599.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 80, 8400],
"data": {
"kind": "cycle",
"values": [0.9, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1]
}
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.25 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [8000, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n1024_c8_k32_ascending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 1024, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 4095.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 8, 1024],
"data": { "kind": "linspace", "start": -0.9, "end": 0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 32 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [256, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n1024_c8_k32_descending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 1024, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 4095.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 8, 1024],
"data": { "kind": "linspace", "start": 0.9, "end": -0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 32 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [256, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n1024_c8_k32_mixed",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 1024, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 4095.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 8, 1024],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 32 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [256, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n2048_c1_k64_ascending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 2048, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 8191.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 1, 2048],
"data": { "kind": "linspace", "start": -0.9, "end": 0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 64 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [64, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n2048_c1_k64_descending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 2048, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 8191.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 1, 2048],
"data": { "kind": "linspace", "start": 0.9, "end": -0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 64 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [64, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n2048_c1_k64_mixed",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 2048, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 8191.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 1, 2048],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 64 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [64, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n2051_c5_k100_ascending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 2051, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 8203.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 2051],
"data": { "kind": "linspace", "start": -0.9, "end": 0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [500, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n2051_c5_k100_descending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 2051, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 8203.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 2051],
"data": { "kind": "linspace", "start": 0.9, "end": -0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [500, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n2051_c5_k100_mixed",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 2051, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 8203.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 5, 2051],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [500, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n4096_c16_k100_ascending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 4096, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 16383.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 4096],
"data": { "kind": "linspace", "start": -0.9, "end": 0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n4096_c16_k100_descending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 4096, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 16383.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 4096],
"data": { "kind": "linspace", "start": 0.9, "end": -0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n4096_c16_k100_mixed",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 4096, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 16383.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 4096],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8192_c16_k100_ascending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8192, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 32767.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 8192],
"data": { "kind": "linspace", "start": -0.9, "end": 0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8192_c16_k100_descending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8192, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 32767.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 8192],
"data": { "kind": "linspace", "start": 0.9, "end": -0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8192_c16_k100_mixed",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8192, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 32767.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 8192],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8193_c16_k100_ascending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8193, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 32771.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 8193],
"data": { "kind": "linspace", "start": -0.9, "end": 0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8193_c16_k100_descending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8193, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 32771.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 8193],
"data": { "kind": "linspace", "start": 0.9, "end": -0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8193_c16_k100_mixed",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8193, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 32771.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 8193],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8400_c80_k100_ascending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8400, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 33599.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 80, 8400],
"data": { "kind": "linspace", "start": -0.9, "end": 0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [8000, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8400_c80_k100_descending",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8400, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 33599.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 80, 8400],
"data": { "kind": "linspace", "start": 0.9, "end": -0.9 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [8000, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8400_c80_k100_mixed",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8400, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 33599.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 80, 8400],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": -2.0 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [8000, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n2048_c1_k64_mixed_filtered",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 2048, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 8191.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 1, 2048],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 64 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.6 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [64, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n4096_c16_k100_mixed_filtered",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 4096, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 16383.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 4096],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.6 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8193_c16_k100_mixed_filtered",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8193, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 32771.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 16, 8193],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.6 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [1600, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "nms_neighbor_n8400_c80_k100_mixed_filtered",
"attrs": { "center_point_box": 0 },
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 8400, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 33599.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 80, 8400],
"data": { "kind": "fillFloat32", "sinStep": 0.713, "cosStep": 0.193, "scale": 0.7 }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 100 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.45 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.6 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [8000, 3] } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "4 * (numel(shapes.boxes) + numel(shapes.scores))" }] }
},
{
"name": "serial_control_n257_k32_ties",
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 257, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 1027.0 }
},
"scores": { "dtype": "float32", "shape": [1, 1, 257], "data": { "kind": "constant", "value": 1.0 } },
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 32 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.5 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.1 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [32, 3] } },
"bench": {
"metrics": [
{
"type": "gflops",
"value": "7 * dim(shapes.boxes, 0) * dim(shapes.scores, 1) * dim(shapes.boxes, 1) * dim(shapes.boxes, 1)"
}
]
}
},
{
"name": "serial_control_n257_k32_ascending",
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 257, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 1027.0 }
},
"scores": { "dtype": "float32", "shape": [1, 1, 257], "data": { "kind": "linspace", "start": 0.2, "end": 1.0 } },
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 32 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.5 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.1 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [32, 3] } },
"bench": {
"metrics": [
{
"type": "gflops",
"value": "7 * dim(shapes.boxes, 0) * dim(shapes.scores, 1) * dim(shapes.boxes, 1) * dim(shapes.boxes, 1)"
}
]
}
},
{
"name": "serial_control_n257_k32_descending",
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 257, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 1027.0 }
},
"scores": { "dtype": "float32", "shape": [1, 1, 257], "data": { "kind": "linspace", "start": 1.0, "end": 0.2 } },
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 32 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.5 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.1 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [32, 3] } },
"bench": {
"metrics": [
{
"type": "gflops",
"value": "7 * dim(shapes.boxes, 0) * dim(shapes.scores, 1) * dim(shapes.boxes, 1) * dim(shapes.boxes, 1)"
}
]
}
},
{
"name": "serial_control_n257_k32_mixed",
"inputs": {
"boxes": {
"dtype": "float32",
"shape": [1, 257, 4],
"data": { "kind": "linspace", "start": 0.0, "end": 1027.0 }
},
"scores": {
"dtype": "float32",
"shape": [1, 1, 257],
"data": { "kind": "cycle", "values": [0.9, 0.25, 0.1, 0.7, 0.3, 0.85, 0.55] }
},
"max_output_boxes_per_class": { "dtype": "uint32", "shape": [], "data": { "kind": "constant", "value": 32 } },
"iou_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.5 } },
"score_threshold": { "dtype": "float32", "shape": [], "data": { "kind": "constant", "value": 0.1 } }
},
"outputs": { "selected_indices": { "dtype": "uint32", "shape": [32, 3] } },
"bench": {
"metrics": [
{
"type": "gflops",
"value": "7 * dim(shapes.boxes, 0) * dim(shapes.scores, 1) * dim(shapes.boxes, 1) * dim(shapes.boxes, 1)"
}
]
}
}
]
}