[TOPI][CUDA] Add faster-rcnn proposal op

topi/python/topi/cuda/__init__.py

@@ -18,3 +18,4 @@
 from . import ssd
 from .ssd import *
 from .nms import *
+from .rcnn import *

topi/python/topi/cuda/nms.py

@@ -35,7 +35,7 @@ def sort_ir(data, index, output):
     p_index = ib.buffer_ptr(index)
     p_out = ib.buffer_ptr(output)
     nthread_tx = max_threads
-    nthread_bx = num_anchors // max_threads + 1
+    nthread_bx = (num_anchors + 1) // 2 // max_threads + 1
     tx = tvm.thread_axis("threadIdx.x")
     bx = tvm.thread_axis("vthread")
     ib.scope_attr(tx, "thread_extent", nthread_tx)
@@ -46,8 +46,10 @@ def sort_ir(data, index, output):
 
     with ib.for_range(0, batch, for_type="unroll") as b:
         start = b * num_anchors
-        with ib.if_scope(tid < num_anchors):
-            p_out[start + tid] = tid
+        for i in range(2):
+            bbox_id = tid * 2 + i
+            with ib.if_scope(bbox_id < num_anchors):
+                p_out[start + bbox_id] = bbox_id
         # OddEvenTransposeSort
         with ib.for_range(0, p_index[b]) as k:
             with ib.if_scope(tid < (p_index[b] + 1) // 2):

topi/python/topi/cuda/rcnn/__init__.py

@@ -0,0 +1,3 @@
+# pylint: disable=wildcard-import
+"""Faster R-CNN and Mask R-CNN operators"""
+from .proposal import *

topi/python/topi/cuda/rcnn/proposal.py

@@ -0,0 +1,356 @@
+# pylint: disable=invalid-name, singleton-comparison
+"""Proposal operator"""
+import math
+import tvm
+from ...vision.rcnn import proposal, generate_anchor, reg_bbox, reg_iou
+from ...util import get_const_tuple, get_const_int
+
+
+def predict_bbox_ir(cls_prob_buf, bbox_pred_buf, im_info_buf, out_buf, scales, ratios,
+                    feature_stride, rpn_min_size, iou_loss):
+    """Predict bounding boxes based on anchors, scores and deltas.
+
+    Parameters
+    ----------
+    cls_prob_buf : tvm.schedule.Buffer
+        4-D with shape [batch, 2 * num_anchors, height, width]
+
+    bbox_pred_buf : tvm.schedule.Buffer
+        4-D with shape [batch, 4 * num_anchors, height, width]
+
+    im_info_buf : tvm.schedule.Buffer
+        2-D with shape [batch, 3]
+
+    out_buf : tvm.schedule.Buffer
+        3-D with shape [batch, num_bbox, 5]
+        The last dimension is in format of [w_start, h_start, w_end, h_end, score]
+
+    scales : list/tuple of float
+        Scales of anchor windoes.
+
+    ratios : list/tuple of float
+        Ratios of anchor windoes.
+
+    feature_stride : int
+        The size of the receptive field each unit in the convolution layer of the rpn, for example
+        the product of all stride's prior to this layer.
+
+    rpn_min_size : int
+        Minimum height or width in proposal.
+
+    iou_loss : bool
+        Usage of IoU loss.
+
+    Returns
+    -------
+    stmt : Stmt
+        The result IR statement.
+    """
+    batch, num_anchors, height, width = get_const_tuple(cls_prob_buf.shape)
+    num_anchors //= 2
+    max_threads = int(tvm.target.current_target(allow_none=False).max_num_threads)
+    nthread_tx = max_threads
+    nthread_bx = (batch * height * width) // max_threads + 1
+    tx = tvm.thread_axis("threadIdx.x")
+    bx = tvm.thread_axis("blockIdx.x")
+    tid = bx * max_threads + tx
+    ib = tvm.ir_builder.create()
+    ib.scope_attr(tx, "thread_extent", nthread_tx)
+    ib.scope_attr(bx, "thread_extent", nthread_bx)
+
+    p_score = ib.buffer_ptr(cls_prob_buf)
+    p_delta = ib.buffer_ptr(bbox_pred_buf)
+    p_im_info = ib.buffer_ptr(im_info_buf)
+    p_out = ib.buffer_ptr(out_buf)
+
+    with ib.if_scope(tid < batch * height * width):
+        w = tid % width
+        h = (tid // width) % height
+        b = tid // width // height
+
+        for k in range(num_anchors):
+            out_index = tid * num_anchors + k
+            ratio = ratios[k // len(scales)]
+            scale = scales[k % len(scales)]
+            anchor = generate_anchor(ratio, scale, feature_stride)
+            im_height = p_im_info[b * 3]
+            im_width = p_im_info[b * 3 + 1]
+            x1 = anchor[0] + w * feature_stride
+            y1 = anchor[1] + h * feature_stride
+            x2 = anchor[2] + w * feature_stride
+            y2 = anchor[3] + h * feature_stride
+
+            delta = [p_delta[((((b * num_anchors + k) * 4 + i) * height + h) * width + w)]
+                     for i in range(4)]
+            regression_func = reg_iou if iou_loss else reg_bbox
+            pred_x1, pred_y1, pred_x2, pred_y2 = regression_func(x1, y1, x2, y2, *delta)
+
+            pred_x1 = tvm.max(tvm.min(pred_x1, im_width - 1.0), 0.0)
+            pred_y1 = tvm.max(tvm.min(pred_y1, im_height - 1.0), 0.0)
+            pred_x2 = tvm.max(tvm.min(pred_x2, im_width - 1.0), 0.0)
+            pred_y2 = tvm.max(tvm.min(pred_y2, im_height - 1.0), 0.0)
+
+            real_height = (im_height / feature_stride).astype('int32')
+            real_width = (im_width / feature_stride).astype('int32')
+
+            bbox_w = pred_x2 - pred_x1 + 1.0
+            bbox_h = pred_y2 - pred_y1 + 1.0
+            min_size = p_im_info[b * 3 + 2] * rpn_min_size
+
+            pred_score = p_score[((b * num_anchors * 2 + num_anchors + k) * height + h) * width + w]
+            pred_score = tvm.expr.Select(tvm.any(h >= real_height, w >= real_width),
+                                         -1.0, pred_score)
+            p_out[out_index * 5 + 0] = pred_x1
+            p_out[out_index * 5 + 1] = pred_y1
+            p_out[out_index * 5 + 2] = pred_x2
+            p_out[out_index * 5 + 3] = pred_y2
+            p_out[out_index * 5 + 4] = pred_score
+
+            with ib.if_scope(tvm.any(bbox_w < min_size, bbox_h < min_size)):
+                p_out[out_index * 5 + 0] -= min_size / 2.0
+                p_out[out_index * 5 + 1] -= min_size / 2.0
+                p_out[out_index * 5 + 2] += min_size / 2.0
+                p_out[out_index * 5 + 3] += min_size / 2.0
+                p_out[out_index * 5 + 4] = -1.0
+
+    return ib.get()
+
+
+def argsort_ir(data_buf, out_index_buf):
+    """Batched odd-even transposition sort.
+
+    Parameters
+    ----------
+    data_buf : tvm.schedule.Buffer
+        2-D with shape [batch, num_bbox]
+
+    out_index_buf : tvm.schedule.Buffer
+        2-D with shape [batch, num_bbox]. Indices of data in sorted order.
+
+    Returns
+    -------
+    stmt : Stmt
+        The result IR statement.
+    """
+    batch, num_bbox = get_const_tuple(data_buf.shape)
+    max_threads = int(tvm.target.current_target(allow_none=False).max_num_threads)
+    ib = tvm.ir_builder.create()
+    p_data = ib.buffer_ptr(data_buf)
+    index_out = ib.buffer_ptr(out_index_buf)
+    nthread_tx = max_threads
+    nthread_bx = (num_bbox + 1) // 2 // max_threads + 1
+    tx = tvm.thread_axis("threadIdx.x")
+    bx = tvm.thread_axis("vthread")
+    ib.scope_attr(tx, "thread_extent", nthread_tx)
+    ib.scope_attr(bx, "virtual_thread", nthread_bx)
+    tid = bx * nthread_tx + tx
+    temp_data = ib.allocate("float32", (1,), name="temp_data", scope="local")
+    temp_index = ib.allocate("int32", (1,), name="temp_index", scope="local")
+
+    with ib.for_range(0, batch, for_type="unroll") as b:
+        start = b * num_bbox
+        for i in range(2):
+            bbox_id = tid * 2 + i
+            with ib.if_scope(bbox_id < num_bbox):
+                index_out[start + bbox_id] = bbox_id
+        with ib.for_range(0, num_bbox) as k:
+            offset = start + 2 * tid + (k % 2)
+            with ib.if_scope(
+                tvm.all(offset + 1 < num_bbox, p_data[offset] < p_data[offset + 1])):
+                temp_data[0] = p_data[offset]
+                p_data[offset] = p_data[offset + 1]
+                p_data[offset + 1] = temp_data[0]
+                temp_index[0] = index_out[offset]
+                index_out[offset] = index_out[offset + 1]
+                index_out[offset + 1] = temp_index[0]
+            ib.emit(tvm.make.Call(None, 'tvm_storage_sync',
+                                  tvm.convert(['shared']),
+                                  tvm.expr.Call.Intrinsic, None, 0))
+    return ib.get()
+
+
+def nms_ir(sorted_bbox_buf, out_buf, nms_threshold):
+    """Non-maximum supression.
+
+    Parameters
+    ----------
+    sorted_bbox_buf : tvm.schedule.Buffer
+        3-D with shape [batch, num_bbox, 5]. The last dimension is in format of
+        [w_start, h_start, w_end, h_end, score].
+
+    out_buf : tvm.schedule.Buffer
+        2-D with shape [batch, num_bbox]. Boolean mask of whether a bounding box should be removed.
+
+    nms_threshold : float
+        Non-maximum suppression threshold.
+
+    Returns
+    -------
+    stmt : Stmt
+        The result IR statement.
+    """
+    def calculate_overlap(out_tensor, box_a_idx, box_b_idx):
+        """Calculate overlap of two boxes.
+        """
+        w = tvm.max(0.0, tvm.min(out_tensor[box_a_idx + 2], out_tensor[box_b_idx + 2])
+                    - tvm.max(out_tensor[box_a_idx], out_tensor[box_b_idx]) + 1.0)
+        h = tvm.max(0.0, tvm.min(out_tensor[box_a_idx + 3], out_tensor[box_b_idx + 3])
+                    - tvm.max(out_tensor[box_a_idx + 1], out_tensor[box_b_idx + 1]) + 1.0)
+        i = w * h
+        u = (out_tensor[box_a_idx + 2] - out_tensor[box_a_idx] + 1.0) * \
+            (out_tensor[box_a_idx + 3] - out_tensor[box_a_idx + 1] + 1.0) + \
+            (out_tensor[box_b_idx + 2] - out_tensor[box_b_idx] + 1.0) * \
+            (out_tensor[box_b_idx + 3] - out_tensor[box_b_idx + 1] + 1.0) - i
+        return i / u
+
+    batch, num_bbox = get_const_tuple(out_buf.shape)
+    max_threads = int(math.sqrt(tvm.target.current_target(allow_none=False).max_num_threads))
+    tx = tvm.thread_axis("threadIdx.x")
+    bx = tvm.thread_axis("blockIdx.x")
+    ib = tvm.ir_builder.create()
+    p_data = ib.buffer_ptr(sorted_bbox_buf)
+    p_out = ib.buffer_ptr(out_buf)
+    nthread_tx = max_threads
+    nthread_bx = num_bbox // max_threads + 1
+    ib.scope_attr(tx, "thread_extent", nthread_tx)
+    ib.scope_attr(bx, "thread_extent", nthread_bx)
+    i = bx * max_threads + tx
+    with ib.for_range(0, batch, for_type="unroll", name="n") as b:
+        base_idx = b * num_bbox
+        with ib.if_scope(i < num_bbox):
+            p_out[base_idx + i] = False
+        with ib.for_range(0, num_bbox - 1) as l:
+            with ib.if_scope(tvm.all(i < num_bbox, i > l, p_out[base_idx + l] == False)):
+                iou = calculate_overlap(p_data, (base_idx + l) * 5, (base_idx + i) * 5)
+                with ib.if_scope(iou > nms_threshold):
+                    p_out[base_idx + i] = True
+    return ib.get()
+
+
+def prepare_output_ir(sorted_bbox_buf, remove_mask_buf, out_buf):
+    """Copy output after applying nms to continuous memory.
+
+    Parameters
+    ----------
+    sorted_bbox_buf : tvm.schedule.Buffer
+        3-D with shape [batch, num_bbox, 5]. The last dimension is in format of
+        [w_start, h_start, w_end, h_end, score].
+
+    remove_mask_buf : tvm.schedule.Buffer
+        2-D with shape [batch, num_bbox]. Boolean mask of whether a bounding box should be removed.
+
+    out_buf : tvm.schedule.Buffer
+        2-D with shape [batch * rpn_post_nms_top_n, 5]. The last dimension is in format of
+        [batch_index, w_start, h_start, w_end, h_end].
+
+    Returns
+    -------
+    stmt : Stmt
+        The result IR statement.
+    """
+    batch, num_bbox, _ = get_const_tuple(sorted_bbox_buf.shape)
+    rpn_post_nms_top_n = get_const_int(out_buf.shape[0]) // batch
+    nthread_tx = batch
+    tx = tvm.thread_axis("threadIdx.x")
+    ib = tvm.ir_builder.create()
+    ib.scope_attr(tx, "thread_extent", nthread_tx)
+    i = ib.allocate('int32', (1,), 'i', scope='local')
+    i[0] = 0
+    p_sorted_bbox = ib.buffer_ptr(sorted_bbox_buf)
+    p_remove = ib.buffer_ptr(remove_mask_buf)
+    p_out = ib.buffer_ptr(out_buf)
+    b = tx
+
+    nkeep = ib.allocate('int32', (1,), 'nkeep', scope='local')
+    nkeep[0] = 0 # number of bbox after nms
+
+    with ib.for_range(0, num_bbox) as j:
+        with ib.if_scope(p_remove[b * num_bbox + j] == False):
+            nkeep[0] += 1
+    with ib.if_scope(nkeep[0] > 0):
+        with ib.for_range(0, tvm.ceil(
+            tvm.const(rpn_post_nms_top_n, 'float32') / nkeep[0]).astype('int32')):
+            with ib.for_range(0, num_bbox) as j:
+                offset_j = (b * num_bbox + j) * 5
+                offset_i = (b * rpn_post_nms_top_n + i[0]) * 5
+                with ib.if_scope(tvm.all(i[0] < rpn_post_nms_top_n,
+                                         p_remove[(b*num_bbox+j)] == False)):
+                    p_out[offset_i] = tvm.expr.Cast('float32', b)
+                    with ib.for_range(0, 4, for_type='unroll') as k:
+                        p_out[offset_i + k + 1] = p_sorted_bbox[offset_j + k]
+                    i[0] = i[0] + 1
+
+    body = ib.get()
+    return body
+
+
+@proposal.register("cuda")
+def proposal_cuda(cls_prob, bbox_pred, im_info, scales, ratios, feature_stride, threshold,
+                  rpn_pre_nms_top_n, rpn_post_nms_top_n, rpn_min_size, iou_loss):
+    """Proposal operator.
+
+    Parameters
+    ----------
+    cls_prob : tvm.Tensor
+        4-D with shape [batch, 2 * num_anchors, height, width]
+
+    bbox_pred : tvm.Tensor
+        4-D with shape [batch, 4 * num_anchors, height, width]
+
+    im_info : tvm.Tensor
+        2-D with shape [batch, 3]
+
+    scales : list/tuple of float
+        Scales of anchor windoes.
+
+    ratios : list/tuple of float
+        Ratios of anchor windoes.
+
+    feature_stride : int
+        The size of the receptive field each unit in the convolution layer of the rpn, for example
+        the product of all stride's prior to this layer.
+
+    threshold : float
+        Non-maximum suppression threshold.
+
+    rpn_pre_nms_top_n : int
+        Number of top scoring boxes to apply NMS. -1 to use all boxes.
+
+    rpn_post_nms_top_n : int
+        Number of top scoring boxes to keep after applying NMS to RPN proposals.
+
+    rpn_min_size : int
+        Minimum height or width in proposal.
+
+    iou_loss : bool
+        Usage of IoU loss.
+
+    Returns
+    -------
+    out : tvm.Tensor
+        2-D tensor with shape [batch * rpn_post_nms_top_n, 5]. The last dimension is in format of
+        [batch_index, w_start, h_start, w_end, h_end].
+    """
+
+    batch, _, height, width = get_const_tuple(cls_prob.shape)
+    num_anchors = len(scales) * len(ratios)
+    num_bbox = height * width * num_anchors
+    rpn_pre_nms_top_n = min(rpn_pre_nms_top_n, num_bbox) if rpn_pre_nms_top_n > 0 else num_bbox
+
+    bbox = tvm.extern((batch, num_bbox, 5), [cls_prob, bbox_pred, im_info], lambda ins, outs:
+                      predict_bbox_ir(ins[0], ins[1], ins[2], outs[0], scales, ratios,
+                                      feature_stride, rpn_min_size, iou_loss),
+                      dtype=bbox_pred.dtype)
+    score = tvm.compute((batch, num_bbox), lambda b, i: bbox[b, i, 4], tag='bbox_score')
+    sorted_index = tvm.extern([score.shape], [score],
+                              lambda ins, outs: argsort_ir(ins[0], outs[0]),
+                              dtype='int32')
+    sorted_bbox = tvm.compute((batch, rpn_pre_nms_top_n, 5),
+                              lambda b, i, j: bbox[b, sorted_index[b, i], j], tag='sorted_bbox')
+    nms_remove_mask = tvm.extern((batch, rpn_pre_nms_top_n), [sorted_bbox],
+                                 lambda ins, outs: nms_ir(ins[0], outs[0], threshold),
+                                 dtype='bool')
+    nms_out = tvm.extern((batch * rpn_post_nms_top_n, 5), [sorted_bbox, nms_remove_mask],
+                         lambda ins, outs: prepare_output_ir(ins[0], ins[1], outs[0]),
+                         dtype=sorted_bbox.dtype)
+    return nms_out