Skip to content

Commit

Permalink
[Feature](bangc-ops):Roiaware_pool3d specify 'paramcheck' (#848)
Browse files Browse the repository at this point in the history
  • Loading branch information
@chqy99
chqy99 authored Oct 10, 2023
1 parent 5d89cb4 commit 7395734
Showing 1 changed file with 90 additions and 69 deletions.
159 changes: 90 additions & 69 deletions bangc-ops/kernels/roiaware_pool3d/roiaware_pool3d.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -102,24 +102,17 @@ static mluOpStatus_t transposeTensor(
const void *input, const int *permute,
const mluOpTensorDescriptor_t workspace_dst_desc, void *workspace_dst,
void *transpose_workspace) {
const std::string API = "[mluOpRoiawarePool3dForward]";
int input_dim = input_desc->dim;
mluOpTransposeDescriptor_t trans_desc = NULL;
size_t transpose_workspace_size = 0;
PARAM_CHECK(
API, MLUOP_STATUS_SUCCESS == mluOpCreateTransposeDescriptor(&trans_desc));
PARAM_CHECK(API, MLUOP_STATUS_SUCCESS == mluOpSetTransposeDescriptor(
trans_desc, input_dim, permute));
PARAM_CHECK(API, MLUOP_STATUS_SUCCESS == mluOpGetTransposeWorkspaceSize(
handle, input_desc, trans_desc,
&transpose_workspace_size));
PARAM_CHECK(API, MLUOP_STATUS_SUCCESS ==
mluOpTranspose_v2(handle, trans_desc, input_desc, input,
workspace_dst_desc, workspace_dst,
transpose_workspace,
transpose_workspace_size));
PARAM_CHECK(
API, MLUOP_STATUS_SUCCESS == mluOpDestroyTransposeDescriptor(trans_desc));
MLUOP_CHECK(mluOpCreateTransposeDescriptor(&trans_desc));
MLUOP_CHECK(mluOpSetTransposeDescriptor(trans_desc, input_dim, permute));
MLUOP_CHECK(mluOpGetTransposeWorkspaceSize(handle, input_desc, trans_desc,
&transpose_workspace_size));
MLUOP_CHECK(mluOpTranspose_v2(handle, trans_desc, input_desc, input,
workspace_dst_desc, workspace_dst,
transpose_workspace, transpose_workspace_size));
MLUOP_CHECK(mluOpDestroyTransposeDescriptor(trans_desc));
return MLUOP_STATUS_SUCCESS;
}

Expand Down Expand Up @@ -224,17 +217,37 @@ mluOpStatus_t MLUOP_WIN_API mluOpRoiawarePool3dForward(
/* boxes_num or channels is the y- or z-dimension in mmcv(cuda),
Maximum y- or z-dimension of a grid of thread blocks
should be less than 65536 in cuda. */
PARAM_CHECK(API, boxes_num < THRESHOLD_OF_BOXES_NUM_AND_CHANNELS);
PARAM_CHECK(API, channels < THRESHOLD_OF_BOXES_NUM_AND_CHANNELS);
if (boxes_num >= THRESHOLD_OF_BOXES_NUM_AND_CHANNELS) {
LOG(ERROR) << API << " Check failed: "
<< "boxes_num should be less than "
<< THRESHOLD_OF_BOXES_NUM_AND_CHANNELS << ".";
return MLUOP_STATUS_NOT_SUPPORTED;
}
if (channels >= THRESHOLD_OF_BOXES_NUM_AND_CHANNELS) {
LOG(ERROR) << API << " Check failed: "
<< "pts_num should be less than "
<< THRESHOLD_OF_BOXES_NUM_AND_CHANNELS << ".";
return MLUOP_STATUS_NOT_SUPPORTED;
}

/* max_pts_each_voxel affects the allocation of NRAM memory space,
so it's limited by the size of NRAM memory space. */
if (rois_desc->dtype == MLUOP_DTYPE_FLOAT) {
PARAM_CHECK(API, max_pts_each_voxel <=
THRESHOLD_OF_MAX_PTS_EACH_VOXEL_FLOAT_FORWARD);
if (max_pts_each_voxel > THRESHOLD_OF_MAX_PTS_EACH_VOXEL_FLOAT_FORWARD) {
LOG(ERROR) << API << " Check failed: "
<< "When the data type is float, "
"max_pts_each_voxel cannot be greater than "
<< THRESHOLD_OF_MAX_PTS_EACH_VOXEL_FLOAT_FORWARD << ".";
return MLUOP_STATUS_NOT_SUPPORTED;
}
} else {
PARAM_CHECK(API, max_pts_each_voxel <=
THRESHOLD_OF_MAX_PTS_EACH_VOXEL_HALF_FORWARD);
if (max_pts_each_voxel > THRESHOLD_OF_MAX_PTS_EACH_VOXEL_HALF_FORWARD) {
LOG(ERROR) << API << " Check failed: "
<< "When the data type is half, "
"max_pts_each_voxel cannot be greater than "
<< THRESHOLD_OF_MAX_PTS_EACH_VOXEL_HALF_FORWARD << ".";
return MLUOP_STATUS_NOT_SUPPORTED;
}
}

const uint64_t tensor_rois_num = mluOpGetTensorElementNum(rois_desc);
Expand All @@ -256,16 +269,15 @@ mluOpStatus_t MLUOP_WIN_API mluOpRoiawarePool3dForward(
// product of boxes_num and pts_num < 2^31
size_t product_boxesNum_ptsNum = (size_t)boxes_num * (size_t)pts_num;
if (product_boxesNum_ptsNum > (size_t)INT32_MAX) {
LOG(ERROR)
<< API
<< ": product of boxes_num and pts_num should be less than 2^31.";
return MLUOP_STATUS_BAD_PARAM;
LOG(ERROR) << API << " Check failed: "
<< "product of boxes_num and pts_num should be less than 2^31.";
return MLUOP_STATUS_NOT_SUPPORTED;
}

// check workspace
if (workspace_size > 0 && workspace == NULL) {
LOG(ERROR) << API
<< ": workspace shouldn't be null when workspace_size > 0.";
LOG(ERROR) << API << " Check failed: "
<< "workspace shouldn't be null when workspace_size > 0.";
return MLUOP_STATUS_BAD_PARAM;
}

Expand Down Expand Up @@ -346,16 +358,16 @@ mluOpStatus_t MLUOP_WIN_API mluOpRoiawarePool3dForward(
}
mluOpTensorDescriptor_t pts_desc_tmp = NULL;
MLUOP_CHECK(mluOpCreateTensorDescriptor(&pts_desc_tmp));
PARAM_CHECK(API,
MLUOP_STATUS_SUCCESS ==
mluOpSetTensorDescriptor(pts_desc_tmp, MLUOP_LAYOUT_ARRAY,
data_dtype, pts_dim, pts_tmp_dims));
PARAM_CHECK(
API, MLUOP_STATUS_SUCCESS ==
transposeTensor(handle, pts_desc, pts, pts_permute, pts_desc_tmp,
pts_workspace, transpose_workspace));
PARAM_CHECK(
API, MLUOP_STATUS_SUCCESS == mluOpDestroyTensorDescriptor(pts_desc_tmp));
MLUOP_CHECK(mluOpSetTensorDescriptor(pts_desc_tmp, MLUOP_LAYOUT_ARRAY,
data_dtype, pts_dim, pts_tmp_dims));

auto ret = transposeTensor(handle, pts_desc, pts, pts_permute, pts_desc_tmp,
pts_workspace, transpose_workspace);
if (ret != MLUOP_STATUS_SUCCESS) {
return ret;
}

MLUOP_CHECK(mluOpDestroyTensorDescriptor(pts_desc_tmp));
VLOG(5) << "[mluOpRoiawarePool3dForward] mluOpTranspose pts end.";

VLOG(5) << "[mluOpRoiawarePool3dForward] mluOpTranspose pts_feature start.";
Expand All @@ -367,35 +379,32 @@ mluOpStatus_t MLUOP_WIN_API mluOpRoiawarePool3dForward(
}
mluOpTensorDescriptor_t pts_feature_desc_tmp = NULL;
MLUOP_CHECK(mluOpCreateTensorDescriptor(&pts_feature_desc_tmp));
PARAM_CHECK(API, MLUOP_STATUS_SUCCESS ==
mluOpSetTensorDescriptor(
pts_feature_desc_tmp, MLUOP_LAYOUT_ARRAY, data_dtype,
pts_feature_dim, pts_feature_tmp_dims));
PARAM_CHECK(API,
MLUOP_STATUS_SUCCESS ==
transposeTensor(handle, pts_feature_desc, pts_feature,
pts_feature_permute, pts_feature_desc_tmp,
pts_feature_workspace, transpose_workspace));
PARAM_CHECK(API, MLUOP_STATUS_SUCCESS ==
mluOpDestroyTensorDescriptor(pts_feature_desc_tmp));
MLUOP_CHECK(mluOpSetTensorDescriptor(pts_feature_desc_tmp, MLUOP_LAYOUT_ARRAY,
data_dtype, pts_feature_dim,
pts_feature_tmp_dims));

ret = transposeTensor(handle, pts_feature_desc, pts_feature,
pts_feature_permute, pts_feature_desc_tmp,
pts_feature_workspace, transpose_workspace);
if (ret != MLUOP_STATUS_SUCCESS) {
return ret;
}

MLUOP_CHECK(mluOpDestroyTensorDescriptor(pts_feature_desc_tmp));
VLOG(5) << "[mluOpRoiawarePool3dForward] mluOpTranspose pts_feature end.";

VLOG(5) << "[mluOpRoiawarePool3dForward] mluOpFill_v3 host pointer start.";
int argmax_initial_value = (pool_method == 0) ? -1 : 0;
PARAM_CHECK(
API, MLUOP_STATUS_SUCCESS == mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST,
&argmax_initial_value,
argmax_desc, argmax));
MLUOP_CHECK(mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST,
&argmax_initial_value, argmax_desc, argmax));
int pts_idx_initial_value = 0;
PARAM_CHECK(API, MLUOP_STATUS_SUCCESS ==
mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST,
&pts_idx_initial_value,
pts_idx_of_voxels_desc, pts_idx_of_voxels));
MLUOP_CHECK(mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST,
&pts_idx_initial_value, pts_idx_of_voxels_desc,
pts_idx_of_voxels));
int pooled_features_initial_value = 0;
PARAM_CHECK(API, MLUOP_STATUS_SUCCESS ==
mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST,
&pooled_features_initial_value,
pooled_features_desc, pooled_features));
MLUOP_CHECK(mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST,
&pooled_features_initial_value, pooled_features_desc,
pooled_features));
VLOG(5) << "[mluOpRoiawarePool3dForward] mluOpFill host pointer end.";

cnrtDim3_t k_dim;
Expand Down Expand Up @@ -522,13 +531,27 @@ mluOpStatus_t MLUOP_WIN_API mluOpRoiawarePool3dBackward(
/* boxes_num or channels is the y- or z-dimension in mmcv(cuda),
Maximum y- or z-dimension of a grid of thread blocks
should be less than 65536 in cuda. */
PARAM_CHECK(API, boxes_num < THRESHOLD_OF_BOXES_NUM_AND_CHANNELS);
PARAM_CHECK(API, channels < THRESHOLD_OF_BOXES_NUM_AND_CHANNELS);
if (boxes_num >= THRESHOLD_OF_BOXES_NUM_AND_CHANNELS) {
LOG(ERROR) << API << " Check failed: "
<< "boxes_num should be less than "
<< THRESHOLD_OF_BOXES_NUM_AND_CHANNELS << ".";
return MLUOP_STATUS_NOT_SUPPORTED;
}
if (channels >= THRESHOLD_OF_BOXES_NUM_AND_CHANNELS) {
LOG(ERROR) << API << " Check failed: "
<< "pts_num should be less than "
<< THRESHOLD_OF_BOXES_NUM_AND_CHANNELS << ".";
return MLUOP_STATUS_NOT_SUPPORTED;
}

/* max_pts_each_voxel affects the allocation of NRAM memory space,
so it's limited by the size of NRAM memory space. */
PARAM_CHECK(API, max_pts_each_voxel <=
THRESHOLD_OF_MAX_PTS_EACH_VOXEL_BACKWARD);
if (max_pts_each_voxel > THRESHOLD_OF_MAX_PTS_EACH_VOXEL_BACKWARD) {
LOG(ERROR) << API << " Check failed: "
<< "max_pts_each_voxel cannot be greater than "
<< THRESHOLD_OF_MAX_PTS_EACH_VOXEL_BACKWARD << ".";
return MLUOP_STATUS_NOT_SUPPORTED;
}

const uint64_t tensor_pts_idx_of_voxels_num =
mluOpGetTensorElementNum(pts_idx_of_voxels_desc);
Expand All @@ -553,7 +576,7 @@ mluOpStatus_t MLUOP_WIN_API mluOpRoiawarePool3dBackward(

// check arch
if (handle->arch < MLUOP_MLU370) {
LOG(ERROR) << API
LOG(ERROR) << API << " Check failed: "
<< " The operator does not match the current architecture.";
return MLUOP_STATUS_ARCH_MISMATCH;
}
Expand Down Expand Up @@ -589,10 +612,8 @@ mluOpStatus_t MLUOP_WIN_API mluOpRoiawarePool3dBackward(
// generate mluOpRoiawarePool3dBackward prototxt end!

int grad_in_initial_value = 0;
PARAM_CHECK(
API, MLUOP_STATUS_SUCCESS == mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST,
&grad_in_initial_value,
grad_in_desc, grad_in));
MLUOP_CHECK(mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST,
&grad_in_initial_value, grad_in_desc, grad_in));
VLOG(5)
<< "[mluOpRoiawarePool3dBackward] Initialize output space successfully.";

Expand Down

0 comments on commit 7395734

Please sign in to comment.