Skip to content

Commit

Permalink
[Fix](bangc-ops): support grad_voxel_feats zero element & fix constra… (
Browse files Browse the repository at this point in the history 
#713)

Co-authored-by: xuminjie <xuminjie@cambricon.com>
  • Loading branch information
@baicaixmj @cambriconxuminjie
baicaixmj and cambriconxuminjie committed Jun 2, 2023
1 parent 18651a1 commit 87b89d7
Show file tree
Hide file tree
Showing 4 changed files with 109 additions and 95 deletions.
2 changes: 1 addition & 1 deletion bangc-ops/kernel_depends.toml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,7 @@ roipoint_pool3d = ["transpose"]
nms = ["transpose"]
carafe = ["fill", "tensor_stride_process"]
ms_deform_attn_forward = ["fill"]
dynamic_point_to_voxel_backward = ["scatter_nd"]
dynamic_point_to_voxel_backward = ["scatter_nd", "fill"]
dynamic_point_to_voxel_forward = ["fill", "unique"]

[bangc-ops.gtest]
Expand Down
184 changes: 103 additions & 81 deletions bangc-ops/kernels/dynamic_point_to_voxel_backward/dynamic_point_to_voxel_backward.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -135,18 +135,25 @@ static mluOpStatus_t DynamicPointToVoxelBackwardParamCheck(
<< ", should meet constraint : "
"5*feats_desc->dims[0]+20*feats_desc->dims[1]+sizeof(int) <= "
<< handle->nram_size;
return MLUOP_STATUS_BAD_PARAM;
}

// 0-element check, after dim and shape check
if (grad_voxel_feats_element_num == 0 || feats_element_num == 0) {
if (mluOpGetTensorElementNum(grad_feats_desc) == 0) {
zero_element = true;
return MLUOP_STATUS_SUCCESS;
}
PARAM_CHECK(interface_name, grad_voxel_feats != NULL);
if (grad_voxel_feats_element_num != 0) {
PARAM_CHECK(interface_name, grad_voxel_feats != NULL);
}
PARAM_CHECK(interface_name, feats != NULL);
PARAM_CHECK(interface_name, voxel_feats != NULL);
if (mluOpGetTensorElementNum(voxel_feats_desc) != 0) {
PARAM_CHECK(interface_name, voxel_feats != NULL);
}
PARAM_CHECK(interface_name, point2voxel_map != NULL);
PARAM_CHECK(interface_name, voxel_points_count != NULL);
if (mluOpGetTensorElementNum(voxel_points_count_desc) != 0) {
PARAM_CHECK(interface_name, voxel_points_count != NULL);
}
PARAM_CHECK(interface_name, voxel_num != NULL);
PARAM_CHECK(interface_name, grad_feats != NULL);
if (workspace_size != 0) {
Expand All @@ -157,42 +164,38 @@ static mluOpStatus_t DynamicPointToVoxelBackwardParamCheck(

static void policyFunc(const mluOpHandle_t handle, cnrtDim3_t *k_dim,
cnrtFunctionType_t *k_type, int N) {
size_t cluster_num = mluop::runtime::getClusterLimitCapability(handle);
size_t core_per_cluster =
mluop::runtime::getCoreNumOfEachUnionCapability(handle);
// dimx equals to num of ipu cores in each cluster

// M <= N, we can not get M's value on host, so use N instead.
size_t cluster_num_need = DIV_UP(N, core_per_cluster);
size_t eight = 8;
size_t cluster_num_real = std::min(cluster_num, eight);
size_t need_cluster = std::min(cluster_num_need, cluster_num_real);

size_t k_dim_x = 0;
// kernel1 need sync all task before kernel2, so UNIONX type is need.
if (need_cluster == 1) {
*k_type = CNRT_FUNC_TYPE_UNION1;
k_dim_x = 1 * core_per_cluster;
} else if (need_cluster == 2) {
*k_type = CNRT_FUNC_TYPE_UNION2;
k_dim_x = 2 * core_per_cluster;
} else if (need_cluster > 2 && need_cluster <= 4) {
*k_type = CNRT_FUNC_TYPE_UNION4;
k_dim_x = 4 * core_per_cluster;
} else if (need_cluster > 4 && need_cluster <= 8) {
*k_type = CNRT_FUNC_TYPE_UNION8;
k_dim_x = 8 * core_per_cluster;
int max_core_num = mluop::runtime::getCoreNumOfJobLimitCapability(handle);
size_t core_num = handle->core_num_per_cluster;
if (N > max_core_num) {
k_dim->x = max_core_num;
*k_type = mluop::runtime::getJobLimitCapabilityCnrtFuncType(handle);
} else {
LOG(ERROR) << "[mluOpDynamicPointToVoxelBackward]: failed to choose kernel "
"to launch";
return;
if (N <= 4) {
k_dim->x = core_num * 1;
*k_type = CNRT_FUNC_TYPE_UNION1;
} else if (N <= 8) {
k_dim->x = core_num * 2;
*k_type = CNRT_FUNC_TYPE_UNION2;
} else if (N <= 16) {
k_dim->x = core_num * 4;
*k_type = CNRT_FUNC_TYPE_UNION4;
} else if (N <= 32) {
k_dim->x = core_num * 8;
*k_type = CNRT_FUNC_TYPE_UNION8;
} else if (N <= 64) {
k_dim->x = core_num * 16;
*k_type = CNRT_FUNC_TYPE_UNION16;
} else {
LOG(ERROR)
<< "[mluOpDynamicPointToVoxelBackward]: failed to choose kernel "
"to launch";
return;
}
}

k_dim->x = k_dim_x;
k_dim->y = 1;
k_dim->z = 1;
VLOG(5) << "Launch Kernel MLUKernelDynamicPointToVoxelBackward in UNION"
<< k_dim_x / core_per_cluster << " type";
<< *k_type / 4 << " type";
}

mluOpStatus_t MLUOP_WIN_API mluOpDynamicPointToVoxelBackward(
Expand Down Expand Up @@ -244,55 +247,74 @@ mluOpStatus_t MLUOP_WIN_API mluOpDynamicPointToVoxelBackward(

const int N = feats_desc->dims[0];
const int C = feats_desc->dims[1];
cnrtDim3_t k_dim;
cnrtFunctionType_t k_type;
policyFunc(handle, &k_dim, &k_type, N);
// 1. get scatter indices
KERNEL_CHECK((KernelDynamicPointToVoxelBackward(
k_dim, k_type, handle->queue, feats, voxel_feats, grad_feats, workspace,
point2voxel_map, voxel_num, N, C)));
// 2. scatter
const auto grad_voxel_feats_element_num =
mluOpGetTensorElementNum(grad_voxel_feats_desc);
const auto grad_feats_element_num = mluOpGetTensorElementNum(grad_feats_desc);
mluOpScatterNdMode_t scatter_mode = MLUOP_SCATTERND_ADD;
mluOpTensorDescriptor_t indices_desc;
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpCreateTensorDescriptor(&indices_desc));
int indices_dims[2] = {(int)grad_voxel_feats_element_num, 1};
INTERNAL_CHECK(interface_name, MLUOP_STATUS_SUCCESS ==
mluOpSetTensorDescriptor(
indices_desc, MLUOP_LAYOUT_ARRAY,
MLUOP_DTYPE_INT32, 2, indices_dims));
mluOpTensorDescriptor_t updates_desc;
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpCreateTensorDescriptor(&updates_desc));
int updates_dims[1] = {(int)grad_voxel_feats_element_num};
INTERNAL_CHECK(interface_name, MLUOP_STATUS_SUCCESS ==
mluOpSetTensorDescriptor(
updates_desc, MLUOP_LAYOUT_ARRAY,
MLUOP_DTYPE_FLOAT, 1, updates_dims));
mluOpTensorDescriptor_t output_desc;
INTERNAL_CHECK(interface_name, MLUOP_STATUS_SUCCESS ==
mluOpCreateTensorDescriptor(&output_desc));
int output_dims[1] = {(int)grad_feats_element_num};
INTERNAL_CHECK(interface_name, MLUOP_STATUS_SUCCESS ==
mluOpSetTensorDescriptor(
output_desc, MLUOP_LAYOUT_ARRAY,
MLUOP_DTYPE_FLOAT, 1, output_dims));
MLUOP_CHECK((mluOpScatterNd_v2(handle, scatter_mode, indices_desc, workspace,
updates_desc, grad_voxel_feats, NULL, NULL,
output_desc, grad_feats)));
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpDestroyTensorDescriptor(indices_desc));
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpDestroyTensorDescriptor(updates_desc));
INTERNAL_CHECK(interface_name, MLUOP_STATUS_SUCCESS ==
mluOpDestroyTensorDescriptor(output_desc));
VLOG(5) << interface_name << " N = " << N << ", C = " << C
<< ", grad_voxel_feats_element_num=" << grad_voxel_feats_element_num
<< ", grad_feats_element_num=" << grad_feats_element_num;
// 1. init output
uint64_t fill_0 = 0x0;
INTERNAL_CHECK(interface_name,
MLUOP_STATUS_SUCCESS ==
mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST, &fill_0,
grad_feats_desc, grad_feats));
cnrtDim3_t k_dim;
cnrtFunctionType_t k_type;
policyFunc(handle, &k_dim, &k_type, N);
if (grad_voxel_feats_element_num != 0) {
// 2. init workspace
mluOpTensorDescriptor_t indices_desc;
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpCreateTensorDescriptor(&indices_desc));
int indices_dims[2] = {(int)grad_voxel_feats_element_num, 1};
INTERNAL_CHECK(interface_name, MLUOP_STATUS_SUCCESS ==
mluOpSetTensorDescriptor(
indices_desc, MLUOP_LAYOUT_ARRAY,
MLUOP_DTYPE_INT32, 2, indices_dims));
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpFill_v3(handle, MLUOP_POINTER_MODE_HOST,
&grad_feats_element_num,
indices_desc, workspace));
// 3. get scatter indices
KERNEL_CHECK((KernelDynamicPointToVoxelBackward(
k_dim, k_type, handle->queue, feats, voxel_feats, grad_feats, workspace,
point2voxel_map, voxel_num, N, C)));
// 4. scatter
mluOpScatterNdMode_t scatter_mode = MLUOP_SCATTERND_ADD;
mluOpTensorDescriptor_t updates_desc;
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpCreateTensorDescriptor(&updates_desc));
int updates_dims[1] = {(int)grad_voxel_feats_element_num};
INTERNAL_CHECK(interface_name, MLUOP_STATUS_SUCCESS ==
mluOpSetTensorDescriptor(
updates_desc, MLUOP_LAYOUT_ARRAY,
MLUOP_DTYPE_FLOAT, 1, updates_dims));
mluOpTensorDescriptor_t output_desc;
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpCreateTensorDescriptor(&output_desc));
int output_dims[1] = {(int)grad_feats_element_num};
INTERNAL_CHECK(interface_name, MLUOP_STATUS_SUCCESS ==
mluOpSetTensorDescriptor(
output_desc, MLUOP_LAYOUT_ARRAY,
MLUOP_DTYPE_FLOAT, 1, output_dims));
MLUOP_CHECK((mluOpScatterNd_v2(handle, scatter_mode, indices_desc,
workspace, updates_desc, grad_voxel_feats,
NULL, NULL, output_desc, grad_feats)));
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpDestroyTensorDescriptor(updates_desc));
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpDestroyTensorDescriptor(output_desc));
INTERNAL_CHECK(
interface_name,
MLUOP_STATUS_SUCCESS == mluOpDestroyTensorDescriptor(indices_desc));
}
GEN_CASE_END();
return MLUOP_STATUS_SUCCESS;
}
Expand Down
14 changes: 3 additions & 11 deletions bangc-ops/kernels/dynamic_point_to_voxel_backward/dynamic_point_to_voxel_backward.mlu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -138,21 +138,13 @@ __mlu_global__ void MLUKernelMaxReduceTracebackScatterIdx(
const int *point2voxel_map, const int *voxel_num, const int N,
const int C) {
const int M = *voxel_num;
if (M == 0) {
return;
}
int size_input = N * sizeof(int);
int size_reduced_flag = M * sizeof(bool);
int size_feats = C * sizeof(T);
int size_feats_idx = C * sizeof(int);
// init workspace and output
int n_per_core = N / taskDim;
int rem = N % taskDim;
n_per_core += (int)(taskId < rem);
int offset_per_core = (taskId * n_per_core + ((taskId < rem) ? 0 : rem)) * C;
if (n_per_core > 0) {
// voxel_from [M,C] fill 'N * C' ([M:N,C] is redundent)
__gdramset(voxel_from + offset_per_core, n_per_core * C, N * C);
// grad_feats [N,C] fill '0'
__gdramset(grad_feats + offset_per_core, n_per_core * C, T(0));
}

int nram_size = MAX_NRAM_SIZE;
int n_limit = (nram_size - size_input - size_reduced_flag - size_feats_idx) /
Expand Down
4 changes: 2 additions & 2 deletions bangc-ops/mlu_op.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3647,9 +3647,9 @@ mluOpGetDynamicPointToVoxelBackwardWorkspaceSize(const mluOpHandle_t handle,
* tensor must be equal to \b grad_voxel_feats[1].
* - The first dimension of \b voxel_num tensor is one.
* - The shape of \b feats is [N, C]:
* - 2C * sizeof(datatype of \b feats) + (N + 2C + 1) * sizeof(int) + N
* - 2C * sizeof(datatype of \b feats) + (N + 3C + 1) * sizeof(int) + N
* must be less than or equal to 640KB on MLU300 series.
* - 2C * sizeof(datatype of \b feats) + (N + 2C + 1) * sizeof(int) + N
* - 2C * sizeof(datatype of \b feats) + (N + 3C + 1) * sizeof(int) + N
* must be less than or equal to 380KB on MLU500 series.
*
* @par API Dependency
Expand Down

0 comments on commit 87b89d7

Please sign in to comment.