Skip to content

Commit

Permalink
Merge pull request #15 from qingshui/paddlebox
Browse files Browse the repository at this point in the history 
1. add fused_concat op, 2. batch_fc remove cpu zero， 3. support embedx two length
  • Loading branch information
@qingshui
qingshui committed Aug 10, 2021
2 parents 2146a1a + db2a495 commit f7d08b4
Show file tree
Hide file tree
Showing 10 changed files with 1,141 additions and 95 deletions.
2 changes: 1 addition & 1 deletion cmake/external/box_ps.cmake
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,7 @@ IF((NOT DEFINED BOX_PS_VER) OR (NOT DEFINED BOX_PS_URL))
SET(BOX_PS_VER "0.1.1" CACHE STRING "" FORCE)
SET(BOX_PS_NAME "box_ps" CACHE STRING "" FORCE)
#SET(BOX_PS_URL "http://box-ps.gz.bcebos.com/box_ps.tar.gz" CACHE STRING "" FORCE)
SET(BOX_PS_URL "data-im.baidu.com:/home/work/var/CI_DATA/im/static/box_ps.tar.gz/box_ps.tar.gz.15" CACHE STRING "" FORCE)
SET(BOX_PS_URL "data-im.baidu.com:/home/work/var/CI_DATA/im/static/box_ps.tar.gz/box_ps.tar.gz.16" CACHE STRING "" FORCE)
ENDIF()
MESSAGE(STATUS "BOX_PS_NAME: ${BOX_PS_NAME}, BOX_PS_URL: ${BOX_PS_URL}")
SET(BOX_PS_SOURCE_DIR "${THIRD_PARTY_PATH}/box_ps")
Expand Down
53 changes: 33 additions & 20 deletions paddle/fluid/framework/fleet/box_wrapper.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -374,6 +374,12 @@ void BoxWrapper::CheckEmbedSizeIsValid(int embedx_dim, int expand_embed_dim) {
"embedx_dim % expand_embed_dim shoule be 0"));

embedx_dim = embedx_dim / expand_embed_dim;
} else if (feature_type_ == static_cast<int>(boxps::FEATURE_VARIABLE)) {
PADDLE_ENFORCE_EQ(expand_embed_dim_, (expand_embed_dim - cvm_offset_),
platform::errors::InvalidArgument(
"SetInstance(): invalid expand_embed_dim. When "
"expand_embed_dim = %d, but got %d.",
expand_embed_dim_, expand_embed_dim));
} else {
PADDLE_ENFORCE_EQ(expand_embed_dim_, expand_embed_dim,
platform::errors::InvalidArgument(
Expand All @@ -396,7 +402,7 @@ void BoxWrapper::PullSparse(const paddle::platform::Place& place,
#define EMBEDX_CASE(i, ...) \
case i: { \
constexpr size_t EmbedxDim = i; \
switch (expand_embed_dim) { \
switch (expand_embed_dim_) { \
__VA_ARGS__ \
default: \
PADDLE_THROW(platform::errors::InvalidArgument( \
Expand All @@ -418,6 +424,9 @@ void BoxWrapper::PullSparse(const paddle::platform::Place& place,
feature_type_ == static_cast<int>(boxps::FEATURE_SHOWCLK)) { \
PullSparseCase<boxps::FeaturePullValueGpuQuant<EmbedxDim, ExpandDim>>( \
place, keys, values, slot_lengths, hidden_size, expand_embed_dim); \
} else if (feature_type_ == static_cast<int>(boxps::FEATURE_VARIABLE)) { \
PullSparseCase<boxps::FeatureVarPullValueGpu<EmbedxDim, ExpandDim>>( \
place, keys, values, slot_lengths, hidden_size, expand_embed_dim); \
} else { \
PullSparseCase<boxps::FeaturePullValueGpu<EmbedxDim, ExpandDim>>( \
place, keys, values, slot_lengths, hidden_size, expand_embed_dim); \
Expand Down Expand Up @@ -457,32 +466,36 @@ void BoxWrapper::PushSparseGrad(const paddle::platform::Place& place,
#define EMBEDX_CASE(i, ...) \
case i: { \
constexpr size_t EmbedxDim = i; \
switch (expand_embed_dim) { \
switch (expand_embed_dim_) { \
__VA_ARGS__ \
default: \
PADDLE_THROW(platform::errors::InvalidArgument( \
"Unsupport this expand embedding size [%d]", expand_embed_dim)); \
} \
} break

#define PUSHSPARSE_CASE(i, ...) \
case i: { \
constexpr size_t ExpandDim = i; \
if (feature_type_ == static_cast<int>(boxps::FEATURE_SHARE_EMBEDDING)) { \
PushSparseGradCase< \
boxps::FeaturePushValueGpuShareEmbedding<EmbedxDim, ExpandDim>>( \
place, keys, grad_values, slot_lengths, hidden_size, \
expand_embed_dim, batch_size); \
} else if (feature_type_ == static_cast<int>(boxps::FEATURE_PCOC)) { \
PushSparseGradCase< \
boxps::FeaturePushValueGpuPCOC<EmbedxDim, ExpandDim>>( \
place, keys, grad_values, slot_lengths, hidden_size, \
expand_embed_dim, batch_size); \
} else { \
PushSparseGradCase<boxps::FeaturePushValueGpu<EmbedxDim, ExpandDim>>( \
place, keys, grad_values, slot_lengths, hidden_size, \
expand_embed_dim, batch_size); \
} \
#define PUSHSPARSE_CASE(i, ...) \
case i: { \
constexpr size_t ExpandDim = i; \
if (feature_type_ == static_cast<int>(boxps::FEATURE_SHARE_EMBEDDING)) { \
PushSparseGradCase< \
boxps::FeaturePushValueGpuShareEmbedding<EmbedxDim, ExpandDim>>( \
place, keys, grad_values, slot_lengths, hidden_size, \
expand_embed_dim, batch_size); \
} else if (feature_type_ == static_cast<int>(boxps::FEATURE_PCOC)) { \
PushSparseGradCase< \
boxps::FeaturePushValueGpuPCOC<EmbedxDim, ExpandDim>>( \
place, keys, grad_values, slot_lengths, hidden_size, \
expand_embed_dim, batch_size); \
} else if (feature_type_ == static_cast<int>(boxps::FEATURE_VARIABLE)) { \
PushSparseGradCase<boxps::FeatureVarPushValueGpu<EmbedxDim, ExpandDim>>( \
place, keys, grad_values, slot_lengths, hidden_size, \
expand_embed_dim, batch_size); \
} else { \
PushSparseGradCase<boxps::FeaturePushValueGpu<EmbedxDim, ExpandDim>>( \
place, keys, grad_values, slot_lengths, hidden_size, \
expand_embed_dim, batch_size); \
} \
} break

CheckEmbedSizeIsValid(hidden_size - cvm_offset_, expand_embed_dim);
Expand Down
179 changes: 177 additions & 2 deletions paddle/fluid/framework/fleet/box_wrapper.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -178,7 +178,67 @@ __global__ void PullCopyExpandNNCross(
}
} // end kernel loop
}
//========================== feature var pull ========================
template <typename FEATURE_VALUE_GPU_TYPE>
__global__ void PullCopyBaseVariable(float** dest,
const FEATURE_VALUE_GPU_TYPE* src,
const int hidden, const int expand_hidden,
const int total_len, uint64_t** keys,
int* total_dims, const int64_t* slot_lens,
const int slot_num, const int* key2slot,
const int cvm_offset) {
CUDA_KERNEL_LOOP(i, total_len) {
int x = key2slot[i];
int y = i - slot_lens[x];

auto& src_val = src[i];
float* dest_ptr = 0;
if (dest[x + slot_num] != 0) {
dest_ptr = dest[x + slot_num] + y * expand_hidden;
total_dims[i] = (static_cast<int>(src_val.embedding_size > 0) << 1);
} else {
dest_ptr = dest[x] + y * hidden;
total_dims[i] = static_cast<int>(src_val.embedding_size > 0);
}
assert(dest_ptr != 0);
const float* src_ptr = reinterpret_cast<const float*>(&src_val.show);
for (int k = 0; k < cvm_offset; ++k) {
dest_ptr[k] = src_ptr[k];
}
} // end kernel loop
}
template <typename FEATURE_VALUE_GPU_TYPE>
__global__ void PullCopyExpandVariable(
float** dest, const FEATURE_VALUE_GPU_TYPE* src, const int total_embedx_dim,
const int embedx_dim, const int expand_dim, const int total_len,
const int* total_dims, const int64_t* slot_lens, const int slot_num,
const int* key2slot, const float scale, const int cvm_offset) {
CUDA_KERNEL_LOOP(i, total_len) {
int idx = i / total_embedx_dim;
int col = i % total_embedx_dim;

int x = key2slot[idx];
int y = idx - slot_lens[x];

auto& src_val = src[idx];
if (dest[x + slot_num] != 0) { // expand
int offset = y * (expand_dim + cvm_offset) + cvm_offset + col;
if (total_dims[idx] & 0x02) {
*(dest[x + slot_num] + offset) = src_val.embedx[col] * scale;
} else {
*(dest[x + slot_num] + offset) = 0;
}
} else if (dest[x] != 0 && col < embedx_dim) { // embedx
int offset = y * (embedx_dim + cvm_offset) + cvm_offset + col;
if (total_dims[idx] & 0x01) {
*(dest[x] + offset) = src_val.embedx[col] * scale;
} else {
*(dest[x] + offset) = 0;
}
}
} // end kernel loop
}
//========================== end ==================================
__global__ void FillKey2Slot(const int total_len, const int64_t* slot_lens,
const int slot_num, int* key2slots) {
CUDA_KERNEL_LOOP(i, total_len) {
Expand Down Expand Up @@ -385,6 +445,60 @@ __global__ void PushCopyExpandNNCross(
}
}
}
//========================== feature variable push ============================
template <typename FeaturePushValueGpuType>
__global__ void PushCopyBaseVariable(
FeaturePushValueGpuType* dest, float** src, const int hidden,
const int expand_hidden, const int total_len, const int bs,
const int* slot_vector, const int* total_dims, const int64_t* slot_lens,
const int slot_num, const int* key2slot, const int cvm_offset) {
CUDA_KERNEL_LOOP(i, total_len) {
int x = key2slot[i];
int y = i - slot_lens[x];

auto& dest_val = dest[i];
dest_val.slot = slot_vector[x];
float* optr = reinterpret_cast<float*>(&dest_val.show);
float* src_val = 0;
if (src[x + slot_num] != 0) {
src_val = reinterpret_cast<float*>(src[x + slot_num] + y * expand_hidden);
} else {
src_val = reinterpret_cast<float*>(src[x] + y * hidden);
}
assert(src_val != 0);
for (int k = 0; k < cvm_offset; ++k) {
optr[k] = src_val[k]; // support variable length
}
dest_val.embed_g *= -1. * bs;
}
}
template <typename FeaturePushValueGpuType>
__global__ void PushCopyExpandVariable(
FeaturePushValueGpuType* dest, float** src, const int total_embedx_dim,
const int embedx_dim, const int expand_dim, const int total_len,
const int bs, const int* slot_vector, const int* total_dims,
const int64_t* slot_lens, const int slot_num, const int* key2slot,
const int cvm_offset) {
CUDA_KERNEL_LOOP(i, total_len) {
int idx = i / total_embedx_dim;
int col = i % total_embedx_dim;

int x = key2slot[idx];
int y = idx - slot_lens[x];

auto& dest_val = dest[idx];
if ((total_dims[idx] & 0x02) && src[x + slot_num] != 0) { // expand
int offset = y * (expand_dim + cvm_offset) + cvm_offset + col;
dest_val.embedx_g[col] = *(src[x + slot_num] + offset) * -1. * bs;
} else if ((total_dims[idx] & 0x01) && src[x] != 0 &&
col < embedx_dim) { // embedx
int offset = y * (embedx_dim + cvm_offset) + cvm_offset + col;
dest_val.embedx_g[col] = *(src[x] + offset) * -1. * bs;
} else {
dest_val.embedx_g[col] = 0;
}
}
}

__device__ void add_calculator_value(const int table_size, const float pred,
const int64_t label, const int idx,
Expand Down Expand Up @@ -479,6 +593,31 @@ void FeaturePullCopyNNCross(cudaStream_t stream, uint64_t** gpu_keys,
key2slot, scale, cvm_offset);
}

template <typename FeaturePullValueType>
void FeaturePullCopyVariable(cudaStream_t stream, uint64_t** gpu_keys,
float** gpu_values, void* src,
const int hidden_size, const size_t embedx_dim,
const size_t expand_dim, const int total_length,
int* total_dims, const int64_t* slot_lens,
const int slot_num, const int* key2slot,
const float scale, const int cvm_offset) {
FeaturePullValueType* pull_values_gpu =
reinterpret_cast<FeaturePullValueType*>(src);
PullCopyBaseVariable<
FeaturePullValueType><<<(total_length + 512 - 1) / 512, 512, 0, stream>>>(
gpu_values, pull_values_gpu, hidden_size, (expand_dim + cvm_offset),
total_length, gpu_keys, total_dims, slot_lens, slot_num, key2slot,
cvm_offset);
// embedx or expand_embedx
int max_embedx_dim = (embedx_dim > expand_dim) ? embedx_dim : expand_dim;
int embedx_total_length = total_length * max_embedx_dim;
PullCopyExpandVariable<FeaturePullValueType><<<
(embedx_total_length + 512 - 1) / 512, 512, 0, stream>>>(
gpu_values, pull_values_gpu, max_embedx_dim, embedx_dim, expand_dim,
embedx_total_length, total_dims, slot_lens, slot_num, key2slot, scale,
cvm_offset);
}

void BoxWrapper::CopyForPull(const paddle::platform::Place& place,
uint64_t** gpu_keys, float** gpu_values,
void* total_values_gpu, const int64_t* slot_lens,
Expand All @@ -492,7 +631,7 @@ void BoxWrapper::CopyForPull(const paddle::platform::Place& place,
#define EMBEDX_CASE(i, ...) \
case i: { \
constexpr size_t EmbedxDim = i; \
switch (expand_embed_dim) { \
switch (expand_embed_dim_) { \
__VA_ARGS__ \
default: \
PADDLE_THROW(platform::errors::InvalidArgument( \
Expand Down Expand Up @@ -538,6 +677,12 @@ void BoxWrapper::CopyForPull(const paddle::platform::Place& place,
stream, gpu_keys, gpu_values, total_values_gpu, hidden_size, \
EmbedxDim, ExpandDim, total_length, total_dims, slot_lens, slot_num, \
key2slot, pull_embedx_scale_, cvm_offset_); \
} else if (feature_type_ == static_cast<int>(boxps::FEATURE_VARIABLE)) { \
FeaturePullCopyVariable< \
boxps::FeatureVarPullValueGpu<EmbedxDim, ExpandDim>>( \
stream, gpu_keys, gpu_values, total_values_gpu, hidden_size, \
EmbedxDim, ExpandDim, total_length, total_dims, slot_lens, slot_num, \
key2slot, 1.0, cvm_offset_); \
} else { \
FeaturePullCopyNNCross< \
boxps::FeaturePullValueGpu<EmbedxDim, ExpandDim>>( \
Expand Down Expand Up @@ -665,6 +810,30 @@ void FeaturePushCopyShareEmbedding(
batch_size, slot_vector, total_dims, slot_lens, slot_num, key2slot,
cvm_offset);
}
template <typename FeaturePushValueGpuType>
void FeaturePushCopyVariable(cudaStream_t stream, void* dest,
float** grad_values, const int hidden_size,
const int embedx_dim, const int expand_dim,
const int total_length, const int batch_size,
const int* slot_vector, const int* total_dims,
const int64_t* slot_lens, const int slot_num,
const int* key2slot, const int cvm_offset) {
FeaturePushValueGpuType* push_grad_values =
reinterpret_cast<FeaturePushValueGpuType*>(dest);
PushCopyBaseVariable<FeaturePushValueGpuType><<<
(total_length + 512 - 1) / 512, 512, 0, stream>>>(
push_grad_values, grad_values, hidden_size, expand_dim + cvm_offset,
total_length, batch_size, slot_vector, total_dims, slot_lens, slot_num,
key2slot, cvm_offset);

int max_embedx_dim = (embedx_dim > expand_dim) ? embedx_dim : expand_dim;
int embedx_total_length = total_length * max_embedx_dim;
PushCopyExpandVariable<FeaturePushValueGpuType><<<
(embedx_total_length + 512 - 1) / 512, 512, 0, stream>>>(
push_grad_values, grad_values, max_embedx_dim, embedx_dim, expand_dim,
embedx_total_length, batch_size, slot_vector, total_dims, slot_lens,
slot_num, key2slot, cvm_offset);
}
void BoxWrapper::CopyForPush(const paddle::platform::Place& place,
float** grad_values, void* total_grad_values_gpu,
const int* d_slot_vector, const int64_t* slot_lens,
Expand All @@ -679,7 +848,7 @@ void BoxWrapper::CopyForPush(const paddle::platform::Place& place,
#define EMBEDX_CASE(i, ...) \
case i: { \
constexpr size_t EmbedxDim = i; \
switch (expand_embed_dim) { \
switch (expand_embed_dim_) { \
__VA_ARGS__ \
default: \
PADDLE_THROW(platform::errors::InvalidArgument( \
Expand Down Expand Up @@ -712,6 +881,12 @@ void BoxWrapper::CopyForPush(const paddle::platform::Place& place,
stream, total_grad_values_gpu, grad_values, hidden_size, EmbedxDim, \
ExpandDim, total_length, batch_size, d_slot_vector, total_dims, \
slot_lens, slot_num, key2slot, cvm_offset_); \
} else if (feature_type_ == static_cast<int>(boxps::FEATURE_VARIABLE)) { \
FeaturePushCopyVariable< \
boxps::FeatureVarPushValueGpu<EmbedxDim, ExpandDim>>( \
stream, total_grad_values_gpu, grad_values, hidden_size, EmbedxDim, \
ExpandDim, total_length, batch_size, d_slot_vector, total_dims, \
slot_lens, slot_num, key2slot, cvm_offset_); \
} else { \
FeaturePushCopyNNCross< \
boxps::FeaturePushValueGpu<EmbedxDim, ExpandDim>>( \
Expand Down
Loading

0 comments on commit f7d08b4

Please sign in to comment.