Skip to content

fix se_atten tabulate when exclude_types is given #2679

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
wanghan-iapcm merged 6 commits into from
Jul 15, 2023
Merged

fix se_atten tabulate when exclude_types is given #2679

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
af3b59c
fix se_atten tabulate when exclude_types is given
njzjz Jul 14, 2023
d9f2d21
fix tests
njzjz Jul 15, 2023
52561b7
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Jul 15, 2023
1cbe1a8
fix TabulateFusionSeAGradGradOp
njzjz Jul 15, 2023
a373376
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Jul 15, 2023
7f7324e
fetch attr in the OP
njzjz Jul 15, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions deepmd/descriptor/se_atten.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1122,6 +1122,7 @@ def _filter_lower(
tf.reshape(inputs_i, [natom, shape_i[1] // 4, 4]),
two_embd,
last_layer_size=outputs_size[-1],
is_sorted=len(self.exclude_types) == 0,
)

if (not self.uniform_seed) and (self.seed is not None):
Expand Down
27 changes: 18 additions & 9 deletions source/lib/include/tabulate.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,8 @@ void tabulate_fusion_se_a_cpu(FPTYPE* out,
const FPTYPE* two_embed,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted = true);

template <typename FPTYPE>
void tabulate_fusion_se_a_grad_cpu(FPTYPE* dy_dem_x,
Expand All @@ -25,7 +26,8 @@ void tabulate_fusion_se_a_grad_cpu(FPTYPE* dy_dem_x,
const FPTYPE* dy,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted = true);

template <typename FPTYPE>
void tabulate_fusion_se_a_grad_grad_cpu(FPTYPE* dz_dy,
Expand All @@ -37,7 +39,8 @@ void tabulate_fusion_se_a_grad_grad_cpu(FPTYPE* dz_dy,
const FPTYPE* dz_dy_dem,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted = true);

template <typename FPTYPE>
void tabulate_fusion_se_t_cpu(FPTYPE* out,
Expand Down Expand Up @@ -115,7 +118,8 @@ void tabulate_fusion_se_a_gpu_cuda(FPTYPE* out,
const FPTYPE* two_embed,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted = true);

template <typename FPTYPE>
void tabulate_fusion_se_a_grad_gpu_cuda(FPTYPE* dy_dem_x,
Expand All @@ -128,7 +132,8 @@ void tabulate_fusion_se_a_grad_gpu_cuda(FPTYPE* dy_dem_x,
const FPTYPE* dy,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted = true);

template <typename FPTYPE>
void tabulate_fusion_se_a_grad_grad_gpu_cuda(FPTYPE* dz_dy,
Expand All @@ -140,7 +145,8 @@ void tabulate_fusion_se_a_grad_grad_gpu_cuda(FPTYPE* dz_dy,
const FPTYPE* dz_dy_dem,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted = true);

template <typename FPTYPE>
void tabulate_fusion_se_t_gpu_cuda(FPTYPE* out,
Expand Down Expand Up @@ -219,7 +225,8 @@ void tabulate_fusion_se_a_gpu_rocm(FPTYPE* out,
const FPTYPE* two_embed,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted = true);

template <typename FPTYPE>
void tabulate_fusion_se_a_grad_gpu_rocm(FPTYPE* dy_dem_x,
Expand All @@ -232,7 +239,8 @@ void tabulate_fusion_se_a_grad_gpu_rocm(FPTYPE* dy_dem_x,
const FPTYPE* dy,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted = true);

template <typename FPTYPE>
void tabulate_fusion_se_a_grad_grad_gpu_rocm(FPTYPE* dz_dy,
Expand All @@ -244,7 +252,8 @@ void tabulate_fusion_se_a_grad_grad_gpu_rocm(FPTYPE* dz_dy,
const FPTYPE* dz_dy_dem,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted = true);

template <typename FPTYPE>
void tabulate_fusion_se_t_gpu_rocm(FPTYPE* out,
Expand Down
53 changes: 33 additions & 20 deletions source/lib/src/cuda/tabulate.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -128,7 +128,8 @@ __global__ void tabulate_fusion_se_a_fifth_order_polynomial(
const FPTYPE stride0,
const FPTYPE stride1,
const int nnei,
const int last_layer_size) {
const int last_layer_size,
const bool is_sorted) {
bool enable_se_atten = two_embed != nullptr;
const int_64 block_idx = blockIdx.x; // nloc
const int thread_idx = threadIdx.x; // last_layer_size
Expand All @@ -141,7 +142,7 @@ __global__ void tabulate_fusion_se_a_fifth_order_polynomial(
FPTYPE var[6];
for (int ii = 0; ii < nnei; ii++) {
FPTYPE xx = em_x[block_idx * nnei + ii];
if (xx == ago) {
if (xx == ago && is_sorted) {
unloop = true;
breakpoint = ii;
}
Expand Down Expand Up @@ -191,7 +192,8 @@ __global__ void tabulate_fusion_se_a_grad_fifth_order_polynomial(
const FPTYPE stride0,
const FPTYPE stride1,
const int nnei,
const int last_layer_size) {
const int last_layer_size,
const bool is_sorted) {
bool enable_se_atten = two_embed != nullptr;
extern __shared__ int _data[];
const int_64 block_idx = blockIdx.x; // nloc
Expand All @@ -211,7 +213,7 @@ __global__ void tabulate_fusion_se_a_grad_fifth_order_polynomial(
FPTYPE ago = __shfl_sync(0xffffffff, em_x[block_idx * nnei + nnei - 1], 0);
for (int ii = warp_idx; ii < nnei; ii += KTILE) {
FPTYPE xx = em_x[block_idx * nnei + ii];
if (ago == xx) {
if (ago == xx && is_sorted) {
unloop = true;
breakpoint = ii;
}
Expand Down Expand Up @@ -286,7 +288,8 @@ __global__ void tabulate_fusion_se_a_grad_grad_fifth_order_polynomial(
const FPTYPE stride0,
const FPTYPE stride1,
const int nnei,
const int last_layer_size) {
const int last_layer_size,
const bool is_sorted) {
extern __shared__ int _data[];
const int_64 block_idx = blockIdx.x; // nloc
const int thread_idx = threadIdx.x; // last_layer_size
Expand All @@ -304,7 +307,7 @@ __global__ void tabulate_fusion_se_a_grad_grad_fifth_order_polynomial(
for (int ii = 0; ii < nnei; ii++) {
FPTYPE xx = em_x[block_idx * nnei + ii];
FPTYPE dz_xx = dz_dy_dem_x[block_idx * nnei + ii];
if (xx == ago) {
if (xx == ago && is_sorted) {
unloop = true;
breakpoint = ii;
}
Expand Down Expand Up @@ -626,16 +629,18 @@ void tabulate_fusion_se_a_gpu_cuda(FPTYPE* out,
const FPTYPE* two_embed,
const int nloc,
const int nnei,
const int last_layer_size) {
const int last_layer_size,
const bool is_sorted) {
if (nloc <= 0) {
return;
}
DPErrcheck(cudaGetLastError());
DPErrcheck(cudaDeviceSynchronize());
tabulate_fusion_se_a_fifth_order_polynomial<FPTYPE, MM, KK>
<<<nloc, last_layer_size>>>(
out, table, em_x, em, two_embed, table_info[0], table_info[1],
table_info[2], table_info[3], table_info[4], nnei, last_layer_size);
<<<nloc, last_layer_size>>>(out, table, em_x, em, two_embed,
table_info[0], table_info[1], table_info[2],
table_info[3], table_info[4], nnei,
last_layer_size, is_sorted);
DPErrcheck(cudaGetLastError());
DPErrcheck(cudaDeviceSynchronize());
}
Expand All @@ -651,7 +656,8 @@ void tabulate_fusion_se_a_grad_gpu_cuda(FPTYPE* dy_dem_x,
const FPTYPE* dy,
const int nloc,
const int nnei,
const int last_layer_size) {
const int last_layer_size,
const bool is_sorted) {
if (nloc <= 0) {
return;
}
Expand All @@ -664,7 +670,7 @@ void tabulate_fusion_se_a_grad_gpu_cuda(FPTYPE* dy_dem_x,
<<<nloc, KK * WARP_SIZE, sizeof(FPTYPE) * MM * last_layer_size>>>(
dy_dem_x, dy_dem, table, em_x, em, two_embed, dy, table_info[0],
table_info[1], table_info[2], table_info[3], table_info[4], nnei,
last_layer_size);
last_layer_size, is_sorted);
DPErrcheck(cudaGetLastError());
DPErrcheck(cudaDeviceSynchronize());
}
Expand All @@ -679,7 +685,8 @@ void tabulate_fusion_se_a_grad_grad_gpu_cuda(FPTYPE* dz_dy,
const FPTYPE* dz_dy_dem,
const int nloc,
const int nnei,
const int last_layer_size) {
const int last_layer_size,
const bool is_sorted) {
if (nloc <= 0) {
return;
}
Expand All @@ -690,7 +697,7 @@ void tabulate_fusion_se_a_grad_grad_gpu_cuda(FPTYPE* dz_dy,
<<<nloc, last_layer_size, sizeof(FPTYPE) * MM * last_layer_size>>>(
dz_dy, table, em_x, em, dz_dy_dem_x, dz_dy_dem, table_info[0],
table_info[1], table_info[2], table_info[3], table_info[4], nnei,
last_layer_size);
last_layer_size, is_sorted);
DPErrcheck(cudaGetLastError());
DPErrcheck(cudaDeviceSynchronize());
}
Expand Down Expand Up @@ -852,7 +859,8 @@ template void tabulate_fusion_se_a_gpu_cuda<float>(float* out,
const float* two_embed,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted);
template void tabulate_fusion_se_a_gpu_cuda<double>(double* out,
const double* table,
const double* table_info,
Expand All @@ -861,7 +869,8 @@ template void tabulate_fusion_se_a_gpu_cuda<double>(double* out,
const double* two_embed,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted);
template void tabulate_fusion_se_a_grad_gpu_cuda<float>(
float* dy_dem_x,
float* dy_dem,
Expand All @@ -873,7 +882,8 @@ template void tabulate_fusion_se_a_grad_gpu_cuda<float>(
const float* dy,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted);
template void tabulate_fusion_se_a_grad_gpu_cuda<double>(
double* dy_dem_x,
double* dy_dem,
Expand All @@ -885,7 +895,8 @@ template void tabulate_fusion_se_a_grad_gpu_cuda<double>(
const double* dy,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted);
template void tabulate_fusion_se_a_grad_grad_gpu_cuda<float>(
float* dz_dy,
const float* table,
Expand All @@ -896,7 +907,8 @@ template void tabulate_fusion_se_a_grad_grad_gpu_cuda<float>(
const float* dz_dy_dem,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted);
template void tabulate_fusion_se_a_grad_grad_gpu_cuda<double>(
double* dz_dy,
const double* table,
Expand All @@ -907,7 +919,8 @@ template void tabulate_fusion_se_a_grad_grad_gpu_cuda<double>(
const double* dz_dy_dem,
const int nloc,
const int nnei,
const int last_layer_size);
const int last_layer_size,
const bool is_sorted);

template void tabulate_fusion_se_t_gpu_cuda<float>(float* out,
const float* table,
Expand Down
Loading