Skip to content
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

Al/fix sm size #1440

Closed
wants to merge 2 commits into from
Closed

Al/fix sm size #1440

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cca90cf
chore(gpu): do no reset shared memory size for tree_add_chunks
agnesLeroy Jul 31, 2024
8749773
chore(gpu): do not reset max shared memory for the pbs
agnesLeroy Jul 31, 2024
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
31 changes: 0 additions & 31 deletions backends/tfhe-cuda-backend/cuda/src/integer/multiplication.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -187,21 +187,6 @@ __host__ void scratch_cuda_integer_sum_ciphertexts_vec_kb(
uint32_t num_blocks_in_radix, uint32_t max_num_radix_in_vec,
int_radix_params params, bool allocate_gpu_memory) {

size_t sm_size = (params.big_lwe_dimension + 1) * sizeof(Torus);
if (sm_size < cuda_get_max_shared_memory(gpu_indexes[0])) {
check_cuda_error(cudaFuncSetAttribute(
tree_add_chunks<Torus, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, sm_size));
cudaFuncSetCacheConfig(tree_add_chunks<Torus, FULLSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
} else {
check_cuda_error(
cudaFuncSetAttribute(tree_add_chunks<Torus, NOSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, 0));
cudaFuncSetCacheConfig(tree_add_chunks<Torus, NOSM>, cudaFuncCachePreferL1);
check_cuda_error(cudaGetLastError());
}
*mem_ptr = new int_sum_ciphertexts_vec_memory<Torus>(
streams, gpu_indexes, gpu_count, params, num_blocks_in_radix,
max_num_radix_in_vec, allocate_gpu_memory);
Expand Down Expand Up @@ -576,22 +561,6 @@ __host__ void scratch_cuda_integer_mult_radix_ciphertext_kb(
cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
int_mul_memory<Torus> **mem_ptr, uint32_t num_radix_blocks,
int_radix_params params, bool allocate_gpu_memory) {
size_t sm_size = (params.big_lwe_dimension + 1) * sizeof(Torus);
if (sm_size < cuda_get_max_shared_memory(gpu_indexes[0])) {
check_cuda_error(cudaFuncSetAttribute(
tree_add_chunks<Torus, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, sm_size));
cudaFuncSetCacheConfig(tree_add_chunks<Torus, FULLSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
} else {
check_cuda_error(
cudaFuncSetAttribute(tree_add_chunks<Torus, NOSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, 0));
cudaFuncSetCacheConfig(tree_add_chunks<Torus, NOSM>, cudaFuncCachePreferL1);
check_cuda_error(cudaGetLastError());
}

*mem_ptr = new int_mul_memory<Torus>(streams, gpu_indexes, gpu_count, params,
num_radix_blocks, allocate_gpu_memory);
}
Expand Down
16 changes: 0 additions & 16 deletions backends/tfhe-cuda-backend/cuda/src/integer/scalar_mul.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,22 +33,6 @@ __host__ void scratch_cuda_integer_radix_scalar_mul_kb(
int_scalar_mul_buffer<T> **mem_ptr, uint32_t num_radix_blocks,
int_radix_params params, bool allocate_gpu_memory) {

size_t sm_size = (params.big_lwe_dimension + 1) * sizeof(T);
if (sm_size < cuda_get_max_shared_memory(gpu_indexes[0])) {
check_cuda_error(cudaFuncSetAttribute(
tree_add_chunks<T, FULLSM>, cudaFuncAttributeMaxDynamicSharedMemorySize,
sm_size));
cudaFuncSetCacheConfig(tree_add_chunks<T, FULLSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
} else {
check_cuda_error(
cudaFuncSetAttribute(tree_add_chunks<T, NOSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, 0));
cudaFuncSetCacheConfig(tree_add_chunks<T, NOSM>, cudaFuncCachePreferL1);
check_cuda_error(cudaGetLastError());
}

*mem_ptr =
new int_scalar_mul_buffer<T>(streams, gpu_indexes, gpu_count, params,
num_radix_blocks, allocate_gpu_memory);
Expand Down
109 changes: 2 additions & 107 deletions backends/tfhe-cuda-backend/cuda/src/pbs/bootstrapping_key.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,16 +112,10 @@ void cuda_convert_lwe_programmable_bootstrap_key(cudaStream_t stream,

cuda_memcpy_async_to_gpu(d_bsk, h_bsk, buffer_size, stream, gpu_index);

double2 *buffer = (double2 *)cuda_malloc_async(0, stream, gpu_index);
double2 *buffer = nullptr;
switch (polynomial_size) {
case 256:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
check_cuda_error(cudaFuncSetAttribute(
batch_NSMFFT<FFTDegree<AmortizedDegree<256>, ForwardFFT>, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_NSMFFT<FFTDegree<AmortizedDegree<256>, ForwardFFT>, FULLSM>,
cudaFuncCachePreferShared));
batch_NSMFFT<FFTDegree<AmortizedDegree<256>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(d_bsk, dest,
buffer);
Expand All @@ -134,12 +128,6 @@ void cuda_convert_lwe_programmable_bootstrap_key(cudaStream_t stream,
break;
case 512:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
check_cuda_error(cudaFuncSetAttribute(
batch_NSMFFT<FFTDegree<AmortizedDegree<512>, ForwardFFT>, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_NSMFFT<FFTDegree<AmortizedDegree<512>, ForwardFFT>, FULLSM>,
cudaFuncCachePreferShared));
batch_NSMFFT<FFTDegree<AmortizedDegree<512>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(d_bsk, dest,
buffer);
Expand All @@ -152,12 +140,6 @@ void cuda_convert_lwe_programmable_bootstrap_key(cudaStream_t stream,
break;
case 1024:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
check_cuda_error(cudaFuncSetAttribute(
batch_NSMFFT<FFTDegree<AmortizedDegree<1024>, ForwardFFT>, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_NSMFFT<FFTDegree<AmortizedDegree<1024>, ForwardFFT>, FULLSM>,
cudaFuncCachePreferShared));
batch_NSMFFT<FFTDegree<AmortizedDegree<1024>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(d_bsk, dest,
buffer);
Expand All @@ -170,12 +152,6 @@ void cuda_convert_lwe_programmable_bootstrap_key(cudaStream_t stream,
break;
case 2048:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
check_cuda_error(cudaFuncSetAttribute(
batch_NSMFFT<FFTDegree<AmortizedDegree<2048>, ForwardFFT>, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_NSMFFT<FFTDegree<AmortizedDegree<2048>, ForwardFFT>, FULLSM>,
cudaFuncCachePreferShared));
batch_NSMFFT<FFTDegree<AmortizedDegree<2048>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(d_bsk, dest,
buffer);
Expand All @@ -188,12 +164,6 @@ void cuda_convert_lwe_programmable_bootstrap_key(cudaStream_t stream,
break;
case 4096:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
check_cuda_error(cudaFuncSetAttribute(
batch_NSMFFT<FFTDegree<AmortizedDegree<4096>, ForwardFFT>, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_NSMFFT<FFTDegree<AmortizedDegree<4096>, ForwardFFT>, FULLSM>,
cudaFuncCachePreferShared));
batch_NSMFFT<FFTDegree<AmortizedDegree<4096>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(d_bsk, dest,
buffer);
Expand All @@ -206,12 +176,6 @@ void cuda_convert_lwe_programmable_bootstrap_key(cudaStream_t stream,
break;
case 8192:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
check_cuda_error(cudaFuncSetAttribute(
batch_NSMFFT<FFTDegree<AmortizedDegree<8192>, ForwardFFT>, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_NSMFFT<FFTDegree<AmortizedDegree<8192>, ForwardFFT>, FULLSM>,
cudaFuncCachePreferShared));
batch_NSMFFT<FFTDegree<AmortizedDegree<8192>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(d_bsk, dest,
buffer);
Expand All @@ -224,12 +188,6 @@ void cuda_convert_lwe_programmable_bootstrap_key(cudaStream_t stream,
break;
case 16384:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
check_cuda_error(cudaFuncSetAttribute(
batch_NSMFFT<FFTDegree<AmortizedDegree<16384>, ForwardFFT>, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_NSMFFT<FFTDegree<AmortizedDegree<16384>, ForwardFFT>, FULLSM>,
cudaFuncCachePreferShared));
batch_NSMFFT<FFTDegree<AmortizedDegree<16384>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(d_bsk, dest,
buffer);
Expand Down Expand Up @@ -265,19 +223,10 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
int gridSize = total_polynomials;
int blockSize = polynomial_size / choose_opt_amortized(polynomial_size);

double2 *buffer;
double2 *buffer = nullptr;
switch (polynomial_size) {
case 256:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
buffer = (double2 *)cuda_malloc_async(0, stream, gpu_index);
check_cuda_error(cudaFuncSetAttribute(
batch_polynomial_mul<FFTDegree<AmortizedDegree<256>, ForwardFFT>,
FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_polynomial_mul<FFTDegree<AmortizedDegree<256>, ForwardFFT>,
FULLSM>,
cudaFuncCachePreferShared));
batch_polynomial_mul<FFTDegree<AmortizedDegree<256>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(input1, input2,
output, buffer);
Expand All @@ -290,15 +239,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
break;
case 512:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
buffer = (double2 *)cuda_malloc_async(0, stream, gpu_index);
check_cuda_error(cudaFuncSetAttribute(
batch_polynomial_mul<FFTDegree<AmortizedDegree<521>, ForwardFFT>,
FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_polynomial_mul<FFTDegree<AmortizedDegree<512>, ForwardFFT>,
FULLSM>,
cudaFuncCachePreferShared));
batch_polynomial_mul<FFTDegree<AmortizedDegree<512>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(input1, input2,
output, buffer);
Expand All @@ -311,15 +251,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
break;
case 1024:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
buffer = (double2 *)cuda_malloc_async(0, stream, gpu_index);
check_cuda_error(cudaFuncSetAttribute(
batch_polynomial_mul<FFTDegree<AmortizedDegree<1024>, ForwardFFT>,
FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_polynomial_mul<FFTDegree<AmortizedDegree<1024>, ForwardFFT>,
FULLSM>,
cudaFuncCachePreferShared));
batch_polynomial_mul<FFTDegree<AmortizedDegree<1024>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(input1, input2,
output, buffer);
Expand All @@ -332,15 +263,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
break;
case 2048:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
buffer = (double2 *)cuda_malloc_async(0, stream, gpu_index);
check_cuda_error(cudaFuncSetAttribute(
batch_polynomial_mul<FFTDegree<AmortizedDegree<2048>, ForwardFFT>,
FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_polynomial_mul<FFTDegree<AmortizedDegree<2048>, ForwardFFT>,
FULLSM>,
cudaFuncCachePreferShared));
batch_polynomial_mul<FFTDegree<AmortizedDegree<2048>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(input1, input2,
output, buffer);
Expand All @@ -353,15 +275,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
break;
case 4096:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
buffer = (double2 *)cuda_malloc_async(0, stream, gpu_index);
check_cuda_error(cudaFuncSetAttribute(
batch_polynomial_mul<FFTDegree<AmortizedDegree<4096>, ForwardFFT>,
FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_polynomial_mul<FFTDegree<AmortizedDegree<4096>, ForwardFFT>,
FULLSM>,
cudaFuncCachePreferShared));
batch_polynomial_mul<FFTDegree<AmortizedDegree<4096>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(input1, input2,
output, buffer);
Expand All @@ -374,15 +287,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
break;
case 8192:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
buffer = (double2 *)cuda_malloc_async(0, stream, gpu_index);
check_cuda_error(cudaFuncSetAttribute(
batch_polynomial_mul<FFTDegree<AmortizedDegree<8192>, ForwardFFT>,
FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_polynomial_mul<FFTDegree<AmortizedDegree<8192>, ForwardFFT>,
FULLSM>,
cudaFuncCachePreferShared));
batch_polynomial_mul<FFTDegree<AmortizedDegree<8192>, ForwardFFT>, FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(input1, input2,
output, buffer);
Expand All @@ -395,15 +299,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
break;
case 16384:
if (shared_memory_size <= cuda_get_max_shared_memory(gpu_index)) {
buffer = (double2 *)cuda_malloc_async(0, stream, gpu_index);
check_cuda_error(cudaFuncSetAttribute(
batch_polynomial_mul<FFTDegree<AmortizedDegree<16384>, ForwardFFT>,
FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
check_cuda_error(cudaFuncSetCacheConfig(
batch_polynomial_mul<FFTDegree<AmortizedDegree<16384>, ForwardFFT>,
FULLSM>,
cudaFuncCachePreferShared));
batch_polynomial_mul<FFTDegree<AmortizedDegree<16384>, ForwardFFT>,
FULLSM>
<<<gridSize, blockSize, shared_memory_size, stream>>>(input1, input2,
Expand Down
22 changes: 0 additions & 22 deletions backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_amortized.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -260,28 +260,6 @@ __host__ void scratch_programmable_bootstrap_amortized(
uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
bool allocate_gpu_memory) {

cudaSetDevice(gpu_index);
uint64_t full_sm =
get_buffer_size_full_sm_programmable_bootstrap_amortized<Torus>(
polynomial_size, glwe_dimension);
uint64_t partial_sm =
get_buffer_size_partial_sm_programmable_bootstrap_amortized<Torus>(
polynomial_size);
if (max_shared_memory >= partial_sm && max_shared_memory < full_sm) {
cudaFuncSetAttribute(
device_programmable_bootstrap_amortized<Torus, params, PARTIALSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, partial_sm);
cudaFuncSetCacheConfig(
device_programmable_bootstrap_amortized<Torus, params, PARTIALSM>,
cudaFuncCachePreferShared);
} else if (max_shared_memory >= partial_sm) {
check_cuda_error(cudaFuncSetAttribute(
device_programmable_bootstrap_amortized<Torus, params, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, full_sm));
check_cuda_error(cudaFuncSetCacheConfig(
device_programmable_bootstrap_amortized<Torus, params, FULLSM>,
cudaFuncCachePreferShared));
}
if (allocate_gpu_memory) {
uint64_t buffer_size =
get_buffer_size_programmable_bootstrap_amortized<Torus>(
Expand Down
24 changes: 0 additions & 24 deletions backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_cg_classic.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -164,30 +164,6 @@ __host__ void scratch_programmable_bootstrap_cg(
uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
bool allocate_gpu_memory) {

cudaSetDevice(gpu_index);
uint64_t full_sm =
get_buffer_size_full_sm_programmable_bootstrap_cg<Torus>(polynomial_size);
uint64_t partial_sm =
get_buffer_size_partial_sm_programmable_bootstrap_cg<Torus>(
polynomial_size);
if (max_shared_memory >= partial_sm && max_shared_memory < full_sm) {
check_cuda_error(cudaFuncSetAttribute(
device_programmable_bootstrap_cg<Torus, params, PARTIALSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, partial_sm));
cudaFuncSetCacheConfig(
device_programmable_bootstrap_cg<Torus, params, PARTIALSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
} else if (max_shared_memory >= partial_sm) {
check_cuda_error(cudaFuncSetAttribute(
device_programmable_bootstrap_cg<Torus, params, FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, full_sm));
cudaFuncSetCacheConfig(
device_programmable_bootstrap_cg<Torus, params, FULLSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
}

*buffer = new pbs_buffer<Torus, CLASSICAL>(
stream, gpu_index, glwe_dimension, polynomial_size, level_count,
input_lwe_ciphertext_count, PBS_VARIANT::CG, allocate_gpu_memory);
Expand Down
Loading
Loading