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

fix(gpu): remove all resettings of shared memory size #1553

Merged
merged 2 commits into from
Sep 24, 2024
Merged

fix(gpu): remove all resettings of shared memory size #1553

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
7104c24
fix(gpu): add back cuda device synchronize in Drop
agnesLeroy Sep 23, 2024
460648b
fix(gpu): remove all resettings of shared memory size
agnesLeroy Sep 18, 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
9 changes: 0 additions & 9 deletions backends/tfhe-cuda-backend/cuda/src/device.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -247,14 +247,5 @@ int cuda_get_max_shared_memory(uint32_t gpu_index) {
cudaDeviceGetAttribute(&max_shared_memory, cudaDevAttrMaxSharedMemoryPerBlock,
gpu_index);
check_cuda_error(cudaGetLastError());
#if CUDA_ARCH == 900
max_shared_memory = 226000;
#elif CUDA_ARCH == 890
max_shared_memory = 127000;
#elif CUDA_ARCH == 800
max_shared_memory = 163000;
#elif CUDA_ARCH == 700
max_shared_memory = 95000;
#endif
return max_shared_memory;
}
98 changes: 0 additions & 98 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 @@ -116,12 +116,6 @@ void cuda_convert_lwe_programmable_bootstrap_key(cudaStream_t stream,
switch (polynomial_size) {
case 256:
if (shared_memory_size <= cuda_get_max_shared_memory(0)) {
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(0)) {
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(0)) {
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(0)) {
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(0)) {
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(0)) {
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(0)) {
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 @@ -270,14 +228,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
case 256:
if (shared_memory_size <= cuda_get_max_shared_memory(0)) {
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 @@ -291,14 +241,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
case 512:
if (shared_memory_size <= cuda_get_max_shared_memory(0)) {
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 @@ -312,14 +254,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
case 1024:
if (shared_memory_size <= cuda_get_max_shared_memory(0)) {
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 @@ -333,14 +267,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
case 2048:
if (shared_memory_size <= cuda_get_max_shared_memory(0)) {
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 @@ -354,14 +280,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
case 4096:
if (shared_memory_size <= cuda_get_max_shared_memory(0)) {
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 @@ -375,14 +293,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
case 8192:
if (shared_memory_size <= cuda_get_max_shared_memory(0)) {
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 @@ -396,14 +306,6 @@ void cuda_fourier_polynomial_mul(cudaStream_t stream, uint32_t gpu_index,
case 16384:
if (shared_memory_size <= cuda_get_max_shared_memory(0)) {
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 @@ -258,28 +258,6 @@ __host__ void scratch_programmable_bootstrap_amortized(
uint32_t glwe_dimension, uint32_t polynomial_size,
uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory) {

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);
int max_shared_memory = cuda_get_max_shared_memory(0);
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 @@ -194,30 +194,6 @@ __host__ void scratch_programmable_bootstrap_cg(
uint32_t polynomial_size, uint32_t level_count,
uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory) {

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);
int max_shared_memory = cuda_get_max_shared_memory(0);
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
63 changes: 0 additions & 63 deletions backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_cg_multibit.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -213,69 +213,6 @@ __host__ void scratch_cg_multi_bit_programmable_bootstrap(
uint32_t polynomial_size, uint32_t level_count,
uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory) {

uint64_t full_sm_keybundle =
get_buffer_size_full_sm_multibit_programmable_bootstrap_keybundle<Torus>(
polynomial_size);
uint64_t full_sm_cg_accumulate =
get_buffer_size_full_sm_cg_multibit_programmable_bootstrap<Torus>(
polynomial_size);
uint64_t partial_sm_cg_accumulate =
get_buffer_size_partial_sm_cg_multibit_programmable_bootstrap<Torus>(
polynomial_size);

int max_shared_memory = cuda_get_max_shared_memory(0);
if (max_shared_memory < full_sm_keybundle) {
check_cuda_error(cudaFuncSetAttribute(
device_multi_bit_programmable_bootstrap_keybundle<Torus, params, NOSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, 0));
cudaFuncSetCacheConfig(
device_multi_bit_programmable_bootstrap_keybundle<Torus, params, NOSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
} else {
check_cuda_error(cudaFuncSetAttribute(
device_multi_bit_programmable_bootstrap_keybundle<Torus, params,
FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, full_sm_keybundle));
cudaFuncSetCacheConfig(
device_multi_bit_programmable_bootstrap_keybundle<Torus, params,
FULLSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
}

if (max_shared_memory < partial_sm_cg_accumulate) {
check_cuda_error(cudaFuncSetAttribute(
device_multi_bit_programmable_bootstrap_cg_accumulate<Torus, params,
NOSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, 0));
cudaFuncSetCacheConfig(
device_multi_bit_programmable_bootstrap_cg_accumulate<Torus, params,
NOSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
} else if (max_shared_memory < full_sm_cg_accumulate) {
check_cuda_error(cudaFuncSetAttribute(
device_multi_bit_programmable_bootstrap_cg_accumulate<Torus, params,
PARTIALSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, partial_sm_cg_accumulate));
cudaFuncSetCacheConfig(
device_multi_bit_programmable_bootstrap_cg_accumulate<Torus, params,
PARTIALSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
} else {
check_cuda_error(cudaFuncSetAttribute(
device_multi_bit_programmable_bootstrap_cg_accumulate<Torus, params,
FULLSM>,
cudaFuncAttributeMaxDynamicSharedMemorySize, full_sm_cg_accumulate));
cudaFuncSetCacheConfig(
device_multi_bit_programmable_bootstrap_cg_accumulate<Torus, params,
FULLSM>,
cudaFuncCachePreferShared);
check_cuda_error(cudaGetLastError());
}

auto lwe_chunk_size = get_lwe_chunk_size<Torus, params>(
gpu_index, input_lwe_ciphertext_count, polynomial_size);
*buffer = new pbs_buffer<Torus, MULTI_BIT>(
Expand Down
Loading
Loading