Skip to content

fix(gpu): fix memory leak and buffer sizing in f128 FFT host functions #3825

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

Open
pdroalves wants to merge 4 commits into
base: main
Choose a base branch
from
Open

fix(gpu): fix memory leak and buffer sizing in f128 FFT host functions #3825

Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f7c1f55
fix(gpu): free scratch buffer in f128 FFT host functions
pdroalves Aug 13, 2026
3f002e5
fix(gpu): allocate f128 FFT scratch buffer in bytes instead of elements
pdroalves Aug 13, 2026
496784b
fix(gpu): size f128 FFT host transform buffers for the whole batch
pdroalves Aug 13, 2026
fe85240
fix(gpu): allocate NOSM FFT scratch in bytes in 128-bit BSK conversion
pdroalves Aug 13, 2026
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
191 changes: 108 additions & 83 deletions backends/tfhe-cuda-backend/cuda/src/fft128/fft128.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -769,34 +769,44 @@ __host__ void host_fourier_transform_forward_as_integer_f128(
double *im0, double *im1, const __uint128_t *standard, const uint32_t N,
const uint32_t number_of_samples) {

// allocate device buffers
double *d_re0 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
double *d_re1 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
double *d_im0 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
double *d_im1 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
__uint128_t *d_standard = (__uint128_t *)cuda_malloc_async(
safe_mul_sizeof<__uint128_t>(N), stream, gpu_index);
// allocate device buffers; each kernel block processes its own sample, so
// every buffer is sized for the whole batch
size_t re_im_size = safe_mul_sizeof<double>(
static_cast<size_t>(number_of_samples), static_cast<size_t>(N / 2));
size_t standard_size = safe_mul_sizeof<__uint128_t>(
static_cast<size_t>(number_of_samples), static_cast<size_t>(N));
double *d_re0 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
double *d_re1 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
double *d_im0 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
double *d_im1 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
__uint128_t *d_standard = static_cast<__uint128_t *>(
cuda_malloc_async(standard_size, stream, gpu_index));

// copy input into device
cuda_memcpy_async_to_gpu(d_standard, standard,
safe_mul_sizeof<__uint128_t>(N), stream, gpu_index);
cuda_memcpy_async_to_gpu(d_standard, standard, standard_size, stream,
gpu_index);

// setup launch parameters
size_t required_shared_memory_size =
safe_mul_sizeof<double>((size_t)(N / 2), (size_t)4);
size_t required_shared_memory_size = safe_mul_sizeof<double>(
static_cast<size_t>(N / 2), static_cast<size_t>(4));
int grid_size = number_of_samples;
int block_size = params::degree / params::opt;
bool full_sm =
(required_shared_memory_size <= cuda_get_max_shared_memory(gpu_index));
size_t buffer_size =
full_sm ? 0
: safe_mul((size_t)number_of_samples, (size_t)(N / 2), (size_t)4);
size_t shared_memory_size = full_sm ? required_shared_memory_size : 0;
double *buffer = (double *)cuda_malloc_async(buffer_size, stream, gpu_index);
// global scratch is only needed by the NOSM kernel variant, when the FFT
// working set does not fit in shared memory; each block uses one
// shared-memory-sized slice of it
double *buffer = full_sm
? nullptr
: static_cast<double *>(cuda_malloc_async(
safe_mul(static_cast<size_t>(number_of_samples),
required_shared_memory_size),
stream, gpu_index));

// configure shared memory for batch fft kernel
if (full_sm) {
Expand Down Expand Up @@ -826,20 +836,17 @@ __host__ void host_fourier_transform_forward_as_integer_f128(
}
check_cuda_error(cudaGetLastError());

cuda_memcpy_async_to_cpu(re0, d_re0, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_cpu(re1, d_re1, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_cpu(im0, d_im0, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_cpu(im1, d_im1, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_cpu(re0, d_re0, re_im_size, stream, gpu_index);
cuda_memcpy_async_to_cpu(re1, d_re1, re_im_size, stream, gpu_index);
cuda_memcpy_async_to_cpu(im0, d_im0, re_im_size, stream, gpu_index);
cuda_memcpy_async_to_cpu(im1, d_im1, re_im_size, stream, gpu_index);

cuda_drop_async(d_standard, stream, gpu_index);
cuda_drop_async(d_re0, stream, gpu_index);
cuda_drop_async(d_re1, stream, gpu_index);
cuda_drop_async(d_im0, stream, gpu_index);
cuda_drop_async(d_im1, stream, gpu_index);
cuda_drop_async(buffer, stream, gpu_index);
}

template <class params>
Expand All @@ -848,34 +855,44 @@ __host__ void host_fourier_transform_forward_as_torus_f128(
double *im0, double *im1, const __uint128_t *standard, const uint32_t N,
const uint32_t number_of_samples) {

// allocate device buffers
double *d_re0 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
double *d_re1 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
double *d_im0 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
double *d_im1 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
__uint128_t *d_standard = (__uint128_t *)cuda_malloc_async(
safe_mul_sizeof<__uint128_t>(N), stream, gpu_index);
// allocate device buffers; each kernel block processes its own sample, so
// every buffer is sized for the whole batch
size_t re_im_size = safe_mul_sizeof<double>(
static_cast<size_t>(number_of_samples), static_cast<size_t>(N / 2));
size_t standard_size = safe_mul_sizeof<__uint128_t>(
static_cast<size_t>(number_of_samples), static_cast<size_t>(N));
double *d_re0 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
double *d_re1 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
double *d_im0 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
double *d_im1 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
__uint128_t *d_standard = static_cast<__uint128_t *>(
cuda_malloc_async(standard_size, stream, gpu_index));

// copy input into device
cuda_memcpy_async_to_gpu(d_standard, standard,
safe_mul_sizeof<__uint128_t>(N), stream, gpu_index);
cuda_memcpy_async_to_gpu(d_standard, standard, standard_size, stream,
gpu_index);

// setup launch parameters
size_t required_shared_memory_size =
safe_mul_sizeof<double>((size_t)(N / 2), (size_t)4);
size_t required_shared_memory_size = safe_mul_sizeof<double>(
static_cast<size_t>(N / 2), static_cast<size_t>(4));
int grid_size = number_of_samples;
int block_size = params::degree / params::opt;
bool full_sm =
(required_shared_memory_size <= cuda_get_max_shared_memory(gpu_index));
size_t buffer_size =
full_sm ? 0
: safe_mul((size_t)number_of_samples, (size_t)(N / 2), (size_t)4);
size_t shared_memory_size = full_sm ? required_shared_memory_size : 0;
double *buffer = (double *)cuda_malloc_async(buffer_size, stream, gpu_index);
// global scratch is only needed by the NOSM kernel variant, when the FFT
// working set does not fit in shared memory; each block uses one
// shared-memory-sized slice of it
double *buffer = full_sm
? nullptr
: static_cast<double *>(cuda_malloc_async(
safe_mul(static_cast<size_t>(number_of_samples),
required_shared_memory_size),
stream, gpu_index));

// configure shared memory for batch fft kernel
if (full_sm) {
Expand All @@ -891,6 +908,7 @@ __host__ void host_fourier_transform_forward_as_torus_f128(
batch_convert_u128_to_f128_as_torus<params>
<<<grid_size, block_size, 0, stream>>>(d_re0, d_re1, d_im0, d_im1,
d_standard);
check_cuda_error(cudaGetLastError());

// call negacyclic 128 bit forward fft.
if (full_sm) {
Expand All @@ -902,21 +920,19 @@ __host__ void host_fourier_transform_forward_as_torus_f128(
<<<grid_size, block_size, shared_memory_size, stream>>>(
d_re0, d_re1, d_im0, d_im1, d_re0, d_re1, d_im0, d_im1, buffer);
}
check_cuda_error(cudaGetLastError());

cuda_memcpy_async_to_cpu(re0, d_re0, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_cpu(re1, d_re1, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_cpu(im0, d_im0, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_cpu(im1, d_im1, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_cpu(re0, d_re0, re_im_size, stream, gpu_index);
cuda_memcpy_async_to_cpu(re1, d_re1, re_im_size, stream, gpu_index);
cuda_memcpy_async_to_cpu(im0, d_im0, re_im_size, stream, gpu_index);
cuda_memcpy_async_to_cpu(im1, d_im1, re_im_size, stream, gpu_index);

cuda_drop_async(d_standard, stream, gpu_index);
cuda_drop_async(d_re0, stream, gpu_index);
cuda_drop_async(d_re1, stream, gpu_index);
cuda_drop_async(d_im0, stream, gpu_index);
cuda_drop_async(d_im1, stream, gpu_index);
cuda_drop_async(buffer, stream, gpu_index);
}

template <class params>
Expand All @@ -925,40 +941,46 @@ __host__ void host_fourier_transform_backward_as_torus_f128(
double const *re0, double const *re1, double const *im0, double const *im1,
const uint32_t N, const uint32_t number_of_samples) {

// allocate device buffers
double *d_re0 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
double *d_re1 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
double *d_im0 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
double *d_im1 = (double *)cuda_malloc_async(safe_mul_sizeof<double>(N / 2),
stream, gpu_index);
__uint128_t *d_standard = (__uint128_t *)cuda_malloc_async(
safe_mul_sizeof<__uint128_t>(N), stream, gpu_index);

// // copy input into device
cuda_memcpy_async_to_gpu(d_re0, re0, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_gpu(d_re1, re1, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_gpu(d_im0, im0, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
cuda_memcpy_async_to_gpu(d_im1, im1, safe_mul_sizeof<double>(N / 2), stream,
gpu_index);
// allocate device buffers; each kernel block processes its own sample, so
// every buffer is sized for the whole batch
size_t re_im_size = safe_mul_sizeof<double>(
static_cast<size_t>(number_of_samples), static_cast<size_t>(N / 2));
size_t standard_size = safe_mul_sizeof<__uint128_t>(
static_cast<size_t>(number_of_samples), static_cast<size_t>(N));
double *d_re0 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
double *d_re1 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
double *d_im0 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
double *d_im1 =
static_cast<double *>(cuda_malloc_async(re_im_size, stream, gpu_index));
__uint128_t *d_standard = static_cast<__uint128_t *>(
cuda_malloc_async(standard_size, stream, gpu_index));

// copy input into device
cuda_memcpy_async_to_gpu(d_re0, re0, re_im_size, stream, gpu_index);
cuda_memcpy_async_to_gpu(d_re1, re1, re_im_size, stream, gpu_index);
cuda_memcpy_async_to_gpu(d_im0, im0, re_im_size, stream, gpu_index);
cuda_memcpy_async_to_gpu(d_im1, im1, re_im_size, stream, gpu_index);

// setup launch parameters
size_t required_shared_memory_size =
safe_mul_sizeof<double>((size_t)(N / 2), (size_t)4);
size_t required_shared_memory_size = safe_mul_sizeof<double>(
static_cast<size_t>(N / 2), static_cast<size_t>(4));
int grid_size = number_of_samples;
int block_size = params::degree / params::opt;
bool full_sm =
(required_shared_memory_size <= cuda_get_max_shared_memory(gpu_index));
size_t buffer_size =
full_sm ? 0
: safe_mul((size_t)number_of_samples, (size_t)(N / 2), (size_t)4);
size_t shared_memory_size = full_sm ? required_shared_memory_size : 0;
double *buffer = (double *)cuda_malloc_async(buffer_size, stream, gpu_index);
// global scratch is only needed by the NOSM kernel variant, when the FFT
// working set does not fit in shared memory; each block uses one
// shared-memory-sized slice of it
double *buffer = full_sm
? nullptr
: static_cast<double *>(cuda_malloc_async(
safe_mul(static_cast<size_t>(number_of_samples),
required_shared_memory_size),
stream, gpu_index));

// configure shared memory for batch fft kernel
if (full_sm) {
Expand All @@ -976,18 +998,21 @@ __host__ void host_fourier_transform_backward_as_torus_f128(
<<<grid_size, block_size, shared_memory_size, stream>>>(
d_re0, d_re1, d_im0, d_im1, d_re0, d_re1, d_im0, d_im1, buffer);
}
check_cuda_error(cudaGetLastError());

batch_convert_f128_to_u128_as_torus<params>
<<<grid_size, block_size, 0, stream>>>(d_standard, d_re0, d_re1, d_im0,
d_im1);
check_cuda_error(cudaGetLastError());

cuda_memcpy_async_to_cpu(standard, d_standard,
safe_mul_sizeof<__uint128_t>(N), stream, gpu_index);
cuda_memcpy_async_to_cpu(standard, d_standard, standard_size, stream,
gpu_index);
cuda_drop_async(d_standard, stream, gpu_index);
cuda_drop_async(d_re0, stream, gpu_index);
cuda_drop_async(d_re1, stream, gpu_index);
cuda_drop_async(d_im0, stream, gpu_index);
cuda_drop_async(d_im1, stream, gpu_index);
cuda_drop_async(buffer, stream, gpu_index);
}

#undef NEG_TWID
Expand Down