github/tfhe-rs
1
1
Fork 0
mirror of https://github.com/zama-ai/tfhe-rs.git synced 2026-01-08 06:13:58 -05:00
Code Issues Packages Projects Releases Wiki Activity

fix(gpu): fix some CPU memory leaks due to the use of new without delete

Browse Source
This commit is contained in:
Agnes Leroy
2025-12-10 17:13:13 +01:00
committed by Agnès Leroy
parent 4f14343668
commit 1f0a83e4bb
4 changed files with 41 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
backends/tfhe-cuda-backend/cuda/src/integer/bitwise_ops.cuh
View File

@@ -102,21 +102,21 @@ __host__ void host_boolean_bitop(CudaStreams streams,
return false;
};
CudaRadixCiphertextFFI *lwe_array_left = new CudaRadixCiphertextFFI;
CudaRadixCiphertextFFI *lwe_array_right = new CudaRadixCiphertextFFI;
CudaRadixCiphertextFFI lwe_array_left;
CudaRadixCiphertextFFI lwe_array_right;
if (needs_noise_reduction(lwe_array_1)) {
copy_radix_ciphertext_slice_async<Torus>(
streams.stream(0), streams.gpu_index(0), mem_ptr->tmp_lwe_left, 0,
lwe_array_1->num_radix_blocks, lwe_array_1, 0,
lwe_array_1->num_radix_blocks);
as_radix_ciphertext_slice<Torus>(lwe_array_left, mem_ptr->tmp_lwe_left, 0,
as_radix_ciphertext_slice<Torus>(&lwe_array_left, mem_ptr->tmp_lwe_left, 0,
lwe_array_1->num_radix_blocks);
integer_radix_apply_univariate_lookup_table<Torus>(
streams, lwe_array_left, lwe_array_left, bsks, ksks,
mem_ptr->message_extract_lut, lwe_array_left->num_radix_blocks);
streams, &lwe_array_left, &lwe_array_left, bsks, ksks,
mem_ptr->message_extract_lut, lwe_array_left.num_radix_blocks);
} else {
as_radix_ciphertext_slice<Torus>(lwe_array_left, lwe_array_1, 0,
as_radix_ciphertext_slice<Torus>(&lwe_array_left, lwe_array_1, 0,
lwe_array_1->num_radix_blocks);
}
@@ -125,37 +125,37 @@ __host__ void host_boolean_bitop(CudaStreams streams,
streams.stream(0), streams.gpu_index(0), mem_ptr->tmp_lwe_right, 0,
lwe_array_2->num_radix_blocks, lwe_array_2, 0,
lwe_array_2->num_radix_blocks);
as_radix_ciphertext_slice<Torus>(lwe_array_right, mem_ptr->tmp_lwe_right, 0,
lwe_array_2->num_radix_blocks);
as_radix_ciphertext_slice<Torus>(&lwe_array_right, mem_ptr->tmp_lwe_right,
0, lwe_array_2->num_radix_blocks);
integer_radix_apply_univariate_lookup_table<Torus>(
streams, lwe_array_right, lwe_array_right, bsks, ksks,
mem_ptr->message_extract_lut, lwe_array_right->num_radix_blocks);
streams, &lwe_array_right, &lwe_array_right, bsks, ksks,
mem_ptr->message_extract_lut, lwe_array_right.num_radix_blocks);
} else {
as_radix_ciphertext_slice<Torus>(lwe_array_right, lwe_array_2, 0,
as_radix_ciphertext_slice<Torus>(&lwe_array_right, lwe_array_2, 0,
lwe_array_2->num_radix_blocks);
}
auto lut = mem_ptr->lut;
uint64_t degrees[lwe_array_left->num_radix_blocks];
uint64_t degrees[lwe_array_left.num_radix_blocks];
if (mem_ptr->op == BITOP_TYPE::BITAND) {
update_degrees_after_bitand(degrees, lwe_array_left->degrees,
lwe_array_right->degrees,
lwe_array_left->num_radix_blocks);
update_degrees_after_bitand(degrees, lwe_array_left.degrees,
lwe_array_right.degrees,
lwe_array_left.num_radix_blocks);
} else if (mem_ptr->op == BITOP_TYPE::BITOR) {
update_degrees_after_bitor(degrees, lwe_array_left->degrees,
lwe_array_right->degrees,
lwe_array_left->num_radix_blocks);
update_degrees_after_bitor(degrees, lwe_array_left.degrees,
lwe_array_right.degrees,
lwe_array_left.num_radix_blocks);
} else if (mem_ptr->op == BITOP_TYPE::BITXOR) {
update_degrees_after_bitxor(degrees, lwe_array_left->degrees,
lwe_array_right->degrees,
lwe_array_left->num_radix_blocks);
update_degrees_after_bitxor(degrees, lwe_array_left.degrees,
lwe_array_right.degrees,
lwe_array_left.num_radix_blocks);
}
// shift argument is hardcoded as 2 here, because natively message modulus for
// boolean block should be 2. lookup table is generated with same factor.
integer_radix_apply_bivariate_lookup_table<Torus>(
streams, lwe_array_out, lwe_array_left, lwe_array_right, bsks, ksks, lut,
lwe_array_out->num_radix_blocks, 2);
streams, lwe_array_out, &lwe_array_left, &lwe_array_right, bsks, ksks,
lut, lwe_array_out->num_radix_blocks, 2);
memcpy(lwe_array_out->degrees, degrees,
lwe_array_out->num_radix_blocks * sizeof(uint64_t));

13
backends/tfhe-cuda-backend/cuda/src/integer/div_rem.cuh
View File

@@ -169,21 +169,21 @@ __host__ void host_unsigned_integer_div_rem_block_by_block_2_2(
CudaRadixCiphertextFFI *comparison_blocks,
CudaRadixCiphertextFFI *d,
int_comparison_buffer<Torus> *comparison_buffer) {
CudaRadixCiphertextFFI *d_msb = new CudaRadixCiphertextFFI;
CudaRadixCiphertextFFI d_msb;
uint32_t slice_start = num_blocks - block_index;
uint32_t slice_end = d->num_radix_blocks;
as_radix_ciphertext_slice<Torus>(d_msb, d, slice_start, slice_end);
comparison_blocks->num_radix_blocks = d_msb->num_radix_blocks;
if (d_msb->num_radix_blocks == 0) {
as_radix_ciphertext_slice<Torus>(&d_msb, d, slice_start, slice_end);
comparison_blocks->num_radix_blocks = d_msb.num_radix_blocks;
if (d_msb.num_radix_blocks == 0) {
cuda_memset_async(
(Torus *)out_boolean_block->ptr, 0,
sizeof(Torus) * (out_boolean_block->lwe_dimension + 1),
streams.stream(gpu_index), streams.gpu_index(gpu_index));
} else {
host_compare_blocks_with_zero<Torus>(
streams.get_ith(gpu_index), comparison_blocks, d_msb,
streams.get_ith(gpu_index), comparison_blocks, &d_msb,
comparison_buffer, &bsks[gpu_index], &ksks[gpu_index],
d_msb->num_radix_blocks, comparison_buffer->is_zero_lut);
d_msb.num_radix_blocks, comparison_buffer->is_zero_lut);
are_all_comparisons_block_true(
streams.get_ith(gpu_index), out_boolean_block, comparison_blocks,
comparison_buffer, &bsks[gpu_index], &ksks[gpu_index],
@@ -202,7 +202,6 @@ __host__ void host_unsigned_integer_div_rem_block_by_block_2_2(
(Torus *)out_boolean_block->ptr, (Torus *)out_boolean_block->ptr,
encoded_scalar, radix_params.big_lwe_dimension, 1);
}
delete d_msb;
};
for (uint j = 0; j < 3; j++) {

14
backends/tfhe-cuda-backend/cuda/src/integer/rerand.cuh
View File

@@ -63,14 +63,14 @@ void rerand_inplace(
// Add ks output to ct
// Check sizes
auto lwes_ffi = new CudaRadixCiphertextFFI;
into_radix_ciphertext(lwes_ffi, lwe_array, num_lwes, output_dimension);
auto ksed_zero_lwes_ffi = new CudaRadixCiphertextFFI;
into_radix_ciphertext(ksed_zero_lwes_ffi, ksed_zero_lwes, num_lwes,
CudaRadixCiphertextFFI lwes_ffi;
into_radix_ciphertext(&lwes_ffi, lwe_array, num_lwes, output_dimension);
CudaRadixCiphertextFFI ksed_zero_lwes_ffi;
into_radix_ciphertext(&ksed_zero_lwes_ffi, ksed_zero_lwes, num_lwes,
output_dimension);
host_addition<Torus>(streams.stream(0), streams.gpu_index(0), lwes_ffi,
lwes_ffi, ksed_zero_lwes_ffi, num_lwes, message_modulus,
carry_modulus);
host_addition<Torus>(streams.stream(0), streams.gpu_index(0), &lwes_ffi,
&lwes_ffi, &ksed_zero_lwes_ffi, num_lwes,
message_modulus, carry_modulus);
}
template <typename Torus>

10
backends/tfhe-cuda-backend/cuda/src/zk/zk.cuh
View File

@@ -91,12 +91,12 @@ __host__ void host_expand_without_verification(
cuda_memset_async(lwe_array_out, 0,
(lwe_dimension + 1) * num_lwes * 2 * sizeof(Torus),
streams.stream(0), streams.gpu_index(0));
auto output = new CudaRadixCiphertextFFI;
into_radix_ciphertext(output, lwe_array_out, 2 * num_lwes, lwe_dimension);
auto input = new CudaRadixCiphertextFFI;
into_radix_ciphertext(input, lwe_array_input, 2 * num_lwes, lwe_dimension);
CudaRadixCiphertextFFI output;
into_radix_ciphertext(&output, lwe_array_out, 2 * num_lwes, lwe_dimension);
CudaRadixCiphertextFFI input;
into_radix_ciphertext(&input, lwe_array_input, 2 * num_lwes, lwe_dimension);
integer_radix_apply_univariate_lookup_table<Torus>(
streams, output, input, bsks, ksks, message_and_carry_extract_luts,
streams, &output, &input, bsks, ksks, message_and_carry_extract_luts,
2 * num_lwes);
}