Files
tfhe-rs/backends/tfhe-cuda-backend/cuda/include/pbs/pbs_utilities.h
2026-01-05 09:48:00 +01:00

539 lines
23 KiB
C++

#ifndef CUDA_BOOTSTRAP_UTILITIES_H
#define CUDA_BOOTSTRAP_UTILITIES_H
#include "device.h"
#include "pbs_enums.h"
#include "vector_types.h"
#include <stdint.h>
template <typename Torus>
uint64_t get_buffer_size_full_sm_programmable_bootstrap_step_one(
uint32_t polynomial_size) {
size_t double_count = (sizeof(Torus) == 16) ? 2 : 1;
return sizeof(Torus) * polynomial_size + // accumulator_rotated
sizeof(double) * 2 * double_count * polynomial_size /
2; // accumulator fft
}
template <typename Torus>
uint64_t get_buffer_size_full_sm_programmable_bootstrap_step_two(
uint32_t polynomial_size) {
size_t double_count = (sizeof(Torus) == 16) ? 2 : 1;
return sizeof(Torus) * polynomial_size + // accumulator
sizeof(double) * 2 * double_count * polynomial_size /
2; // accumulator fft
}
template <typename Torus>
uint64_t
get_buffer_size_partial_sm_programmable_bootstrap(uint32_t polynomial_size) {
size_t double_count = (sizeof(Torus) == 16) ? 2 : 1;
return sizeof(double) * 2 * double_count * polynomial_size /
2; // accumulator fft
}
template <typename Torus>
uint64_t
get_buffer_size_full_sm_programmable_bootstrap_tbc(uint32_t polynomial_size) {
return sizeof(Torus) * polynomial_size + // accumulator_rotated
sizeof(Torus) * polynomial_size + // accumulator
sizeof(double2) * polynomial_size / 2; // accumulator fft
}
template <typename Torus>
uint64_t get_buffer_size_partial_sm_programmable_bootstrap_tbc(
uint32_t polynomial_size) {
return sizeof(double2) * polynomial_size / 2; // accumulator fft mask & body
}
template <typename Torus>
uint64_t get_buffer_size_sm_dsm_plus_tbc_classic_programmable_bootstrap(
uint32_t polynomial_size) {
return sizeof(double2) * polynomial_size / 2; // tbc
}
template <typename Torus>
uint64_t get_buffer_size_full_sm_programmable_bootstrap_tbc_2_2_params(
uint32_t polynomial_size) {
// In the first implementation with 2-2 params, we need up to 5 polynomials in
// shared memory we can optimize this later
return sizeof(Torus) * polynomial_size * 5;
}
template <typename Torus>
uint64_t
get_buffer_size_full_sm_programmable_bootstrap_cg(uint32_t polynomial_size) {
size_t double_count = (sizeof(Torus) == 16) ? 2 : 1;
return sizeof(Torus) * polynomial_size + // accumulator_rotated
sizeof(Torus) * polynomial_size + // accumulator
sizeof(double) * polynomial_size / 2 * 2 *
double_count; // accumulator fft
}
template <typename Torus>
uint64_t
get_buffer_size_partial_sm_programmable_bootstrap_cg(uint32_t polynomial_size) {
size_t double_count = (sizeof(Torus) == 16) ? 2 : 1;
return sizeof(double) * polynomial_size / 2 * 2 *
double_count; // accumulator fft mask & body
}
template <typename Torus>
bool supports_distributed_shared_memory_on_classic_programmable_bootstrap(
uint32_t polynomial_size, uint32_t max_shared_memory);
template <typename Torus, PBS_TYPE pbs_type> struct pbs_buffer;
template <typename Torus> struct pbs_buffer<Torus, PBS_TYPE::CLASSICAL> {
int8_t *d_mem;
Torus *global_accumulator;
double2 *global_join_buffer;
PBS_VARIANT pbs_variant;
PBS_MS_REDUCTION_T noise_reduction_type;
bool gpu_memory_allocated;
pbs_buffer(cudaStream_t stream, uint32_t gpu_index, uint32_t lwe_dimension,
uint32_t glwe_dimension, uint32_t polynomial_size,
uint32_t level_count, uint32_t input_lwe_ciphertext_count,
PBS_VARIANT pbs_variant, bool allocate_gpu_memory,
PBS_MS_REDUCTION_T noise_reduction_type, uint64_t &size_tracker)
: noise_reduction_type(noise_reduction_type) {
gpu_memory_allocated = allocate_gpu_memory;
cuda_set_device(gpu_index);
this->pbs_variant = pbs_variant;
auto max_shared_memory = cuda_get_max_shared_memory(gpu_index);
switch (pbs_variant) {
case PBS_VARIANT::DEFAULT: {
uint64_t full_sm_step_one =
get_buffer_size_full_sm_programmable_bootstrap_step_one<Torus>(
polynomial_size);
uint64_t full_sm_step_two =
get_buffer_size_full_sm_programmable_bootstrap_step_two<Torus>(
polynomial_size);
uint64_t partial_sm =
get_buffer_size_partial_sm_programmable_bootstrap<Torus>(
polynomial_size);
uint64_t partial_dm_step_one = full_sm_step_one - partial_sm;
uint64_t partial_dm_step_two = full_sm_step_two - partial_sm;
uint64_t full_dm = full_sm_step_one;
uint64_t device_mem = 0;
if (max_shared_memory < partial_sm) {
device_mem = full_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
} else if (max_shared_memory < full_sm_step_two) {
device_mem = (partial_dm_step_two + partial_dm_step_one * level_count) *
input_lwe_ciphertext_count * (glwe_dimension + 1);
} else if (max_shared_memory < full_sm_step_one) {
device_mem = partial_dm_step_one * input_lwe_ciphertext_count *
level_count * (glwe_dimension + 1);
}
// Otherwise, both kernels run all in shared memory
d_mem = (int8_t *)cuda_malloc_with_size_tracking_async(
device_mem, stream, gpu_index, size_tracker, allocate_gpu_memory);
global_join_buffer = (double2 *)cuda_malloc_with_size_tracking_async(
(glwe_dimension + 1) * level_count * input_lwe_ciphertext_count *
(polynomial_size / 2) * sizeof(double2),
stream, gpu_index, size_tracker, allocate_gpu_memory);
global_accumulator = (Torus *)cuda_malloc_with_size_tracking_async(
(glwe_dimension + 1) * input_lwe_ciphertext_count * polynomial_size *
sizeof(Torus),
stream, gpu_index, size_tracker, allocate_gpu_memory);
} break;
case PBS_VARIANT::CG: {
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);
uint64_t partial_dm = full_sm - partial_sm;
uint64_t full_dm = full_sm;
uint64_t device_mem = 0;
if (max_shared_memory < partial_sm) {
device_mem = full_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
} else if (max_shared_memory < full_sm) {
device_mem = partial_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
}
// Otherwise, both kernels run all in shared memory
d_mem = (int8_t *)cuda_malloc_with_size_tracking_async(
device_mem, stream, gpu_index, size_tracker, allocate_gpu_memory);
global_join_buffer = (double2 *)cuda_malloc_with_size_tracking_async(
(glwe_dimension + 1) * level_count * input_lwe_ciphertext_count *
polynomial_size / 2 * sizeof(double2),
stream, gpu_index, size_tracker, allocate_gpu_memory);
} break;
#if CUDA_ARCH >= 900
case PBS_VARIANT::TBC: {
bool supports_dsm =
supports_distributed_shared_memory_on_classic_programmable_bootstrap<
Torus>(polynomial_size, max_shared_memory);
uint64_t full_sm =
get_buffer_size_full_sm_programmable_bootstrap_tbc<Torus>(
polynomial_size);
uint64_t partial_sm =
get_buffer_size_partial_sm_programmable_bootstrap_tbc<Torus>(
polynomial_size);
uint64_t minimum_sm_tbc = 0;
if (supports_dsm)
minimum_sm_tbc =
get_buffer_size_sm_dsm_plus_tbc_classic_programmable_bootstrap<
Torus>(polynomial_size);
uint64_t partial_dm = full_sm - partial_sm;
uint64_t full_dm = full_sm;
uint64_t device_mem = 0;
// There is a minimum amount of memory we need to run the TBC PBS, which
// is minimum_sm_tbc. We know that minimum_sm_tbc bytes are available
// because otherwise the previous check would have redirected
// computation to some other variant. If over that we don't have more
// partial_sm bytes, TBC PBS will run on NOSM. If we have partial_sm but
// not full_sm bytes, it will run on PARTIALSM. Otherwise, FULLSM.
//
// NOSM mode actually requires minimum_sm_tbc shared memory bytes.
if (max_shared_memory < partial_sm + minimum_sm_tbc) {
device_mem = full_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
} else if (max_shared_memory < full_sm + minimum_sm_tbc) {
device_mem = partial_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
}
// Otherwise, both kernels run all in shared memory
d_mem = (int8_t *)cuda_malloc_with_size_tracking_async(
device_mem, stream, gpu_index, size_tracker, allocate_gpu_memory);
global_join_buffer = (double2 *)cuda_malloc_with_size_tracking_async(
(glwe_dimension + 1) * level_count * input_lwe_ciphertext_count *
polynomial_size / 2 * sizeof(double2),
stream, gpu_index, size_tracker, allocate_gpu_memory);
} break;
#endif
default:
PANIC("Cuda error (PBS): unsupported implementation variant.")
}
}
void release(cudaStream_t stream, uint32_t gpu_index) {
cuda_drop_with_size_tracking_async(d_mem, stream, gpu_index,
gpu_memory_allocated);
cuda_drop_with_size_tracking_async(global_join_buffer, stream, gpu_index,
gpu_memory_allocated);
if (pbs_variant == DEFAULT)
cuda_drop_with_size_tracking_async(global_accumulator, stream, gpu_index,
gpu_memory_allocated);
}
};
template <typename Torus, PBS_TYPE pbs_type> struct pbs_buffer_128;
template <typename InputTorus>
struct pbs_buffer_128<InputTorus, PBS_TYPE::CLASSICAL> {
int8_t *d_mem;
__uint128_t *global_accumulator;
double *global_join_buffer;
PBS_VARIANT pbs_variant;
PBS_MS_REDUCTION_T noise_reduction_type;
bool gpu_memory_allocated;
pbs_buffer_128(cudaStream_t stream, uint32_t gpu_index,
uint32_t lwe_dimension, uint32_t glwe_dimension,
uint32_t polynomial_size, uint32_t level_count,
uint32_t input_lwe_ciphertext_count, PBS_VARIANT pbs_variant,
bool allocate_gpu_memory,
PBS_MS_REDUCTION_T noise_reduction_type,
uint64_t &size_tracker)
: noise_reduction_type(noise_reduction_type) {
gpu_memory_allocated = allocate_gpu_memory;
cuda_set_device(gpu_index);
this->pbs_variant = pbs_variant;
auto max_shared_memory = cuda_get_max_shared_memory(gpu_index);
size_t global_join_buffer_size = (glwe_dimension + 1) * level_count *
input_lwe_ciphertext_count *
polynomial_size / 2 * sizeof(double) * 4;
switch (pbs_variant) {
case PBS_VARIANT::DEFAULT: {
uint64_t full_sm_step_one =
get_buffer_size_full_sm_programmable_bootstrap_step_one<__uint128_t>(
polynomial_size);
uint64_t full_sm_step_two =
get_buffer_size_full_sm_programmable_bootstrap_step_two<__uint128_t>(
polynomial_size);
uint64_t partial_sm =
get_buffer_size_partial_sm_programmable_bootstrap<__uint128_t>(
polynomial_size);
uint64_t partial_dm_step_one = full_sm_step_one - partial_sm;
uint64_t partial_dm_step_two = full_sm_step_two - partial_sm;
uint64_t full_dm = full_sm_step_one;
uint64_t device_mem = 0;
if (max_shared_memory < partial_sm) {
device_mem = full_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
} else if (max_shared_memory < full_sm_step_two) {
device_mem = (partial_dm_step_two + partial_dm_step_one * level_count) *
input_lwe_ciphertext_count * (glwe_dimension + 1);
} else if (max_shared_memory < full_sm_step_one) {
device_mem = partial_dm_step_one * input_lwe_ciphertext_count *
level_count * (glwe_dimension + 1);
}
// Otherwise, both kernels run all in shared memory
d_mem = (int8_t *)cuda_malloc_with_size_tracking_async(
device_mem, stream, gpu_index, size_tracker, allocate_gpu_memory);
global_join_buffer = (double *)cuda_malloc_with_size_tracking_async(
global_join_buffer_size, stream, gpu_index, size_tracker,
allocate_gpu_memory);
global_accumulator = (__uint128_t *)cuda_malloc_with_size_tracking_async(
(glwe_dimension + 1) * input_lwe_ciphertext_count * polynomial_size *
sizeof(__uint128_t),
stream, gpu_index, size_tracker, allocate_gpu_memory);
} break;
case PBS_VARIANT::CG: {
uint64_t full_sm =
get_buffer_size_full_sm_programmable_bootstrap_cg<__uint128_t>(
polynomial_size);
uint64_t partial_sm =
get_buffer_size_partial_sm_programmable_bootstrap_cg<__uint128_t>(
polynomial_size);
uint64_t partial_dm = full_sm - partial_sm;
uint64_t full_dm = full_sm;
uint64_t device_mem = 0;
if (max_shared_memory < partial_sm) {
device_mem = full_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
} else if (max_shared_memory < full_sm) {
device_mem = partial_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
}
// Otherwise, both kernels run all in shared memory
d_mem = (int8_t *)cuda_malloc_with_size_tracking_async(
device_mem, stream, gpu_index, size_tracker, allocate_gpu_memory);
global_join_buffer = (double *)cuda_malloc_with_size_tracking_async(
global_join_buffer_size, stream, gpu_index, size_tracker,
allocate_gpu_memory);
} break;
#if CUDA_ARCH >= 900
case PBS_VARIANT::TBC: {
bool supports_dsm =
supports_distributed_shared_memory_on_classic_programmable_bootstrap<
__uint128_t>(polynomial_size, max_shared_memory);
uint64_t full_sm =
get_buffer_size_full_sm_programmable_bootstrap_tbc<__uint128_t>(
polynomial_size);
uint64_t partial_sm =
get_buffer_size_partial_sm_programmable_bootstrap_tbc<__uint128_t>(
polynomial_size);
uint64_t minimum_sm_tbc = 0;
if (supports_dsm)
minimum_sm_tbc =
get_buffer_size_sm_dsm_plus_tbc_classic_programmable_bootstrap<
__uint128_t>(polynomial_size);
uint64_t partial_dm = full_sm - partial_sm;
uint64_t full_dm = full_sm;
uint64_t device_mem = 0;
// There is a minimum amount of memory we need to run the TBC PBS, which
// is minimum_sm_tbc. We know that minimum_sm_tbc bytes are available
// because otherwise the previous check would have redirected
// computation to some other variant. If over that we don't have more
// partial_sm bytes, TBC PBS will run on NOSM. If we have partial_sm but
// not full_sm bytes, it will run on PARTIALSM. Otherwise, FULLSM.
//
// NOSM mode actually requires minimum_sm_tbc shared memory bytes.
if (max_shared_memory < partial_sm + minimum_sm_tbc) {
device_mem = full_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
} else if (max_shared_memory < full_sm + minimum_sm_tbc) {
device_mem = partial_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
}
// Otherwise, both kernels run all in shared memory
d_mem = (int8_t *)cuda_malloc_with_size_tracking_async(
device_mem, stream, gpu_index, size_tracker, allocate_gpu_memory);
global_join_buffer = (double *)cuda_malloc_with_size_tracking_async(
global_join_buffer_size, stream, gpu_index, size_tracker,
allocate_gpu_memory);
} break;
#endif
default:
PANIC("Cuda error (PBS): unsupported implementation variant.")
}
}
void release(cudaStream_t stream, uint32_t gpu_index) {
cuda_drop_with_size_tracking_async(d_mem, stream, gpu_index,
gpu_memory_allocated);
cuda_drop_with_size_tracking_async(global_join_buffer, stream, gpu_index,
gpu_memory_allocated);
if (pbs_variant == DEFAULT)
cuda_drop_with_size_tracking_async(global_accumulator, stream, gpu_index,
gpu_memory_allocated);
}
};
template <typename Torus>
uint64_t get_buffer_size_programmable_bootstrap_cg(
uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
uint32_t input_lwe_ciphertext_count, uint32_t max_shared_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);
uint64_t partial_dm = full_sm - partial_sm;
uint64_t full_dm = full_sm;
uint64_t device_mem = 0;
if (max_shared_memory < partial_sm) {
device_mem = full_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
} else if (max_shared_memory < full_sm) {
device_mem = partial_dm * input_lwe_ciphertext_count * level_count *
(glwe_dimension + 1);
}
uint64_t buffer_size = device_mem + (glwe_dimension + 1) * level_count *
input_lwe_ciphertext_count *
polynomial_size / 2 * sizeof(double2);
return buffer_size + buffer_size % sizeof(double2);
}
template <typename InputTorus, typename Torus>
bool has_support_to_cuda_programmable_bootstrap_cg(uint32_t glwe_dimension,
uint32_t polynomial_size,
uint32_t level_count,
uint32_t num_samples,
uint32_t max_shared_memory);
template <typename InputTorus, typename Torus>
void cuda_programmable_bootstrap_cg_lwe_ciphertext_vector(
void *stream, uint32_t gpu_index, Torus *lwe_array_out,
Torus const *lwe_output_indexes, Torus const *lut_vector,
Torus const *lut_vector_indexes, InputTorus const *lwe_array_in,
Torus const *lwe_input_indexes, double2 const *bootstrapping_key,
pbs_buffer<Torus, CLASSICAL> *buffer, uint32_t lwe_dimension,
uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t base_log,
uint32_t level_count, uint32_t num_samples, uint32_t num_many_lut,
uint32_t lut_stride);
template <typename InputTorus, typename Torus>
void cuda_programmable_bootstrap_lwe_ciphertext_vector(
void *stream, uint32_t gpu_index, Torus *lwe_array_out,
Torus const *lwe_output_indexes, Torus const *lut_vector,
Torus const *lut_vector_indexes, InputTorus const *lwe_array_in,
Torus const *lwe_input_indexes, double2 const *bootstrapping_key,
pbs_buffer<Torus, CLASSICAL> *buffer, uint32_t lwe_dimension,
uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t base_log,
uint32_t level_count, uint32_t num_samples, uint32_t num_many_lut,
uint32_t lut_stride);
#if (CUDA_ARCH >= 900)
template <typename InputTorus, typename Torus>
void cuda_programmable_bootstrap_tbc_lwe_ciphertext_vector(
void *stream, uint32_t gpu_index, Torus *lwe_array_out,
Torus const *lwe_output_indexes, Torus const *lut_vector,
Torus const *lut_vector_indexes, InputTorus const *lwe_array_in,
Torus const *lwe_input_indexes, double2 const *bootstrapping_key,
pbs_buffer<Torus, CLASSICAL> *buffer, uint32_t lwe_dimension,
uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t base_log,
uint32_t level_count, uint32_t num_samples, uint32_t num_many_lut,
uint32_t lut_stride);
template <typename Torus>
uint64_t scratch_cuda_programmable_bootstrap_tbc(
void *stream, uint32_t gpu_index, pbs_buffer<Torus, CLASSICAL> **pbs_buffer,
uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
uint32_t level_count, uint32_t input_lwe_ciphertext_count,
bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type);
#endif
template <typename InputTorus, typename Torus>
uint64_t scratch_cuda_programmable_bootstrap_cg(
void *stream, uint32_t gpu_index, pbs_buffer<Torus, CLASSICAL> **pbs_buffer,
uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
uint32_t level_count, uint32_t input_lwe_ciphertext_count,
bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type);
template <typename InputTorus, typename Torus>
uint64_t scratch_cuda_programmable_bootstrap(
void *stream, uint32_t gpu_index, pbs_buffer<Torus, CLASSICAL> **buffer,
uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
uint32_t level_count, uint32_t input_lwe_ciphertext_count,
bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type);
template <typename Torus>
bool has_support_to_cuda_programmable_bootstrap_tbc(uint32_t num_samples,
uint32_t glwe_dimension,
uint32_t polynomial_size,
uint32_t level_count,
uint32_t max_shared_memory);
bool has_support_to_cuda_programmable_bootstrap_128_cg(
uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
uint32_t num_samples, uint32_t max_shared_memory);
#ifdef __CUDACC__
__device__ inline int get_start_ith_ggsw(int i, uint32_t polynomial_size,
int glwe_dimension,
uint32_t level_count);
template <typename T>
__device__ const T *get_ith_mask_kth_block(const T *ptr, int i, int k,
int level, uint32_t polynomial_size,
int glwe_dimension,
uint32_t level_count);
template <typename T>
__device__ T *get_ith_mask_kth_block(T *ptr, int i, int k, int level,
uint32_t polynomial_size,
int glwe_dimension, uint32_t level_count);
template <typename T, uint32_t polynomial_size, uint32_t glwe_dimension,
uint32_t level_count, uint32_t level_id>
__device__ const T *get_ith_mask_kth_block_2_2_params(const T *ptr,
int iteration, int k);
template <typename T>
__device__ T *get_ith_body_kth_block(T *ptr, int i, int k, int level,
uint32_t polynomial_size,
int glwe_dimension, uint32_t level_count);
template <typename T>
__device__ const T *get_multi_bit_ith_lwe_gth_group_kth_block(
const T *ptr, int g, int i, int k, int level, uint32_t grouping_factor,
uint32_t polynomial_size, uint32_t glwe_dimension, uint32_t level_count);
#endif
#endif // CUDA_BOOTSTRAP_UTILITIES_H