diff --git a/test/test_csprng.py b/test/test_csprng.py index b4c5443..21f0642 100644 --- a/test/test_csprng.py +++ b/test/test_csprng.py @@ -354,5 +354,51 @@ def test_const_generator(self): second = torch.empty(self.size, dtype=dtype, device=device).random_(generator=const_gen) self.assertTrue((first - second).max().abs() == 0) + def test_encrypt_decrypt(self): + key_size_bytes = 16 + block_size_bytes = 16 + + def sizeof(dtype): + if dtype == torch.bool: + return 1 + elif dtype.is_floating_point: + return torch.finfo(dtype).bits // 8 + else: + return torch.iinfo(dtype).bits // 8 + + for device in self.all_devices: + for key_dtype in self.all_dtypes: + key_size = key_size_bytes // sizeof(key_dtype) + key = torch.empty(key_size, dtype=key_dtype, device=device).random_() + for initial_dtype in self.all_dtypes: + for encrypted_dtype in self.all_dtypes: + for decrypted_dtype in self.all_dtypes: + for initial_size in [0, 4, 8, 15, 16, 23, 42]: + for mode in ["ecb", "ctr"]: + encrypted_size = (initial_size * sizeof(initial_dtype) + block_size_bytes - 1) // block_size_bytes * block_size_bytes // sizeof(encrypted_dtype) + decrypted_size = (encrypted_size * sizeof(encrypted_dtype) + block_size_bytes - 1) // block_size_bytes * block_size_bytes // sizeof(decrypted_dtype) + + initial = torch.empty(initial_size, dtype=initial_dtype, device=device).random_() + encrypted = torch.empty(encrypted_size, dtype=encrypted_dtype, device=device).random_() + decrypted = torch.empty(decrypted_size, dtype=decrypted_dtype, device=device).random_() + + initial_np = initial.cpu().numpy().view(np.int8) + decrypted_np = decrypted.cpu().numpy().view(np.int8) + padding_size_bytes = initial_size * sizeof(initial_dtype) - decrypted_size * sizeof(decrypted_dtype) + if padding_size_bytes != 0: + decrypted_np = decrypted_np[:padding_size_bytes] + + csprng.encrypt(initial, encrypted, key, "aes128", mode) + + if initial_size > 8: + self.assertFalse(np.array_equal(initial_np, decrypted_np)) + + csprng.decrypt(encrypted, decrypted, key, "aes128", mode) + decrypted_np = decrypted.cpu().numpy().view(np.int8) + if padding_size_bytes != 0: + decrypted_np = decrypted_np[:padding_size_bytes] + + self.assertTrue(np.array_equal(initial_np, decrypted_np)) + if __name__ == '__main__': unittest.main() diff --git a/torchcsprng/csrc/aes.h b/torchcsprng/csrc/aes.h index 09596f1..c4ece07 100644 --- a/torchcsprng/csrc/aes.h +++ b/torchcsprng/csrc/aes.h @@ -63,15 +63,7 @@ namespace aes { #define Nr 10 // The number of rounds in AES Cipher. #endif -#if !defined(__CUDACC__) && !defined(__HIPCC__) -struct ulonglong2 // TODO: should have something like `__builtin_align__(16)` -{ - unsigned long long int x, y; -}; -#endif - -typedef ulonglong2 block_t; -constexpr size_t block_t_size = sizeof(block_t); +constexpr size_t block_t_size = 16; typedef uint8_t state_t[4][4]; @@ -97,6 +89,24 @@ TORCH_CSPRNG_CONSTANT const uint8_t sbox[256] = { 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94, 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf, 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68, 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16 }; +TORCH_CSPRNG_CONSTANT const uint8_t rsbox[256] = { + 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38, 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb, + 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87, 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb, + 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d, 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e, + 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2, 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25, + 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16, 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92, + 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda, 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84, + 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a, 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06, + 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02, 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b, + 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea, 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73, + 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85, 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e, + 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89, 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b, + 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20, 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4, + 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31, 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f, + 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d, 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef, + 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0, 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61, + 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26, 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d }; + // The round constant word array, Rcon[i], contains the values given by // x to the power (i-1) being powers of x (x is denoted as {02}) in the field GF(2^8) TORCH_CSPRNG_CONSTANT const uint8_t Rcon[11] = { @@ -104,6 +114,8 @@ TORCH_CSPRNG_CONSTANT const uint8_t Rcon[11] = { #define getSBoxValue(num) (sbox[(num)]) +#define getSBoxInvert(num) (rsbox[(num)]) + // This function produces Nb(Nr+1) round keys. The round keys are used in each round to decrypt the states. TORCH_CSPRNG_HOST_DEVICE void KeyExpansion(uint8_t* RoundKey, const uint8_t* Key){ unsigned int i, j, k; @@ -257,6 +269,78 @@ TORCH_CSPRNG_HOST_DEVICE void MixColumns(state_t* state) } } +TORCH_CSPRNG_HOST_DEVICE uint8_t Multiply(uint8_t x, uint8_t y) +{ + return (((y & 1) * x) ^ + ((y>>1 & 1) * xtime(x)) ^ + ((y>>2 & 1) * xtime(xtime(x))) ^ + ((y>>3 & 1) * xtime(xtime(xtime(x)))) ^ + ((y>>4 & 1) * xtime(xtime(xtime(xtime(x)))))); /* this last call to xtime() can be omitted */ +} + +// MixColumns function mixes the columns of the state matrix. +// The method used to multiply may be difficult to understand for the inexperienced. +// Please use the references to gain more information. +TORCH_CSPRNG_HOST_DEVICE void InvMixColumns(state_t* state) +{ + int i; + uint8_t a, b, c, d; + for (i = 0; i < 4; ++i) + { + a = (*state)[i][0]; + b = (*state)[i][1]; + c = (*state)[i][2]; + d = (*state)[i][3]; + + (*state)[i][0] = Multiply(a, 0x0e) ^ Multiply(b, 0x0b) ^ Multiply(c, 0x0d) ^ Multiply(d, 0x09); + (*state)[i][1] = Multiply(a, 0x09) ^ Multiply(b, 0x0e) ^ Multiply(c, 0x0b) ^ Multiply(d, 0x0d); + (*state)[i][2] = Multiply(a, 0x0d) ^ Multiply(b, 0x09) ^ Multiply(c, 0x0e) ^ Multiply(d, 0x0b); + (*state)[i][3] = Multiply(a, 0x0b) ^ Multiply(b, 0x0d) ^ Multiply(c, 0x09) ^ Multiply(d, 0x0e); + } +} + +// The SubBytes Function Substitutes the values in the +// state matrix with values in an S-box. +TORCH_CSPRNG_HOST_DEVICE void InvSubBytes(state_t* state) +{ + uint8_t i, j; + for (i = 0; i < 4; ++i) + { + for (j = 0; j < 4; ++j) + { + (*state)[j][i] = getSBoxInvert((*state)[j][i]); + } + } +} + +TORCH_CSPRNG_HOST_DEVICE void InvShiftRows(state_t* state) +{ + uint8_t temp; + + // Rotate first row 1 columns to right + temp = (*state)[3][1]; + (*state)[3][1] = (*state)[2][1]; + (*state)[2][1] = (*state)[1][1]; + (*state)[1][1] = (*state)[0][1]; + (*state)[0][1] = temp; + + // Rotate second row 2 columns to right + temp = (*state)[0][2]; + (*state)[0][2] = (*state)[2][2]; + (*state)[2][2] = temp; + + temp = (*state)[1][2]; + (*state)[1][2] = (*state)[3][2]; + (*state)[3][2] = temp; + + // Rotate third row 3 columns to right + temp = (*state)[0][3]; + (*state)[0][3] = (*state)[1][3]; + (*state)[1][3] = (*state)[2][3]; + (*state)[2][3] = (*state)[3][3]; + (*state)[3][3] = temp; +} + TORCH_CSPRNG_HOST_DEVICE void encrypt(uint8_t* state, const uint8_t* key) { uint8_t RoundKey[176]; KeyExpansion(RoundKey, key); @@ -284,4 +368,29 @@ TORCH_CSPRNG_HOST_DEVICE void encrypt(uint8_t* state, const uint8_t* key) { AddRoundKey(Nr, (state_t*)state, RoundKey); } +TORCH_CSPRNG_HOST_DEVICE void decrypt(uint8_t* state, const uint8_t* key) { + uint8_t RoundKey[176]; + KeyExpansion(RoundKey, key); + + uint8_t round = 0; + + // Add the First round key to the state before starting the rounds. + AddRoundKey(Nr, (state_t*)state, RoundKey); + + // There will be Nr rounds. + // The first Nr-1 rounds are identical. + // These Nr rounds are executed in the loop below. + // Last one without InvMixColumn() + for (round = (Nr - 1); ; --round) + { + InvShiftRows((state_t*)state); + InvSubBytes((state_t*)state); + AddRoundKey(round, (state_t*)state, RoundKey); + if (round == 0) { + break; + } + InvMixColumns((state_t*)state); + } +} + }}} diff --git a/torchcsprng/csrc/block_cipher.h b/torchcsprng/csrc/block_cipher.h index 2ca313f..8418e3d 100644 --- a/torchcsprng/csrc/block_cipher.h +++ b/torchcsprng/csrc/block_cipher.h @@ -30,144 +30,173 @@ namespace torch { namespace csprng { -// Generates `block_t_size`-bytes random key Tensor on CPU -// using `generator`, which must be an instance of `at::CPUGeneratorImpl` -// and passes it to the `device`. -template -at::Tensor key_tensor(size_t block_t_size, c10::optional generator) { - std::lock_guard lock(generator->mutex()); - auto gen = at::check_generator(generator); - if (gen->key().defined()) { - return gen->key().clone(); - } - auto t = torch::empty({static_cast(block_t_size)}, torch::kUInt8); - using random_t = uint32_t; - constexpr size_t random_t_size = sizeof(random_t); - for (size_t i = 0; i < block_t_size / random_t_size; i++) { - const auto rand = gen->random(); - for (size_t j = 0; j < random_t_size; j++) { - size_t k = i * random_t_size + j; - t[k] = static_cast((rand >> (j * 8)) & 0xff); +template +TORCH_CSPRNG_HOST_DEVICE static void copy_input_to_block(int64_t idx, uint8_t* block, int block_size, + void* input_ptr, int64_t input_numel, int input_type_size, input_index_calc_t input_index_calc) { + for (auto i = 0; i < block_size / input_type_size; ++i) { + const auto linear_index = idx * (block_size / input_type_size) + i; + if (linear_index < input_numel) { + std::memcpy( + block + i * input_type_size, + &(reinterpret_cast(input_ptr)[input_index_calc(linear_index)]), + input_type_size + ); } } - return t; } -// A simple container for random state sub-blocks that implements RNG interface -// with random() and random64() methods, that are used by transformation function -template -struct RNGValues { - TORCH_CSPRNG_HOST_DEVICE RNGValues(uint64_t* vals) { - memcpy(&vals_, vals, size * sizeof(uint64_t)); - } - uint32_t TORCH_CSPRNG_HOST_DEVICE random() { auto res = static_cast(vals_[index]); index++; return res; } - uint64_t TORCH_CSPRNG_HOST_DEVICE random64() { auto res = vals_[index]; index++; return res; } -private: - uint64_t vals_[size]; - int index = 0; -}; - -// Runs a block cipher in a counter mode in approximately `numel / (block_t_size / sizeof(uint_t) / N)` CUDA threads, -// without any assumption about target tensor layout. It uses `index_calc` to find memory locations of -// the tensor elements. -// `scalar_t` is a scalar type equivalent of target tensor dtype -// `uint_t` is an unsigned integral type of sub-blocks that random state is divided to -// (e.g, 16 bytes random state block can be divided into 16 uint8_t sub-blocks -// or 8 uint16_t sub-block or 4 uint32_t sub-block or 2 uint64_t sub-blocks) -// `N` is a number of sub-block which is used by `transform_func` -// to generate a random value of specific distribution (e.g. `normal` uses 2) -// `numel` is a number of elements in target tensor -// `block_t_size` is a number of bytes in cipher's block (e.g. 16 for AES128) -// `cipher` is a callable that receives a counter `idx` and returns an encrypted block -// `transform_func` is a callable that converts N `uint_t` random state sub-blocks passed in RNGValues into target dtype `scalar_t` -template -TORCH_CSPRNG_HOST_DEVICE static void block_cipher_kernel_helper(int idx, scalar_t* data, int64_t numel, size_t block_t_size, cipher_t cipher, transform_t transform_func, index_calc_t index_calc) { - const int unroll_factor = block_t_size / sizeof(uint_t) / N; - if (unroll_factor * idx < numel) { - auto block = cipher(idx); - UNROLL_IF_CUDA - for (auto i = 0; i < unroll_factor; ++i) { - const auto li = unroll_factor * idx + i; - if (li < numel) { - uint64_t vals[N]; - UNROLL_IF_CUDA - for (size_t j = 0; j < N; j++) { - vals[j] = (reinterpret_cast(&block))[N * i + j]; - } - RNGValues rng(vals); - data[index_calc(li)] = transform_func(&rng); - } +template +TORCH_CSPRNG_HOST_DEVICE static void copy_block_to_output(int64_t idx, uint8_t* block, int output_elem_per_block, + void* output_ptr, int64_t output_numel, int output_type_size, output_index_calc_t output_index_calc) { + for (auto i = 0; i < output_elem_per_block; ++i) { + const auto linear_index = idx * output_elem_per_block + i; + if (linear_index < output_numel) { + std::memcpy( + &(reinterpret_cast(output_ptr)[output_index_calc(linear_index)]), + block + i * output_type_size, + output_type_size + ); } } } +template +TORCH_CSPRNG_HOST_DEVICE static void block_cipher_kernel_helper( + int64_t idx, cipher_t cipher, int output_elem_per_block, + void* input_ptr, int64_t input_numel, int input_type_size, input_index_calc_t input_index_calc, + void* output_ptr, int64_t output_numel, int output_type_size, output_index_calc_t output_index_calc, + transform_t transform) { + uint8_t block[block_size]; + std::memset(&block, 0, block_size); // is it ok to use zeros as padding? + if (input_ptr != nullptr) { + copy_input_to_block(idx, block, block_size, input_ptr, input_numel, input_type_size, input_index_calc); + } + cipher(idx, block); + transform(block); + copy_block_to_output(idx, block, output_elem_per_block, output_ptr, output_numel, output_type_size, output_index_calc); +} + #if defined(__CUDACC__) || defined(__HIPCC__) -template -__global__ static void block_cipher_kernel_cuda(scalar_t* data, int64_t numel, int block_t_size, cipher_t cipher, transform_t transform_func, index_calc_t index_calc) { +template +__global__ static void block_cipher_kernel_cuda(cipher_t cipher, int output_elem_per_block, + void* input_ptr, int64_t input_numel, int input_type_size, input_index_calc_t input_index_calc, + void* output_ptr, int64_t output_numel, int output_type_size, output_index_calc_t output_index_calc, + transform_t transform) { const auto idx = blockIdx.x * blockDim.x + threadIdx.x; - block_cipher_kernel_helper(idx, data, numel, block_t_size, cipher, transform_func, index_calc); + block_cipher_kernel_helper(idx, cipher, output_elem_per_block, + input_ptr, input_numel, input_type_size, input_index_calc, + output_ptr, output_numel, output_type_size, output_index_calc, + transform); } #endif -template -static void block_cipher_kernel_cpu_serial(int64_t begin, int64_t end, scalar_t* data, int64_t numel, int block_t_size, cipher_t cipher, transform_t transform_func, index_calc_t index_calc) { +template +static void block_cipher_kernel_cpu_serial(int64_t begin, int64_t end, cipher_t cipher, int output_elem_per_block, + void* input_ptr, int64_t input_numel, int input_type_size, input_index_calc_t input_index_calc, + void* output_ptr, int64_t output_numel, int output_type_size, output_index_calc_t output_index_calc, + transform_t transform) { for (auto idx = begin; idx < end; ++idx) { - block_cipher_kernel_helper(idx, data, numel, block_t_size, cipher, transform_func, index_calc); + block_cipher_kernel_helper(idx, cipher, output_elem_per_block, + input_ptr, input_numel, input_type_size, input_index_calc, + output_ptr, output_numel, output_type_size, output_index_calc, + transform); } } -template -static void block_cipher_kernel_cpu(int64_t total, scalar_t* data, int64_t numel, int block_t_size, cipher_t cipher, transform_t transform_func, index_calc_t index_calc) { +template +static void block_cipher_kernel_cpu(int64_t total, cipher_t cipher, int output_elem_per_block, + void* input_ptr, int64_t input_numel, int input_type_size, input_index_calc_t input_index_calc, + void* output_ptr, int64_t output_numel, int output_type_size, output_index_calc_t output_index_calc, + transform_t transform_func) { if (total < at::internal::GRAIN_SIZE || at::get_num_threads() == 1) { - block_cipher_kernel_cpu_serial(0, total, data, numel, block_t_size, cipher, transform_func, index_calc); + block_cipher_kernel_cpu_serial(0, total, cipher, output_elem_per_block, + input_ptr, input_numel, input_type_size, input_index_calc, + output_ptr, output_numel, output_type_size, output_index_calc, + transform_func); } else { at::parallel_for(0, total, at::internal::GRAIN_SIZE, [&](int64_t begin, int64_t end) { - block_cipher_kernel_cpu_serial(begin, end, data, numel, block_t_size, cipher, transform_func, index_calc); + block_cipher_kernel_cpu_serial(begin, end, cipher, output_elem_per_block, + input_ptr, input_numel, input_type_size, input_index_calc, + output_ptr, output_numel, output_type_size, output_index_calc, + transform_func); }); } } -// Runs a block cipher in a counter mode in approximately `numel / (block_t_size / sizeof(uint_t) / N)` CUDA threads. -// Each CUDA thread generates `block_t_size`-bytes random state and divides it into `block_t_size / sizeof(uint_t)` sub-blocks. -// Then `transform_func` transforms `N` random state sub-blocks passed in a `RNGValues` to final random values of type `scalar_t`. -template -void block_cipher_ctr_mode(at::TensorIterator& iter, int block_t_size, cipher_t cipher, transform_t transform_func) { - const auto numel = iter.numel(); - if (numel == 0) { +template +void block_cipher( + void* input_ptr, int64_t input_numel, int input_type_size, input_index_calc_t input_index_calc, + void* output_ptr, int64_t output_numel, int output_type_size, output_index_calc_t output_index_calc, + Device device, cipher_t cipher, int output_elem_per_block, transform_t transform_func) { + if (output_ptr == nullptr || output_numel == 0) { return; } - const int unroll_factor = block_t_size / sizeof(uint_t) / N; - const auto block = 256; - const auto grid = (numel + (block * unroll_factor) - 1) / (block * unroll_factor); - scalar_t* data = (scalar_t*)iter.data_ptr(0); - auto offset_calc = make_offset_calculator<1>(iter); - auto index_calc_identity = [] TORCH_CSPRNG_HOST_DEVICE (int li) -> int { return li; }; - auto index_calc_offset = [offset_calc] TORCH_CSPRNG_HOST_DEVICE (int li) -> int { return offset_calc.get(li)[0] / sizeof(scalar_t); }; - if (iter.device_type() == at::kCPU) { - if (iter.output(0).is_contiguous()) { - block_cipher_kernel_cpu( - grid * block, data, numel, block_t_size, cipher, transform_func, index_calc_identity); - } else { - block_cipher_kernel_cpu( - grid * block, data, numel, block_t_size, cipher, transform_func, index_calc_offset); - } - } else if (iter.device_type() == at::kCUDA) { + + if (device.type() == at::kCPU) { + const auto total = (output_numel + output_elem_per_block - 1) / output_elem_per_block; + block_cipher_kernel_cpu(total, + cipher, output_elem_per_block, + input_ptr, input_numel, input_type_size, input_index_calc, + output_ptr, output_numel, output_type_size, output_index_calc, + transform_func + ); + } else if (device.type() == at::kCUDA) { #if defined(__CUDACC__) || defined(__HIPCC__) + const auto threads = 256; + const auto grid = (output_numel + (threads * output_elem_per_block) - 1) / (threads * output_elem_per_block); auto stream = at::cuda::getCurrentCUDAStream(); - if (iter.output(0).is_contiguous()) { - block_cipher_kernel_cuda<<>>( - data, numel, block_t_size, cipher, transform_func, index_calc_identity); - } else { - block_cipher_kernel_cuda<<>>( - data, numel, block_t_size, cipher, transform_func, index_calc_offset); - } + block_cipher_kernel_cuda<<>>( + cipher, output_elem_per_block, + input_ptr, input_numel, input_type_size, input_index_calc, + output_ptr, output_numel, output_type_size, output_index_calc, + transform_func + ); AT_CUDA_CHECK(cudaGetLastError()); #else - TORCH_CHECK(false, "csprng was compiled without CUDA support"); + TORCH_CHECK(false, "torchcsprng was compiled without CUDA support"); #endif } else { - TORCH_CHECK(false, "block_cipher_ctr_mode supports only CPU and CUDA devices"); + TORCH_CHECK(false, "block_cipher supports only CPU and CUDA devices"); } } +template +void block_cipher(Tensor input, Tensor output, cipher_t cipher) { + const auto input_ptr = input.data_ptr(); + const auto input_numel = input.numel(); + + // Otherwise OffsetCalculator/IntDivider crashes with integer division by zero + if (input_ptr == nullptr || input_numel == 0) { + return; + } + + const auto input_type_size = input.element_size(); + const auto input_offset_calc = make_offset_calculator<1>(TensorIterator::nullary_op(input)); + const auto input_index_calc = [input_offset_calc] TORCH_CSPRNG_HOST_DEVICE (uint32_t li) -> uint32_t { + return input_offset_calc.get(li)[0]; + }; + + const auto output_ptr = output.data_ptr(); + const auto output_numel = output.numel(); + + // Otherwise OffsetCalculator/IntDivider crashes with integer division by zero + if (output_ptr == nullptr || output_numel == 0) { + return; + } + + const auto output_type_size = output.element_size(); + const auto output_offset_calc = make_offset_calculator<1>(TensorIterator::nullary_op(output)); + const auto output_index_calc = [output_offset_calc] TORCH_CSPRNG_HOST_DEVICE (uint32_t li) -> uint32_t { + return output_offset_calc.get(li)[0]; + }; + + const auto device = output.device(); + + block_cipher( + input_ptr, input_numel, input_type_size, input_index_calc, + output_ptr, output_numel, output_type_size, output_index_calc, + device, cipher, block_size / output_type_size, + [] TORCH_CSPRNG_HOST_DEVICE (uint8_t* x) {}); +} + }} diff --git a/torchcsprng/csrc/csprng.h b/torchcsprng/csrc/csprng.h index f1fa0f9..cf31c10 100644 --- a/torchcsprng/csrc/csprng.h +++ b/torchcsprng/csrc/csprng.h @@ -57,6 +57,29 @@ struct CSPRNGGeneratorImpl : public c10::GeneratorImpl { Tensor key_; }; +// Generates `block_t_size`-bytes random key Tensor on CPU +// using `generator`, which must be an instance of `at::CPUGeneratorImpl` +// and passes it to the `device`. +template +at::Tensor key_tensor(size_t block_t_size, c10::optional generator) { + std::lock_guard lock(generator->mutex()); + auto gen = at::check_generator(generator); + if (gen->key().defined()) { + return gen->key().clone(); + } + auto key = torch::empty({static_cast(block_t_size)}, torch::kUInt8); + using random_t = typename std::result_of::type; + constexpr size_t random_t_size = sizeof(random_t); + for (size_t i = 0; i < block_t_size / random_t_size; i++) { + const auto rand = gen->random(); + for (size_t j = 0; j < random_t_size; j++) { + size_t k = i * random_t_size + j; + key[k] = static_cast((rand >> (j * 8)) & 0xff); + } + } + return key; +} + template Tensor aes128_key_tensor(Generator generator) { return key_tensor(aes::block_t_size, generator); @@ -64,6 +87,20 @@ Tensor aes128_key_tensor(Generator generator) { // ==================================================================================================================== +// A simple container for random state sub-blocks that implements RNG interface +// with random() and random64() methods, that are used by transformation function +template +struct RNGValues { + TORCH_CSPRNG_HOST_DEVICE RNGValues(uint64_t* vals) { + memcpy(&vals_, vals, size * sizeof(uint64_t)); + } + uint32_t TORCH_CSPRNG_HOST_DEVICE random() { auto res = static_cast(vals_[index]); index++; return res; } + uint64_t TORCH_CSPRNG_HOST_DEVICE random64() { auto res = vals_[index]; index++; return res; } +private: + uint64_t vals_[size]; + int index = 0; +}; + // Applies AES in CTR mode with the `key` for passed TensorIterator iter. // `scalar_t` is a scalar type equivalent of target tensor dtype // `uint_t` is an unsigned integral type of sub-blocks that random state is divided to @@ -74,16 +111,37 @@ Tensor aes128_key_tensor(Generator generator) { // `key` is a CUDA pointer to random key memory block // `transform_func` is a callable that converts N `uint_t` random state sub-blocks passed in RNGValues into target dtype `scalar_t` template -void aes_helper(TensorIterator& iter, const uint8_t* key, transform_t transform_func) { - block_cipher_ctr_mode(iter, aes::block_t_size, - [key] TORCH_CSPRNG_HOST_DEVICE (unsigned int idx) -> aes::block_t { - aes::block_t block; - memset(&block, 0, aes::block_t_size); - block.x = idx; - aes::encrypt(reinterpret_cast(&block), key); - return block; +void aes_helper(TensorIterator& iter, const uint8_t* key_bytes, transform_t transform_func) { + auto output = iter.tensor(0); + const auto output_offset_calc = make_offset_calculator<1>(TensorIterator::nullary_op(output)); + const auto output_index_calc = [output_offset_calc] TORCH_CSPRNG_HOST_DEVICE (uint32_t li) -> uint32_t { + return output_offset_calc.get(li)[0]; + }; + block_cipher( + nullptr, 0, 0, output_index_calc, + output.data_ptr(), output.numel(), output.element_size(), output_index_calc, + iter.device_type(), + [key_bytes] TORCH_CSPRNG_HOST_DEVICE (int64_t idx, uint8_t* block) -> void { + uint8_t idx_block[aes::block_t_size]; + std::memset(&idx_block, 0, aes::block_t_size); + *(reinterpret_cast(idx_block)) = idx; + aes::encrypt(idx_block, key_bytes); + for (size_t i = 0; i < aes::block_t_size; i++) { + block[i] ^= idx_block[i]; + } }, - transform_func + aes::block_t_size / (N * sizeof(uint_t)), + [transform_func] TORCH_CSPRNG_HOST_DEVICE (uint8_t* block) { + const auto n = aes::block_t_size / (N * sizeof(uint_t)); + for (size_t i = 0; i < n; ++i) { + uint64_t vals[N]; + for (size_t j = 0; j < N; ++j) { + vals[j] = (reinterpret_cast(block))[N * i + j]; + } + RNGValues rng(vals); + reinterpret_cast(block)[i] = transform_func(&rng); + } + } ); } @@ -151,7 +209,7 @@ struct RandomFromToKernel { std::is_same::value || std::is_same::value || std::is_same::value || - std::is_same::value) && range >= 1ULL << 32) + std::is_same::value)/* TODO: && range >= 1ULL << 32*/) { random_from_to_kernel_helper(iter, range, base, key); } else { @@ -416,6 +474,89 @@ Tensor& randperm_generator_out(Tensor& result, int64_t n, c10::optional(input, output, + [key_bytes] TORCH_CSPRNG_HOST_DEVICE (int64_t idx, uint8_t* block) -> void { + aes::encrypt(block, key_bytes); + } + ); +} + +void aes_ecb_decrypt(Tensor input, Tensor output, uint8_t* key_bytes) { + block_cipher(input, output, + [key_bytes] TORCH_CSPRNG_HOST_DEVICE (int64_t idx, uint8_t* block) -> void { + aes::decrypt(block, key_bytes); + } + ); +} + +void aes_ctr_encrypt(Tensor input, Tensor output, uint8_t* key_bytes) { + block_cipher(input, output, + [key_bytes] TORCH_CSPRNG_HOST_DEVICE (int64_t idx, uint8_t* block) -> void { + uint8_t idx_block[aes::block_t_size]; + std::memset(&idx_block, 0, aes::block_t_size); + *(reinterpret_cast(idx_block)) = idx; + aes::encrypt(idx_block, key_bytes); + for (size_t i = 0; i < aes::block_t_size; i++) { + block[i] ^= idx_block[i]; + } + } + ); +} + +void aes_ctr_decrypt(Tensor input, Tensor output, uint8_t* key_bytes) { + aes_ctr_encrypt(input, output, key_bytes); +} + +Tensor encrypt_pybind(Tensor input, Tensor output, Tensor key, const std::string& cipher, const std::string& mode) { + TORCH_CHECK(input.device() == output.device() && input.device() == key.device(), "input, output and key tensors must have the same device"); + const auto output_size_bytes = output.numel() * output.itemsize(); + const auto input_size_bytes = input.numel() * input.itemsize(); + const auto input_size_bytes_rounded = (input_size_bytes + aes::block_t_size - 1) / aes::block_t_size * aes::block_t_size; + TORCH_CHECK(output_size_bytes == input_size_bytes_rounded, + "output size in bytes(", output_size_bytes, + ") is not equal to input size in bytes rounded to block size(", + input_size_bytes_rounded, ")"); + check_cipher(cipher, key); + const auto key_bytes = reinterpret_cast(key.contiguous().data_ptr()); + if (mode == "ecb") { + aes_ecb_encrypt(input, output, key_bytes); + } else if (mode == "ctr") { + aes_ctr_encrypt(input, output, key_bytes); + } else { + TORCH_CHECK(false, "encrypt/decrypt supports \"ecb\" and \"ctr\" modes, \"", mode, "\" is not supported."); + } + return output; +} + +Tensor decrypt_pybind(Tensor input, Tensor output, Tensor key, const std::string& cipher, const std::string& mode) { + TORCH_CHECK(input.device() == output.device() && input.device() == key.device(), "input, output and key tensors must have the same device"); + const auto output_size_bytes = output.numel() * output.itemsize(); + const auto input_size_bytes = input.numel() * input.itemsize(); + TORCH_CHECK(output_size_bytes == input_size_bytes, "input and output tensors must have the same size in byte"); + TORCH_CHECK(input_size_bytes % aes::block_t_size == 0, "input tensor size in bytes must divisible by cipher block size in bytes"); + check_cipher(cipher, key); + const auto key_bytes = reinterpret_cast(key.contiguous().data_ptr()); + if (mode == "ecb") { + aes_ecb_decrypt(input, output, key_bytes); + } else if (mode == "ctr") { + aes_ctr_decrypt(input, output, key_bytes); + } else { + TORCH_CHECK(false, "encrypt/decrypt supports \"ecb\" and \"ctr\" modes, \"", mode, "\" is not supported."); + } + return output; +} + +// ==================================================================================================================== + Generator create_random_device_generator(c10::optional token = c10::nullopt) { if (token.has_value()) { return make_generator(*token); @@ -481,4 +622,6 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { m.def("create_mt19937_generator", &create_mt19937_generator, py::arg("seed") = nullptr); m.def("aes128_key_tensor", &aes128_key_tensor_pybind); m.def("create_const_generator", &create_const_generator); + m.def("encrypt", &encrypt_pybind); + m.def("decrypt", &decrypt_pybind); }