提交 fddffed3 编写于 作者: J Jiawei Wang 提交者: GitHub

Merge pull request #646 from Mycaster/optimize-quantization-tool

quant embedding table with multi threads
......@@ -12,13 +12,23 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <sys/time.h>
#include <fstream>
#include <iostream>
#include <memory>
#include <thread>
#include "core/predictor/framework.pb.h"
#include "quant.h"
#include "seq_file.h"
inline uint64_t time_diff(const struct timeval &start_time,
const struct timeval &end_time) {
return (end_time.tv_sec - start_time.tv_sec) * 1000000 +
(end_time.tv_usec - start_time.tv_usec);
}
using paddle::framework::proto::VarType;
std::map<int, size_t> var_type_size;
void reg_var_types() {
......@@ -100,8 +110,8 @@ int dump_parameter(const char *input_file, const char *output_file) {
char *value_buf = new char[value_buf_len];
size_t offset = 0;
for (int64_t i = 0; i < dims[0]; ++i) {
// std::cout << "key_len " << key_len << " value_len " << value_buf_len <<
// std::endl;
// std::cout << "key_len " << key_len << " value_len " << value_buf_len
// << std::endl;
memcpy(value_buf, tensor_buf + offset, value_buf_len);
seq_file_writer.write((char *)&i, sizeof(i), value_buf, value_buf_len);
offset += value_buf_len;
......@@ -109,14 +119,14 @@ int dump_parameter(const char *input_file, const char *output_file) {
return 0;
}
int compress_parameter(const char *file1, const char *file2, int bits) {
float *read_embedding_table(const char *file1, std::vector<int64_t> &dims) {
std::ifstream is(file1);
// Step 1: is read version, os write version
uint32_t version;
is.read(reinterpret_cast<char *>(&version), sizeof(version));
if (version != 0) {
std::cout << "Version number " << version << " not supported" << std::endl;
return -1;
return NULL;
}
std::cout << "Version size: " << sizeof(version) << std::endl;
// Step 2: is read LoD level, os write LoD level
......@@ -138,7 +148,7 @@ int compress_parameter(const char *file1, const char *file2, int bits) {
is.read(reinterpret_cast<char *>(&version), sizeof(version));
if (version != 0) {
std::cout << "Version number " << version << " not supported" << std::endl;
return -1;
return NULL;
}
// Step 4: is read Tensor Data, os write min/max/quant data
......@@ -149,10 +159,10 @@ int compress_parameter(const char *file1, const char *file2, int bits) {
is.read(reinterpret_cast<char *>(buf.get()), size);
if (!desc.ParseFromArray(buf.get(), size)) {
std::cout << "Cannot parse tensor desc" << std::endl;
return -1;
return NULL;
}
// read tensor
std::vector<int64_t> dims;
// std::vector<int64_t> dims;
dims.reserve(static_cast<size_t>(desc.dims().size()));
std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims));
......@@ -164,7 +174,7 @@ int compress_parameter(const char *file1, const char *file2, int bits) {
if (dims.size() != 2) {
std::cout << "Parameter dims not 2D" << std::endl;
return -1;
return NULL;
}
size_t numel = 1;
......@@ -176,47 +186,96 @@ int compress_parameter(const char *file1, const char *file2, int bits) {
char *tensor_buf = new char[buf_size];
is.read(static_cast<char *>(tensor_buf), buf_size);
float *tensor_float_buf = reinterpret_cast<float *>(tensor_buf);
size_t per_line_size = dims[1] * 1 + 2 * sizeof(float);
char *tensor_out = new char[per_line_size * dims[0]];
return tensor_float_buf;
}
float loss = 0;
float all_loss = 0;
int compress_parameter_parallel(const char *file1,
const char *file2,
int bits,
int n_threads) {
#define MIN_THREADS (1)
#define MAX_THREADS (80)
std::vector<int64_t> dims;
float *emb_table = read_embedding_table(file1, dims);
if (emb_table == NULL || dims.size() != 2) {
return -1;
}
// int64_t dict_size = dims[0]/100000000;
int64_t dict_size = dims[0];
int64_t emb_size = dims[1];
size_t per_line_size = emb_size * 1 + 2 * sizeof(float);
n_threads = std::min(std::max(MIN_THREADS, n_threads), MAX_THREADS);
int64_t step = dict_size / n_threads;
std::vector<char *> result;
result.reserve(dict_size + 1);
double pow2bits = pow(2, bits);
std::cout << "Start Quant" << std::endl;
std::vector<std::thread> threads;
for (int i = 0; i < n_threads + 1; ++i) {
threads.push_back(std::thread([=, &result]() {
int64_t start = i * step;
int64_t end = (i + 1) * step;
if (i == n_threads) {
if (start == dict_size) {
return;
}
end = dict_size;
}
printf("THREAD[%d], index [%ld, %ld), start Quant table...\n",
i,
start,
end);
struct timeval quant_start;
gettimeofday(&(quant_start), NULL);
for (int64_t k = start; k < end; ++k) {
float xmin = 0, xmax = 0, loss = 0;
char *tensor_temp = new char[per_line_size];
greedy_search(
emb_table + k * emb_size, xmin, xmax, loss, emb_size, bits);
// 得出 loss 最小的时候的 scale
float scale = (xmax - xmin) * (pow2bits - 1);
char *min_ptr = tensor_temp;
char *max_ptr = tensor_temp + sizeof(float);
memcpy(min_ptr, &xmin, sizeof(float));
memcpy(max_ptr, &xmax, sizeof(float));
for (size_t e = 0; e < emb_size; ++e) {
float x = *(emb_table + k * emb_size + e);
int val = round((x - xmin) / scale);
val = std::max(0, val);
val = std::min((int)pow2bits - 1, val);
*(tensor_temp + 2 * sizeof(float) + e) = val;
}
result[k] = tensor_temp;
if ((k - start) % 10000 == 0) {
printf("THREAD[%d], handle line: %ld\n", i, k - start);
}
}
struct timeval quant_end;
gettimeofday(&(quant_end), NULL);
printf("THREAD[%d], Quantization finished, cost: %lu us!!!\n",
i,
time_diff(quant_start, quant_end));
}));
}
for (auto &thread : threads) {
thread.join();
}
SeqFileWriter seq_file_writer(file2);
size_t offset = 0;
for (int64_t i = 0; i < dims[0]; ++i) {
float xmin = 0, xmax = 0, loss = 0;
size_t scale = dims[1];
char *tensor_temp = new char[per_line_size];
greedy_search(
tensor_float_buf + i * dims[1], xmin, xmax, loss, scale, bits);
for (size_t e = 0; e < dims[1]; ++e) {
float x = *(tensor_float_buf + i * dims[1] + e);
int val = round((x - xmin) / (xmax - xmin) * (pow(2, bits) - 1));
val = std::max(0, val);
val = std::min((int)pow(2, bits) - 1, val);
char *min_ptr = tensor_temp;
char *max_ptr = tensor_temp + sizeof(float);
memcpy(min_ptr, &xmin, sizeof(float));
memcpy(max_ptr, &xmax, sizeof(float));
*(tensor_temp + 2 * sizeof(float) + e) = val;
float unit = (xmax - xmin) / pow(2, bits);
float trans_val = unit * val + xmin;
}
seq_file_writer.write((char *)&i, sizeof(i), tensor_temp, per_line_size);
for (int64_t i = 0; i < dict_size; i++) {
seq_file_writer.write((char *)&i, sizeof(i), result[i], per_line_size);
}
return 0;
}
int main(int argc, char **argv) {
if (argc < 3 || argc > 4) {
std::cout << "Usage: if no compress, please follow:" << std::endl;
std::cout << "seq_generator PARAMETER_FILE OUTPUT_FILE\n" << std::endl;
if (argc < 3 || argc > 5) {
std::cout << "Usage:" << std::endl;
std::cout << "if no compress, please follow:" << std::endl;
std::cout << " seq_generator PARAMETER_FILE OUTPUT_FILE\n" << std::endl;
std::cout << "if compress, please follow: " << std::endl;
std::cout << "seq_generator PARAMETER_FILE OUTPUT_FILE QUANT_BITS"
std::cout << " seq_generator PARAMETER_FILE OUTPUT_FILE QUANT_BITS "
"[N_THREADS]"
<< std::endl;
std::cout << "Now it only support 8 bit." << std::endl;
std::cout << " Now it only support 8 bit." << std::endl;
return -1;
}
reg_var_types();
......@@ -227,7 +286,13 @@ int main(int argc, char **argv) {
}
if (argc == 4) {
std::cout << "generate compressed sparse param sequence file" << std::endl;
compress_parameter(argv[1], argv[2], atoi(argv[3]));
compress_parameter_parallel(argv[1], argv[2], atoi(argv[3]), 1);
return 0;
}
if (argc == 5) {
std::cout << "parallel generate compressed sparse param sequence file"
<< std::endl;
compress_parameter_parallel(argv[1], argv[2], atoi(argv[3]), atoi(argv[4]));
return 0;
}
}
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册