Skip to content

1. add int8/int16 feature #40

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

Merged
TONYSKYZENG merged 1 commit into from
Jun 9, 2023
Merged

1. add int8/int16 feature #40

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
23 changes: 17 additions & 6 deletions src/CPPAlgos/INT8CPPAlgo.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,6 +38,16 @@ torch::Tensor AMMBench::INT8CPPAlgo::fp32amm(torch::Tensor tensor1, torch::Tenso

return resultTensor.clone();
}
static float getScaleingFactor( float scalingBase,std::vector<float> matrix1Float)
{
float maxa=*std::max_element(matrix1Float.begin(), matrix1Float.end());
float minaAbs= abs(*std::min_element(matrix1Float.begin(), matrix1Float.end()));
if (minaAbs>maxa)
{
maxa=minaAbs;
}
return scalingBase/maxa;
}
torch::Tensor AMMBench::INT8CPPAlgo::int4amm(torch::Tensor tensor1, torch::Tensor tensor2)
{
auto A_size = tensor1.sizes();
Expand All @@ -54,7 +64,8 @@ torch::Tensor AMMBench::INT8CPPAlgo::int4amm(torch::Tensor tensor1, torch::Tenso
std::vector<float> matrix1Float(tensor1.data_ptr<float>(), tensor1.data_ptr<float>() + rows1 * cols1);
// Convert the input matrices to int8
std::vector<int8_t> matrix1(rows1 * cols1);
float scale1 = 7.0 / *std::max_element(matrix1Float.begin(), matrix1Float.end());

float scale1 = getScaleingFactor(7.0,matrix1Float);
for (int i = 0; i < rows1 * cols1; ++i) {
matrix1[i] = static_cast<int8_t>(matrix1Float[i] * scale1);
}
Expand All @@ -65,7 +76,7 @@ torch::Tensor AMMBench::INT8CPPAlgo::int4amm(torch::Tensor tensor1, torch::Tenso
*/
std::vector<float> matrix2Float(tensor2.data_ptr<float>(), tensor2.data_ptr<float>() + cols1 * cols2);
std::vector<int8_t> matrix2(cols1 * cols2);
float scale2 = 7.0 / *std::max_element(matrix2Float.begin(), matrix2Float.end());
float scale2 = getScaleingFactor(7.0,matrix2Float);
for (int i = 0; i < cols1 * cols2; ++i) {
matrix2[i] = static_cast<int8_t>(matrix2Float[i] * scale2);
}
Expand Down Expand Up @@ -138,7 +149,7 @@ torch::Tensor AMMBench::INT8CPPAlgo::int8amm(torch::Tensor tensor1, torch::Tenso
std::vector<float> matrix1Float(tensor1.data_ptr<float>(), tensor1.data_ptr<float>() + rows1 * cols1);
// Convert the input matrices to int8
std::vector<int8_t> matrix1(rows1 * cols1);
float scale1 = 127.0 / *std::max_element(matrix1Float.begin(), matrix1Float.end());
float scale1 = getScaleingFactor(127.0,matrix1Float);
for (int i = 0; i < rows1 * cols1; ++i) {
matrix1[i] = static_cast<int8_t>(matrix1Float[i] * scale1);
}
Expand All @@ -149,7 +160,7 @@ torch::Tensor AMMBench::INT8CPPAlgo::int8amm(torch::Tensor tensor1, torch::Tenso
*/
std::vector<float> matrix2Float(tensor2.data_ptr<float>(), tensor2.data_ptr<float>() + cols1 * cols2);
std::vector<int8_t> matrix2(cols1 * cols2);
float scale2 = 127.0 / *std::max_element(matrix2Float.begin(), matrix2Float.end());
float scale2 = getScaleingFactor(127.0,matrix2Float);
for (int i = 0; i < cols1 * cols2; ++i) {
matrix2[i] = static_cast<int8_t>(matrix2Float[i] * scale2);
}
Expand Down Expand Up @@ -216,7 +227,7 @@ torch::Tensor AMMBench::INT8CPPAlgo::int16amm(torch::Tensor tensor1, torch::Tens
std::vector<float> matrix1Float(tensor1.data_ptr<float>(), tensor1.data_ptr<float>() + rows1 * cols1);
// Convert the input matrices to int8
std::vector<int16_t> matrix1(rows1 * cols1);
float scale1 = 32767.0 / *std::max_element(matrix1Float.begin(), matrix1Float.end());
float scale1 = getScaleingFactor(32767.0,matrix1Float);
for (int i = 0; i < rows1 * cols1; ++i) {
matrix1[i] = static_cast<int16_t>(matrix1Float[i] * scale1);
}
Expand All @@ -227,7 +238,7 @@ torch::Tensor AMMBench::INT8CPPAlgo::int16amm(torch::Tensor tensor1, torch::Tens
*/
std::vector<float> matrix2Float(tensor2.data_ptr<float>(), tensor2.data_ptr<float>() + cols1 * cols2);
std::vector<int16_t> matrix2(cols1 * cols2);
float scale2 = 32767.0/ *std::max_element(matrix2Float.begin(), matrix2Float.end());
float scale2 =getScaleingFactor(32767.0,matrix2Float);
for (int i = 0; i < cols1 * cols2; ++i) {
matrix2[i] = static_cast<int16_t>(matrix2Float[i] * scale2);
}
Expand Down