vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

metal : add POOL2D and fix IM2COL (#9943)

Browse source 
* add pool_2d

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>

* fix im2col and add unittest for N>=1024

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>

* add tests for N % 1024 != 0

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>

* remove trailing whitespaces

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>

* apply suggestions

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>

* apply more optimization

- original IM2COL kernel + _ext with MIN()

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>

* apply review: change kernel name of pool_2d

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>

* apply review

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>

* fix more formatting and enhance readability

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>

---------

Signed-off-by: Junhee Yoo <junhee.yoo@navercorp.com>
This commit is contained in:
Jun Hee Yoo 2024-10-23 19:33:45 +09:00 • committed by GitHub
parent 873279b159
commit 4c9388fb96
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 298 additions and 18 deletions
Download patch file Download diff file Expand all files Collapse all files

10
tests/test-backend-ops.cpp
View file

@ -3316,6 +3316,16 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
// test cases for 2D im2col
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 32}, {3, 3, 1, 32}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 32}, {3, 3, 2, 32}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 1024}, {3, 3, 1, 1024}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 1024}, {3, 3, 2, 1024}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2048}, {3, 3, 1, 2048}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2048}, {3, 3, 2, 2048}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2560}, {3, 3, 1, 2560}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2560}, {3, 3, 2, 2560}, 1, 1, 1, 1, 1, 1, true));
// sycl backend will limit task global_range < MAX_INT
// test cases for 2D im2col with large input W and H (occurs in stable-diffusion)
// however these cases need to alloc more memory which may fail in some devices (Intel Arc770, etc.)