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

[SYCL] Fix SYCL im2col and convert Overflow with Large Dims (#9052)

Browse source 
* sycl: fix im2col overflow and sync with cuda

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* sycl: fix convert overflow

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* sycl: fix convert and dequantize

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* sycl: fix ib in dmmv

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* sycl:refine convert

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* sycl: move downsample global_range into common

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* test: add im2col and convert test cases

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* test: make new cases only in sycl

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* test: comment new test_cases for only local testing

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

---------

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>
This commit is contained in:
zhentaoyu 2024-08-20 23:06:51 +08:00 • committed by GitHub
parent 90db8146d5
commit 4f8d19ff17
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
11 changed files with 333 additions and 262 deletions
Download patch file Download diff file Expand all files Collapse all files

13
tests/test-backend-ops.cpp
View file

@ -2145,6 +2145,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
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.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));
// 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.)
// these cases are verified (pass) in Intel(R) Data Center GPU Max 1100 (sycl backend) and NV A30 (cuda backend)
// test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {1024, 1024, 256, 1}, {3, 3, 256, 1}, 1, 1, 1, 1, 1, 1, true));
// test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {1024, 1024, 256, 1}, {3, 3, 256, 1}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_conv_transpose_1d());
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 2, 0, 1));
@ -2287,6 +2294,12 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 64, 45, 128, { 8, 1}, {4, 1}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45, 64, { 8, 1}, {4, 1}));
// sycl backend will limit task global_range < MAX_INT
// test case for f16-type-convert-to-fp32 kernel with large k under fp32 compute dtype (occurs in stable-diffusion)
// however this case needs to alloc more memory which may fail in some devices (Intel Arc770, etc.)
// this case is verified (pass) in Intel(R) Data Center GPU Max 1100 (sycl backend) and NV A30 (cuda backend)
// test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 512, 262144, 9216, {1, 1}, {1, 1}));
for (ggml_type type_a : base_types) {
for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
for (int n_mats : {4, 8}) {