happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Fix parameter order in conv2d_implicit and add comprehensive test cases for 2D convolution

This commit is contained in:
bssrdf 2025-09-03 15:45:09 -04:00
parent 3877608dc0
commit 6d84cbb5ab
2 changed files with 25 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cuda/conv2d-implicit.cu
View File

@ -355,7 +355,7 @@ void ggml_cuda_op_conv2d_implicit(ggml_backend_cuda_context & ctx, ggml_tensor *
const int64_t total = B * OC * OH * OW;
param_t params = { B, IC, IH, IW, OC, KH, KW, ST_X, ST_Y, PD_X, PD_Y, DL_X, DL_Y, OH, OW };
param_t params = { B, IC, IH, IW, OC, KH, KW, ST_Y, ST_X, PD_Y, PD_X, DL_Y, DL_X, OH, OW };
if (kernel->type == GGML_TYPE_F16) {
conv2d_implicit_cuda_f16(X_D, (half *) K_D, Y_D, params, st);

24
tests/test-backend-ops.cpp
View File

@ -5790,6 +5790,30 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
}
}
for (uint32_t s0 : { 1, 3 }) {
for (uint32_t p1 : { 2, 5 }) {
for (uint32_t Cin : { 1, 25 }) {
for (uint32_t Cout : { 1, 12 }) {
for (uint32_t KH : { 1, 2, 3, 11 }) {
for (uint32_t KW : { 1, 2, 3, 11 }) {
for (uint32_t H : { 1, 133 }) {
for (uint32_t W : { 1, 141 }) {
if (calc_conv_output_size(W, KW, s0, p0, d0) > 0 &&
calc_conv_output_size(H, KH, s1, p1, d1) > 0) {
for (auto kernel_type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
test_cases.emplace_back(new test_conv_2d_implicit(
{ W, H, Cin, 2 }, { KW, KH, Cin, Cout }, kernel_type, s0, s1, p0, p1, d0, d1, 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.)