ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

CUDA: add conv_2d_transpose (#14287)

Browse source 
* CUDA: add conv_2d_transpose

* remove direct include of cuda_fp16

* Review: add brackets for readability, remove ggml_set_param and add asserts
This commit is contained in:
Aman Gupta 2025-06-20 22:48:24 +08:00 • committed by GitHub
parent 22015b2092
commit c959f462a0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 134 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

34
tests/test-backend-ops.cpp
View file

@ -2725,6 +2725,35 @@ struct test_conv_transpose_1d : public test_case {
}
};
// GGML_OP_CONV_TRANSPOSE_2D
struct test_conv_transpose_2d : public test_case {
const std::array<int64_t, 4> ne_input;
const std::array<int64_t, 4> ne_kernel;
const int stride;
std::string vars() override {
return VARS_TO_STR3(ne_input, ne_kernel, stride);
}
test_conv_transpose_2d(std::array<int64_t, 4> ne_input = {10, 10, 3, 1}, // [input_width, input_height, input_channels, 1]
std::array<int64_t, 4> ne_kernel = {3, 3, 3, 1}, // [kernel_width, kernel_height, input_channels, 1]
int stride = 1)
: ne_input(ne_input), ne_kernel(ne_kernel), stride(stride){}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data());
ggml_set_name(input, "input");
ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne_kernel.data());
ggml_set_name(kernel, "kernel");
ggml_tensor * out = ggml_conv_transpose_2d_p0(ctx, kernel, input, stride);
ggml_set_name(out, "out");
return out;
}
};
// GGML_OP_IM2COL
struct test_im2col : public test_case {
const ggml_type type_input;
@ -4050,6 +4079,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1));
test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1));
test_cases.emplace_back(new test_conv_transpose_2d({3, 2, 3, 1}, {2, 2, 1, 3}, 1));
test_cases.emplace_back(new test_conv_transpose_2d({10, 10, 9, 1}, {3, 3, 1, 9}, 2));
test_cases.emplace_back(new test_count_equal(GGML_TYPE_F32, {4, 500, 1, 1}));
test_cases.emplace_back(new test_count_equal(GGML_TYPE_F32, {4, 5000, 1, 1}));
@ -4618,6 +4650,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, false));
test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, true));
test_cases.emplace_back(new test_conv_transpose_2d({256, 256, 256, 1}, {3, 3, 16, 256}, 1));
return test_cases;
}