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

feat: cuda implementation for ggml_conv_transpose_1d (ggml/854)

Browse source 
* conv transpose 1d passing test for 1d input and kernel

* working for different input and output channel counts, added test for variable stride

* initial draft appears to work with stride other than 1

* working with all old and new conv1d  tests

* added a test for large tensors

* removed use cuda hardcoding

* restored test-conv-transpose.c

* removed unused arugments, and fixed bug where test failure would cause subsequent tests to fail

* fixed accumulator bug

* added test to test-backend-ops

* fixed mistake

* addressed review

* fixed includes

* removed blank lines

* style and warning fixes

* return failure when test fails

* fix supports_op

---------

Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
John Balis 2024-07-02 11:09:52 -05:00 • committed by Georgi Gerganov
parent 470939d483
commit fde13b3bb9
4 changed files with 146 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

42
tests/test-backend-ops.cpp
View file

@ -1266,6 +1266,36 @@ struct test_pool2d : public test_case {
}
};
// GGML_OP_CONV_TRANSPOSE_1D
struct test_conv_transpose_1d : public test_case {
const std::array<int64_t, 4> ne_input;
const std::array<int64_t, 4> ne_kernel;
// stride
const int s0;
// padding
const int p0;
// dilation
const int d0;
std::string vars() override {
return VARS_TO_STR5(ne_input, ne_kernel, s0, p0, d0);
}
test_conv_transpose_1d(std::array<int64_t, 4> ne_input = {197, 32, 1, 1}, // [input_width, input_height, input_channels, 1]
std::array<int64_t, 4> ne_kernel = {16, 32, 32, 1}, // [kernel_width, kernel_height, input_channels, 1]
int s0 = 1, int p0 = 0, int d0 = 1)
: ne_input(ne_input), ne_kernel(ne_kernel), s0(s0), p0(p0), d0(d0) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data());
ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_kernel.data());
ggml_tensor * out = ggml_conv_transpose_1d(ctx, kernel, input, s0, p0, d0);
return out;
}
};
// GGML_OP_IM2COL
struct test_im2col : public test_case {
const ggml_type type_input;
@ -1279,7 +1309,7 @@ struct test_im2col : public test_case {
// padding
const int p0;
const int p1;
// dilatation
// dilation
const int d0;
const int d1;
// mode
@ -2098,6 +2128,16 @@ 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_F32));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
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));
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 1, 0, 1));
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 2, 0, 1));
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 1, 0, 1));
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_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 1}));
test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {2, 1, 1, 1}));
test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 2, 1, 1}));