mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 30 Packages Projects Releases Wiki Activity

ggml-webgpu: Support non-contiguous src0 and overlapping src0/src1 in binary ops (#19850)

Browse Source 
* ggml-webgpu: Add binary op support for overlapping and non-contiguous.

* Add newline to binary.wgsl

* Append the test of binary op for src overlapping  to test_bin_bcast.

* Remove unnecessary newline.
This commit is contained in:
Masashi Yoshimura
2026-03-03 00:59:53 +09:00
committed by GitHub
parent feefb92836
commit 36a7a6589c
4 changed files with 131 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-backend-ops.cpp
+20 -5
View File
@@ -2977,6 +2977,7 @@ struct test_bin_bcast : public test_case {
const std::array<int, 4> nr;
int nf; // number of fused ops, nf == 1 -> single op (no fusion)
bool perm1; // permute src1?
bool src_overlap; // src0 and src1 are overlapping views of the same buffer
bool run_whole_graph() override { return nf > 1; }
@@ -2992,8 +2993,8 @@ struct test_bin_bcast : public test_case {
std::array<int64_t, 4> ne = {10, 10, 1, 1},
std::array<int, 4> nr = {1, 2, 1, 1},
int nf = 1,
bool perm1 = false)
: op(op), type(type), ne(ne), nr(nr), nf(nf), perm1(perm1) {}
bool perm1 = false, bool src_overlap = false)
: op(op), type(type), ne(ne), nr(nr), nf(nf), perm1(perm1), src_overlap(src_overlap) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
GGML_ASSERT(nf <= 16);
@@ -3008,6 +3009,8 @@ struct test_bin_bcast : public test_case {
b[i] = ggml_new_tensor_4d(ctx, type, ne[p[0]], ne[p[1]], ne[p[2]], ne[p[3]]);
b[i] = ggml_permute(ctx, b[i], p[0], p[1], p[2], p[3]);
} else if (src_overlap) {
b[i] = ggml_view_4d(ctx, a, ne[0], ne[1], ne[2], 2 * (ne[3] / 3), a->nb[1], a->nb[2], a->nb[3], (ne[3] / 3) * a->nb[3]);
} else {
b[i] = ggml_new_tensor(ctx, type, 4, ne.data());
}
@@ -3021,7 +3024,13 @@ struct test_bin_bcast : public test_case {
ggml_set_param(b[0]);
}
ggml_tensor * out = a;
ggml_tensor *out;
if (src_overlap) {
out = ggml_view_4d(ctx, a, ne[0], ne[1], ne[2], 2 * (ne[3] / 3), a->nb[1], a->nb[2], a->nb[3], 0);
} else {
out = a;
}
for (int i = 0; i < nf; ++i) {
out = op(ctx, out, b[i]);
@@ -7527,9 +7536,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
}
}
auto add_test_bin_bcast = [&](ggml_type type, std::array<int64_t, 4> ne, std::array<int, 4> nr, bool perm1 = false) {
auto add_test_bin_bcast = [&](ggml_type type, std::array<int64_t, 4> ne, std::array<int, 4> nr, bool perm1 = false, bool src_overlap = false) {
for (auto op : {ggml_add, ggml_sub, ggml_mul, ggml_div}) {
test_cases.emplace_back(new test_bin_bcast(op, type, ne, nr, 1, perm1));
test_cases.emplace_back(new test_bin_bcast(op, type, ne, nr, 1, perm1, src_overlap));
}
};
for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
@@ -7549,6 +7558,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
add_test_bin_bcast(type, {10, 5, 4, 3}, {2, 2, 2, 2}, perm1);
}
// src_overlap
add_test_bin_bcast(type, {10, 5, 4, 6}, {1, 1, 1, 1}, false, true);
add_test_bin_bcast(type, {10, 5, 4, 5}, {1, 1, 1, 1}, false, true);
add_test_bin_bcast(type, {1, 1, 120, 120}, {1, 1, 1, 1}, false, true);
add_test_bin_bcast(type, {1, 1, 4, 320}, {1, 1, 1, 1}, false, true);
// test case for k_bin_bcast_unravel in CUDA backend
add_test_bin_bcast(type, {1, 1, 65536, 1}, {256, 1, 1, 1});