Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

sycl : add Kronecker product FWHT support for sizes 384, 640, 768, 1280 (#28016)

Browse Source
This commit is contained in:
Jingxin (Philip) Li
2026-09-01 11:47:08 -04:00
committed by GitHub
parent 8887a48f05
commit 1f3d318734
2 changed files with 283 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-backend-ops.cpp
+111 -13
View File
@@ -4660,6 +4660,51 @@ struct test_mul_mat : public test_case {
}
};
#define P 1.0f
#define N -1.0f
// constant Hadamard matrix via Paley I construction
static constexpr float H12[12][12] = {
{ P, P, P, P, P, P, P, P, P, P, P, P },
{ P, N, P, N, P, P, P, N, N, N, P, N },
{ P, N, N, P, N, P, P, P, N, N, N, P },
{ P, P, N, N, P, N, P, P, P, N, N, N },
{ P, N, P, N, N, P, N, P, P, P, N, N },
{ P, N, N, P, N, N, P, N, P, P, P, N },
{ P, N, N, N, P, N, N, P, N, P, P, P },
{ P, P, N, N, N, P, N, N, P, N, P, P },
{ P, P, P, N, N, N, P, N, N, P, N, P },
{ P, P, P, P, N, N, N, P, N, N, P, N },
{ P, N, P, P, P, N, N, N, P, N, N, P },
{ P, P, N, P, P, P, N, N, N, P, N, N }
};
static constexpr float H20[20][20] = {
{ P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P },
{ P, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N },
{ P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P },
{ P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P },
{ P, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N },
{ P, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N },
{ P, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N },
{ P, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N },
{ P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P },
{ P, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N },
{ P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P },
{ P, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N },
{ P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P },
{ P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P },
{ P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P },
{ P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P },
{ P, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N },
{ P, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N },
{ P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P },
{ P, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N }
};
#undef P
#undef N
// GGML_HINT_SRC0_IS_HADAMARD
struct test_mul_mat_hadamard : public test_mul_mat {
test_mul_mat_hadamard(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
@@ -4684,20 +4729,57 @@ struct test_mul_mat_hadamard : public test_mul_mat {
void initialize_tensors(ggml_context * ctx) override {
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (strcmp(t->name, "a") == 0) {
const int64_t n_cols = t->ne[0];
const int64_t n_rows = ggml_nrows(t);
const int64_t n_cols = t->ne[0];
const int64_t n_rows = ggml_nrows(t);
std::vector<float> data(n_cols * n_rows);
float scale = 1.0f / sqrtf((float)n_cols);
for (int64_t r = 0; r < n_rows; r++) {
float * row_data = data.data() + r * n_cols;
for (int64_t i = 0; i < n_cols; i++) {
int pop = 0;
int64_t val = r & i;
while (val) {
pop += (val & 1);
val >>= 1;
float scale = 1.0f / sqrtf((float) n_cols);
auto is_pow2 = [](const int64_t a) {
return (a > 0) && ((a & (a - 1)) == 0);
};
const bool is_kronecker =
((n_cols % 12 == 0) && is_pow2(n_cols / 12)) || ((n_cols % 20 == 0) && is_pow2(n_cols / 20));
if (is_kronecker) {
const int64_t B = (n_cols % 12 == 0 && is_pow2(n_cols / 12)) ? 12 : 20;
const int64_t M = n_cols / B;
for (int64_t r = 0; r < n_rows; r++) {
float * row_data = data.data() + r * n_cols;
const int64_t r_mod = r % n_cols;
const int64_t r_b = r_mod / B;
const int64_t r_m = r_mod % B;
for (int64_t i = 0; i < n_cols; i++) {
const int64_t c_b = i / B;
const int64_t c_m = i % B;
int pop = 0;
int64_t val = r_b & c_b;
while (val) {
pop += (val & 1);
val >>= 1;
}
const float sign_m = (pop % 2 == 0) ? 1.0f : -1.0f;
const float sign_b = (B == 12) ? H12[c_m][r_m] : H20[c_m][r_m];
row_data[i] = scale * sign_b * sign_m;
}
}
}
else if (is_pow2(n_cols)) {
for (int64_t r = 0; r < n_rows; r++) {
float * row_data = data.data() + r * n_cols;
for (int64_t i = 0; i < n_cols; i++) {
int pop_cnt = 0;
int64_t val = r & i;
while (val) {
pop_cnt += (val & 1);
val >>= 1;
}
row_data[i] = (pop_cnt % 2 == 0) ? scale : -scale;
}
row_data[i] = (pop % 2 == 0) ? scale : -scale;
}
}
ggml_backend_tensor_set(t, data.data(), 0, data.size() * sizeof(float));
@@ -9298,6 +9380,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 256, 512, 256)); // many rows
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 32, 1, 32)); // too small (N<64)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 1024, 1, 1024)); // too big (N>512)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 384, 1, 384)); // m=12 (N=384)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 384, 32, 384)); // m=12 (batch)
test_cases.emplace_back(
new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 384, 4, 384, { 2, 3 })); // m=12 (multi-dim)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 768, 1, 768)); // m=12 (N=768)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 640, 1, 640)); // m=20 (N=640)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 640, 32, 640)); // m=20 (batch)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 1280, 1, 1280)); // m=20 (N=1280)
#if 0
// > 4GB A matrix. Too slow to be enabled by default.
@@ -10500,7 +10590,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 128, 2048, 128));
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 256, 2048, 256));
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 512, 2048, 512));
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 384, 1, 384)); // m=12 (N=384)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 384, 32, 384)); // m=12 (batch)
test_cases.emplace_back(
new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 384, 4, 384, { 2, 3 })); // m=12 (multi-dim)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 768, 1, 768)); // m=12 (N=768)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 640, 1, 640)); // m=20 (N=640)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 640, 32, 640)); // m=20 (batch)
test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 1280, 1, 1280)); // m=20 (N=1280)
test_cases.emplace_back(new test_solve_tri(GGML_TYPE_F32, { 64, 64, 4, 4 }, { 32, 64, 4, 4 }));
test_cases.emplace_back(new test_solve_tri(GGML_TYPE_F32, { 128, 128, 4, 2 }, { 32, 128, 4, 2 }));
// qwen3next with CHUNK_SIZE 64