diff options
| author | Emmanuel Gil Peyrot <linkmauve@linkmauve.fr> | 2023-01-02 17:46:18 +0100 |
|---|---|---|
| committer | Derek Foreman <derek.foreman@collabora.com> | 2023-01-11 20:03:36 +0000 |
| commit | e13c99690b7e1c4175fef7a58200694a2f3509ba (patch) | |
| tree | de236a97b8aa1bdd4274086e986b3fad8283e815 /shared | |
| parent | 102acac6a9c758d47d876ddcecc4ad9a10b0c34b (diff) | |
| download | weston-e13c99690b7e1c4175fef7a58200694a2f3509ba.tar.gz | |
libweston: Optimise matrix multiplication
The previous version used div() to separate the column and row of the
current element, but that function is implemented as a libc call, which
prevented the compiler from vectorising the loop and made matrix
multiplication appear quite high in profiles.
With div() removed, we are down from 64 calls to vfmadd132ss acting on
one float at a time, to just 8 calls to vfmadd132ps when compiled with
AVX2 support (or 16 mulps, 16 addps with SSE2 support only), and the
function isn’t a hot spot any more.
Signed-off-by: Emmanuel Gil Peyrot <linkmauve@linkmauve.fr>
Diffstat (limited to 'shared')
| -rw-r--r-- | shared/matrix.c | 18 |
1 files changed, 9 insertions, 9 deletions
diff --git a/shared/matrix.c b/shared/matrix.c index f301c5fa..72717f37 100644 --- a/shared/matrix.c +++ b/shared/matrix.c @@ -61,16 +61,16 @@ weston_matrix_multiply(struct weston_matrix *m, const struct weston_matrix *n) { struct weston_matrix tmp; const float *row, *column; - div_t d; - int i, j; + int i, j, k; - for (i = 0; i < 16; i++) { - tmp.d[i] = 0; - d = div(i, 4); - row = m->d + d.quot * 4; - column = n->d + d.rem; - for (j = 0; j < 4; j++) - tmp.d[i] += row[j] * column[j * 4]; + for (i = 0; i < 4; i++) { + row = m->d + i * 4; + for (j = 0; j < 4; j++) { + tmp.d[4 * i + j] = 0; + column = n->d + j; + for (k = 0; k < 4; k++) + tmp.d[4 * i + j] += row[k] * column[k * 4]; + } } tmp.type = m->type | n->type; memcpy(m, &tmp, sizeof tmp); |