aboutsummaryrefslogtreecommitdiff
path: root/pixman/test/lowlevel-blt-bench.c
diff options
context:
space:
mode:
authormarha <marha@users.sourceforge.net>2013-01-31 08:07:51 +0100
committermarha <marha@users.sourceforge.net>2013-01-31 08:07:51 +0100
commit436f8246cd5a377596f6e42e6562da186ef8df48 (patch)
treea00e72591f276cac16bc0fd65bc263ba89c93c00 /pixman/test/lowlevel-blt-bench.c
parent69c8cec54b01ed522bf10baf20da70304bac701a (diff)
downloadvcxsrv-436f8246cd5a377596f6e42e6562da186ef8df48.tar.gz
vcxsrv-436f8246cd5a377596f6e42e6562da186ef8df48.tar.bz2
vcxsrv-436f8246cd5a377596f6e42e6562da186ef8df48.zip
libxtrans mesa pixman xkeyboard-config git update 31 jan 2013
libxtrans: bd53f4c8543faf910a7a151241ee07661b4d57ad mesa: a527b2192e3cb4a68af927ab405e38181d2fcf75 pixman: 958bd334b3c17f529c80f2eeef4224f45c62f292 xkeyboard-config: 46cfd7a3b5c6b6e00648407bd4c0dd56ae86d65a
Diffstat (limited to 'pixman/test/lowlevel-blt-bench.c')
-rw-r--r--pixman/test/lowlevel-blt-bench.c31
1 files changed, 25 insertions, 6 deletions
diff --git a/pixman/test/lowlevel-blt-bench.c b/pixman/test/lowlevel-blt-bench.c
index 7336fa0d5..8e80b4280 100644
--- a/pixman/test/lowlevel-blt-bench.c
+++ b/pixman/test/lowlevel-blt-bench.c
@@ -33,6 +33,14 @@
#define L1CACHE_SIZE (8 * 1024)
#define L2CACHE_SIZE (128 * 1024)
+/* This is applied to both L1 and L2 tests - alternatively, you could
+ * parameterise bench_L or split it into two functions. It could be
+ * read at runtime on some architectures, but it only really matters
+ * that it's a number that's an integer divisor of both cacheline
+ * lengths, and further, it only really matters for caches that don't
+ * do allocate0on-write. */
+#define CACHELINE_LENGTH (32) /* bytes */
+
#define WIDTH 1920
#define HEIGHT 1080
#define BUFSIZE (WIDTH * HEIGHT * 4)
@@ -168,18 +176,29 @@ bench_L (pixman_op_t op,
int width,
int lines_count)
{
- int64_t i, j;
+ int64_t i, j, k;
int x = 0;
int q = 0;
volatile int qx;
for (i = 0; i < n; i++)
{
- /* touch destination buffer to fetch it into L1 cache */
- for (j = 0; j < width + 64; j += 16) {
- q += dst[j];
- q += src[j];
- }
+ /* For caches without allocate-on-write, we need to force the
+ * destination buffer back into the cache on each iteration,
+ * otherwise if they are evicted during the test, they remain
+ * uncached. This doesn't matter for tests which read the
+ * destination buffer, or for caches that do allocate-on-write,
+ * but in those cases this loop just adds constant time, which
+ * should be successfully cancelled out.
+ */
+ for (j = 0; j < lines_count; j++)
+ {
+ for (k = 0; k < width + 62; k += CACHELINE_LENGTH / sizeof *dst)
+ {
+ q += dst[j * WIDTH + k];
+ }
+ q += dst[j * WIDTH + width + 62];
+ }
if (++x >= 64)
x = 0;
call_func (func, op, src_img, mask_img, dst_img, x, 0, x, 0, 63 - x, 0, width, lines_count);