benchmarks/gem_blt: Estimate memory bandwidth to improve test runtime

If we autotune the workload to only take 0.1s and then repeat the
measurements over 2s, we can bound the benchmark runtime. (Roughly of
course! Sometimes the dispartity between main memory CPU bandwidth, and
GPU execution bandwidth throws off the runtime, but that's the purpose
of the benchmark!)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

1 files changed, 44 insertions, 22 deletions

diff --git a/benchmarks/gem_blt.c b/benchmarks/gem_blt.c
index 8fe80bc1..820695c9 100644
--- a/benchmarks/gem_blt.c
+++ b/benchmarks/gem_blt.c
@@ -53,6 +53,37 @@
 
 static int has_64bit_reloc;
 
+static double
+elapsed(const struct timespec *start, const struct timespec *end)
+{
+	return (end->tv_sec - start->tv_sec) + 1e-9*(end->tv_nsec - start->tv_nsec);
+}
+
+static int baseline(uint64_t bytes, int milliseconds)
+{
+	struct timespec start, end;
+	const int size = 64*1024*1024;
+	int count = 0;
+	void *mem;
+
+	mem = malloc(size);
+	if (mem == NULL)
+		return 1;
+
+	clock_gettime(CLOCK_MONOTONIC, &start);
+	do {
+		memset(mem, count, size);
+		count++;
+		clock_gettime(CLOCK_MONOTONIC, &end);
+		if (elapsed(&start, &end) > 0.1)
+			break;
+	} while (1);
+
+	free(mem);
+
+	return ceil(1e-3*milliseconds/elapsed(&start, &end) * (count * size) / bytes);
+}
+
 static int gem_linear_blt(int fd,
 			  uint32_t *batch,
 			  int offset,
@@ -147,11 +178,6 @@ static int gem_linear_blt(int fd,
 	return (b+2 - batch) * sizeof(uint32_t);
 }
 
-static double elapsed(const struct timespec *start, const struct timespec *end)
-{
-	return (end->tv_sec - start->tv_sec) + 1e-9*(end->tv_nsec - start->tv_nsec);
-}
-
 static int __gem_execbuf(int fd, struct drm_i915_gem_execbuffer2 *execbuf)
 {
 	int err = 0;
@@ -160,14 +186,14 @@ static int __gem_execbuf(int fd, struct drm_i915_gem_execbuffer2 *execbuf)
 	return err;
 }
 
-static int run(int object, int batch, int count, int set, int reps)
+static int run(int object, int batch, int time, int reps)
 {
 	struct drm_i915_gem_execbuffer2 execbuf;
 	struct drm_i915_gem_exec_object2 exec[3];
 	struct drm_i915_gem_relocation_entry *reloc;
 	uint32_t *buf, handle, src, dst;
 	int fd, len, gen, size, nreloc;
-	int ring;
+	int ring, count;
 
 	size = ALIGN(batch * 64, 4096);
 	reloc = malloc(sizeof(*reloc)*size/32*2);
@@ -237,10 +263,13 @@ static int run(int object, int batch, int count, int set, int reps)
 	if (execbuf.flags & LOCAL_I915_EXEC_HANDLE_LUT)
 		execbuf.flags |= LOCAL_I915_EXEC_NO_RELOC;
 
+	/* Guess how many loops we need for 0.1s */
+	count = baseline((uint64_t)object * batch, 100);
+
 	while (reps--) {
 		double min = HUGE_VAL;
 
-		for (int s = 0; s < set; s++) {
+		for (int s = 0; s <= time / 100; s++) {
 			struct timespec start, end;
 			double t;
 
@@ -265,30 +294,23 @@ static int run(int object, int batch, int count, int set, int reps)
 int main(int argc, char **argv)
 {
 	int size = 1024*1024;
-	int count = 1;
 	int reps = 13;
-	int set = 30;
+	int time = 2000;
 	int batch = 1;
 	int c;
 
-	while ((c = getopt (argc, argv, "c:r:s:b:S:")) != -1) {
+	while ((c = getopt (argc, argv, "s:b:r:t:")) != -1) {
 		switch (c) {
-		case 'c':
-			count = atoi(optarg);
-			if (count < 1)
-				count = 1;
-			break;
-
 		case 's':
 			size = atoi(optarg);
 			if (size < 4096)
 				size = 4096;
 			break;
 
-		case 'S':
-			set = atoi(optarg);
-			if (set < 1)
-				set = 1;
+		case 't':
+			time = atoi(optarg);
+			if (time < 1)
+				time = 1;
 			break;
 
 		case 'r':
@@ -308,5 +330,5 @@ int main(int argc, char **argv)
 		}
 	}
 
-	return run(size, batch, count, set, reps);
+	return run(size, batch, time, reps);
 }