[Intel-gfx] [PATCH i-g-t] tests/pm_rpm tests for set_caching and set_tiling ioctl(s)

[Imre Deak]

On ke, 2015-11-25 at 19:16 +0200, marius.c.vlad at intel.com wrote:
> From: Marius Vlad <marius.c.vlad at intel.com>
> 
> Signed-off-by: Marius Vlad <marius.c.vlad at intel.com>
> ---
>  tests/pm_rpm.c | 120 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 120 insertions(+)
> 
> diff --git a/tests/pm_rpm.c b/tests/pm_rpm.c
> index c4fb19c..157cf29 100644
> --- a/tests/pm_rpm.c
> +++ b/tests/pm_rpm.c
> @@ -1729,6 +1729,120 @@ static void planes_subtest(bool universal, bool dpms)
>  	}
>  }
>  
> +static void pm_test_tiling(void)
> +{
> +	uint32_t *handles;
> +	uint8_t **gem_bufs;
> +
> +	int max_gem_objs = 0;
> +	uint8_t off_bit = 20;
> +	uint32_t gtt_obj_max_size = (16 * 1024 * 1024);
> +
> +	uint32_t i, j, tiling_modes[3] = {
> +		I915_TILING_NONE,
> +		I915_TILING_X,
> +		I915_TILING_Y,
> +	};
> +	uint32_t ti, sw;
> +
> +	/* default value */
> +	uint32_t stride = 1024;
> +
> +	/* calculate how many objects we can map */
> +	for (j = 1 << off_bit; j <= gtt_obj_max_size; j <<= 1, max_gem_objs++)
> +		;

With these sizes we may end up with all objects properly aligned,
that's why I suggested smaller objects. Based on I830_FENCE_START_MASK
we could allocate for example starting from 16kB to 256kB.

> +
> +	gem_bufs = calloc(max_gem_objs, sizeof(uint8_t *));
> +	handles = malloc(sizeof(uint32_t) * max_gem_objs);

Nitpick: sizeof(*ptr) is safer and you could've used calloc in both
cases.

> +
> +	/* map to gtt and store some random data */
> +	for (i = 0, j = 1 << off_bit; j <= gtt_obj_max_size; j <<= 1, i++) {
> +		handles[i] = gem_create(drm_fd, j);
> +		gem_bufs[i] = gem_mmap__gtt(drm_fd, handles[i], j, PROT_WRITE);
> +		memset(gem_bufs[i], 0x65, j);
> +	}
> +
> +	/* try to set different tiling for each handle */
> +	for (i = 0; i < ARRAY_SIZE(tiling_modes); i++) {
> +		disable_all_screens_and_wait(&ms_data);
> +
> +		for (j = 0; j < max_gem_objs; j++) {
> +			gem_set_tiling(drm_fd, handles[j], tiling_modes[i], stride);
> +
> +			gem_get_tiling(drm_fd, handles[j], &ti, &sw);
> +			igt_assert(tiling_modes[i] == ti);
> +		}
> +
> +		enable_one_screen_and_wait(&ms_data);

Ok, but after the second iteration all objects could be properly
aligned, so it's better to close/realloc/memset the objects in each
iteration.

> +	}
> +
> +	for (i = 0, j = 1 << off_bit; j <= gtt_obj_max_size; j <<= 1, i++) {
> +		igt_assert(munmap(gem_bufs[i], j) == 0);
> +		gem_close(drm_fd, handles[i]);
> +	}
> +
> +	free(gem_bufs);
> +	free(handles);
> +}
> +
> +static void pm_test_caching(void)
> +{
> +	uint32_t *handles;
> +	uint8_t **gem_bufs;
> +	int8_t has_caching_display = -1;
> +
> +	uint32_t i, j, got_caching;
> +	uint32_t gtt_obj_max_size = (16 * 1024 * 1024);
> +	uint32_t cache_levels[3] = {
> +		I915_CACHING_NONE,
> +		I915_CACHING_CACHED,		/* LLC caching */
> +		I915_CACHING_DISPLAY,		/* eDRAM caching */
> +	};
> +
> +	int max_gem_objs = 0;
> +	uint8_t off_bit = 20;
> +
> +	for (j = 1 << off_bit; j <= gtt_obj_max_size; j <<= 1, max_gem_objs++)
> +		;

No need to bother about alignment here, so we can just use a single
16kB object for example.

> +
> +	gem_bufs = calloc(max_gem_objs, sizeof(uint8_t *));
> +	handles = malloc(sizeof(uint32_t) * max_gem_objs);
> +
> +	for (i = 0, j = 1 << off_bit; j <= gtt_obj_max_size; j <<= 1, i++) {
> +		handles[i] = gem_create(drm_fd, j);
> +		gem_bufs[i] = gem_mmap__gtt(drm_fd, handles[i], j, PROT_WRITE);
> +		memset(gem_bufs[i], 0x65, j);
> +	}
> +
> +	/* figure out if we have cache display available on the platform */
> +	gem_set_caching(drm_fd, handles[0], I915_CACHING_DISPLAY);
> +	if (gem_get_caching(drm_fd, handles[0]))

No need to hardcode I915_CACHING_NONE here. Also I liked the original
version to check this everywhere better, by accepting both
CACHING_DISPLAY and CACHING_NONE as a result.

> +		has_caching_display++;
> +
> +	for (i = 0; i < ARRAY_SIZE(cache_levels) + has_caching_display; i++) {
> +		disable_all_screens_and_wait(&ms_data);
> +
> +		for (j = 0; j < max_gem_objs; j++) {
> +			gem_set_caching(drm_fd, handles[j], cache_levels[i]);
> +
> +			igt_debug("Verying cache for handle %u, level %u\n", j, i);
> +			got_caching = gem_get_caching(drm_fd, handles[j]);
> +
> +			igt_assert(got_caching == cache_levels[i]);
> +		}
> +
> +		enable_one_screen_and_wait(&ms_data);

The object can be unbound after the IOCTL so you need to do a memset at
the begin of each iteration.

> +	}
> +
> +	for (i = 0, j = 1 << off_bit; j <= gtt_obj_max_size; j <<= 1, i++) {
> +		igt_assert(munmap(gem_bufs[i], j) == 0);
> +		gem_close(drm_fd, handles[i]);
> +	}
> +
> +	free(handles);
> +	free(gem_bufs);
> +}
> +
>  static void fences_subtest(bool dpms)
>  {
>  	int i;
> @@ -1927,6 +2041,12 @@ int main(int argc, char *argv[])
>  	igt_subtest("gem-execbuf-stress-extra-wait")
>  		gem_execbuf_stress_subtest(rounds, WAIT_STATUS | WAIT_EXTRA);
>  
> +	/* power-wake reference tests */
> +	igt_subtest("pm-tiling")
> +		pm_test_tiling();
> +	igt_subtest("pm-caching")
> +		pm_test_caching();
> +
>  	igt_fixture
>  		teardown_environment();
>