On 08/24/2016 12:42 PM, Chris Wilson wrote:
> From about kernel 4.9, GTT mmaps are virtually unlimited. A new
> parameter, I915_PARAM_MMAP_GTT_VERSION, is added to advertise the
> feature so query it and use it to avoid limiting tiled allocations to
> only fit within the mappable aperture.
>
> Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> Cc: Kenneth Graunke <kenneth@xxxxxxxxxxxxx>
> ---
> src/mesa/drivers/dri/i965/brw_context.c | 17 ++---------------
> src/mesa/drivers/dri/i965/brw_context.h | 2 +-
> src/mesa/drivers/dri/i965/intel_screen.c | 32 ++++++++++++++++++++++++++++++++
> src/mesa/drivers/dri/i965/intel_screen.h | 2 ++
> 4 files changed, 37 insertions(+), 16 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_context.c b/src/mesa/drivers/dri/i965/brw_context.c
> index b0f9063..79dba1e 100644
> --- a/src/mesa/drivers/dri/i965/brw_context.c
> +++ b/src/mesa/drivers/dri/i965/brw_context.c
> @@ -1033,21 +1033,6 @@ brwCreateContext(gl_api api,
> brw->urb.max_ds_entries = devinfo->urb.max_ds_entries;
> brw->urb.max_gs_entries = devinfo->urb.max_gs_entries;
>
> - /* Estimate the size of the mappable aperture into the GTT. There's an
> - * ioctl to get the whole GTT size, but not one to get the mappable subset.
> - * It turns out it's basically always 256MB, though some ancient hardware
> - * was smaller.
> - */
> - uint32_t gtt_size = 256 * 1024 * 1024;
> -
> - /* We don't want to map two objects such that a memcpy between them would
> - * just fault one mapping in and then the other over and over forever. So
> - * we would need to divide the GTT size by 2. Additionally, some GTT is
> - * taken up by things like the framebuffer and the ringbuffer and such, so
> - * be more conservative.
> - */
> - brw->max_gtt_map_object_size = gtt_size / 4;
> -
> if (brw->gen == 6)
> brw->urb.gs_present = false;
>
> @@ -1058,6 +1043,8 @@ brwCreateContext(gl_api api,
>
> brw->predicate.state = BRW_PREDICATE_STATE_RENDER;
>
> + brw->max_gtt_map_object_size = screen->max_gtt_map_object_size;
> +
> brw->use_resource_streamer = screen->has_resource_streamer &&
> (env_var_as_boolean("INTEL_USE_HW_BT", false) ||
> env_var_as_boolean("INTEL_USE_GATHER", false));
> diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
> index f2dd164..523f36c 100644
> --- a/src/mesa/drivers/dri/i965/brw_context.h
> +++ b/src/mesa/drivers/dri/i965/brw_context.h
> @@ -861,7 +861,7 @@ struct brw_context
> */
> bool perf_debug;
>
> - uint32_t max_gtt_map_object_size;
> + uint64_t max_gtt_map_object_size;
>
> int gen;
> int gt;
> diff --git a/src/mesa/drivers/dri/i965/intel_screen.c b/src/mesa/drivers/dri/i965/intel_screen.c
> index 98f1c76..62eacba 100644
> --- a/src/mesa/drivers/dri/i965/intel_screen.c
> +++ b/src/mesa/drivers/dri/i965/intel_screen.c
> @@ -1006,6 +1006,17 @@ intel_get_boolean(struct intel_screen *screen, int param)
> return (intel_get_param(screen, param, &value) == 0) && value;
> }
>
> +static int
> +intel_get_integer(struct intel_screen *screen, int param)
> +{
> + int value = -1;
> +
> + if (intel_get_param(screen, param, &value) == 0)
> + return value;
> +
> + return -1;
> +}
> +
> static void
> intelDestroyScreen(__DRIscreen * sPriv)
> {
> @@ -1576,6 +1587,27 @@ __DRIconfig **intelInitScreen2(__DRIscreen *psp)
> if (INTEL_DEBUG & DEBUG_AUB)
> drm_intel_bufmgr_gem_set_aub_dump(intelScreen->bufmgr, true);
>
> +#define I915_PARAM_MMAP_GTT_VERSION 40 /* XXX delete me with new libdrm */
> + if (intel_get_integer(intelScreen, I915_PARAM_MMAP_GTT_VERSION) >= 1) {
> + /* Theorectically unlimited! */
> + intelScreen->max_gtt_map_object_size = UINT64_MAX;
Well... not quite unlimited, right? Isn't the actual VMA space less
than 64-bits? I thought it was more like 48 bits.
> + } else {
> + /* Estimate the size of the mappable aperture into the GTT. There's an
> + * ioctl to get the whole GTT size, but not one to get the mappable subset.
> + * It turns out it's basically always 256MB, though some ancient hardware
> + * was smaller.
> + */
> + uint32_t gtt_size = 256 * 1024 * 1024;
> +
> + /* We don't want to map two objects such that a memcpy between them would
> + * just fault one mapping in and then the other over and over forever. So
> + * we would need to divide the GTT size by 2. Additionally, some GTT is
> + * taken up by things like the framebuffer and the ringbuffer and such, so
> + * be more conservative.
> + */
> + intelScreen->max_gtt_map_object_size = gtt_size / 4;
> + }
> +
> intelScreen->hw_has_swizzling = intel_detect_swizzling(intelScreen);
> intelScreen->hw_has_timestamp = intel_detect_timestamp(intelScreen);
>
> diff --git a/src/mesa/drivers/dri/i965/intel_screen.h b/src/mesa/drivers/dri/i965/intel_screen.h
> index 35cd5df..fa806a3 100644
> --- a/src/mesa/drivers/dri/i965/intel_screen.h
> +++ b/src/mesa/drivers/dri/i965/intel_screen.h
> @@ -44,6 +44,8 @@ struct intel_screen
>
> __DRIscreen *driScrnPriv;
>
> + uint64_t max_gtt_map_object_size;
> +
> bool no_hw;
> bool hw_has_swizzling;
> bool has_fence_fd;
>
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
