On 20/03/18 00:08, Kenneth Graunke wrote:
On Wednesday, March 14, 2018 10:19:11 AM PDT Lionel Landwerlin wrote:
There are a couple of ways we can get the fusing information from the
kernel :
- Through DRM_I915_GETPARAM with the SLICE_MASK/SUBSLICE_MASK
parameters
- Through the new DRM_IOCTL_I915_QUERY by requesting the
DRM_I915_QUERY_TOPOLOGY_INFO
The second method is more accurate and also gives us the EUs fusing
masks. It's also a requirement for CNL as this platform has asymetric
subslices and the first method SUBSLICE_MASK value is assumed uniform
across slices.
Signed-off-by: Lionel Landwerlin <lionel.g.landwer...@intel.com>
---
src/intel/dev/gen_device_info.c | 129 ++++++++++++++++++++++++++++++++++++++++
src/intel/dev/gen_device_info.h | 11 ++++
2 files changed, 140 insertions(+)
diff --git a/src/intel/dev/gen_device_info.c b/src/intel/dev/gen_device_info.c
index c1bdc997f2c..a8c9f7738b2 100644
--- a/src/intel/dev/gen_device_info.c
+++ b/src/intel/dev/gen_device_info.c
@@ -28,8 +28,11 @@
#include <unistd.h>
#include "gen_device_info.h"
#include "compiler/shader_enums.h"
+#include "util/bitscan.h"
#include "util/macros.h"
+#include <i915_drm.h>
+
/**
* Get the PCI ID for the device name.
*
@@ -913,6 +916,132 @@ fill_masks(struct gen_device_info *devinfo)
}
}
+static void
+reset_masks(struct gen_device_info *devinfo)
+{
+ devinfo->subslice_slice_stride =
+ devinfo->eu_subslice_stride =
+ devinfo->eu_slice_stride = 0;
+
+ devinfo->num_slices =
+ devinfo->num_eu_per_subslice = 0;
We tend to avoid chained assignments like this, can we just do:
devinfo->subslice_slice_stride = 0;
devinfo->eu_subslice_stride = 0;
devinfo->eu_slice_stride = 0;
devinfo->num_slices = 0;
devinfo->num_eu_per_subslice = 0;
With it already being on multiple lines, it's no more code, and easier
to read IMHO.
Sure.
+ memset(devinfo->num_subslices, 0, sizeof(devinfo->num_subslices));
+
+ memset(&devinfo->slice_masks, 0, sizeof(devinfo->slice_masks));
+ memset(devinfo->subslice_masks, 0, sizeof(devinfo->subslice_masks));
+ memset(devinfo->eu_masks, 0, sizeof(devinfo->eu_masks));
+}
+
+void
+gen_device_info_update_from_masks(struct gen_device_info *devinfo,
+ uint32_t slice_mask,
+ uint32_t subslice_mask,
+ uint32_t n_eus)
I wonder if it would be better to simply populate a
drm_i915_query_topology_info structure from the masks, if we don't
have the topology query uABI. Then, we could just have one codepath
for filling out devinfo, with a "whoops, no kernel support" shim.
I'm not sure whether it would be more or less code, but it might
drop some of the duplicated complexity?
Yeah, make sense.
+{
+ reset_masks(devinfo);
+
+ assert((slice_mask & 0xff) == slice_mask);
+
+ devinfo->slice_masks = slice_mask;
+ devinfo->num_slices = __builtin_popcount(devinfo->slice_masks);
_mesa_bitcount() here and elsewhere.
Only issue with this is that it pulls in the headers from what I
consider to be the GL headers : src/mesa/main/imports.h
+
+ uint32_t max_slices = util_last_bit(slice_mask);
+ uint32_t max_subslices = util_last_bit(subslice_mask);
+ devinfo->subslice_slice_stride = DIV_ROUND_UP(max_subslices, 8);
+ uint32_t n_subslices = 0;
+ for (int s = 0; s < util_last_bit(slice_mask); s++) {
for (int s = 0; s < max_slices; s++) {
Thanks!
+ if ((slice_mask & (1UL << s)) == 0)
1u << s
Perhaps you like my topology-struct idea and will make a v3...if not...
Will do :)
Reviewed-by: Kenneth Graunke <kenn...@whitecape.org>
+ continue;
+
+ for (int b = 0; b < devinfo->subslice_slice_stride; b++) {
+ int subslice_offset = s * devinfo->subslice_slice_stride + b;
+
+ devinfo->subslice_masks[subslice_offset] =
+ (subslice_mask >> (b * 8)) & 0xff;
+ devinfo->num_subslices[s] +=
+ __builtin_popcount(devinfo->subslice_masks[subslice_offset]);
+ }
+
+ n_subslices += devinfo->num_subslices[s];
+ }
+
+ /* We expect the total number of EUs to be uniformly distributed throughout
+ * the subslices.
+ */
+ assert((n_eus % n_subslices) == 0);
+ devinfo->num_eu_per_subslice = n_eus / n_subslices;
+
+ devinfo->eu_subslice_stride = DIV_ROUND_UP(devinfo->num_eu_per_subslice, 8);
+ devinfo->eu_slice_stride = devinfo->eu_subslice_stride * max_subslices;
+
+ for (int s = 0; s < max_slices; s++) {
+ if ((slice_mask & (1UL << s)) == 0)
+ continue;
+
+ for (int ss = 0; ss < max_subslices; ss++) {
+ if ((subslice_mask & (1UL << ss)) == 0)
+ continue;
+
+ for (int b = 0; b < devinfo->eu_subslice_stride; b++) {
+ int eus_offset = s * devinfo->eu_slice_stride +
+ ss * devinfo->eu_subslice_stride + b;
+
+ devinfo->eu_masks[eus_offset] =
+ (((1UL << devinfo->num_eu_per_subslice) - 1) >> (b * 8)) & 0xff;
+ }
+ }
+ }
+}
+
+void
+gen_device_info_update_from_topology(struct gen_device_info *devinfo,
+ const struct drm_i915_query_topology_info
*topology)
+{
+ reset_masks(devinfo);
+
+ devinfo->subslice_slice_stride = topology->subslice_stride;
+
+ devinfo->eu_subslice_stride = DIV_ROUND_UP(topology->max_eus_per_subslice,
8);
+ devinfo->eu_slice_stride = topology->max_subslices *
devinfo->eu_subslice_stride;
+
+ assert(sizeof(devinfo->slice_masks) >= DIV_ROUND_UP(topology->max_slices,
8));
+ memcpy(&devinfo->slice_masks, topology->data,
DIV_ROUND_UP(topology->max_slices, 8));
+ devinfo->num_slices = __builtin_popcount(devinfo->slice_masks);
+
+ uint32_t subslice_mask_len =
+ topology->max_slices * topology->subslice_stride;
+ assert(sizeof(devinfo->subslice_masks) >= subslice_mask_len);
+ memcpy(devinfo->subslice_masks, &topology->data[topology->subslice_offset],
+ subslice_mask_len);
+
+ uint32_t n_subslices = 0;
+ for (int s = 0; s < topology->max_slices; s++) {
+ if ((devinfo->slice_masks & (1UL << s)) == 0)
+ continue;
+
+ for (int b = 0; b < devinfo->subslice_slice_stride; b++) {
+ devinfo->num_subslices[s] +=
+ __builtin_popcount(devinfo->subslice_masks[b]);
+ }
+ n_subslices += devinfo->num_subslices[s];
+ }
+
+ uint32_t eu_mask_len =
+ topology->eu_stride * topology->max_subslices * topology->max_slices;
+ assert(sizeof(devinfo->eu_masks) >= eu_mask_len);
+ memcpy(devinfo->eu_masks, &topology->data[topology->eu_offset],
eu_mask_len);
+
+ uint32_t n_eus = 0;
+ for (int b = 0; b < eu_mask_len; b++)
+ n_eus += __builtin_popcount(devinfo->eu_masks[b]);
+
+ /* We expect the total number of EUs to be uniformly distributed throughout
+ * the subslices.
+ */
+ assert((n_eus % n_subslices) == 0);
+ devinfo->num_eu_per_subslice = n_eus / n_subslices;
+}
+
bool
gen_get_device_info(int devid, struct gen_device_info *devinfo)
{
diff --git a/src/intel/dev/gen_device_info.h b/src/intel/dev/gen_device_info.h
index 793ce094850..41aa54ab424 100644
--- a/src/intel/dev/gen_device_info.h
+++ b/src/intel/dev/gen_device_info.h
@@ -34,6 +34,8 @@
extern "C" {
#endif
+struct drm_i915_query_topology_info;
+
#define GEN_DEVICE_MAX_SLICES (6) /* Maximum on gen10 */
#define GEN_DEVICE_MAX_SUBSLICES (8) /* Maximum on gen11 */
#define GEN_DEVICE_MAX_EUS_PER_SUBSLICE (10) /* Maximum on Haswell */
@@ -248,6 +250,15 @@ int gen_device_name_to_pci_device_id(const char *name);
bool gen_get_device_info(int devid, struct gen_device_info *devinfo);
const char *gen_get_device_name(int devid);
+/* Used with SLICE_MASK/SUBSLICE_MASK values from DRM_I915_GETPARAM. */
+void gen_device_info_update_from_masks(struct gen_device_info *devinfo,
+ uint32_t slice_mask,
+ uint32_t subslice_mask,
+ uint32_t n_eus);
+/* Used with DRM_IOCTL_I915_QUERY & DRM_I915_QUERY_TOPOLOGY_INFO. */
+void gen_device_info_update_from_topology(struct gen_device_info *devinfo,
+ const struct
drm_i915_query_topology_info *topology);
+
#ifdef __cplusplus
}
#endif
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
