On Thu, Aug 05, 2021 at 09:36:42AM -0700, Matt Roper wrote:
From: Stuart Summers <stuart.summers@xxxxxxxxx>
Starting in XeHP, the concept of slice has been removed in favor of
DSS (Dual-Subslice) masks for various workload types. These workloads have
been divided into those enabled for geometry and those enabled for compute.
i915 currently maintains a single set of S/SS/EU masks for the device.
The goal of this patch set is to minimize the amount of impact to prior
generations while still giving the user maximum flexibility.
v2:
- Generalize a comment about uapi access to geometry/compute masks; the
proposed uapi has changed since the comment was first written, and
will show up in a future series once the userspace code is published.
(Lucas)
Bspec: 33117, 33118, 20376
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@xxxxxxxxx>
Cc: Matt Roper <matthew.d.roper@xxxxxxxxx>
Cc: Lucas De Marchi <lucas.demarchi@xxxxxxxxx>
Signed-off-by: Stuart Summers <stuart.summers@xxxxxxxxx>
Signed-off-by: Steve Hampson <steven.t.hampson@xxxxxxxxx>
Signed-off-by: Matt Roper <matthew.d.roper@xxxxxxxxx>
---
drivers/gpu/drm/i915/gt/intel_sseu.c | 66 +++++++++++++++++++++-------
drivers/gpu/drm/i915/gt/intel_sseu.h | 5 ++-
drivers/gpu/drm/i915/i915_reg.h | 3 +-
include/uapi/drm/i915_drm.h | 3 --
4 files changed, 55 insertions(+), 22 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_sseu.c b/drivers/gpu/drm/i915/gt/intel_sseu.c
index bbd272943c3f..9cf157a2454f 100644
--- a/drivers/gpu/drm/i915/gt/intel_sseu.c
+++ b/drivers/gpu/drm/i915/gt/intel_sseu.c
@@ -46,11 +46,11 @@ u32 intel_sseu_get_subslices(const struct sseu_dev_info *sseu, u8 slice)
}
void intel_sseu_set_subslices(struct sseu_dev_info *sseu, int slice,
- u32 ss_mask)
+ u8 *subslice_mask, u32 ss_mask)
{
int offset = slice * sseu->ss_stride;
- memcpy(&sseu->subslice_mask[offset], &ss_mask, sseu->ss_stride);
+ memcpy(&subslice_mask[offset], &ss_mask, sseu->ss_stride);
}
unsigned int
@@ -100,14 +100,24 @@ static u16 compute_eu_total(const struct sseu_dev_info *sseu)
return total;
}
-static void gen11_compute_sseu_info(struct sseu_dev_info *sseu,
- u8 s_en, u32 ss_en, u16 eu_en)
+static u32 get_ss_stride_mask(struct sseu_dev_info *sseu, u8 s, u32 ss_en)
+{
+ u32 ss_mask;
+
+ ss_mask = ss_en >> (s * sseu->max_subslices);
+ ss_mask &= GENMASK(sseu->max_subslices - 1, 0);
+
+ return ss_mask;
+}
+
+static void gen11_compute_sseu_info(struct sseu_dev_info *sseu, u8 s_en,
+ u32 g_ss_en, u32 c_ss_en, u16 eu_en)
{
int s, ss;
- /* ss_en represents entire subslice mask across all slices */
+ /* g_ss_en/c_ss_en represent entire subslice mask across all slices */
GEM_BUG_ON(sseu->max_slices * sseu->max_subslices >
- sizeof(ss_en) * BITS_PER_BYTE);
+ sizeof(g_ss_en) * BITS_PER_BYTE);
for (s = 0; s < sseu->max_slices; s++) {
if ((s_en & BIT(s)) == 0)
@@ -115,7 +125,22 @@ static void gen11_compute_sseu_info(struct sseu_dev_info *sseu,
sseu->slice_mask |= BIT(s);
- intel_sseu_set_subslices(sseu, s, ss_en);
+ /*
+ * XeHP introduces the concept of compute vs geometry DSS. To
+ * reduce variation between GENs around subslice usage, store a
+ * mask for both the geometry and compute enabled masks since
+ * userspace will need to be able to query these masks
+ * independently. Also compute a total enabled subslice count
+ * for the purposes of selecting subslices to use in a
+ * particular GEM context.
+ */
+ intel_sseu_set_subslices(sseu, s, sseu->compute_subslice_mask,
+ get_ss_stride_mask(sseu, s, c_ss_en));
+ intel_sseu_set_subslices(sseu, s, sseu->geometry_subslice_mask,
+ get_ss_stride_mask(sseu, s, g_ss_en));
+ intel_sseu_set_subslices(sseu, s, sseu->subslice_mask,
+ get_ss_stride_mask(sseu, s,
+ g_ss_en | c_ss_en));
for (ss = 0; ss < sseu->max_subslices; ss++)
if (intel_sseu_has_subslice(sseu, s, ss))
@@ -129,7 +154,7 @@ static void gen12_sseu_info_init(struct intel_gt *gt)
{
struct sseu_dev_info *sseu = >->info.sseu;
struct intel_uncore *uncore = gt->uncore;
- u32 dss_en;
+ u32 g_dss_en, c_dss_en = 0;
u16 eu_en = 0;
u8 eu_en_fuse;
u8 s_en;
@@ -145,10 +170,12 @@ static void gen12_sseu_info_init(struct intel_gt *gt)
* across the entire device. Then calculate out the DSS for each
* workload type within that software slice.
*/
- if (IS_DG2(gt->i915) || IS_XEHPSDV(gt->i915))
+ if (IS_DG2(gt->i915) || IS_XEHPSDV(gt->i915)) {
intel_sseu_set_info(sseu, 1, 32, 16);
- else
+ sseu->has_compute_dss = 1;
+ } else {
intel_sseu_set_info(sseu, 1, 6, 16);
+ }
sseu->has_compute_dss doesn't make much sense. It will only be used
locally.
Even after the UAPI change to add a query, it still doesn't make sense
as the query could just be checking the compute mask != 0 to decide if
it has compute dss or not.
So, I'd ditch this sseu->has_compute_dss and replace it with a local
variable.
/*
* As mentioned above, Xe_HP does not have the concept of a slice.
@@ -160,7 +187,9 @@ static void gen12_sseu_info_init(struct intel_gt *gt)
s_en = intel_uncore_read(uncore, GEN11_GT_SLICE_ENABLE) &
GEN11_GT_S_ENA_MASK;
- dss_en = intel_uncore_read(uncore, GEN12_GT_DSS_ENABLE);
+ g_dss_en = intel_uncore_read(uncore, GEN12_GT_GEOMETRY_DSS_ENABLE);
+ if (sseu->has_compute_dss)
+ c_dss_en = intel_uncore_read(uncore, GEN12_GT_COMPUTE_DSS_ENABLE);
/* one bit per pair of EUs */
if (GRAPHICS_VER_FULL(gt->i915) >= IP_VER(12, 50))
@@ -173,7 +202,7 @@ static void gen12_sseu_info_init(struct intel_gt *gt)
if (eu_en_fuse & BIT(eu))
eu_en |= BIT(eu * 2) | BIT(eu * 2 + 1);
- gen11_compute_sseu_info(sseu, s_en, dss_en, eu_en);
+ gen11_compute_sseu_info(sseu, s_en, g_dss_en, c_dss_en, eu_en);
/* TGL only supports slice-level power gating */
sseu->has_slice_pg = 1;
@@ -199,7 +228,7 @@ static void gen11_sseu_info_init(struct intel_gt *gt)
eu_en = ~(intel_uncore_read(uncore, GEN11_EU_DISABLE) &
GEN11_EU_DIS_MASK);
- gen11_compute_sseu_info(sseu, s_en, ss_en, eu_en);
+ gen11_compute_sseu_info(sseu, s_en, ss_en, 0, eu_en);
/* ICL has no power gating restrictions. */
sseu->has_slice_pg = 1;
@@ -240,7 +269,7 @@ static void cherryview_sseu_info_init(struct intel_gt *gt)
sseu_set_eus(sseu, 0, 1, ~disabled_mask);
}
- intel_sseu_set_subslices(sseu, 0, subslice_mask);
+ intel_sseu_set_subslices(sseu, 0, sseu->subslice_mask, subslice_mask);
sseu->eu_total = compute_eu_total(sseu);
@@ -296,7 +325,8 @@ static void gen9_sseu_info_init(struct intel_gt *gt)
/* skip disabled slice */
continue;
- intel_sseu_set_subslices(sseu, s, subslice_mask);
+ intel_sseu_set_subslices(sseu, s, sseu->subslice_mask,
+ subslice_mask);
eu_disable = intel_uncore_read(uncore, GEN9_EU_DISABLE(s));
for (ss = 0; ss < sseu->max_subslices; ss++) {
@@ -408,7 +438,8 @@ static void bdw_sseu_info_init(struct intel_gt *gt)
/* skip disabled slice */
continue;
- intel_sseu_set_subslices(sseu, s, subslice_mask);
+ intel_sseu_set_subslices(sseu, s, sseu->subslice_mask,
+ subslice_mask);
for (ss = 0; ss < sseu->max_subslices; ss++) {
u8 eu_disabled_mask;
@@ -506,7 +537,8 @@ static void hsw_sseu_info_init(struct intel_gt *gt)
sseu->eu_per_subslice);
for (s = 0; s < sseu->max_slices; s++) {
- intel_sseu_set_subslices(sseu, s, subslice_mask);
+ intel_sseu_set_subslices(sseu, s, sseu->subslice_mask,
+ subslice_mask);
for (ss = 0; ss < sseu->max_subslices; ss++) {
sseu_set_eus(sseu, s, ss,
diff --git a/drivers/gpu/drm/i915/gt/intel_sseu.h b/drivers/gpu/drm/i915/gt/intel_sseu.h
index 0270acdcc157..61944829d069 100644
--- a/drivers/gpu/drm/i915/gt/intel_sseu.h
+++ b/drivers/gpu/drm/i915/gt/intel_sseu.h
@@ -32,6 +32,8 @@ struct drm_printer;
struct sseu_dev_info {
u8 slice_mask;
u8 subslice_mask[GEN_MAX_SLICES * GEN_MAX_SUBSLICE_STRIDE];
+ u8 geometry_subslice_mask[GEN_MAX_SLICES * GEN_MAX_SUBSLICE_STRIDE];
+ u8 compute_subslice_mask[GEN_MAX_SLICES * GEN_MAX_SUBSLICE_STRIDE];
u8 eu_mask[GEN_MAX_SLICES * GEN_MAX_SUBSLICES * GEN_MAX_EU_STRIDE];
u16 eu_total;
u8 eu_per_subslice;
@@ -41,6 +43,7 @@ struct sseu_dev_info {
u8 has_slice_pg:1;
u8 has_subslice_pg:1;
u8 has_eu_pg:1;
+ u8 has_compute_dss:1;
as mentioned above, this should be gone
/* Topology fields */
u8 max_slices;
@@ -104,7 +107,7 @@ intel_sseu_subslices_per_slice(const struct sseu_dev_info *sseu, u8 slice);
u32 intel_sseu_get_subslices(const struct sseu_dev_info *sseu, u8 slice);
void intel_sseu_set_subslices(struct sseu_dev_info *sseu, int slice,
- u32 ss_mask);
+ u8 *subslice_mask, u32 ss_mask);
void intel_sseu_info_init(struct intel_gt *gt);
diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h
index 8bfd646fc403..f8d3cd11eced 100644
--- a/drivers/gpu/drm/i915/i915_reg.h
+++ b/drivers/gpu/drm/i915/i915_reg.h
@@ -3151,7 +3151,8 @@ static inline bool i915_mmio_reg_valid(i915_reg_t reg)
#define GEN11_GT_SUBSLICE_DISABLE _MMIO(0x913C)
-#define GEN12_GT_DSS_ENABLE _MMIO(0x913C)
+#define GEN12_GT_GEOMETRY_DSS_ENABLE _MMIO(0x913C)
+#define GEN12_GT_COMPUTE_DSS_ENABLE _MMIO(0x9144)
#define XEHP_EU_ENABLE _MMIO(0x9134)
#define XEHP_EU_ENA_MASK 0xFF
diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
index bde5860b3686..fb36e897cc86 100644
--- a/include/uapi/drm/i915_drm.h
+++ b/include/uapi/drm/i915_drm.h
@@ -2609,9 +2609,6 @@ struct drm_i915_query {
* Z / 8] >> (Z % 8)) & 1
*/
struct drm_i915_query_topology_info {
- /*
- * Unused for now. Must be cleared to zero.
- */
without uapi changes we shouldn't change the uapi headers.
thanks
Lucas De Marchi
__u16 flags;
__u16 max_slices;
--
2.25.4
