Re: [PATCH v12 3/4] drm/i915: Add WABB blit for Wa_16018031267 / Wa_16018063123

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




On 10/18/2023 3:24 PM, Andrzej Hajda wrote:
On 20.09.2023 23:07, Jonathan Cavitt wrote:
Apply WABB blit for Wa_16018031267 / Wa_16018063123.
Additionally, update the lrc selftest to exercise the new
WABB changes.

Co-developed-by: Nirmoy Das <nirmoy.das@xxxxxxxxx>
Signed-off-by: Jonathan Cavitt <jonathan.cavitt@xxxxxxxxx> > ---
  drivers/gpu/drm/i915/gt/intel_engine_regs.h |   3 +
  drivers/gpu/drm/i915/gt/intel_gt.h          |   4 +
  drivers/gpu/drm/i915/gt/intel_gt_types.h    |   2 +
  drivers/gpu/drm/i915/gt/intel_lrc.c         | 100 +++++++++++++++++++-
  drivers/gpu/drm/i915/gt/selftest_lrc.c      |  65 +++++++++----
  5 files changed, 153 insertions(+), 21 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/intel_engine_regs.h b/drivers/gpu/drm/i915/gt/intel_engine_regs.h
index fdd4ddd3a978a..b8618ee3e3041 100644
--- a/drivers/gpu/drm/i915/gt/intel_engine_regs.h
+++ b/drivers/gpu/drm/i915/gt/intel_engine_regs.h
@@ -118,6 +118,9 @@
  #define   CCID_EXTENDED_STATE_RESTORE        BIT(2)
  #define   CCID_EXTENDED_STATE_SAVE        BIT(3)
  #define RING_BB_PER_CTX_PTR(base)        _MMIO((base) + 0x1c0) /* gen8+ */
+#define   PER_CTX_BB_FORCE            BIT(2)
+#define   PER_CTX_BB_VALID            BIT(0)
+
  #define RING_INDIRECT_CTX(base)            _MMIO((base) + 0x1c4) /* gen8+ */   #define RING_INDIRECT_CTX_OFFSET(base)        _MMIO((base) + 0x1c8) /* gen8+ */
  #define ECOSKPD(base)                _MMIO((base) + 0x1d0)
diff --git a/drivers/gpu/drm/i915/gt/intel_gt.h b/drivers/gpu/drm/i915/gt/intel_gt.h
index 239848bcb2a42..40cc0005dd735 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt.h
+++ b/drivers/gpu/drm/i915/gt/intel_gt.h
@@ -83,6 +83,10 @@ struct drm_printer;
            ##__VA_ARGS__);                    \
  } while (0)
  +#define NEEDS_FASTCOLOR_BLT_WABB(engine) ( \
+    IS_GFX_GT_IP_RANGE(engine->gt, IP_VER(12, 55), IP_VER(12, 71)) && \
+    engine->class == COPY_ENGINE_CLASS)
+
  static inline bool gt_is_root(struct intel_gt *gt)
  {
      return !gt->info.id;
diff --git a/drivers/gpu/drm/i915/gt/intel_gt_types.h b/drivers/gpu/drm/i915/gt/intel_gt_types.h
index def7dd0eb6f19..4917633f299dd 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt_types.h
+++ b/drivers/gpu/drm/i915/gt/intel_gt_types.h
@@ -307,6 +307,8 @@ enum intel_gt_scratch_field {
        /* 8 bytes */
      INTEL_GT_SCRATCH_FIELD_COHERENTL3_WA = 256,
+
+    INTEL_GT_SCRATCH_FIELD_DUMMY_BLIT = 384,
  };
    #define intel_gt_support_legacy_fencing(gt) ((gt)->ggtt->num_fences > 0) diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
index eaf66d9031665..db7088abeef38 100644
--- a/drivers/gpu/drm/i915/gt/intel_lrc.c
+++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
@@ -828,6 +828,18 @@ lrc_ring_indirect_offset_default(const struct intel_engine_cs *engine)
      return 0;
  }
  +static void
+lrc_setup_bb_per_ctx(u32 *regs,
+             const struct intel_engine_cs *engine,
+             u32 ctx_bb_ggtt_addr)
+{
+    GEM_BUG_ON(lrc_ring_wa_bb_per_ctx(engine) == -1);
+    regs[lrc_ring_wa_bb_per_ctx(engine) + 1] =
+        ctx_bb_ggtt_addr |
+        PER_CTX_BB_FORCE |
+        PER_CTX_BB_VALID;
+}
+
  static void
  lrc_setup_indirect_ctx(u32 *regs,
                 const struct intel_engine_cs *engine,
@@ -1020,7 +1032,13 @@ static u32 context_wa_bb_offset(const struct intel_context *ce)
      return PAGE_SIZE * ce->wa_bb_page;
  }
  -static u32 *context_indirect_bb(const struct intel_context *ce)
+/*
+ * per_ctx below determines which WABB section is used.
+ * When true, the function returns the location of the
+ * PER_CTX_BB.  When false, the function returns the
+ * location of the INDIRECT_CTX.
+ */
+static u32 *context_wabb(const struct intel_context *ce, bool per_ctx)
  {
      void *ptr;
  @@ -1029,6 +1047,7 @@ static u32 *context_indirect_bb(const struct intel_context *ce)
      ptr = ce->lrc_reg_state;
      ptr -= LRC_STATE_OFFSET; /* back to start of context image */
      ptr += context_wa_bb_offset(ce);
+    ptr += per_ctx ? PAGE_SIZE : 0;
        return ptr;
  }
@@ -1105,7 +1124,8 @@ __lrc_alloc_state(struct intel_context *ce, struct intel_engine_cs *engine)
        if (GRAPHICS_VER(engine->i915) >= 12) {
          ce->wa_bb_page = context_size / PAGE_SIZE;
-        context_size += PAGE_SIZE;
+        /* INDIRECT_CTX and PER_CTX_BB need separate pages. */
+        context_size += PAGE_SIZE * 2;
      }
        if (intel_context_is_parent(ce) && intel_engine_uses_guc(engine)) { @@ -1407,12 +1427,85 @@ gen12_emit_indirect_ctx_xcs(const struct intel_context *ce, u32 *cs)
      return gen12_emit_aux_table_inv(ce->engine, cs);
  }
  +static u32 *xehp_emit_fastcolor_blt_wabb(const struct intel_context *ce, u32 *cs)
+{
+    struct intel_gt *gt = ce->engine->gt;
+    int mocs = gt->mocs.uc_index << 1;
+
+    /**
+     * Wa_16018031267 / Wa_16018063123 requires that SW forces the
+     * main copy engine arbitration into round robin mode.  We
+     * additionally need to submit the following WABB blt command
+     * to produce 4 subblits with each subblit generating 0 byte

It is been a while I looked into the WA. I don't think this can be 0 byte.


+     * write requests as WABB:
+     *
+     * XY_FASTCOLOR_BLT
+     *  BG0    -> 5100000E
+     *  BG1    -> 0000003F (Dest pitch)
+     *  BG2    -> 00000000 (X1, Y1) = (0, 0)
+     *  BG3    -> 00040001 (X2, Y2) = (1, 4)
+     *  BG4    -> scratch
+     *  BG5    -> scratch
+     *  BG6-12 -> 00000000
+     *  BG13   -> 20004004 (Surf. Width= 2,Surf. Height = 5 )
+     *  BG14   -> 00000010 (Qpitch = 4)
+     *  BG15   -> 00000000
+     */
+    *cs++ = XY_FAST_COLOR_BLT_CMD | (16 - 2);
+    *cs++ = FIELD_PREP(XY_FAST_COLOR_BLT_MOCS_MASK, mocs) | 0x3f;
+    *cs++ = 0;
+    *cs++ = 4 << 16 | 1;
+    *cs++ = lower_32_bits(ce->vm->rsvd.start);
+    *cs++ = upper_32_bits(ce->vm->rsvd.start);


If my above doubts holds then this address should exist so the copy engine can do a copy operation. Now sure how this will

work out if this is with NULL_PTE.

Andrzej, could you please validate this once.


Regards,

Nirmoy


+    *cs++ = 0;
+    *cs++ = 0;
+    *cs++ = 0;
+    *cs++ = 0;
+    *cs++ = 0;
+    *cs++ = 0;
+    *cs++ = 0;
+    *cs++ = 0x20004004;
+    *cs++ = 0x10;
+    *cs++ = 0;
+
+    return cs;
+}
+
+static u32 *
+xehp_emit_per_ctx_bb(const struct intel_context *ce, u32 *cs)
+{
+    /* Wa_16018031267, Wa_16018063123 */
+    if (NEEDS_FASTCOLOR_BLT_WABB(ce->engine))
+        cs = xehp_emit_fastcolor_blt_wabb(ce, cs);
+
+    return cs;
+}
+
+static void
+setup_per_ctx_bb(const struct intel_context *ce,
+         const struct intel_engine_cs *engine,
+         u32 *(*emit)(const struct intel_context *, u32 *))
+{
+    /* Place PER_CTX_BB on next page after INDIRECT_CTX */
+    u32 * const start = context_wabb(ce, true);
+    u32 *cs;
+
+    cs = emit(ce, start);
+
+    /* PER_CTX_BB must manually terminate */
+    *cs++ = MI_BATCH_BUFFER_END;
+
+    GEM_BUG_ON(cs - start > I915_GTT_PAGE_SIZE / sizeof(*cs));
+    lrc_setup_bb_per_ctx(ce->lrc_reg_state, engine,
+                 lrc_indirect_bb(ce) + PAGE_SIZE);
+}
+
  static void
  setup_indirect_ctx_bb(const struct intel_context *ce,
                const struct intel_engine_cs *engine,
                u32 *(*emit)(const struct intel_context *, u32 *))
  {
-    u32 * const start = context_indirect_bb(ce);
+    u32 * const start = context_wabb(ce, false);
      u32 *cs;
        cs = emit(ce, start);
@@ -1511,6 +1604,7 @@ u32 lrc_update_regs(const struct intel_context *ce,
          /* Mutually exclusive wrt to global indirect bb */
          GEM_BUG_ON(engine->wa_ctx.indirect_ctx.size);
          setup_indirect_ctx_bb(ce, engine, fn);
+        setup_per_ctx_bb(ce, engine, xehp_emit_per_ctx_bb);
      }
        return lrc_descriptor(ce) | CTX_DESC_FORCE_RESTORE;
diff --git a/drivers/gpu/drm/i915/gt/selftest_lrc.c b/drivers/gpu/drm/i915/gt/selftest_lrc.c

Looks OK.
Shall selftest be in the same commit, ie is spilting breaking sth?
If not maybe separate commit would be better.

Regards
Andrzej


index 5f826b6dcf5d6..e17b8777d21dc 100644
--- a/drivers/gpu/drm/i915/gt/selftest_lrc.c
+++ b/drivers/gpu/drm/i915/gt/selftest_lrc.c
@@ -1555,7 +1555,7 @@ static int live_lrc_isolation(void *arg)
      return err;
  }
  -static int indirect_ctx_submit_req(struct intel_context *ce)
+static int wabb_ctx_submit_req(struct intel_context *ce)
  {
      struct i915_request *rq;
      int err = 0;
@@ -1579,7 +1579,8 @@ static int indirect_ctx_submit_req(struct intel_context *ce)
  #define CTX_BB_CANARY_INDEX  (CTX_BB_CANARY_OFFSET / sizeof(u32))
    static u32 *
-emit_indirect_ctx_bb_canary(const struct intel_context *ce, u32 *cs)
+emit_wabb_ctx_canary(const struct intel_context *ce,
+             u32 *cs, bool per_ctx)
  {
      *cs++ = MI_STORE_REGISTER_MEM_GEN8 |
          MI_SRM_LRM_GLOBAL_GTT |
@@ -1587,26 +1588,43 @@ emit_indirect_ctx_bb_canary(const struct intel_context *ce, u32 *cs)
      *cs++ = i915_mmio_reg_offset(RING_START(0));
      *cs++ = i915_ggtt_offset(ce->state) +
          context_wa_bb_offset(ce) +
-        CTX_BB_CANARY_OFFSET;
+        CTX_BB_CANARY_OFFSET +
+        (per_ctx ? PAGE_SIZE : 0);
      *cs++ = 0;
        return cs;
  }
  +static u32 *
+emit_indirect_ctx_bb_canary(const struct intel_context *ce, u32 *cs)
+{
+    return emit_wabb_ctx_canary(ce, cs, false);
+}
+
+static u32 *
+emit_per_ctx_bb_canary(const struct intel_context *ce, u32 *cs)
+{
+    return emit_wabb_ctx_canary(ce, cs, true);
+}
+
  static void
-indirect_ctx_bb_setup(struct intel_context *ce)
+wabb_ctx_setup(struct intel_context *ce, bool per_ctx)
  {
-    u32 *cs = context_indirect_bb(ce);
+    u32 *cs = context_wabb(ce, per_ctx);
        cs[CTX_BB_CANARY_INDEX] = 0xdeadf00d;
  -    setup_indirect_ctx_bb(ce, ce->engine, emit_indirect_ctx_bb_canary);
+    if (per_ctx)
+        setup_per_ctx_bb(ce, ce->engine, emit_per_ctx_bb_canary);
+    else
+        setup_indirect_ctx_bb(ce, ce->engine, emit_indirect_ctx_bb_canary);
  }
  -static bool check_ring_start(struct intel_context *ce)
+static bool check_ring_start(struct intel_context *ce, bool per_ctx)
  {
      const u32 * const ctx_bb = (void *)(ce->lrc_reg_state) -
-        LRC_STATE_OFFSET + context_wa_bb_offset(ce);
+        LRC_STATE_OFFSET + context_wa_bb_offset(ce) +
+        (per_ctx ? PAGE_SIZE : 0);
        if (ctx_bb[CTX_BB_CANARY_INDEX] == ce->lrc_reg_state[CTX_RING_START])
          return true;
@@ -1618,21 +1636,21 @@ static bool check_ring_start(struct intel_context *ce)
      return false;
  }
  -static int indirect_ctx_bb_check(struct intel_context *ce)
+static int wabb_ctx_check(struct intel_context *ce, bool per_ctx)
  {
      int err;
  -    err = indirect_ctx_submit_req(ce);
+    err = wabb_ctx_submit_req(ce);
      if (err)
          return err;
  -    if (!check_ring_start(ce))
+    if (!check_ring_start(ce, per_ctx))
          return -EINVAL;
        return 0;
  }
  -static int __live_lrc_indirect_ctx_bb(struct intel_engine_cs *engine)
+static int __lrc_wabb_ctx(struct intel_engine_cs *engine, bool per_ctx)
  {
      struct intel_context *a, *b;
      int err;
@@ -1667,14 +1685,14 @@ static int __live_lrc_indirect_ctx_bb(struct intel_engine_cs *engine)        * As ring start is restored apriori of starting the indirect ctx bb and        * as it will be different for each context, it fits to this purpose.
       */
-    indirect_ctx_bb_setup(a);
-    indirect_ctx_bb_setup(b);
+    wabb_ctx_setup(a, per_ctx);
+    wabb_ctx_setup(b, per_ctx);
  -    err = indirect_ctx_bb_check(a);
+    err = wabb_ctx_check(a, per_ctx);
      if (err)
          goto unpin_b;
  -    err = indirect_ctx_bb_check(b);
+    err = wabb_ctx_check(b, per_ctx);
    unpin_b:
      intel_context_unpin(b);
@@ -1688,7 +1706,7 @@ static int __live_lrc_indirect_ctx_bb(struct intel_engine_cs *engine)
      return err;
  }
  -static int live_lrc_indirect_ctx_bb(void *arg)
+static int lrc_wabb_ctx(void *arg, bool per_ctx)
  {
      struct intel_gt *gt = arg;
      struct intel_engine_cs *engine;
@@ -1697,7 +1715,7 @@ static int live_lrc_indirect_ctx_bb(void *arg)
        for_each_engine(engine, gt, id) {
          intel_engine_pm_get(engine);
-        err = __live_lrc_indirect_ctx_bb(engine);
+        err = __lrc_wabb_ctx(engine, per_ctx);
          intel_engine_pm_put(engine);
            if (igt_flush_test(gt->i915))
@@ -1710,6 +1728,16 @@ static int live_lrc_indirect_ctx_bb(void *arg)
      return err;
  }
  +static int live_lrc_indirect_ctx_bb(void *arg)
+{
+    return lrc_wabb_ctx(arg, false);
+}
+
+static int live_lrc_per_ctx_bb(void *arg)
+{
+    return lrc_wabb_ctx(arg, true);
+}
+
  static void garbage_reset(struct intel_engine_cs *engine,
                struct i915_request *rq)
  {
@@ -1947,6 +1975,7 @@ int intel_lrc_live_selftests(struct drm_i915_private *i915)
          SUBTEST(live_lrc_garbage),
          SUBTEST(live_pphwsp_runtime),
          SUBTEST(live_lrc_indirect_ctx_bb),
+        SUBTEST(live_lrc_per_ctx_bb),
      };
        if (!HAS_LOGICAL_RING_CONTEXTS(i915))




[Index of Archives]     [AMD Graphics]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux