diff mbox series

[v5,04/10] drm/i915/perf: implement active wait for noa configurations

Message ID 20190627080045.8814-5-lionel.g.landwerlin@intel.com (mailing list archive)
State New, archived
Headers show
Series drm/i915: Vulkan performance query support | expand

Commit Message

Lionel Landwerlin June 27, 2019, 8 a.m. UTC
NOA configuration take some amount of time to apply. That amount of
time depends on the size of the GT. There is no documented time for
this. For example, past experimentations with powergating
configuration changes seem to indicate a 60~70us delay. We go with
500us as default for now which should be over the required amount of
time (according to HW architects).

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
---
 drivers/gpu/drm/i915/gt/intel_gpu_commands.h |  24 +++
 drivers/gpu/drm/i915/i915_debugfs.c          |  25 +++
 drivers/gpu/drm/i915/i915_drv.h              |   8 +
 drivers/gpu/drm/i915/i915_perf.c             | 178 ++++++++++++++++++-
 drivers/gpu/drm/i915/i915_reg.h              |   4 +-
 5 files changed, 236 insertions(+), 3 deletions(-)

Comments

Chris Wilson June 27, 2019, 8:27 a.m. UTC | #1
Quoting Lionel Landwerlin (2019-06-27 09:00:39)
> diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
> index 2f1dc9be3bfb..b2f5ba87921c 100644
> --- a/drivers/gpu/drm/i915/i915_perf.c
> +++ b/drivers/gpu/drm/i915/i915_perf.c
> @@ -410,6 +410,8 @@ static int alloc_oa_config_buffer(struct drm_i915_private *i915,
>         size_t config_length = 0;
>         u32 *cs;
>  
> +       lockdep_assert_held(&i915->drm.struct_mutex);

What inside *this* function depends on the struct_mutex? The noa_wait
vma is already pinned and I didn't spot anything else added that might
be contentious.

>         if (oa_config->mux_regs_len > 0) {
>                 config_length += DIV_ROUND_UP(oa_config->mux_regs_len,
>                                               MI_LOAD_REGISTER_IMM_MAX_REGS) * 4;
> @@ -425,7 +427,7 @@ static int alloc_oa_config_buffer(struct drm_i915_private *i915,
>                                               MI_LOAD_REGISTER_IMM_MAX_REGS) * 4;
>                 config_length += oa_config->flex_regs_len * 8;
>         }
> -       config_length += 4; /* MI_BATCH_BUFFER_END */
> +       config_length += 12; /* MI_BATCH_BUFFER_START into noa_wait loop */
>         config_length = ALIGN(config_length, I915_GTT_PAGE_SIZE);
>  
>         bo = i915_gem_object_create_shmem(i915, config_length);
> @@ -442,7 +444,12 @@ static int alloc_oa_config_buffer(struct drm_i915_private *i915,
>         cs = write_cs_mi_lri(cs, oa_config->b_counter_regs, oa_config->b_counter_regs_len);
>         cs = write_cs_mi_lri(cs, oa_config->flex_regs, oa_config->flex_regs_len);
>  
> -       *cs++ = MI_BATCH_BUFFER_END;
> +
> +       /* Jump into the NOA wait busy loop. */
> +       *cs++ = (INTEL_GEN(i915) < 8 ?
> +                MI_BATCH_BUFFER_START : MI_BATCH_BUFFER_START_GEN8);
> +       *cs++ = i915->perf.oa.noa_wait->node.start;

i915_ggtt_offset(i915->perf.oa.noa_wait);
double checks that you are in the ggtt and so assuming u32 is valid.

This is operating inside the user context right? (If I remember
correctly, you plug into the user dispatch) If so, where do you save and
restore the CS_GPR overwritten?
-Chris
Lionel Landwerlin June 27, 2019, 8:52 a.m. UTC | #2
On 27/06/2019 11:27, Chris Wilson wrote:
> Quoting Lionel Landwerlin (2019-06-27 09:00:39)
>> diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
>> index 2f1dc9be3bfb..b2f5ba87921c 100644
>> --- a/drivers/gpu/drm/i915/i915_perf.c
>> +++ b/drivers/gpu/drm/i915/i915_perf.c
>> @@ -410,6 +410,8 @@ static int alloc_oa_config_buffer(struct drm_i915_private *i915,
>>          size_t config_length = 0;
>>          u32 *cs;
>>   
>> +       lockdep_assert_held(&i915->drm.struct_mutex);
> What inside *this* function depends on the struct_mutex? The noa_wait
> vma is already pinned and I didn't spot anything else added that might
> be contentious.


Thanks, will drop this.


>
>>          if (oa_config->mux_regs_len > 0) {
>>                  config_length += DIV_ROUND_UP(oa_config->mux_regs_len,
>>                                                MI_LOAD_REGISTER_IMM_MAX_REGS) * 4;
>> @@ -425,7 +427,7 @@ static int alloc_oa_config_buffer(struct drm_i915_private *i915,
>>                                                MI_LOAD_REGISTER_IMM_MAX_REGS) * 4;
>>                  config_length += oa_config->flex_regs_len * 8;
>>          }
>> -       config_length += 4; /* MI_BATCH_BUFFER_END */
>> +       config_length += 12; /* MI_BATCH_BUFFER_START into noa_wait loop */
>>          config_length = ALIGN(config_length, I915_GTT_PAGE_SIZE);
>>   
>>          bo = i915_gem_object_create_shmem(i915, config_length);
>> @@ -442,7 +444,12 @@ static int alloc_oa_config_buffer(struct drm_i915_private *i915,
>>          cs = write_cs_mi_lri(cs, oa_config->b_counter_regs, oa_config->b_counter_regs_len);
>>          cs = write_cs_mi_lri(cs, oa_config->flex_regs, oa_config->flex_regs_len);
>>   
>> -       *cs++ = MI_BATCH_BUFFER_END;
>> +
>> +       /* Jump into the NOA wait busy loop. */
>> +       *cs++ = (INTEL_GEN(i915) < 8 ?
>> +                MI_BATCH_BUFFER_START : MI_BATCH_BUFFER_START_GEN8);
>> +       *cs++ = i915->perf.oa.noa_wait->node.start;
> i915_ggtt_offset(i915->perf.oa.noa_wait);
> double checks that you are in the ggtt and so assuming u32 is valid.


Right, will do.


>
> This is operating inside the user context right? (If I remember
> correctly, you plug into the user dispatch) If so, where do you save and
> restore the CS_GPR overwritten?


Duh! In a previous iteration I was saving this in the scratch BO as a 
given location.

Should I alloc my own buffer?


Thanks a lot,


-Lionel


> -Chris
>
Chris Wilson June 27, 2019, 10:10 a.m. UTC | #3
Quoting Lionel Landwerlin (2019-06-27 09:52:34)
> On 27/06/2019 11:27, Chris Wilson wrote:
> > This is operating inside the user context right? (If I remember
> > correctly, you plug into the user dispatch) If so, where do you save and
> > restore the CS_GPR overwritten?
> 
> 
> Duh! In a previous iteration I was saving this in the scratch BO as a 
> given location.
> 
> Should I alloc my own buffer?

You only need a few temporaries right? You can squeeze into
engine->gt->scratch. We need to mark out reserved fields in there!

Pick an offset in the upper 2k as a starting point.
-Chris
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
index e7eff9db343e..4a66af38c87b 100644
--- a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
+++ b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
@@ -151,6 +151,7 @@ 
 #define   MI_BATCH_GTT		    (2<<6) /* aliased with (1<<7) on gen4 */
 #define MI_BATCH_BUFFER_START_GEN8	MI_INSTR(0x31, 1)
 #define   MI_BATCH_RESOURCE_STREAMER (1<<10)
+#define   MI_BATCH_PREDICATE         (1 << 15) /* HSW+ on RCS only*/
 
 /*
  * 3D instructions used by the kernel
@@ -226,6 +227,29 @@ 
 #define   PIPE_CONTROL_DEPTH_CACHE_FLUSH		(1<<0)
 #define   PIPE_CONTROL_GLOBAL_GTT (1<<2) /* in addr dword */
 
+#define MI_MATH(x) MI_INSTR(0x1a, (x)-1)
+#define   MI_ALU_OP(op, src1, src2) (((op) << 20) | ((src1) << 10) | (src2))
+/* operands */
+#define   MI_ALU_OP_NOOP     0
+#define   MI_ALU_OP_LOAD     128
+#define   MI_ALU_OP_LOADINV  1152
+#define   MI_ALU_OP_LOAD0    129
+#define   MI_ALU_OP_LOAD1    1153
+#define   MI_ALU_OP_ADD      256
+#define   MI_ALU_OP_SUB      257
+#define   MI_ALU_OP_AND      258
+#define   MI_ALU_OP_OR       259
+#define   MI_ALU_OP_XOR      260
+#define   MI_ALU_OP_STORE    384
+#define   MI_ALU_OP_STOREINV 1408
+/* sources */
+#define   MI_ALU_SRC_REG(x)  (x) /* 0 -> 15 */
+#define   MI_ALU_SRC_SRCA    32
+#define   MI_ALU_SRC_SRCB    33
+#define   MI_ALU_SRC_ACCU    49
+#define   MI_ALU_SRC_ZF      50
+#define   MI_ALU_SRC_CF      51
+
 /*
  * Commands used only by the command parser
  */
diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
index eeecdad0e3ca..6b49fda145e7 100644
--- a/drivers/gpu/drm/i915/i915_debugfs.c
+++ b/drivers/gpu/drm/i915/i915_debugfs.c
@@ -3646,6 +3646,30 @@  DEFINE_SIMPLE_ATTRIBUTE(i915_wedged_fops,
 			i915_wedged_get, i915_wedged_set,
 			"%llu\n");
 
+static int
+i915_perf_noa_delay_set(void *data, u64 val)
+{
+	struct drm_i915_private *i915 = data;
+
+	atomic64_set(&i915->perf.oa.noa_programming_delay, val);
+	return 0;
+}
+
+static int
+i915_perf_noa_delay_get(void *data, u64 *val)
+{
+	struct drm_i915_private *i915 = data;
+
+	*val = atomic64_read(&i915->perf.oa.noa_programming_delay);
+	return 0;
+}
+
+DEFINE_SIMPLE_ATTRIBUTE(i915_perf_noa_delay_fops,
+			i915_perf_noa_delay_get,
+			i915_perf_noa_delay_set,
+			"%llu\n");
+
+
 #define DROP_UNBOUND	BIT(0)
 #define DROP_BOUND	BIT(1)
 #define DROP_RETIRE	BIT(2)
@@ -4411,6 +4435,7 @@  static const struct i915_debugfs_files {
 	const char *name;
 	const struct file_operations *fops;
 } i915_debugfs_files[] = {
+	{"i915_perf_noa_delay", &i915_perf_noa_delay_fops},
 	{"i915_wedged", &i915_wedged_fops},
 	{"i915_cache_sharing", &i915_cache_sharing_fops},
 	{"i915_gem_drop_caches", &i915_drop_caches_fops},
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 333cc3c7690c..0323c8182a67 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1835,6 +1835,14 @@  struct drm_i915_private {
 
 			struct i915_oa_ops ops;
 			const struct i915_oa_format *oa_formats;
+
+			/**
+			 * A batch buffer doing a wait on the GPU for the NOA
+			 * logic to be reprogrammed.
+			 */
+			struct i915_vma *noa_wait;
+
+			atomic64_t noa_programming_delay;
 		} oa;
 	} perf;
 
diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
index 2f1dc9be3bfb..b2f5ba87921c 100644
--- a/drivers/gpu/drm/i915/i915_perf.c
+++ b/drivers/gpu/drm/i915/i915_perf.c
@@ -410,6 +410,8 @@  static int alloc_oa_config_buffer(struct drm_i915_private *i915,
 	size_t config_length = 0;
 	u32 *cs;
 
+	lockdep_assert_held(&i915->drm.struct_mutex);
+
 	if (oa_config->mux_regs_len > 0) {
 		config_length += DIV_ROUND_UP(oa_config->mux_regs_len,
 					      MI_LOAD_REGISTER_IMM_MAX_REGS) * 4;
@@ -425,7 +427,7 @@  static int alloc_oa_config_buffer(struct drm_i915_private *i915,
 					      MI_LOAD_REGISTER_IMM_MAX_REGS) * 4;
 		config_length += oa_config->flex_regs_len * 8;
 	}
-	config_length += 4; /* MI_BATCH_BUFFER_END */
+	config_length += 12; /* MI_BATCH_BUFFER_START into noa_wait loop */
 	config_length = ALIGN(config_length, I915_GTT_PAGE_SIZE);
 
 	bo = i915_gem_object_create_shmem(i915, config_length);
@@ -442,7 +444,12 @@  static int alloc_oa_config_buffer(struct drm_i915_private *i915,
 	cs = write_cs_mi_lri(cs, oa_config->b_counter_regs, oa_config->b_counter_regs_len);
 	cs = write_cs_mi_lri(cs, oa_config->flex_regs, oa_config->flex_regs_len);
 
-	*cs++ = MI_BATCH_BUFFER_END;
+
+	/* Jump into the NOA wait busy loop. */
+	*cs++ = (INTEL_GEN(i915) < 8 ?
+		 MI_BATCH_BUFFER_START : MI_BATCH_BUFFER_START_GEN8);
+	*cs++ = i915->perf.oa.noa_wait->node.start;
+	*cs++ = 0;
 
 	i915_gem_object_flush_map(bo);
 	i915_gem_object_unpin_map(bo);
@@ -1465,6 +1472,7 @@  static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
 	mutex_lock(&dev_priv->drm.struct_mutex);
 	dev_priv->perf.oa.exclusive_stream = NULL;
 	dev_priv->perf.oa.ops.disable_metric_set(dev_priv);
+	i915_vma_unpin_and_release(&dev_priv->perf.oa.noa_wait, 0);
 	mutex_unlock(&dev_priv->drm.struct_mutex);
 
 	free_oa_buffer(dev_priv);
@@ -1651,6 +1659,156 @@  static int alloc_oa_buffer(struct drm_i915_private *dev_priv)
 	return ret;
 }
 
+static int alloc_noa_wait(struct drm_i915_private *i915)
+{
+	struct drm_i915_gem_object *bo;
+	struct i915_vma *vma;
+	u64 delay_ns = atomic64_read(&i915->perf.oa.noa_programming_delay), delay_ticks;
+	u32 *batch, *cs, *jump;
+	int ret;
+
+	bo = i915_gem_object_create_shmem(i915, 4096);
+	if (IS_ERR(bo)) {
+		DRM_ERROR("Failed to allocate NOA wait batchbuffer\n");
+		return PTR_ERR(bo);
+	}
+
+	/*
+	 * We pin in GGTT because we jump into this buffer now because
+	 * multiple OA config BOs will have a jump to this address and it
+	 * needs to be fixed during the lifetime of the i915/perf stream.
+	 */
+	vma = i915_gem_object_ggtt_pin(bo, NULL, 0, 4096, 0);
+	if (IS_ERR(vma)) {
+		ret = PTR_ERR(vma);
+		goto err_unref;
+	}
+
+	batch = cs = i915_gem_object_pin_map(bo, I915_MAP_WB);
+	if (IS_ERR(batch)) {
+		ret = PTR_ERR(batch);
+		goto err_unpin;
+	}
+
+	/*
+	 * Initial snapshot of the timestamp register to implement the wait.
+	 * We work with 32b values, so clear out the top 32b bits of the
+	 * register because the ALU works 64bits.
+	 */
+	*cs++ = MI_LOAD_REGISTER_IMM(1);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(0)) + 4;
+	*cs++ = 0;
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(RENDER_RING_BASE));
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(0));
+
+	/*
+	 * This is the location we're going to jump back into until the
+	 * required amount of time has passed.
+	 */
+	jump = cs;
+
+	/*
+	 * Take another snapshot of the timestamp register. Take care to clear
+	 * up the top 32bits of CS_GPR(1) as we're using it for other
+	 * operations below.
+	 */
+	*cs++ = MI_LOAD_REGISTER_IMM(1);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(1)) + 4;
+	*cs++ = 0;
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(RENDER_RING_BASE));
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(1));
+
+	/*
+	 * Do a diff between the 2 timestamps and store the result back into
+	 * CS_GPR(1).
+	 */
+	*cs++ = MI_MATH(5);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_LOAD, MI_ALU_SRC_SRCA, MI_ALU_SRC_REG(1));
+	*cs++ = MI_ALU_OP(MI_ALU_OP_LOAD, MI_ALU_SRC_SRCB, MI_ALU_SRC_REG(0));
+	*cs++ = MI_ALU_OP(MI_ALU_OP_SUB, 0, 0);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_STORE, MI_ALU_SRC_REG(2), MI_ALU_SRC_ACCU);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_STORE, MI_ALU_SRC_REG(3), MI_ALU_SRC_CF);
+
+	/*
+	 * Transfer the carry flag (set to 1 if ts1 < ts0, meaning the
+	 * timestamp have rolled over the 32bits) into the predicate register
+	 * to be used for the predicated jump.
+	 */
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(3));
+	*cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1);
+
+	/* Restart from the beginning if we had timestamps roll over. */
+	*cs++ = (INTEL_GEN(i915) < 8 ?
+		 MI_BATCH_BUFFER_START : MI_BATCH_BUFFER_START_GEN8) |
+		MI_BATCH_PREDICATE;
+	*cs++ = vma->node.start;
+	*cs++ = 0;
+
+	/*
+	 * Now add the diff between to previous timestamps and add it to :
+	 *      (((1 * << 64) - 1) - delay_ns)
+	 *
+	 * When the Carry Flag contains 1 this means the elapsed time is
+	 * longer than the expected delay, and we can exit the wait loop.
+	 */
+	delay_ticks = 0xffffffffffffffff -
+		DIV64_U64_ROUND_UP(delay_ns *
+				   RUNTIME_INFO(i915)->cs_timestamp_frequency_khz,
+				   1000000ull);
+	*cs++ = MI_LOAD_REGISTER_IMM(2);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(4));
+	*cs++ = lower_32_bits(delay_ticks);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(4)) + 4;
+	*cs++ = upper_32_bits(delay_ticks);
+
+	*cs++ = MI_MATH(4);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_LOAD, MI_ALU_SRC_SRCA, MI_ALU_SRC_REG(2));
+	*cs++ = MI_ALU_OP(MI_ALU_OP_LOAD, MI_ALU_SRC_SRCB, MI_ALU_SRC_REG(4));
+	*cs++ = MI_ALU_OP(MI_ALU_OP_ADD, 0, 0);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_STOREINV, MI_ALU_SRC_REG(5), MI_ALU_SRC_CF);
+
+	/*
+	 * Transfer the result into the predicate register to be used for the
+	 * predicated jump.
+	 */
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(5));
+	*cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1);
+
+	/* Predicate the jump.  */
+	*cs++ = (INTEL_GEN(i915) < 8 ?
+		 MI_BATCH_BUFFER_START : MI_BATCH_BUFFER_START_GEN8) |
+		MI_BATCH_PREDICATE;
+	*cs++ = vma->node.start + (jump - batch) * 4;
+	*cs++ = 0;
+
+	/* Clear the predicate register */
+	*cs++ = MI_LOAD_REGISTER_IMM(1);
+	*cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1);
+	*cs++ = 0;
+
+	/* And return to the ring. */
+	*cs++ = MI_BATCH_BUFFER_END;
+
+	i915_gem_object_flush_map(bo);
+	i915_gem_object_unpin_map(bo);
+
+	i915->perf.oa.noa_wait = vma;
+
+	return 0;
+
+err_unpin:
+	__i915_vma_unpin(vma);
+
+err_unref:
+	i915_gem_object_put(bo);
+
+	return ret;
+}
+
 static void config_oa_regs(struct drm_i915_private *dev_priv,
 			   const struct i915_oa_reg *regs,
 			   u32 n_regs)
@@ -2219,6 +2377,12 @@  static int i915_oa_stream_init(struct i915_perf_stream *stream,
 		goto err_config;
 	}
 
+	ret = alloc_noa_wait(dev_priv);
+	if (ret) {
+		DRM_DEBUG("Unable to allocate NOA wait batch buffer\n");
+		goto err_noa_wait_alloc;
+	}
+
 	/* PRM - observability performance counters:
 	 *
 	 *   OACONTROL, performance counter enable, note:
@@ -2271,6 +2435,13 @@  static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	intel_uncore_forcewake_put(&dev_priv->uncore, FORCEWAKE_ALL);
 	intel_runtime_pm_put(&dev_priv->runtime_pm, stream->wakeref);
 
+	mutex_lock(&dev_priv->drm.struct_mutex);
+	i915_vma_unpin_and_release(&dev_priv->perf.oa.noa_wait, 0);
+	mutex_unlock(&dev_priv->drm.struct_mutex);
+
+err_noa_wait_alloc:
+	i915_oa_config_put(stream->oa_config);
+
 err_config:
 	if (stream->ctx)
 		oa_put_render_ctx_id(stream);
@@ -3642,6 +3813,9 @@  void i915_perf_init(struct drm_i915_private *dev_priv)
 		mutex_init(&dev_priv->perf.metrics_lock);
 		idr_init(&dev_priv->perf.metrics_idr);
 
+		atomic64_set(&dev_priv->perf.oa.noa_programming_delay,
+			     500 * 1000 /* 500us */);
+
 		dev_priv->perf.initialized = true;
 	}
 }
diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h
index 7e6009cefb18..e12b2fccef70 100644
--- a/drivers/gpu/drm/i915/i915_reg.h
+++ b/drivers/gpu/drm/i915/i915_reg.h
@@ -567,7 +567,9 @@  static inline bool i915_mmio_reg_valid(i915_reg_t reg)
 #define MI_PREDICATE_SRC0_UDW	_MMIO(0x2400 + 4)
 #define MI_PREDICATE_SRC1	_MMIO(0x2408)
 #define MI_PREDICATE_SRC1_UDW	_MMIO(0x2408 + 4)
-
+#define MI_PREDICATE_DATA       _MMIO(0x2410)
+#define MI_PREDICATE_RESULT     _MMIO(0x2418)
+#define MI_PREDICATE_RESULT_1   _MMIO(0x241c)
 #define MI_PREDICATE_RESULT_2	_MMIO(0x2214)
 #define  LOWER_SLICE_ENABLED	(1 << 0)
 #define  LOWER_SLICE_DISABLED	(0 << 0)