Re: [Freedreno] [PATCH] drm/msm/dpu: re-introduce dpu core revision to the catalog

Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx> · Tue, 6 Jun 2023 04:03:24 +0300

On 06/06/2023 03:55, Abhinav Kumar wrote:


On 6/3/2023 7:21 PM, Dmitry Baryshkov wrote:
On 31/05/2023 21:25, Abhinav Kumar wrote:


On 5/31/2023 3:07 AM, Dmitry Baryshkov wrote:
On 31/05/2023 06:05, Abhinav Kumar wrote:


On 5/30/2023 7:53 PM, Dmitry Baryshkov wrote:
On Wed, 31 May 2023 at 03:54, Abhinav Kumar 
<quic_abhinavk@xxxxxxxxxxx> wrote:

With [1] dpu core revision was dropped in favor of using the
compatible string from the device tree to select the dpu catalog
being used in the device.

This approach works well however also necessitates adding catalog
entries for small register level details as dpu capabilities and/or
features bloating the catalog unnecessarily. Examples include but
are not limited to data_compress, interrupt register set, widebus 
etc.

Introduce the dpu core revision back as an entry to the catalog 
so that
we can just use dpu revision checks and enable those bits which
should be enabled unconditionally and not controlled by a catalog
and also simplify the changes to do something like:

if (dpu_core_revision > xxxxx && dpu_core_revision < xxxxx)
         enable the bit;

Also, add some of the useful macros back to be able to use dpu core
revision effectively.

[1]: 
https://patchwork.freedesktop.org/patch/530891/?series=113910&rev=4

Signed-off-by: Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx>
---
  .../msm/disp/dpu1/catalog/dpu_3_0_msm8998.h   |  1 +
  .../msm/disp/dpu1/catalog/dpu_4_0_sdm845.h    |  1 +
  .../msm/disp/dpu1/catalog/dpu_5_0_sm8150.h    |  1 +
  .../msm/disp/dpu1/catalog/dpu_5_1_sc8180x.h   |  1 +
  .../msm/disp/dpu1/catalog/dpu_6_0_sm8250.h    |  1 +
  .../msm/disp/dpu1/catalog/dpu_6_2_sc7180.h    |  1 +
  .../msm/disp/dpu1/catalog/dpu_6_3_sm6115.h    |  1 +
  .../msm/disp/dpu1/catalog/dpu_6_5_qcm2290.h   |  1 +
  .../msm/disp/dpu1/catalog/dpu_7_0_sm8350.h    |  1 +
  .../msm/disp/dpu1/catalog/dpu_7_2_sc7280.h    |  1 +
  .../msm/disp/dpu1/catalog/dpu_8_0_sc8280xp.h  |  1 +
  .../msm/disp/dpu1/catalog/dpu_8_1_sm8450.h    |  1 +
  .../msm/disp/dpu1/catalog/dpu_9_0_sm8550.h    |  1 +
  .../gpu/drm/msm/disp/dpu1/dpu_hw_catalog.h    | 31 
++++++++++++++++++-
  14 files changed, 43 insertions(+), 1 deletion(-)


[skipped catalog changes]

diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_catalog.h 
b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_catalog.h
index 677048cc3b7d..cc4aa75a1219 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_catalog.h
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_hw_catalog.h
@@ -19,6 +19,33 @@
   */
  #define MAX_BLOCKS    12

+#define DPU_HW_VER(MAJOR, MINOR, STEP)\
+                 ((((unsigned int)MAJOR & 0xF) << 28) |\
+                 ((MINOR & 0xFFF) << 16) |\
+                 (STEP & 0xFFFF))
+
+#define DPU_HW_MAJOR(rev)((rev) >> 28)
+#define DPU_HW_MINOR(rev)(((rev) >> 16) & 0xFFF)
+#define DPU_HW_STEP(rev)((rev) & 0xFFFF)
+#define DPU_HW_MAJOR_MINOR(rev)((rev) >> 16)
+
+#define IS_DPU_MAJOR_MINOR_SAME(rev1, rev2)   \
+(DPU_HW_MAJOR_MINOR((rev1)) == DPU_HW_MAJOR_MINOR((rev2)))
+
+#define DPU_HW_VER_300 DPU_HW_VER(3, 0, 0) /* 8998 v1.0 */
+#define DPU_HW_VER_400 DPU_HW_VER(4, 0, 0) /* sdm845 v1.0 */
+#define DPU_HW_VER_500 DPU_HW_VER(5, 0, 0) /* sm8150 v1.0 */
+#define DPU_HW_VER_510 DPU_HW_VER(5, 1, 1) /* sc8180 */
+#define DPU_HW_VER_600 DPU_HW_VER(6, 0, 0) /* sm8250 */
+#define DPU_HW_VER_620 DPU_HW_VER(6, 2, 0) /* sc7180 v1.0 */
+#define DPU_HW_VER_630 DPU_HW_VER(6, 3, 0) /* sm6115|sm4250 */
+#define DPU_HW_VER_650 DPU_HW_VER(6, 5, 0) /* qcm2290|sm4125 */
+#define DPU_HW_VER_700 DPU_HW_VER(7, 0, 0) /* sm8350 */
+#define DPU_HW_VER_720 DPU_HW_VER(7, 2, 0) /* sc7280 */
+#define DPU_HW_VER_800 DPU_HW_VER(8, 0, 0) /* sc8280xp */
+#define DPU_HW_VER_810 DPU_HW_VER(8, 1, 0) /* sm8450 */
+#define DPU_HW_VER_900 DPU_HW_VER(9, 0, 0) /* sm8550 */

Instead of having defines for all SoCs (which can quickly become
unmanageable) and can cause merge conflicts, I'd suggest inlining all
the defines into respective catalog files.


Sure, that can be done.

Also, I'm not sure that the "step" should be a part of the catalog. I
know that this follows the hardware revision. However, please correct
me if I'm wrong, different step levels are used for revisions of the
same SoC. The original code that was reading the hw revision from the
hardware register, listed both 5.0.0 and 5.0.1 for sm8150.


This is one of the things i noticed while making this change.

Before the catalog rework, we used to handle even steps as we used 
to read that from the register and match it with the mdss_cfg 
handler. But after the rework, we dont handle steps anymore. Yes, 
you are right that different step levels are used for the revisions 
of the same SOC and so with that, i dont expect or atleast am not 
aware of DPU differences between steps but I am not able to rule it 
out.

So are you suggesting we drop step altogether and DPU_HW_VER() 
macro shall only handle major and minor versions? With the current 
chipsets I see, it should not make a difference . Its just that I 
am not sure if that will never happen.

Yes. The goal of this rework would be to drop generic features and 
to replace those checks with DPU-revision lookups. Correct?

Yes thats right.

I think that from this perspective having to handle toe step 
revision is a sign of an overkill. Having to handle the step 
revision is a sign of paltform feature (or mis-feature) rather than 
a generic DPU bit.


Not entirely. Lets not forget that at the moment even dpu_perf_cfg is 
part of the catalog. Even if in terms of major HW blocks steps 
shouldnt change, there is absolutely no guarantee that perf data cannot.

This is what is the sticking point for me which is holding me back 
against dropping step. Thoughts?

We usually do not support ES versions of the chips, only the final 
version. So supporting the perf data for earlier revisions is also not 
required.


ack, we will drop step in that case. and good to know about the ES 
versions.


In fact I suppose that even handling a minor revision would be an 
overkill. Why don't we start with .dpu_major instead of .core_rev? 
We can add .dpu_minor if/when required.


No, unfortunately we cannot drop minor version for sure. I am seeing 
examples in downstream code where some of the features are available 
after a minor verion as well.

Can you please give an example?


Yes, watchdog timer, intf reset counter are available only after DPU HW 
version 8.1 (not major version 8).

Hmm, IIRC, wd timer was available for ages. Was it moved together with 
the introduction of MDSS_PERIPH_0_REMOVED?

But anyway, I see your point. Let's have major and minor. I'd probably 
still ask for the separate major and minor fields, if you don't mind.

--
With best wishes
Dmitry







