MOCS versioning
Subsystems (3)
cover MOCS versioning
Copying the kernel commit message:
Starting with GEN9, Memory Object Control State (MOCS) becomes an index into a table as opposed to the direct programming within the command. The table has 62 usable entries (ie 6 bits can represent all settings), and each buffer type may use one of these 62 entries to describe cacheability type, and age (and some other less useful fields).
Because we hadn't dealt with MOCS settings like this, we didn't think ahead too well and have ended up with a mess for GEN9 (and soon GEN10) platform. The plan for for future platforms is that the ideal MOCS settings will be determined, defined, and written in the public PRMs. After this point, the i915.ko will absorb these settings and sometime afterwards flip the alpha switch. All driver releases without the final MOCS table must be considered alpha. Here on, userspace can assume the MOCS table is definitively done. There will be some reserved entries for 'oh shit' scenarios. This avoids versioning the MOCS table which leaves somewhat of a mess in userspace trying to handle arbitrarily many MOCS versions.
But we do have a mess on GEN9. In the beginning, the MOCS table entries were pre-populated by the hardware based on estimations made prior to tapeout and we could just use that. Subsequently much performance tuning was done to determine optimal settings that the i915 driver should load on top of the hardware defaults. That was posted last as v6 of the original per-engine MOCS settings:
https://patchwork.freedesktop.org/patch/53237/. Since the MOCS table is not context saved/restored, it isn't feasible to let userspace upload its own MOCS table. After a good amount of debate, it was decided that we'd utilize only the minimal set of entires in mesa anyway, and so we took only those entries for our MOCS entries.
Now we've come to the realization that indeed there are other MOCS entries which are more optimal for various buffer types and workloads. The problem is that the meaning of the indices is ABI (we assume index 0 is the uncached entry, and that there are only 3 entries total).
What this patch [simply] aims to do is expose a parameter to inform userspace which "version" of the table was loaded by i915. Upon sufficient data, new entries can be added, and the version can be bumped. For example, from my original mesa mocs branch:
commit c9b0481bce24af032386701de0266eb5bc24e988
Author: Ben Widawsky <[email protected]> Date: Fri Apr 8 10:21:16 2016 -0700
i965: Use PTE mocs
src/mesa/drivers/dri/i965/brw_mocs.c +6 -0
@@ -14,6 +14,9 @@/* Skylake: MOCS is now an index into an array of 62 different caching * configurations programmed by the kernel. */ + +/* TC=PTE, LeCC=PTE, LRUM=3, L3CC=WB */ +#define SKL_MOCS_PTE_PTE (3 << 1)/* TC=LLC/eLLC, LeCC=WB, LRUM=3, L3CC=WB */ #define SKL_MOCS_WB (2 << 1)/* TC=LLC/eLLC, LeCC=PTE, LRUM=3, L3CC=WB */
@@ -26,6 +29,9 @@ brw_mocs_get_control_state(const struct brw_context *brw,switch (brw->gen) {default :case 9 : +if (brw->intelScreen->mocs_version >1 ) +return SKL_MOCS_PTE_PTE ; +return type ==INTEL_MOCS_PTE ?SKL_MOCS_PTE :SKL_MOCS_WB ;case 8 :return type ==INTEL_MOCS_PTE ?BDW_MOCS_PTE :BDW_MOCS_WB ;
tl;dr: A versioned MOCS table will allow userspace to be aware of new and potentially interesting cacheability settings. Next GEN platforms will not be considered production worthy until the MOCS table is finalized.
Ben Widawsky (1):
drm/i915: Version the MOCS settings
drivers/gpu/drm/i915/i915_drv.c | 3 +++ drivers/gpu/drm/i915/i915_drv.h | 2 ++ drivers/gpu/drm/i915/i915_pci.c | 13 +++++++++---- include/uapi/drm/i915_drm.h | 8 ++++++++ 4 files changed, 22 insertions(+), 4 deletions(-)
Ben Widawsky (2):
intel: Merge latest i915 uapi
intel: Make driver aware of MOCS table version
src/intel/drm/i915_drm.h | 8 ++++++++ src/intel/vulkan/anv_device.c | 12 ++++++++++++ src/intel/vulkan/anv_private.h | 2 ++ src/mesa/drivers/dri/i915/intel_context.c | 7 ++++++- src/mesa/drivers/dri/i965/intel_screen.c | 14 ++++++++++++++ src/mesa/drivers/dri/i965/intel_screen.h | 2 ++ 6 files changed, 44 insertions(+), 1 deletion(-)
Signature · 5 lines
2.13.2 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH 1/1 drm/i915: Version the MOCS settings
From: Ben Widawsky <[email protected]>
Starting with GEN9, Memory Object Control State (MOCS) becomes an index into a table as opposed to the direct programming within the command. The table has 62 usable entries (ie 6 bits can represent all settings), and each buffer type may use one of these 62 entries to describe cacheability type, and age (and some other less useful fields).
Because we hadn't dealt with MOCS settings like this, we didn't think ahead too well and have ended up with a mess for GEN9 (and soon GEN10) platform. The plan for for future platforms is that the ideal MOCS settings will be determined, defined, and written in the public PRMs. After this point, the i915.ko will absorb these settings and sometime afterwards flip the alpha switch. All driver releases without the final MOCS table must be considered alpha. Here on, userspace can assume the MOCS table is definitively done. There will be some reserved entries for 'oh shit' scenarios. This avoids versioning the MOCS table which leaves somewhat of a mess in userspace trying to handle arbitrarily many MOCS versions.
But we do have a mess on GEN9. In the beginning, the MOCS table entries were pre-populated by the hardware based on estimations made prior to tapeout and we could just use that. Subsequently much performance tuning was done to determine optimal settings that the i915 driver should load on top of the hardware defaults. That was posted last as v6 of the original per-engine MOCS settings:
https://patchwork.freedesktop.org/patch/53237/. Since the MOCS table is not context saved/restored, it isn't feasible to let userspace upload its own MOCS table. After a good amount of debate, it was decided that we'd utilize only the minimal set of entires in mesa anyway, and so we took only those entries for our MOCS entries.
Now we've come to the realization that indeed there are other MOCS entries which are more optimal for various buffer types and workloads. The problem is that the meaning of the indices is ABI (we assume index 0 is the uncached entry, and that there are only 3 entries total).
What this patch [simply] aims to do is expose a parameter to inform userspace which "version" of the table was loaded by i915. Upon sufficient data, new entries can be added, and the version can be bumped. For example, from my original mesa mocs branch:
commit c9b0481bce24af032386701de0266eb5bc24e988
Author: Ben Widawsky <[email protected]> Date: Fri Apr 8 10:21:16 2016 -0700
i965: Use PTE mocs
src/mesa/drivers/dri/i965/brw_mocs.c +6 -0
@@ -14,6 +14,9 @@/* Skylake: MOCS is now an index into an array of 62 different caching * configurations programmed by the kernel. */ + +/* TC=PTE, LeCC=PTE, LRUM=3, L3CC=WB */ +#define SKL_MOCS_PTE_PTE (3 << 1)/* TC=LLC/eLLC, LeCC=WB, LRUM=3, L3CC=WB */ #define SKL_MOCS_WB (2 << 1)/* TC=LLC/eLLC, LeCC=PTE, LRUM=3, L3CC=WB */
@@ -26,6 +29,9 @@ brw_mocs_get_control_state(const struct brw_context *brw,switch (brw->gen) {default :case 9 : +if (brw->intelScreen->mocs_version >1 ) +return SKL_MOCS_PTE_PTE ; +return type ==INTEL_MOCS_PTE ?SKL_MOCS_PTE :SKL_MOCS_WB ;case 8 :return type ==INTEL_MOCS_PTE ?BDW_MOCS_PTE :BDW_MOCS_WB ;
tl;dr: A versioned MOCS table will allow userspace to be aware of new and potentially interesting cacheability settings. Next GEN platforms will not be considered production worthy until the MOCS table is finalized.
v2: Update 1.5 year old patch. Add comments. Update commit message.
--- drivers/gpu/drm/i915/i915_drv.c | 3 +++ drivers/gpu/drm/i915/i915_drv.h | 2 ++ drivers/gpu/drm/i915/i915_pci.c | 13 +++++++++---- include/uapi/drm/i915_drm.h | 8 ++++++++ 4 files changed, 22 insertions(+), 4 deletions(-)
drivers/gpu/drm/i915/i915_drv.c +3 -0
@@ -401,6 +401,9 @@ static int i915_getparam(struct drm_device *dev, void *data,if (!value)return -ENODEV ;break ; +case I915_PARAM_MOCS_TABLE_VERSION : + value =INTEL_INFO (dev_priv)->mocs_version; +break ;default :DRM_DEBUG ("Unknown parameter %d\n" , param->param);return -EINVAL ;
drivers/gpu/drm/i915/i915_drv.h +2 -0
@@ -859,6 +859,8 @@ struct intel_device_info {u16 degamma_lut_size;u16 gamma_lut_size; } color; + +u8 mocs_version; };struct intel_display_error_state ;
drivers/gpu/drm/i915/i915_pci.c +9 -4
@@ -356,7 +356,8 @@ static const struct intel_device_info intel_cherryview_info = { .platform = INTEL_SKYLAKE, \ .has_csr = 1, \ .has_guc = 1, \ - .ddb_size = 896 + .ddb_size = 896, \ + .mocs_version = MOCS_TABLE_VERSION static const struct intel_device_info intel_skylake_info = {SKL_PLATFORM ,
@@ -390,6 +391,7 @@ static const struct intel_device_info intel_skylake_gt3_info = { .has_full_ppgtt = 1, \ .has_full_48bit_ppgtt = 1, \ .has_reset_engine = 1, \ + .mocs_version =MOCS_TABLE_VERSION , \GEN_DEFAULT_PIPEOFFSETS , \IVB_CURSOR_OFFSETS , \BDW_COLORS
@@ -413,7 +415,8 @@ static const struct intel_device_info intel_geminilake_info = { .platform = INTEL_KABYLAKE, \ .has_csr = 1, \ .has_guc = 1, \ - .ddb_size = 896 + .ddb_size = 896, \ + .mocs_version = MOCS_TABLE_VERSION static const struct intel_device_info intel_kabylake_info = {KBL_PLATFORM ,
@@ -431,7 +434,8 @@ static const struct intel_device_info intel_kabylake_gt3_info = { .platform = INTEL_COFFEELAKE, \ .has_csr = 1, \ .has_guc = 1, \ - .ddb_size = 896 + .ddb_size = 896, \ + .mocs_version = MOCS_TABLE_VERSION static const struct intel_device_info intel_coffeelake_info = {CFL_PLATFORM ,
@@ -448,7 +452,8 @@ static const struct intel_device_info intel_cannonlake_info = { .platform = INTEL_CANNONLAKE, .gen = 10, .ddb_size = 1024, - .has_csr = 1, + .has_csr = 1, \ + .mocs_version = MOCS_TABLE_VERSION };/*
include/uapi/drm/i915_drm.h +8 -0
@@ -431,6 +431,14 @@ typedef struct drm_i915_irq_wait {*/ #define I915_PARAM_HAS_EXEC_BATCH_FIRST 48 +/* What version of the MOCS table we have. For GEN9 GPUs, the PRM defined +* non-optimal settings for the MOCS table. As a result, we were required to use a +* small subset, and later add new settings. This param allows userspace to +* determine which settings are there. +*/ +#define MOCS_TABLE_VERSION 1/* Build time MOCS table version */ +#define I915_PARAM_MOCS_TABLE_VERSION 49 + typedef struct drm_i915_getparam { __s32 param; /*
Signature · 6 lines
2.13.2 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH 1/1] drm/i915: Version the MOCS settings
Quoting Ben Widawsky (2017-07-07 00:27:01)
Quoted 11 lines
drivers/gpu/drm/i915/i915_drv.c | 3 +++ drivers/gpu/drm/i915/i915_drv.h | 2 ++ drivers/gpu/drm/i915/i915_pci.c | 13 +++++++++---- include/uapi/drm/i915_drm.h | 8 ++++++++ 4 files changed, 22 insertions(+), 4 deletions(-)drivers/gpu/drm/i915/i915_drv.c +0
-0
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data, · line 405
Quoted 5 lines
drivers/gpu/drm/i915/i915_drv.c +2
-0@@ -401,6 +401,9 @@ static int i915_getparam(struct drm_device *dev, void *data,if (!value)return -ENODEV ;break ; +case I915_PARAM_MOCS_TABLE_VERSION : + value =INTEL_INFO (dev_priv)->mocs_version;
If we use intel_mocs_get_table_version() we can put this magic number in intel_mocs.c next to the tables, where we can keep its history and hopefully be able to remember to update it.
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data,
Quoted 6 lines
drivers/gpu/drm/i915/i915_drv.c +6
-0+/* What version of the MOCS table we have. For GEN9 GPUs, the PRM defined +* non-optimal settings for the MOCS table. As a result, we were required to use a +* small subset, and later add new settings. This param allows userspace to +* determine which settings are there. +*/ +#define MOCS_TABLE_VERSION 1/* Build time MOCS table version */
How are you planing to share this? When we update we bump this number, and then mesa copies it across and uses it after verifying it as 0,1 on an old kernel.
I don't think you want to expose the updated constant here, but symbolic names for each version? (What would be the point?)
Next question, why a version number and not just the number of entries defined? Each index is defined by ABI once assigned, so the number of entries still operates as a version number and allows easy checking.
if (advanced_cacheing_idx < kernel_max_mocs) return advanced_cacheing_idx; if (default_cacheing_idx < kernel_max_mocs) return default_cacheing_idx; return follow_pte_idx;
give or take the smarts to choose the preferred indices for any particular scenario.
In the future, if we finally get user defined mocs, the table_size will then give the start of the user modifiable indices (presming they want to keep the predefined entries for compatibility?))
-Chris
_______________________________________________
mesa-dev mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH 1/1] drm/i915: Version the MOCS settings
Quoted 33 lines · Chris Wilson
Quoting Ben Widawsky (2017-07-07 00:27:01)
Quoted 11 lines
drivers/gpu/drm/i915/i915_drv.c | 3 +++ drivers/gpu/drm/i915/i915_drv.h | 2 ++ drivers/gpu/drm/i915/i915_pci.c | 13 +++++++++---- include/uapi/drm/i915_drm.h | 8 ++++++++ 4 files changed, 22 insertions(+), 4 deletions(-)drivers/gpu/drm/i915/i915_drv.c +0
-0drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data, · line 405
Quoted 5 lines
drivers/gpu/drm/i915/i915_drv.c +2
-0@@ -401,6 +401,9 @@ static int i915_getparam(struct drm_device *dev, void *data,if (!value)return -ENODEV ;break ; +case I915_PARAM_MOCS_TABLE_VERSION : + value =INTEL_INFO (dev_priv)->mocs_version;If we use
intel_mocs_get_table_version()we can put this magic number inintel_mocs.cnext to the tables, where we can keep its history and hopefully be able to remember to update it.drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data,
Quoted 6 lines
drivers/gpu/drm/i915/i915_drv.c +6
-0+/* What version of the MOCS table we have. For GEN9 GPUs, the PRM defined +* non-optimal settings for the MOCS table. As a result, we were required to use a +* small subset, and later add new settings. This param allows userspace to +* determine which settings are there. +*/ +#define MOCS_TABLE_VERSION 1/* Build time MOCS table version */ How are you planing to share this? When we update we bump this number, and then mesa copies it across and uses it after verifying it as 0,1 on an old kernel.
Quoted 3 lines
I don't think you want to expose the updated constant here, but symbolic names for each version? (What would be the point?)
FWIW I have to agree with Chris here - having the value is of limited use. Furthermore it mostly confuses people when writing the user space parts.
For example:
Mesa implements v1 and uses the define. Kernel headers get updated to v2 and Mesa supporting v1 gets rebuild against them.
Mesa stores/treats as the MOCS version has "v2" when the actual hardware/kernel supports "v1".
The expected issues vary depending on the implementation, but I suspect it won't be fun :-)
IMHO it's better if user space is explicit on the versions it supports and kernel should avoid exposing such defines unless really needed.
HTH
Emil
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Mesa-dev] [PATCH 1/1] drm/i915: Version the MOCS settings
In reply to Chris WilsonQuoted 30 lines · Chris Wilson
Quoting Ben Widawsky (2017-07-07 00:27:01)
Quoted 2 lines
drivers/gpu/drm/i915/i915_drv.c | 3 +++ drivers/gpu/drm/i915/i915_drv.h | 2 ++drivers/gpu/drm/i915/i915_pci.c | 13 +++++++++----Review
Quoted 4 lines
include/uapi/drm/i915_drm.h | 8 ++++++++ 4 files changed, 22 insertions(+), 4 deletions(-)(quoted diff) +0
-0
b/drivers/gpu/drm/i915/i915_drv.cdrivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev,
Quoted 4 lines
index
9167a73f3c69..26c27b6ae814100644drivers/gpu/drm/i915/i915_drv.c +0
-0@@ -401,6 +401,9 @@ static int i915_getparam(struct drm_device *dev,void *data,
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data, · line 405
Quoted 5 lines
drivers/gpu/drm/i915/i915_drv.c +2
-0if (!value)return -ENODEV ;break ; +case I915_PARAM_MOCS_TABLE_VERSION : + value =INTEL_INFO (dev_priv)->mocs_version;If we use
intel_mocs_get_table_version()we can put this magic number inintel_mocs.cnext to the tables, where we can keep its history and hopefully be able to remember to update it.drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, · line 406
Quoted 1 line
drivers/gpu/drm/i915/i915_drv.c +1
-0+/* What version of the MOCS table we have. For GEN9 GPUs, the PRM defined
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, · line 407
Quoted 1 line
drivers/gpu/drm/i915/i915_drv.c +1
-0+* non-optimal settings for the MOCS table. As a result, we were required to use a
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, · line 408
Quoted 1 line
drivers/gpu/drm/i915/i915_drv.c +1
-0+* small subset, and later add new settings. This param allows userspace to
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev,
Quoted 3 lines
drivers/gpu/drm/i915/i915_drv.c +3
-0+* determine which settings are there. +*/ +#define MOCS_TABLE_VERSION 1/* Build time MOCS table version */
How are you planing to share this? When we update we bump this number, and then mesa copies it across and uses it after verifying it as 0,1 on an old kernel.
Agreed. I don't see how having a #define for compile-time mocs version is useful. The compile-time version doesn't really matter and we wouldn't want to use that in i965/anv anyway (more on that in the other patch).
Quoted 14 lines
I don't think you want to expose the updated constant here, but symbolic names for each version? (What would be the point?)
Next question, why a version number and not just the number of entries defined? Each index is defined by ABI once assigned, so the number of entries still operates as a version number and allows easy checking.
if (advanced_cacheing_idx < kernel_max_mocs) return advanced_cacheing_idx; if (default_cacheing_idx < kernel_max_mocs) return default_cacheing_idx; return follow_pte_idx;
Quoted 3 lines
give or take the smarts to choose the preferred indices for any particular scenario.
I'll have to think about it a bit more but this sounds like a fairly good idea. I see two major benefits:
1. The kernel can return ARRAY_SIZE(mocs_table_for_your_gen) and we will
never forget to update it.
2. It makes the "does this MOCS value exist" check much easier. I imagine
future userspace code which chooses mocs values having some sort of "try and fall back" approach to making MOCS choices and this would be convenient.
That said, having it be a version may have it's advantages, I just don't know what they are yet.
--Jason
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Mesa-dev] [PATCH 1/1] drm/i915: Version the MOCS settings
Quoted 38 lines · Jason Ekstrand
Quoted 35 lines · Chris Wilson
Quoting Ben Widawsky (2017-07-07 00:27:01)
Quoted 2 lines
drivers/gpu/drm/i915/i915_drv.c | 3 +++ drivers/gpu/drm/i915/i915_drv.h | 2 ++drivers/gpu/drm/i915/i915_pci.c | 13 +++++++++----Review
Quoted 4 lines
include/uapi/drm/i915_drm.h | 8 ++++++++ 4 files changed, 22 insertions(+), 4 deletions(-)(quoted diff) +0
-0
b/drivers/gpu/drm/i915/i915_drv.cdrivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev,
Quoted 4 lines
index
9167a73f3c69..26c27b6ae814100644drivers/gpu/drm/i915/i915_drv.c +0
-0@@ -401,6 +401,9 @@ static int i915_getparam(struct drm_device *dev,void *data,
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data, · line 405
Quoted 5 lines
drivers/gpu/drm/i915/i915_drv.c +2
-0if (!value)return -ENODEV ;break ; +case I915_PARAM_MOCS_TABLE_VERSION : + value =INTEL_INFO (dev_priv)->mocs_version;If we use
intel_mocs_get_table_version()we can put this magic number inintel_mocs.cnext to the tables, where we can keep its history and hopefully be able to remember to update it.drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, · line 406
Quoted 1 line
drivers/gpu/drm/i915/i915_drv.c +1
-0+/* What version of the MOCS table we have. For GEN9 GPUs, the PRM defined
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, · line 407
Quoted 1 line
drivers/gpu/drm/i915/i915_drv.c +1
-0+* non-optimal settings for the MOCS table. As a result, we were required to use a
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, · line 408
Quoted 1 line
drivers/gpu/drm/i915/i915_drv.c +1
-0+* small subset, and later add new settings. This param allows userspace to
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev,
Quoted 3 lines
drivers/gpu/drm/i915/i915_drv.c +3
-0+* determine which settings are there. +*/ +#define MOCS_TABLE_VERSION 1/* Build time MOCS table version */
Quoted 10 lines
Quoted 4 lines
How are you planing to share this? When we update we bump this number, and then mesa copies it across and uses it after verifying it as 0,1 on an old kernel.
Agreed. I don't see how having a #define for compile-time mocs version is useful. The compile-time version doesn't really matter and we wouldn't want to use that in i965/anv anyway (more on that in the other patch).
I think we're all agreed here.
Quoted 27 lines
Quoted 14 lines
I don't think you want to expose the updated constant here, but symbolic names for each version? (What would be the point?)
Next question, why a version number and not just the number of entries defined? Each index is defined by ABI once assigned, so the number of entries still operates as a version number and allows easy checking.
if (advanced_cacheing_idx < kernel_max_mocs) return advanced_cacheing_idx; if (default_cacheing_idx < kernel_max_mocs) return default_cacheing_idx; return follow_pte_idx;Quoted 3 lines
give or take the smarts to choose the preferred indices for any particular scenario.
I'll have to think about it a bit more but this sounds like a fairly good idea. I see two major benefits:
1. The kernel can return
ARRAY_SIZE(mocs_table_for_your_gen)and we will never forget to update it.
2. It makes the "does this MOCS value exist" check much easier. I imagine future userspace code which chooses mocs values having some sort of "try and fall back" approach to making MOCS choices and this would be convenient.
Quoted 4 lines
That said, having it be a version may have it's advantages, I just don't know what they are yet.
--Jason
Please direct comments to my response to Chris if you have more. To me it's 6 one way, half dozen the other - and I believe version has a more direct meaning (and the ability to potentially, albeit a terrible idea, rewrite entries).
If people are going to block a review based on this, I will change it, but I'd rather not.
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH 1/1] drm/i915: Version the MOCS settings
In reply to Chris WilsonQuoted 7 lines · Chris Wilson
Quoting Ben Widawsky (2017-07-07 00:27:01)
Quoted 6 lines
drivers/gpu/drm/i915/i915_drv.c | 3 +++ drivers/gpu/drm/i915/i915_drv.h | 2 ++ drivers/gpu/drm/i915/i915_pci.c | 13 +++++++++---- include/uapi/drm/i915_drm.h | 8 ++++++++ 4 files changed, 22 insertions(+), 4 deletions(-)
Quoted 15 lines
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data, · line 405
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.c +2
-0@@ -401,6 +401,9 @@ static int i915_getparam(struct drm_device *dev, void *data,if (!value)return -ENODEV ;break ; +case I915_PARAM_MOCS_TABLE_VERSION : + value =INTEL_INFO (dev_priv)->mocs_version;If we use
intel_mocs_get_table_version()we can put this magic number inintel_mocs.cnext to the tables, where we can keep its history and hopefully be able to remember to update it.
Yeah, that seems like an improvement to me as well.
Quoted 11 lines
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data,
Quoted 6 lines
drivers/gpu/drm/i915/i915_drv.c +6
-0+/* What version of the MOCS table we have. For GEN9 GPUs, the PRM defined +* non-optimal settings for the MOCS table. As a result, we were required to use a +* small subset, and later add new settings. This param allows userspace to +* determine which settings are there. +*/ +#define MOCS_TABLE_VERSION 1/* Build time MOCS table version */ How are you planing to share this? When we update we bump this number, and then mesa copies it across and uses it after verifying it as 0,1 on an old kernel.
Quoted 3 lines
I don't think you want to expose the updated constant here, but symbolic names for each version? (What would be the point?)
At least one thing wrong here is we would need per GEN constants, which is maybe what you meant and I misunderstood. Assuming you had per GEN constants, which I don't like, I believe everything works out fine. So, I'll remove this compile time MOCS versioning.
Quoted 11 lines
Next question, why a version number and not just the number of entries defined? Each index is defined by ABI once assigned, so the number of entries still operates as a version number and allows easy checking.
if (advanced_cacheing_idx < kernel_max_mocs) return advanced_cacheing_idx; if (default_cacheing_idx < kernel_max_mocs) return default_cacheing_idx; return follow_pte_idx;
Quoted 7 lines
give or take the smarts to choose the preferred indices for any particular scenario.
In the future, if we finally get user defined mocs, the
table_sizewill then give the start of the user modifiable indices (presming they want to keep the predefined entries for compatibility?))
-Chris
Yes, I considered this as well. I see no difference really as to one versus the other. In fact, if you're to support multiple table versions, I think it's actually easier with a pure version:
switch (kernel_mocs_version) {
case 3:
return new_best_cacheing_index; case 2: return old_best_cacheing_index; case 1: return naive_best_index; }
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH 1/1] drm/i915: Version the MOCS settings
Quoting Ben Widawsky (2017-07-07 19:42:25)
Quoted 37 lines · Chris Wilson
Quoted 7 lines
Quoting Ben Widawsky (2017-07-07 00:27:01)
Quoted 6 lines
drivers/gpu/drm/i915/i915_drv.c | 3 +++ drivers/gpu/drm/i915/i915_drv.h | 2 ++ drivers/gpu/drm/i915/i915_pci.c | 13 +++++++++---- include/uapi/drm/i915_drm.h | 8 ++++++++ 4 files changed, 22 insertions(+), 4 deletions(-)Quoted 15 lines
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data, · line 405
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.c +2
-0@@ -401,6 +401,9 @@ static int i915_getparam(struct drm_device *dev, void *data,if (!value)return -ENODEV ;break ; +case I915_PARAM_MOCS_TABLE_VERSION : + value =INTEL_INFO (dev_priv)->mocs_version;If we use
intel_mocs_get_table_version()we can put this magic number inintel_mocs.cnext to the tables, where we can keep its history and hopefully be able to remember to update it.Yeah, that seems like an improvement to me as well.
Quoted 11 lines
drivers/gpu/drm/i915/i915_drv.c static int i915_getparam(struct drm_device *dev, void *data,
Quoted 6 lines
drivers/gpu/drm/i915/i915_drv.c +6
-0+/* What version of the MOCS table we have. For GEN9 GPUs, the PRM defined +* non-optimal settings for the MOCS table. As a result, we were required to use a +* small subset, and later add new settings. This param allows userspace to +* determine which settings are there. +*/ +#define MOCS_TABLE_VERSION 1/* Build time MOCS table version */ How are you planing to share this? When we update we bump this number, and then mesa copies it across and uses it after verifying it as 0,1 on an old kernel.
Quoted 8 lines
Quoted 3 lines
I don't think you want to expose the updated constant here, but symbolic names for each version? (What would be the point?)
At least one thing wrong here is we would need per GEN constants, which is maybe what you meant and I misunderstood. Assuming you had per GEN constants, which I don't like, I believe everything works out fine. So, I'll remove this compile time MOCS versioning.
I figured you were going towards per-gen versioning, which is kind of why I liked the idea of table size -- but that only makes sense if somehow the index has the same meaning across gen (which it won't).
Quoted 23 lines
Quoted 11 lines
Next question, why a version number and not just the number of entries defined? Each index is defined by ABI once assigned, so the number of entries still operates as a version number and allows easy checking.
if (advanced_cacheing_idx < kernel_max_mocs) return advanced_cacheing_idx; if (default_cacheing_idx < kernel_max_mocs) return default_cacheing_idx; return follow_pte_idx;Quoted 7 lines
give or take the smarts to choose the preferred indices for any particular scenario.
In the future, if we finally get user defined mocs, the
table_sizewill then give the start of the user modifiable indices (presming they want to keep the predefined entries for compatibility?))
-ChrisYes, I considered this as well. I see no difference really as to one versus the other. In fact, if you're to support multiple table versions, I think it's actually easier with a pure version:
Quoted 8 lines
switch (
kernel_mocs_version) {
case 3:return new_best_cacheing_index; case 2: return old_best_cacheing_index; case 1: return naive_best_index; }
Indeed 6 of one, half a dozen of the other. Whichever you pick, 3 years down the line you wish you picked the other. The big advantage of using an absolute version is that you can just stuff these into tables. Ok, I like that more, a version parameter (that may be per-gen) worksforme. -Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH 2/3 intel: Merge latest i915 uapi
In reply to Ben Widawsky--- src/intel/drm/i915_drm.h | 8 ++++++++ 1 file changed, 8 insertions(+)
src/intel/drm/i915_drm.h +8 -0
@@ -431,6 +431,14 @@ typedef struct drm_i915_irq_wait {*/ #define I915_PARAM_HAS_EXEC_BATCH_FIRST 48 +/* What version of the MOCS table we have. For GEN9 GPUs, the PRM defined +* non-optimal settings for the MOCS table. As a result, we were required to use a +* small subset, and later add new settings. This param allows userspace to +* determine which settings are there. +*/ +#define MOCS_TABLE_VERSION 1/* Build time MOCS table version */ +#define I915_PARAM_MOCS_TABLE_VERSION 49 + typedef struct drm_i915_getparam { __s32 param; /*
Signature · 6 lines
2.13.2 _______________________________________________ mesa-dev mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/mesa-dev
PATCH 3/3 intel: Make driver aware of MOCS table version
In reply to Ben WidawskyWe don't yet have optimal MOCS settings, but we have enough to know how to at least determine when we might have non-optimal settings within our driver.
--- src/intel/vulkan/anv_device.c | 12 ++++++++++++ src/intel/vulkan/anv_private.h | 2 ++ src/mesa/drivers/dri/i915/intel_context.c | 7 ++++++- src/mesa/drivers/dri/i965/intel_screen.c | 14 ++++++++++++++ src/mesa/drivers/dri/i965/intel_screen.h | 2 ++ 5 files changed, 36 insertions(+), 1 deletion(-)
src/intel/vulkan/anv_device.c +12 -0
@@ -368,6 +368,18 @@ anv_physical_device_init(struct anv_physical_device *device, device->info.max_cs_threads = max_cs_threads; } +if (device->info.gen >=9 ) { + device->mocs_version =anv_gem_get_param (fd, +I915_PARAM_MOCS_TABLE_VERSION ); +switch (device->mocs_version) { +default : +anv_perf_warn ("Kernel exposes newer MOCS table\n" ); +case 1 : +case 0 : + device->mocs_version =MOCS_TABLE_VERSION ; + } + } +brw_process_intel_debug_variable (); device->compiler =brw_compiler_create (NULL , &device->info);
src/intel/vulkan/anv_private.h +2 -0
@@ -684,6 +684,8 @@ struct anv_physical_device {uint32_t eu_total;uint32_t subslice_total; +uint8_t mocs_version; +struct {uint32_t type_count;struct anv_memory_type types[VK_MAX_MEMORY_TYPES ];
src/mesa/drivers/dri/i915/intel_context.c +6 -1
@@ -521,8 +521,13 @@ intelInitContext(struct intel_context *intel,INTEL_DEBUG =parse_debug_string (getenv ("INTEL_DEBUG" ), debug_control);if (INTEL_DEBUG &DEBUG_BUFMGR )dri_bufmgr_set_debug (intel->bufmgr, true); -if (INTEL_DEBUG &DEBUG_PERF ) +if (INTEL_DEBUG &DEBUG_PERF ) { intel->perf_debug = true; +if (screen->mocs_version >MOCS_TABLE_VERSION ) { +fprintf (stderr,"Kernel exposes newer MOCS table\n" ); + screen->mocs_version =MOCS_TABLE_VERSION ; + } + }if (INTEL_DEBUG &DEBUG_AUB )drm_intel_bufmgr_gem_set_aub_dump (intel->bufmgr, true);
src/mesa/drivers/dri/i965/intel_screen.c +14 -0
@@ -2301,6 +2301,20 @@ __DRIconfig **intelInitScreen2(__DRIscreen *dri_screen) (ret !=-1 || errno !=EINVAL ); } +if (devinfo->gen >=9 ) { + screen->mocs_version =intel_get_integer (screen, +I915_PARAM_MOCS_TABLE_VERSION ); +switch (screen->mocs_version) { +case 1 : +case 0 : + screen->mocs_version =MOCS_TABLE_VERSION ; +break ; +default : +/* We want to perf debug, but we can't yet */ +break ; + } + } + dri_screen->extensions = !screen->has_context_reset_notification ? screenExtensions : intelRobustScreenExtensions;
src/mesa/drivers/dri/i965/intel_screen.h +2 -0
@@ -112,6 +112,8 @@ struct intel_screenbool mesa_format_supports_texture[MESA_FORMAT_COUNT ];bool mesa_format_supports_render[MESA_FORMAT_COUNT ];enum isl_format mesa_to_isl_render_format[MESA_FORMAT_COUNT ]; + +unsigned mocs_version; };extern void intelDestroyContext (__DRIcontext * driContextPriv);
Signature · 6 lines
2.13.2 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH 3/3] intel: Make driver aware of MOCS table version
Quoted 20 lines · Ben Widawsky
We don't yet have optimal MOCS settings, but we have enough to know how to at least determine when we might have non-optimal settings within our driver.
Signed-off-by Ben Widawsky <[email protected]>--- src/intel/vulkan/anv_device.c | 12 ++++++++++++ src/intel/vulkan/anv_private.h | 2 ++ src/mesa/drivers/dri/i915/intel_context.c | 7 ++++++- src/mesa/drivers/dri/i965/intel_screen.c | 14 ++++++++++++++ src/mesa/drivers/dri/i965/intel_screen.h | 2 ++ 5 files changed, 36 insertions(+), 1 deletion(-)src/intel/vulkan/anv_device.c +0
-0@@ -368,6 +368,18 @@ anv_physical_device_init(struct anv_physical_device *device, device->info.max_cs_threads = max_cs_threads;
src/intel/vulkan/anv_device.c anv_physical_device_init(struct anv_physical_device · line 379
Quoted 10 lines
src/intel/vulkan/anv_device.c +6
-0} +if (device->info.gen >=9 ) { + device->mocs_version =anv_gem_get_param (fd, +I915_PARAM_MOCS_TABLE_VERSION ); +switch (device->mocs_version) { +default : +anv_perf_warn ("Kernel exposes newer MOCS table\n" );
A perf_warn here seems reasonable though it makes more sense to me to make it
if (device->mocs_version > ANV_MAX_KNOWN_MOCS_VERSION)
anv_perf_warn("...");
src/intel/vulkan/anv_device.c anv_physical_device_init(struct anv_physical_device *device, · line 379
Quoted 4 lines
src/intel/vulkan/anv_device.c +3
-0+case 1 : +case 0 : + device->mocs_version =MOCS_TABLE_VERSION ;
Why are we stomping device->mocs_version to MOCS_TABLE_VERSION? Are you just trying to avoid the version 0? If so, why not just have
/* If the MOCS_TABLE_VERSION query fails, assume version 1 */ if (device->mocs_version == 0)
device->mocs_version = 1;
I don't think we want to have it dependent on a #define in an external header file. What if someone updates it for i965 and doesn't update anv or vice-versa?
Quoted 51 lines
src/intel/vulkan/anv_device.c +3
-0+ } + } +brw_process_intel_debug_variable(); device->compiler = brw_compiler_create(NULL, &device->info);src/intel/vulkan/anv_ renamed from src/intel/vulkan/anv_private.h +0
-0
private.h
index573778dad5..b8241a9b22100644src/intel/vulkan/anv_private.h +2
-0@@ -684,6 +684,8 @@ struct anv_physical_device {uint32_t eu_total;uint32_t subslice_total; +uint8_t mocs_version; +struct {uint32_t type_count;struct anv_memory_type types[VK_MAX_MEMORY_TYPES];
src/mesa/drivers/dri/i915/intel_context.c +6
-1@@ -521,8 +521,13 @@ intelInitContext(struct intel_context *intel,INTEL_DEBUG =parse_debug_string (getenv ("INTEL_DEBUG" ), debug_control);if (INTEL_DEBUG &DEBUG_BUFMGR )dri_bufmgr_set_debug (intel->bufmgr, true); -if (INTEL_DEBUG &DEBUG_PERF ) +if (INTEL_DEBUG &DEBUG_PERF ) { intel->perf_debug = true; +if (screen->mocs_version >MOCS_TABLE_VERSION ) { +fprintf (stderr,"Kernel exposes newer MOCS table\n" ); + screen->mocs_version =MOCS_TABLE_VERSION ; + } + }if (INTEL_DEBUG &DEBUG_AUB )drm_intel_bufmgr_gem_set_aub_dump(intel->bufmgr, true);src/mesa/drivers/dri/i965/intel_screen.c +0
-0@@ -2301,6 +2301,20 @@ __DRIconfig **intelInitScreen2(__DRIscreen *dri_screen) (ret !=-1 || errno !=EINVAL ); }
src/mesa/drivers/dri/i965/intel_screen.c __DRIconfig **intelInitScreen2(__DRIscreen · line 2313
Quoted 10 lines
src/mesa/drivers/dri/i965/intel_screen.c +7
-0+if (devinfo->gen >=9 ) { + screen->mocs_version =intel_get_integer (screen, +I915_PARAM_MOCS_TABLE_VERSION ); +switch (screen->mocs_version) { +case 1 : +case 0 : + screen->mocs_version =MOCS_TABLE_VERSION ;
Same comments apply here.
Quoted 27 lines
src/mesa/drivers/dri/i965/intel_screen.c +7
-0+break ; +default : +/* We want to perf debug, but we can't yet */ +break ; + } + } +dri_screen->extensions = !screen->has_context_reset_notification ? screenExtensions : intelRobustScreenExtensions;src/mesa/drivers/dri/i965/intel_screen.h +2
-1@@ -112,6 +112,8 @@ struct intel_screenbool mesa_format_supports_texture[MESA_FORMAT_COUNT ];bool mesa_format_supports_render[MESA_FORMAT_COUNT ];enum isl_format mesa_to_isl_render_format[MESA_FORMAT_COUNT ]; + +unsigned mocs_version; };extern void intelDestroyContext (__DRIcontext * driContextPriv); --2.13.2
Quoted 5 lines
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH 3/3] intel: Make driver aware of MOCS table version
Quoted 36 lines · Jason Ekstrand
Quoted 20 lines · Ben Widawsky
We don't yet have optimal MOCS settings, but we have enough to know how to at least determine when we might have non-optimal settings within our driver.
Signed-off-by Ben Widawsky <[email protected]>--- src/intel/vulkan/anv_device.c | 12 ++++++++++++ src/intel/vulkan/anv_private.h | 2 ++ src/mesa/drivers/dri/i915/intel_context.c | 7 ++++++- src/mesa/drivers/dri/i965/intel_screen.c | 14 ++++++++++++++ src/mesa/drivers/dri/i965/intel_screen.h | 2 ++ 5 files changed, 36 insertions(+), 1 deletion(-)src/intel/vulkan/anv_device.c +0
-0@@ -368,6 +368,18 @@ anv_physical_device_init(struct anv_physical_device *device, device->info.max_cs_threads = max_cs_threads;src/intel/vulkan/anv_device.c anv_physical_device_init(struct anv_physical_device · line 379
Quoted 10 lines
src/intel/vulkan/anv_device.c +6
-0} +if (device->info.gen >=9 ) { + device->mocs_version =anv_gem_get_param (fd, +I915_PARAM_MOCS_TABLE_VERSION ); +switch (device->mocs_version) { +default : +anv_perf_warn ("Kernel exposes newer MOCS table\n" );A
perf_warnhere seems reasonable though it makes more sense to me to make it
Quoted 4 lines
if (
device->mocs_version>ANV_MAX_KNOWN_MOCS_VERSION)
anv_perf_warn("...");
One thing to keep in mind: the max MOCS version can vary by platform (hopefully it doesn't).
Quoted 5 lines
src/intel/vulkan/anv_device.c anv_physical_device_init(struct anv_physical_device *device, · line 379
Quoted 4 lines
src/intel/vulkan/anv_device.c +3
-0+case 1 : +case 0 : + device->mocs_version =MOCS_TABLE_VERSION ;
Quoted 7 lines
Why are we stomping
device->mocs_versiontoMOCS_TABLE_VERSION? Are you just trying to avoid the version 0? If so, why not just have/* If the
MOCS_TABLE_VERSIONquery fails, assume version 1 */ if (device->mocs_version== 0)
device->mocs_version= 1;
I think the switch looks better, especially as the versions increase.
Quoted 5 lines
I don't think we want to have it dependent on a #define in an external header file. What if someone updates it for i965 and doesn't update anv or vice-versa?
Yeah, I am removing that external define as mentioned in the other thread. I think it was a bad idea that I jammed in at the last minute.
Quoted 88 lines
Quoted 51 lines
src/intel/vulkan/anv_device.c +3
-0+ } + } +brw_process_intel_debug_variable(); device->compiler = brw_compiler_create(NULL, &device->info);src/intel/vulkan/anv_ renamed from src/intel/vulkan/anv_private.h +0
-0
private.h
index573778dad5..b8241a9b22100644src/intel/vulkan/anv_private.h +2
-0@@ -684,6 +684,8 @@ struct anv_physical_device {uint32_t eu_total;uint32_t subslice_total; +uint8_t mocs_version; +struct {uint32_t type_count;struct anv_memory_type types[VK_MAX_MEMORY_TYPES];
src/mesa/drivers/dri/i915/intel_context.c +6
-1@@ -521,8 +521,13 @@ intelInitContext(struct intel_context *intel,INTEL_DEBUG =parse_debug_string (getenv ("INTEL_DEBUG" ), debug_control);if (INTEL_DEBUG &DEBUG_BUFMGR )dri_bufmgr_set_debug (intel->bufmgr, true); -if (INTEL_DEBUG &DEBUG_PERF ) +if (INTEL_DEBUG &DEBUG_PERF ) { intel->perf_debug = true; +if (screen->mocs_version >MOCS_TABLE_VERSION ) { +fprintf (stderr,"Kernel exposes newer MOCS table\n" ); + screen->mocs_version =MOCS_TABLE_VERSION ; + } + }if (INTEL_DEBUG &DEBUG_AUB )drm_intel_bufmgr_gem_set_aub_dump(intel->bufmgr, true);src/mesa/drivers/dri/i965/intel_screen.c +0
-0@@ -2301,6 +2301,20 @@ __DRIconfig **intelInitScreen2(__DRIscreen *dri_screen) (ret !=-1 || errno !=EINVAL ); }src/mesa/drivers/dri/i965/intel_screen.c __DRIconfig **intelInitScreen2(__DRIscreen · line 2313
Quoted 10 lines
src/mesa/drivers/dri/i965/intel_screen.c +7
-0+if (devinfo->gen >=9 ) { + screen->mocs_version =intel_get_integer (screen, +I915_PARAM_MOCS_TABLE_VERSION ); +switch (screen->mocs_version) { +case 1 : +case 0 : + screen->mocs_version =MOCS_TABLE_VERSION ;Same comments apply here.
Quoted 23 lines
src/mesa/drivers/dri/i965/intel_screen.c +7
-0+break ; +default : +/* We want to perf debug, but we can't yet */ +break ; + } + } +dri_screen->extensions = !screen->has_context_reset_notification ? screenExtensions : intelRobustScreenExtensions;src/mesa/drivers/dri/i965/intel_screen.h +2
-0@@ -112,6 +112,8 @@ struct intel_screenbool mesa_format_supports_texture[MESA_FORMAT_COUNT ];bool mesa_format_supports_render[MESA_FORMAT_COUNT ];enum isl_format mesa_to_isl_render_format[MESA_FORMAT_COUNT ]; + +unsigned mocs_version; };
Quoted 1 line
src/mesa/drivers/dri/i965/intel_screen.h struct intel_screen · line 119
Quoted 1 line
src/mesa/drivers/dri/i965/intel_screen.h +0
-0extern void intelDestroyContext (__DRIcontext * driContextPriv);
If you feel strongly about moving to if/else, I can do it. Let me know.
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx