Re: [PATCH v4] drm/i915: Use Transparent Hugepages when IOMMU is enabled

[Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>]

On 07/09/2021 12:13, Eero Tamminen wrote:
> Hi,
> 
> For completeness sake, it might be worth mentioning specifically what 
> (synthetic) test-cases regress with THP patch.
> 
> * Skylake GT4e:
>    20-25% SynMark TexMem*
>    (whereas all MemBW GPU tests either improve or are not affected)
> 
> * Broxton J4205:
>    7% MemBW GPU texture
>    2-3% SynMark TexMem*
> 
> * Tigerlake-H:
>    7% MemBW GPU blend

Ah right that makes sense. All the entries marker with asterisk under 
the "with patch" list. Okay if I just add an explanation on what does 
the asterisk mean for them at a single place?

And about the Broxton one. In the bug you put "15-20% MemBW GPU texture" 
and "10% SynMark TexMem*" so from where are these numbers now?

> 
> I have no idea why on GEN9 texture accesses regress, but on GEN12 TGL 
> it's render buffer blend that regresses.
> 
> Blend (read+write) regressing is especially odd, as neither render 
> buffer read nor write regresses.
> 
> Maybe that is a GEN12 specific driver bug similar to Mesa/i965 bug from 
> few years back in how its shaders access render buffer, that had caused 
> SIMD32 accesses to regress memory BW bound test-cases perf a bit 
> compared to SIMD16?
> 
> (Blend test is likely to run nowadays as SIMD32.)

No idea on this one from me, leaving to more qualified people to comment.

Regards,

Tvrtko


> 
>      - Eero
> 
> On 7.9.2021 13.34, Tvrtko Ursulin wrote:
>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>>
>> Usage of Transparent Hugepages was disabled in 9987da4b5dcf
>> ("drm/i915: Disable THP until we have a GPU read BW W/A"), but since it
>> appears majority of performance regressions reported with an enabled 
>> IOMMU
>> can be almost eliminated by turning them on, lets just do that.
>>
>> To err on the side of safety we keep the current default in cases where
>> IOMMU is not active, and only when it is default to the 
>> "huge=within_size"
>> mode. Although there probably would be wins to enable them throughout,
>> more extensive testing across benchmarks and platforms would need to be
>> done.
>>
>> With the patch and IOMMU enabled my local testing on a small Skylake part
>> shows OglVSTangent regression being reduced from ~14% (IOMMU on versus
>> IOMMU off) to ~2% (same comparison but with THP on).
>>
>> More detailed testing done in the below referenced Gitlab issue by Eero:
>>
>> Skylake GT4e:
>>
>> Performance drops from enabling IOMMU:
>>
>>      30-35% SynMark CSDof
>>      20-25% Unigine Heaven, MemBW GPU write, SynMark VSTangent
>>      ~20% GLB Egypt  (1/2 screen window)
>>      10-15% GLB T-Rex (1/2 screen window)
>>      8-10% GfxBench T-Rex, MemBW GPU blit
>>      7-8% SynMark DeferredAA + TerrainFly* + ZBuffer
>>      6-7% GfxBench Manhattan 3.0 + 3.1, SynMark TexMem128 & CSCloth
>>      5-6% GfxBench CarChase, Unigine Valley
>>      3-5% GfxBench Vulkan & GL AztecRuins + ALU2, MemBW GPU texture,
>>           SynMark Fill*, Deferred, TerrainPan*
>>      1-2% Most of the other tests
>>
>> With the patch drops become:
>>
>>      20-25% SynMark TexMem*
>>      15-20% GLB Egypt (1/2 screen window)
>>      10-15% GLB T-Rex (1/2 screen window)
>>      4-7% GfxBench T-Rex, GpuTest Triangle
>>      1-8% GfxBench ALU2 (offscreen 1%, onscreen 8%)
>>      3% GfxBench Manhattan 3.0, SynMark CSDof
>>      2-3% Unigine Heaven + Valley, MemBW GPU texture
>>      1-3 GfxBench Manhattan 3.1 + CarChase + Vulkan & GL AztecRuins
>>
>> Broxton:
>>
>> Performance drops from IOMMU, without patch:
>>
>>      30% MemBW GPU write
>>      25% SynMark ZBuffer + Fill*
>>      20% MemBW GPU blit
>>      15% MemBW GPU blend, GpuTest Triangle
>>      10-15% MemBW GPU texture
>>      10% GLB Egypt, Unigine Heaven (had hangs), SynMark TerrainFly*
>>      7-9% GLB T-Rex, GfxBench Manhattan 3.0 + T-Rex,
>>           SynMark Deferred* + TexMem*
>>      6-8% GfxBench CarChase, Unigine Valley,
>>           SynMark CSCloth + ShMapVsm + TerrainPan*
>>      5-6% GfxBench Manhattan 3.1 + GL AztecRuins,
>>           SynMark CSDof + TexFilterTri
>>      2-4% GfxBench ALU2, SynMark DrvRes + GSCloth + ShMapPcf + 
>> Batch[0-5] +
>>           TexFilterAniso, GpuTest GiMark + 32-bit Julia
>>
>> And with patch:
>>
>>      15-20% MemBW GPU texture
>>      10% SynMark TexMem*
>>      8-9% GLB Egypt (1/2 screen window)
>>      4-5% GLB T-Rex (1/2 screen window)
>>      3-6% GfxBench Manhattan 3.0, GpuTest FurMark,
>>           SynMark Deferred + TexFilterTri
>>      3-4% GfxBench Manhattan 3.1 + T-Rex, SynMark VSInstancing
>>      2-4% GpuTest Triangle, SynMark DeferredAA
>>      2-3% Unigine Heaven + Valley
>>      1-3% SynMark Terrain*
>>      1-2% GfxBench CarChase, SynMark TexFilterAniso + ZBuffer
>>
>> Tigerlake-H:
>>
>>      20-25% MemBW GPU texture
>>      15-20% GpuTest Triangle
>>      13-15% SynMark TerrainFly* + DeferredAA + HdrBloom
>>      8-10% GfxBench Manhattan 3.1, SynMark TerrainPan* + DrvRes
>>      6-7% GfxBench Manhattan 3.0, SynMark TexMem*
>>      4-8% GLB onscreen Fill + T-Rex + Egypt (more in onscreen than
>>           offscreen versions of T-Rex/Egypt)
>>      4-6% GfxBench CarChase + GLES AztecRuins + ALU2, GpuTest 32-bit 
>> Julia,
>>           SynMark CSDof + DrvState
>>      3-5% GfxBench T-Rex + Egypt, Unigine Heaven + Valley, GpuTest Plot3D
>>      1-7% Media tests
>>      2-3% MemBW GPU blit
>>      1-3% Most of the rest of 3D tests
>>
>> With the patch:
>>
>>      6-8% MemBW GPU blend => the only regression in these tests (compared
>>           to IOMMU without THP)
>>      4-6% SynMark DrvState (not impacted) + HdrBloom (improved)
>>      3-4% GLB T-Rex
>>      ~3% GLB Egypt, SynMark DrvRes
>>      1-3% GfxBench T-Rex + Egypt, SynMark TexFilterTri
>>      1-2% GfxBench CarChase + GLES AztecRuins, Unigine Valley,
>>          GpuTest Triangle
>>      ~1% GfxBench Manhattan 3.0/3.1, Unigine Heaven
>>
>> Perf of several tests actually improved with IOMMU + THP, compared to no
>> IOMMU / no THP:
>>
>>      10-15% SynMark Batch[0-3]
>>      5-10% MemBW GPU texture, SynMark ShMapVsm
>>      3-4% SynMark Fill* + Geom*
>>      2-3% SynMark TexMem512 + CSCloth
>>      1-2% SynMark TexMem128 + DeferredAA
>>
>> v2:
>>   * Add Kconfig dependency to transparent hugepages and some help text.
>>   * Move to helper for easier handling of kernel build options.
>>
>> v3:
>>   * Drop Kconfig. (Daniel)
>>
>> v4:
>>   * Add some benchmark results to commit message.
>>
>> References: b901bb89324a ("drm/i915/gemfs: enable THP")
>> References: 9987da4b5dcf ("drm/i915: Disable THP until we have a GPU 
>> read BW W/A")
>> References: https://gitlab.freedesktop.org/drm/intel/-/issues/430
>> Co-developed-by: Chris Wilson <chris@chris-wilson.co.uk>
>> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
>> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
>> Cc: Matthew Auld <matthew.auld@intel.com>
>> Cc: Eero Tamminen <eero.t.tamminen@intel.com>
>> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>> Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
>> Cc: Daniel Vetter <daniel@ffwll.ch>
>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>> Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com> # v1
>> ---
>>   drivers/gpu/drm/i915/gem/i915_gemfs.c | 22 +++++++++++++++++++---
>>   1 file changed, 19 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/i915/gem/i915_gemfs.c 
>> b/drivers/gpu/drm/i915/gem/i915_gemfs.c
>> index 5e6e8c91ab38..dbdbdc344d87 100644
>> --- a/drivers/gpu/drm/i915/gem/i915_gemfs.c
>> +++ b/drivers/gpu/drm/i915/gem/i915_gemfs.c
>> @@ -6,7 +6,6 @@
>>   #include <linux/fs.h>
>>   #include <linux/mount.h>
>> -#include <linux/pagemap.h>
>>   #include "i915_drv.h"
>>   #include "i915_gemfs.h"
>> @@ -15,6 +14,7 @@ int i915_gemfs_init(struct drm_i915_private *i915)
>>   {
>>       struct file_system_type *type;
>>       struct vfsmount *gemfs;
>> +    char *opts;
>>       type = get_fs_type("tmpfs");
>>       if (!type)
>> @@ -26,10 +26,26 @@ int i915_gemfs_init(struct drm_i915_private *i915)
>>        *
>>        * One example, although it is probably better with a per-file
>>        * control, is selecting huge page allocations 
>> ("huge=within_size").
>> -     * Currently unused due to bandwidth issues (slow reads) on 
>> Broadwell+.
>> +     * However, we only do so to offset the overhead of iommu lookups
>> +     * due to bandwidth issues (slow reads) on Broadwell+.
>>        */
>> -    gemfs = kern_mount(type);
>> +    opts = NULL;
>> +    if (intel_vtd_active()) {
>> +        if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
>> +            static char huge_opt[] = "huge=within_size"; /* r/w */
>> +
>> +            opts = huge_opt;
>> +            drm_info(&i915->drm,
>> +                 "Transparent Hugepage mode '%s'\n",
>> +                 opts);
>> +        } else {
>> +            drm_notice(&i915->drm,
>> +                   "Transparent Hugepage support is recommended for 
>> optimal performance when IOMMU is enabled!\n");
>> +        }
>> +    }
>> +
>> +    gemfs = vfs_kern_mount(type, SB_KERNMOUNT, type->name, opts);
>>       if (IS_ERR(gemfs))
>>           return PTR_ERR(gemfs);
>>