[Decode] Use 64K contiguous pages for WA for scanout buffers

* [Decode] Use 64K contiguous pages for WA for scanout buffers

This is kmd restriction for scanout buffers

The surface for display must create with scanout, otherwise leading
display failure.

media_softlet/linux/common/ddi/media_libva_util_next.cpp

@@ -900,16 +900,24 @@ VAStatus MediaLibvaUtilNext::CreateInternalSurface(
     unsigned int patIndex = MosInterface::GetPATIndexFromGmm(mediaDrvCtx->pGmmClientContext, gmmResourceInfo);
     bool isCpuCacheable   = gmmResourceInfo->GetResFlags().Info.Cacheable;
 
+    //This is wa for tile4 + compressed + scanout surface(!isCpuCacheable)
+    bool is64kPageAlignmentNeed =
+        MEDIA_IS_WA(&mediaDrvCtx->WaTable, WaTile4CompressScanoutSurf64KNeed)
+        && gmmResourceInfo->GetResFlags().Info.Tile4
+        && (gmmResourceInfo->GetResFlags().Info.MediaCompressed || gmmResourceInfo->GetResFlags().Info.RenderCompressed)
+        && !isCpuCacheable;
+
     if ( params.tileFormat == TILING_NONE )
     {
         struct mos_drm_bo_alloc alloc;
         alloc.name = "Media";
         alloc.size = gmmSize;
-        alloc.alignment = 4096;
+        alloc.alignment = is64kPageAlignmentNeed ? PAGE_SIZE_64K : PAGE_SIZE_4K;
         alloc.ext.tiling_mode = TILING_NONE;
         alloc.ext.mem_type = params.memType;
         alloc.ext.pat_index = patIndex;
         alloc.ext.cpu_cacheable = isCpuCacheable;
+        alloc.ext.scanout_surf  = !isCpuCacheable;
         bo = mos_bo_alloc(mediaDrvCtx->pDrmBufMgr, &alloc);
         params.pitch = gmmPitch;
     }
@@ -920,10 +928,12 @@ VAStatus MediaLibvaUtilNext::CreateInternalSurface(
         alloc_tiled.x = gmmPitch;
         alloc_tiled.y = (gmmSize + gmmPitch -1)/gmmPitch;
         alloc_tiled.cpp = 1;
+        alloc_tiled.alignment = is64kPageAlignmentNeed ? PAGE_SIZE_64K : PAGE_SIZE_4K;
         alloc_tiled.ext.tiling_mode = params.tileFormat;
         alloc_tiled.ext.mem_type = params.memType;;
         alloc_tiled.ext.pat_index = patIndex;
         alloc_tiled.ext.cpu_cacheable = isCpuCacheable;
+        alloc_tiled.ext.scanout_surf  = !isCpuCacheable;
 
         bo = mos_bo_alloc_tiled(mediaDrvCtx->pDrmBufMgr, &alloc_tiled);
         params.pitch = alloc_tiled.pitch;

media_softlet/linux/common/os/i915/include/mos_bufmgr_api.h

@@ -134,6 +134,7 @@ enum device_type {
 #define MEMZONE_PRIME_START   (MEMZONE_DEVICE_START + MEMZONE_DEVICE_SIZE)
 #define MEMZONE_PRIME_SIZE    (1ull << 40)
 #define MEMZONE_TOTAL         (1ull << 48)
+#define PAGE_SIZE_4K          (1ull << 12)
 #define PAGE_SIZE_64K         (1ull << 16)
 #define PAGE_SIZE_1M          (1ull << 20)
 #define PAGE_SIZE_2M          (1ull << 21)
@@ -211,6 +212,7 @@ struct mos_drm_bo_alloc_ext{
     int mem_type = 0;
     uint16_t pat_index = PAT_INDEX_INVALID;
     bool     cpu_cacheable = true;
+    bool     scanout_surf  = false;
 };
 
 struct mos_drm_bo_alloc {
@@ -238,6 +240,7 @@ struct mos_drm_bo_alloc_tiled {
     int y = 0;
     int cpp = 0;
     unsigned long pitch = 0;
+    unsigned int alignment = 0;
 
     struct mos_drm_bo_alloc_ext ext;
 };

media_softlet/linux/common/os/xe/mos_bufmgr_xe.c

@@ -84,7 +84,6 @@ typedef struct MOS_OCA_EXEC_LIST_INFO mos_oca_exec_list_info;
 
 #include "mos_bufmgr_priv.h"
 
-#define PAGE_SIZE_4K                   (1ull << 12)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 
@@ -1186,12 +1185,12 @@ __mos_bo_set_offset_xe(MOS_LINUX_BO *bo)
         else if (MEMZONE_DEVICE == bo_gem->mem_region)
         {
             alignment = MAX(bufmgr_gem->default_alignment[MOS_XE_MEM_CLASS_VRAM], PAGE_SIZE_64K);
-            offset = __mos_bo_vma_alloc_xe(bo->bufmgr, (enum mos_memory_zone)bo_gem->mem_region, bo->size, PAGE_SIZE_64K);
+            offset = __mos_bo_vma_alloc_xe(bo->bufmgr, (enum mos_memory_zone)bo_gem->mem_region, bo->size, alignment);
         }
         else if (MEMZONE_SYS == bo_gem->mem_region)
         {
             alignment = MAX(bufmgr_gem->default_alignment[MOS_XE_MEM_CLASS_SYSMEM], PAGE_SIZE_64K);
-            offset = __mos_bo_vma_alloc_xe(bo->bufmgr, (enum mos_memory_zone)bo_gem->mem_region, bo->size, PAGE_SIZE_64K);
+            offset = __mos_bo_vma_alloc_xe(bo->bufmgr, (enum mos_memory_zone)bo_gem->mem_region, bo->size, alignment);
         }
         else
         {
@@ -1350,9 +1349,10 @@ mos_bo_alloc_xe(struct mos_bufmgr *bufmgr,
      */
     create.cpu_caching = alloc->ext.cpu_cacheable ? DRM_XE_GEM_CPU_CACHING_WB : DRM_XE_GEM_CPU_CACHING_WC;
 
-    if ((strcmp(alloc->name, "MEDIA") == 0 || strcmp(alloc->name, "Media") == 0)
-        && create.cpu_caching == DRM_XE_GEM_CPU_CACHING_WC)
-            create.flags |= DRM_XE_GEM_CREATE_FLAG_SCANOUT;
+    if (alloc->ext.scanout_surf)
+    {
+        create.flags |= DRM_XE_GEM_CREATE_FLAG_SCANOUT;
+    }
 
     ret = drmIoctl(bufmgr_gem->fd,
         DRM_IOCTL_XE_GEM_CREATE,
@@ -1470,12 +1470,12 @@ mos_bo_alloc_tiled_xe(struct mos_bufmgr *bufmgr,
     unsigned long size, stride;
     uint32_t tiling;
 
-    uint32_t alignment = bufmgr_gem->default_alignment[MOS_XE_MEM_CLASS_SYSMEM];
+    uint32_t alignment = MAX(alloc_tiled->alignment, bufmgr_gem->default_alignment[MOS_XE_MEM_CLASS_SYSMEM]);
 
     if (bufmgr_gem->has_vram &&
        (MOS_MEMPOOL_VIDEOMEMORY == alloc_tiled->ext.mem_type   || MOS_MEMPOOL_DEVICEMEMORY == alloc_tiled->ext.mem_type))
     {
-        alignment = bufmgr_gem->default_alignment[MOS_XE_MEM_CLASS_VRAM];
+        alignment = MAX(alloc_tiled->alignment, bufmgr_gem->default_alignment[MOS_XE_MEM_CLASS_VRAM]);
     }
 
     do {

media_softlet/linux/xe2_hpm/ddi/media_sku_wa_bmg.cpp

@@ -279,7 +279,9 @@ static bool InitBmgMediaWaExt(struct GfxDeviceInfo *devInfo,
       CPU blt call will add/remove padding on the platform*/
     MEDIA_WR_WA(waTable, WaDisableGmmLibOffsetInDeriveImage, 1);
 
-
+    // This requirement comes from XE KMD for tile4 + compress + scanout surface
+    // Currently, it only applies to xe2 dgpu
+    MEDIA_WR_WA(waTable, WaTile4CompressScanoutSurf64KNeed , 1);
 
     /* Turn off MMC for codec, need to remove once turn it on */
     MEDIA_WR_WA(waTable, WaDisableCodecMmc, 0);