#define I915_CACHELINE_SIZE 64
#define I915_CACHELINE_MASK (I915_CACHELINE_SIZE - 1)
-static const uint32_t tileable_formats[] = { DRM_FORMAT_ARGB1555, DRM_FORMAT_ABGR8888,
- DRM_FORMAT_ARGB8888, DRM_FORMAT_RGB565,
- DRM_FORMAT_XBGR8888, DRM_FORMAT_XRGB1555,
- DRM_FORMAT_XRGB8888, DRM_FORMAT_UYVY,
- DRM_FORMAT_YUYV };
+static const uint32_t render_target_formats[] = { DRM_FORMAT_ABGR8888, DRM_FORMAT_ARGB1555,
+ DRM_FORMAT_ARGB8888, DRM_FORMAT_RGB565,
+ DRM_FORMAT_XBGR2101010, DRM_FORMAT_XBGR8888,
+ DRM_FORMAT_XRGB1555, DRM_FORMAT_XRGB2101010,
+ DRM_FORMAT_XRGB8888 };
-static const uint32_t linear_only_formats[] = { DRM_FORMAT_GR88, DRM_FORMAT_R8, DRM_FORMAT_YVU420,
- DRM_FORMAT_YVU420_ANDROID };
+static const uint32_t tileable_texture_source_formats[] = { DRM_FORMAT_GR88, DRM_FORMAT_R8,
+ DRM_FORMAT_UYVY, DRM_FORMAT_YUYV };
+
+static const uint32_t texture_source_formats[] = { DRM_FORMAT_YVU420, DRM_FORMAT_YVU420_ANDROID,
+ DRM_FORMAT_NV12 };
struct i915_device {
uint32_t gen;
* Older hardware can't scanout Y-tiled formats. Newer devices can, and
* report this functionality via format modifiers.
*/
- for (i = 0; i < drv->backend->combos.size; i++) {
- combo = &drv->backend->combos.data[i];
- if (combo->format == item->format) {
- if ((combo->metadata.tiling == I915_TILING_Y &&
- item->modifier == I915_FORMAT_MOD_Y_TILED) ||
- (combo->metadata.tiling == I915_TILING_X &&
- item->modifier == I915_FORMAT_MOD_X_TILED)) {
- combo->metadata.modifier = item->modifier;
- combo->usage |= item->usage;
- } else if (combo->metadata.tiling != I915_TILING_Y) {
- combo->usage |= item->usage;
- }
+ for (i = 0; i < drv_array_size(drv->combos); i++) {
+ combo = (struct combination *)drv_array_at_idx(drv->combos, i);
+ if (combo->format != item->format)
+ continue;
+
+ if (item->modifier == DRM_FORMAT_MOD_LINEAR &&
+ combo->metadata.tiling == I915_TILING_X) {
+ /*
+ * FIXME: drv_query_kms() does not report the available modifiers
+ * yet, but we know that all hardware can scanout from X-tiled
+ * buffers, so let's add this to our combinations, except for
+ * cursor, which must not be tiled.
+ */
+ combo->use_flags |= item->use_flags & ~BO_USE_CURSOR;
}
+
+ if (combo->metadata.modifier == item->modifier)
+ combo->use_flags |= item->use_flags;
}
return 0;
static int i915_add_combinations(struct driver *drv)
{
int ret;
- uint32_t i, num_items;
- struct kms_item *items;
+ uint32_t i;
+ struct drv_array *kms_items;
struct format_metadata metadata;
- uint64_t flags = BO_COMMON_USE_MASK;
+ uint64_t render_use_flags, texture_use_flags;
+
+ render_use_flags = BO_USE_RENDER_MASK;
+ texture_use_flags = BO_USE_TEXTURE_MASK;
metadata.tiling = I915_TILING_NONE;
metadata.priority = 1;
- metadata.modifier = DRM_FORMAT_MOD_NONE;
+ metadata.modifier = DRM_FORMAT_MOD_LINEAR;
- ret = drv_add_combinations(drv, linear_only_formats, ARRAY_SIZE(linear_only_formats),
- &metadata, flags);
- if (ret)
- return ret;
+ drv_add_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
+ &metadata, render_use_flags);
- ret = drv_add_combinations(drv, tileable_formats, ARRAY_SIZE(tileable_formats), &metadata,
- flags);
- if (ret)
- return ret;
+ drv_add_combinations(drv, texture_source_formats, ARRAY_SIZE(texture_source_formats),
+ &metadata, texture_use_flags);
+
+ drv_add_combinations(drv, tileable_texture_source_formats,
+ ARRAY_SIZE(tileable_texture_source_formats), &metadata,
+ texture_use_flags);
drv_modify_combination(drv, DRM_FORMAT_XRGB8888, &metadata, BO_USE_CURSOR | BO_USE_SCANOUT);
drv_modify_combination(drv, DRM_FORMAT_ARGB8888, &metadata, BO_USE_CURSOR | BO_USE_SCANOUT);
- flags &= ~BO_USE_SW_WRITE_OFTEN;
- flags &= ~BO_USE_SW_READ_OFTEN;
- flags &= ~BO_USE_LINEAR;
+ /* IPU3 camera ISP supports only NV12 output. */
+ drv_modify_combination(drv, DRM_FORMAT_NV12, &metadata,
+ BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE);
+ /*
+ * R8 format is used for Android's HAL_PIXEL_FORMAT_BLOB and is used for JPEG snapshots
+ * from camera.
+ */
+ drv_modify_combination(drv, DRM_FORMAT_R8, &metadata,
+ BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE);
+
+ render_use_flags &= ~BO_USE_RENDERSCRIPT;
+ render_use_flags &= ~BO_USE_SW_WRITE_OFTEN;
+ render_use_flags &= ~BO_USE_SW_READ_OFTEN;
+ render_use_flags &= ~BO_USE_LINEAR;
+
+ texture_use_flags &= ~BO_USE_RENDERSCRIPT;
+ texture_use_flags &= ~BO_USE_SW_WRITE_OFTEN;
+ texture_use_flags &= ~BO_USE_SW_READ_OFTEN;
+ texture_use_flags &= ~BO_USE_LINEAR;
metadata.tiling = I915_TILING_X;
metadata.priority = 2;
+ metadata.modifier = I915_FORMAT_MOD_X_TILED;
- ret = drv_add_combinations(drv, tileable_formats, ARRAY_SIZE(tileable_formats), &metadata,
- flags);
- if (ret)
- return ret;
+ drv_add_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
+ &metadata, render_use_flags);
+
+ drv_add_combinations(drv, tileable_texture_source_formats,
+ ARRAY_SIZE(tileable_texture_source_formats), &metadata,
+ texture_use_flags);
metadata.tiling = I915_TILING_Y;
metadata.priority = 3;
+ metadata.modifier = I915_FORMAT_MOD_Y_TILED;
- ret = drv_add_combinations(drv, tileable_formats, ARRAY_SIZE(tileable_formats), &metadata,
- flags);
- if (ret)
- return ret;
+ drv_add_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),
+ &metadata, render_use_flags);
+
+ drv_add_combinations(drv, tileable_texture_source_formats,
+ ARRAY_SIZE(tileable_texture_source_formats), &metadata,
+ texture_use_flags);
- items = drv_query_kms(drv, &num_items);
- if (!items || !num_items)
+ kms_items = drv_query_kms(drv);
+ if (!kms_items)
return 0;
- for (i = 0; i < num_items; i++) {
- ret = i915_add_kms_item(drv, &items[i]);
+ for (i = 0; i < drv_array_size(kms_items); i++) {
+ ret = i915_add_kms_item(drv, (struct kms_item *)drv_array_at_idx(kms_items, i));
if (ret) {
- free(items);
+ drv_array_destroy(kms_items);
return ret;
}
}
- free(items);
+ drv_array_destroy(kms_items);
return 0;
}
break;
}
+ /*
+ * The alignment calculated above is based on the full size luma plane and to have chroma
+ * planes properly aligned with subsampled formats, we need to multiply luma alignment by
+ * subsampling factor.
+ */
+ switch (bo->format) {
+ case DRM_FORMAT_YVU420_ANDROID:
+ case DRM_FORMAT_YVU420:
+ horizontal_alignment *= 2;
+ /* Fall through */
+ case DRM_FORMAT_NV12:
+ vertical_alignment *= 2;
+ break;
+ }
+
*aligned_height = ALIGN(bo->height, vertical_alignment);
if (i915->gen > 3) {
*stride = ALIGN(*stride, horizontal_alignment);
return i915_add_combinations(drv);
}
-static int i915_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
- uint32_t flags)
+static int i915_bo_create_for_modifier(struct bo *bo, uint32_t width, uint32_t height,
+ uint32_t format, uint64_t modifier)
{
int ret;
size_t plane;
struct drm_i915_gem_create gem_create;
struct drm_i915_gem_set_tiling gem_set_tiling;
- if (flags & (BO_USE_CURSOR | BO_USE_LINEAR | BO_USE_SW_READ_OFTEN | BO_USE_SW_WRITE_OFTEN))
+ switch (modifier) {
+ case DRM_FORMAT_MOD_LINEAR:
bo->tiling = I915_TILING_NONE;
- else if (flags & BO_USE_SCANOUT)
+ break;
+ case I915_FORMAT_MOD_X_TILED:
bo->tiling = I915_TILING_X;
- else
+ break;
+ case I915_FORMAT_MOD_Y_TILED:
bo->tiling = I915_TILING_Y;
+ break;
+ }
+
+ bo->format_modifiers[0] = modifier;
stride = drv_stride_from_format(format, width, 0);
- /*
- * Align the Y plane to 128 bytes so the chroma planes would be aligned
- * to 64 byte boundaries. This is an Intel HW requirement.
- */
- if (format == DRM_FORMAT_YVU420 || format == DRM_FORMAT_YVU420_ANDROID) {
- stride = ALIGN(stride, 128);
- bo->tiling = I915_TILING_NONE;
- }
ret = i915_align_dimensions(bo, bo->tiling, &stride, &height);
if (ret)
return ret;
- drv_bo_from_format(bo, stride, height, format);
+ /*
+ * HAL_PIXEL_FORMAT_YV12 requires the buffer height not be aligned, but we need to keep
+ * total size as with aligned height to ensure enough padding space after each plane to
+ * satisfy GPU alignment requirements.
+ *
+ * We do it by first calling drv_bo_from_format() with aligned height and
+ * DRM_FORMAT_YVU420, which allows height alignment, saving the total size it calculates
+ * and then calling it again with requested parameters.
+ *
+ * This relies on the fact that i965 driver uses separate surfaces for each plane and
+ * contents of padding bytes is not affected, as it is only used to satisfy GPU cache
+ * requests.
+ *
+ * This is enforced by Mesa in src/intel/isl/isl_gen8.c, inside
+ * isl_gen8_choose_image_alignment_el(), which is used for GEN9 and GEN8.
+ */
+ if (format == DRM_FORMAT_YVU420_ANDROID) {
+ uint32_t unaligned_height = bo->height;
+ size_t total_size;
+
+ drv_bo_from_format(bo, stride, height, DRM_FORMAT_YVU420);
+ total_size = bo->total_size;
+ drv_bo_from_format(bo, stride, unaligned_height, format);
+ bo->total_size = total_size;
+ } else {
+ drv_bo_from_format(bo, stride, height, format);
+ }
+
+ /*
+ * Quoting Mesa ISL library:
+ *
+ * - For linear surfaces, additional padding of 64 bytes is required at
+ * the bottom of the surface. This is in addition to the padding
+ * required above.
+ */
+ if (bo->tiling == I915_TILING_NONE)
+ bo->total_size += 64;
memset(&gem_create, 0, sizeof(gem_create));
gem_create.size = bo->total_size;
return 0;
}
+static int i915_bo_create(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,
+ uint64_t use_flags)
+{
+ struct combination *combo;
+
+ combo = drv_get_combination(bo->drv, format, use_flags);
+ if (!combo)
+ return -EINVAL;
+
+ return i915_bo_create_for_modifier(bo, width, height, format, combo->metadata.modifier);
+}
+
+static int i915_bo_create_with_modifiers(struct bo *bo, uint32_t width, uint32_t height,
+ uint32_t format, const uint64_t *modifiers, uint32_t count)
+{
+ static const uint64_t modifier_order[] = {
+ I915_FORMAT_MOD_Y_TILED,
+ I915_FORMAT_MOD_X_TILED,
+ DRM_FORMAT_MOD_LINEAR,
+ };
+ uint64_t modifier;
+
+ modifier = drv_pick_modifier(modifiers, count, modifier_order, ARRAY_SIZE(modifier_order));
+
+ return i915_bo_create_for_modifier(bo, width, height, format, modifier);
+}
+
static void i915_close(struct driver *drv)
{
free(drv->priv);
ret = drmIoctl(bo->drv->fd, DRM_IOCTL_I915_GEM_GET_TILING, &gem_get_tiling);
if (ret) {
+ drv_gem_bo_destroy(bo);
fprintf(stderr, "drv: DRM_IOCTL_I915_GEM_GET_TILING failed.");
return ret;
}
return 0;
}
-static void *i915_bo_map(struct bo *bo, struct map_info *data, size_t plane)
+static void *i915_bo_map(struct bo *bo, struct vma *vma, size_t plane, uint32_t map_flags)
{
int ret;
void *addr;
- struct drm_i915_gem_set_domain set_domain;
- memset(&set_domain, 0, sizeof(set_domain));
- set_domain.handle = bo->handles[0].u32;
if (bo->tiling == I915_TILING_NONE) {
struct drm_i915_gem_mmap gem_map;
memset(&gem_map, 0, sizeof(gem_map));
+ if ((bo->use_flags & BO_USE_SCANOUT) && !(bo->use_flags & BO_USE_RENDERSCRIPT))
+ gem_map.flags = I915_MMAP_WC;
+
gem_map.handle = bo->handles[0].u32;
gem_map.offset = 0;
gem_map.size = bo->total_size;
}
addr = (void *)(uintptr_t)gem_map.addr_ptr;
- set_domain.read_domains = I915_GEM_DOMAIN_CPU;
- set_domain.write_domain = I915_GEM_DOMAIN_CPU;
-
} else {
struct drm_i915_gem_mmap_gtt gem_map;
memset(&gem_map, 0, sizeof(gem_map));
return MAP_FAILED;
}
- addr = mmap(0, bo->total_size, PROT_READ | PROT_WRITE, MAP_SHARED, bo->drv->fd,
+ addr = mmap(0, bo->total_size, drv_get_prot(map_flags), MAP_SHARED, bo->drv->fd,
gem_map.offset);
-
- set_domain.read_domains = I915_GEM_DOMAIN_GTT;
- set_domain.write_domain = I915_GEM_DOMAIN_GTT;
}
if (addr == MAP_FAILED) {
return addr;
}
+ vma->length = bo->total_size;
+ return addr;
+}
+
+static int i915_bo_invalidate(struct bo *bo, struct mapping *mapping)
+{
+ int ret;
+ struct drm_i915_gem_set_domain set_domain;
+
+ memset(&set_domain, 0, sizeof(set_domain));
+ set_domain.handle = bo->handles[0].u32;
+ if (bo->tiling == I915_TILING_NONE) {
+ set_domain.read_domains = I915_GEM_DOMAIN_CPU;
+ if (mapping->vma->map_flags & BO_MAP_WRITE)
+ set_domain.write_domain = I915_GEM_DOMAIN_CPU;
+ } else {
+ set_domain.read_domains = I915_GEM_DOMAIN_GTT;
+ if (mapping->vma->map_flags & BO_MAP_WRITE)
+ set_domain.write_domain = I915_GEM_DOMAIN_GTT;
+ }
+
ret = drmIoctl(bo->drv->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain);
if (ret) {
- fprintf(stderr, "drv: DRM_IOCTL_I915_GEM_SET_DOMAIN failed\n");
- return MAP_FAILED;
+ fprintf(stderr, "drv: DRM_IOCTL_I915_GEM_SET_DOMAIN with %d\n", ret);
+ return ret;
}
- data->length = bo->total_size;
- return addr;
+ return 0;
}
-static int i915_bo_unmap(struct bo *bo, struct map_info *data)
+static int i915_bo_flush(struct bo *bo, struct mapping *mapping)
{
struct i915_device *i915 = bo->drv->priv;
if (!i915->has_llc && bo->tiling == I915_TILING_NONE)
- i915_clflush(data->addr, data->length);
+ i915_clflush(mapping->vma->addr, mapping->vma->length);
- return munmap(data->addr, data->length);
+ return 0;
}
-static uint32_t i915_resolve_format(uint32_t format)
+static uint32_t i915_resolve_format(uint32_t format, uint64_t use_flags)
{
switch (format) {
case DRM_FORMAT_FLEX_IMPLEMENTATION_DEFINED:
+ /* KBL camera subsystem requires NV12. */
+ if (use_flags & (BO_USE_CAMERA_READ | BO_USE_CAMERA_WRITE))
+ return DRM_FORMAT_NV12;
/*HACK: See b/28671744 */
return DRM_FORMAT_XBGR8888;
case DRM_FORMAT_FLEX_YCbCr_420_888:
- return DRM_FORMAT_YVU420_ANDROID;
+ /*
+ * KBL camera subsystem requires NV12. Our other use cases
+ * don't care:
+ * - Hardware video supports NV12,
+ * - USB Camera HALv3 supports NV12,
+ * - USB Camera HALv1 doesn't use this format.
+ * Moreover, NV12 is preferred for video, due to overlay
+ * support on SKL+.
+ */
+ return DRM_FORMAT_NV12;
default:
return format;
}
}
-struct backend backend_i915 = {
+const struct backend backend_i915 = {
.name = "i915",
.init = i915_init,
.close = i915_close,
.bo_create = i915_bo_create,
+ .bo_create_with_modifiers = i915_bo_create_with_modifiers,
.bo_destroy = drv_gem_bo_destroy,
.bo_import = i915_bo_import,
.bo_map = i915_bo_map,
- .bo_unmap = i915_bo_unmap,
+ .bo_unmap = drv_bo_munmap,
+ .bo_invalidate = i915_bo_invalidate,
+ .bo_flush = i915_bo_flush,
.resolve_format = i915_resolve_format,
};