[PATCH 3/3] drm/amd/display: Extract 3rd plane from metadata
Bas Nieuwenhuizen
bas at basnieuwenhuizen.nl
Wed Nov 11 02:48:40 UTC 2020
Now that we actually expose the right number of planes we get
problems with modifier aware libva if we use an unsupported modifier.
So this ensures that we as much as possible use a modifier that
radeonsi can actually render to without going to extreme difficulty
to synchronize an extra DCC metadata plane with implicit sync.
Looking into the opaque metadata isn't ideal but this converts to
something mainly for the purpose of the userspace driver and falls
back properly if it isn't what we expect.
Signed-off-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_display.c | 106 ++++++++++++++++++--
1 file changed, 98 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c
index f66ce6ec4843..c9e98499b8a4 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c
@@ -631,6 +631,57 @@ amdgpu_lookup_format_info(u32 format, uint64_t modifier)
return NULL;
}
+
+/*
+ * Tries to extract the renderable DCC offset from the opaque metadata attached
+ * to the buffer.
+ */
+static int
+extract_render_dcc_offset(struct amdgpu_device *adev,
+ struct drm_gem_object *obj,
+ uint64_t *offset)
+{
+ struct amdgpu_bo *rbo;
+ int r = 0;
+ uint32_t metadata[10]; /* Something that fits a descriptor + header. */
+ uint32_t size;
+
+ rbo = gem_to_amdgpu_bo(obj);
+ r = amdgpu_bo_reserve(rbo, false);
+
+ if (unlikely(r)) {
+ /* Don't show error message when returning -ERESTARTSYS */
+ if (r != -ERESTARTSYS)
+ DRM_ERROR("Unable to reserve buffer: %d\n", r);
+ return r;
+ }
+
+ r = amdgpu_bo_get_metadata(rbo, metadata, sizeof(metadata), &size, NULL);
+ amdgpu_bo_unreserve(rbo);
+
+ if (r)
+ return r;
+
+ /*
+ * The first word is the metadata version, and we need space for at least
+ * the version + pci vendor+device id + 8 words for a descriptor.
+ */
+ if (size < 40 || metadata[0] != 1)
+ return -EINVAL;
+
+ if (adev->family >= AMDGPU_FAMILY_NV) {
+ /* resource word 6/7 META_DATA_ADDRESS{_LO} */
+ *offset = ((u64)metadata[9] << 16u) |
+ ((metadata[8] & 0xFF000000u) >> 16);
+ } else {
+ /* resource word 5/7 META_DATA_ADDRESS */
+ *offset = ((u64)metadata[9] << 8u) |
+ ((u64)(metadata[7] & 0x1FE0000u) << 23);
+ }
+
+ return 0;
+}
+
static int convert_tiling_flags_to_modifier(struct amdgpu_framebuffer *afb)
{
struct amdgpu_device *adev = drm_to_adev(afb->base.dev);
@@ -646,6 +697,8 @@ static int convert_tiling_flags_to_modifier(struct amdgpu_framebuffer *afb)
int pipe_xor_bits = 0;
int bank_xor_bits = 0;
int packers = 0;
+ int rb = 0;
+ int pipes = ilog2(adev->gfx.config.gb_addr_config_fields.num_pipes);
uint32_t dcc_offset = AMDGPU_TILING_GET(afb->tiling_flags, DCC_OFFSET_256B);
switch (swizzle >> 2) {
@@ -691,18 +744,17 @@ static int convert_tiling_flags_to_modifier(struct amdgpu_framebuffer *afb)
if (has_xor) {
switch (version) {
case AMD_FMT_MOD_TILE_VER_GFX10_RBPLUS:
- pipe_xor_bits = min(block_size_bits - 8,
- ilog2(adev->gfx.config.gb_addr_config_fields.num_pipes));
+ pipe_xor_bits = min(block_size_bits - 8, pipes);
packers = min(block_size_bits - 8 - pipe_xor_bits,
ilog2(adev->gfx.config.gb_addr_config_fields.num_pkrs));
break;
case AMD_FMT_MOD_TILE_VER_GFX10:
- pipe_xor_bits = min(block_size_bits - 8,
- ilog2(adev->gfx.config.gb_addr_config_fields.num_pipes));
+ pipe_xor_bits = min(block_size_bits - 8, pipes);
break;
case AMD_FMT_MOD_TILE_VER_GFX9:
- pipe_xor_bits = min(block_size_bits - 8,
- ilog2(adev->gfx.config.gb_addr_config_fields.num_pipes) +
+ rb = ilog2(adev->gfx.config.gb_addr_config_fields.num_se) +
+ ilog2(adev->gfx.config.gb_addr_config_fields.num_rb_per_se);
+ pipe_xor_bits = min(block_size_bits - 8, pipes +
ilog2(adev->gfx.config.gb_addr_config_fields.num_se));
bank_xor_bits = min(block_size_bits - 8 - pipe_xor_bits,
ilog2(adev->gfx.config.gb_addr_config_fields.num_banks));
@@ -721,6 +773,7 @@ static int convert_tiling_flags_to_modifier(struct amdgpu_framebuffer *afb)
bool dcc_i64b = AMDGPU_TILING_GET(afb->tiling_flags, DCC_INDEPENDENT_64B) != 0;
bool dcc_i128b = version >= AMD_FMT_MOD_TILE_VER_GFX10_RBPLUS;
const struct drm_format_info *format_info;
+ u64 render_dcc_offset;
/* Enable constant encode on RAVEN2 and later. */
bool dcc_constant_encode = adev->asic_type > CHIP_RAVEN ||
@@ -738,8 +791,45 @@ static int convert_tiling_flags_to_modifier(struct amdgpu_framebuffer *afb)
AMD_FMT_MOD_SET(DCC_MAX_COMPRESSED_BLOCK, max_cblock_size);
afb->base.offsets[1] = dcc_offset * 256 + afb->base.offsets[0];
- afb->base.pitches[1] = AMDGPU_TILING_GET(afb->tiling_flags, DCC_PITCH_MAX) + 1;
-
+ afb->base.pitches[1] =
+ AMDGPU_TILING_GET(afb->tiling_flags, DCC_PITCH_MAX) + 1;
+
+ /*
+ * If the userspace driver uses retiling the tiling flags do not contain
+ * info on the renderable DCC buffer. Luckily the opaque metadata contains
+ * the info so we can try to extract it. The kernel does not use this info
+ * but we should convert it to a modifier plane for getfb2, so the
+ * userspace driver that gets it doesn't have to juggle around another DCC
+ * plane internally.
+ */
+ if (extract_render_dcc_offset(adev, afb->base.obj[0],
+ &render_dcc_offset) == 0 &&
+ render_dcc_offset != 0 &&
+ render_dcc_offset != afb->base.offsets[1] &&
+ render_dcc_offset < UINT_MAX) {
+ uint32_t dcc_block_bits; /* of base surface data */
+
+ modifier |= AMD_FMT_MOD_SET(DCC_RETILE, 1);
+ afb->base.offsets[2] = render_dcc_offset;
+
+ if (adev->family >= AMDGPU_FAMILY_NV) {
+ int extra_pipe = 0;
+
+ if (adev->asic_type >= CHIP_SIENNA_CICHLID &&
+ pipes == packers && pipes > 1)
+ extra_pipe = 1;
+
+ dcc_block_bits = max(20, 16 + pipes + extra_pipe);
+ } else {
+ modifier |= AMD_FMT_MOD_SET(RB, rb) |
+ AMD_FMT_MOD_SET(PIPE, pipes);
+ dcc_block_bits = max(20, 18 + rb);
+ }
+
+ dcc_block_bits -= ilog2(afb->base.format->cpp[0]);
+ afb->base.pitches[2] = ALIGN(afb->base.width,
+ 1u << ((dcc_block_bits + 1) / 2));
+ }
format_info = amdgpu_lookup_format_info(afb->base.format->format,
modifier);
if (!format_info)
--
2.29.2
More information about the amd-gfx
mailing list