From cc8ac112fca0c141eace0e3703fca08ebe0962e0 Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Fri, 6 Aug 2021 20:06:30 -0400 Subject: [PATCH 1/4] nvhost_nvdec_common: Stub UnmapBuffer Ioctl Skip unmapping nvdec buffers to avoid breaking the continuity of the VP9 reference frame addresses, and the risk of invalidating data before the async GPU thread is done with it. --- .../nvdrv/devices/nvhost_nvdec_common.cpp | 27 +++---------------- 1 file changed, 4 insertions(+), 23 deletions(-) diff --git a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp index 1403a39d0..207a5b90d 100644 --- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp +++ b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp @@ -178,30 +178,11 @@ NvResult nvhost_nvdec_common::MapBuffer(const std::vector& input, std::vecto } NvResult nvhost_nvdec_common::UnmapBuffer(const std::vector& input, std::vector& output) { - IoctlMapBuffer params{}; - std::memcpy(¶ms, input.data(), sizeof(IoctlMapBuffer)); - std::vector cmd_buffer_handles(params.num_entries); - SliceVectors(input, cmd_buffer_handles, params.num_entries, sizeof(IoctlMapBuffer)); - - auto& gpu = system.GPU(); - - for (auto& cmd_buffer : cmd_buffer_handles) { - const auto object{nvmap_dev->GetObject(cmd_buffer.map_handle)}; - if (!object) { - LOG_ERROR(Service_NVDRV, "invalid cmd_buffer nvmap_handle={:X}", cmd_buffer.map_handle); - std::memcpy(output.data(), ¶ms, output.size()); - return NvResult::InvalidState; - } - if (const auto size{RemoveBufferMap(object->dma_map_addr)}; size) { - gpu.MemoryManager().Unmap(object->dma_map_addr, *size); - } else { - // This occurs quite frequently, however does not seem to impact functionality - LOG_DEBUG(Service_NVDRV, "invalid offset=0x{:X} dma=0x{:X}", object->addr, - object->dma_map_addr); - } - object->dma_map_addr = 0; - } + // This is intntionally stubbed. + // Skip unmapping buffers here, as to not break the continuity of the VP9 reference frame + // addresses, and risk invalidating data before the async GPU thread is done with it std::memset(output.data(), 0, output.size()); + LOG_DEBUG(Service_NVDRV, "(STUBBED) called"); return NvResult::Success; } From a3f80a97a3f82dd739febc2cee496b2016eb282a Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Sun, 1 Aug 2021 22:26:00 -0400 Subject: [PATCH 2/4] vp9: Fix reference frame refreshes This resolves the artifacting when decoding VP9 streams. --- src/video_core/command_classes/codecs/vp9.cpp | 75 +++++++------------ src/video_core/command_classes/codecs/vp9.h | 2 +- 2 files changed, 30 insertions(+), 47 deletions(-) diff --git a/src/video_core/command_classes/codecs/vp9.cpp b/src/video_core/command_classes/codecs/vp9.cpp index 902bc2a98..b3e3462d6 100644 --- a/src/video_core/command_classes/codecs/vp9.cpp +++ b/src/video_core/command_classes/codecs/vp9.cpp @@ -613,7 +613,7 @@ VpxBitStreamWriter VP9::ComposeUncompressedHeader() { // Reset context prev_frame_probs = default_probs; - swap_next_golden = false; + swap_ref_indices = false; loop_filter_ref_deltas.fill(0); loop_filter_mode_deltas.fill(0); @@ -626,73 +626,57 @@ VpxBitStreamWriter VP9::ComposeUncompressedHeader() { // intra only, meaning the frame can be recreated with no other references current_frame_info.intra_only = true; - } else { - if (!current_frame_info.show_frame) { uncomp_writer.WriteBit(current_frame_info.intra_only); - if (!current_frame_info.last_frame_was_key) { - swap_next_golden = !swap_next_golden; - } } else { current_frame_info.intra_only = false; } if (!current_frame_info.error_resilient_mode) { uncomp_writer.WriteU(0, 2); // Reset frame context. } - - // Last, Golden, Altref frames - std::array ref_frame_index{0, 1, 2}; - - // Set when next frame is hidden - // altref and golden references are swapped - if (swap_next_golden) { - ref_frame_index = std::array{0, 2, 1}; + const auto& curr_offsets = current_frame_info.frame_offsets; + const auto& next_offsets = next_frame.info.frame_offsets; + const bool ref_frames_different = curr_offsets[1] != curr_offsets[2]; + const bool next_references_swap = + (next_offsets[1] == curr_offsets[2]) || (next_offsets[2] == curr_offsets[1]); + const bool needs_ref_swap = ref_frames_different && next_references_swap; + if (needs_ref_swap) { + swap_ref_indices = !swap_ref_indices; } + union { + u32 raw; + BitField<0, 1, u32> refresh_last; + BitField<1, 2, u32> refresh_golden; + BitField<2, 1, u32> refresh_alt; + } refresh_frame_flags; - // update Last Frame - u64 refresh_frame_flags = 1; - - // golden frame may refresh, determined if the next golden frame offset is changed - bool golden_refresh = false; - if (grace_period <= 0) { - for (s32 index = 1; index < 3; ++index) { - if (current_frame_info.frame_offsets[index] != - next_frame.info.frame_offsets[index]) { - current_frame_info.refresh_frame[index] = true; - golden_refresh = true; - grace_period = 3; - } + refresh_frame_flags.raw = 0; + for (u32 index = 0; index < 3; ++index) { + // Refresh indices that use the current frame as an index + if (curr_offsets[3] == next_offsets[index]) { + refresh_frame_flags.raw |= 1u << index; } } - - if (current_frame_info.show_frame && - (!next_frame.info.show_frame || next_frame.info.is_key_frame)) { - // Update golden frame - refresh_frame_flags = swap_next_golden ? 2 : 4; + if (swap_ref_indices) { + const u32 temp = refresh_frame_flags.refresh_golden; + refresh_frame_flags.refresh_golden.Assign(refresh_frame_flags.refresh_alt.Value()); + refresh_frame_flags.refresh_alt.Assign(temp); } - - if (!current_frame_info.show_frame) { - // Update altref - refresh_frame_flags = swap_next_golden ? 2 : 4; - } else if (golden_refresh) { - refresh_frame_flags = 3; - } - if (current_frame_info.intra_only) { uncomp_writer.WriteU(frame_sync_code, 24); - uncomp_writer.WriteU(static_cast(refresh_frame_flags), 8); + uncomp_writer.WriteU(refresh_frame_flags.raw, 8); uncomp_writer.WriteU(current_frame_info.frame_size.width - 1, 16); uncomp_writer.WriteU(current_frame_info.frame_size.height - 1, 16); uncomp_writer.WriteBit(false); // Render and frame size different. } else { - uncomp_writer.WriteU(static_cast(refresh_frame_flags), 8); - - for (s32 index = 1; index < 4; index++) { + const bool swap_indices = needs_ref_swap ^ swap_ref_indices; + const auto ref_frame_index = swap_indices ? std::array{0, 2, 1} : std::array{0, 1, 2}; + uncomp_writer.WriteU(refresh_frame_flags.raw, 8); + for (size_t index = 1; index < 4; index++) { uncomp_writer.WriteU(ref_frame_index[index - 1], 3); uncomp_writer.WriteU(current_frame_info.ref_frame_sign_bias[index], 1); } - uncomp_writer.WriteBit(true); // Frame size with refs. uncomp_writer.WriteBit(false); // Render and frame size different. uncomp_writer.WriteBit(current_frame_info.allow_high_precision_mv); @@ -812,7 +796,6 @@ const std::vector& VP9::ComposeFrameHeader(const NvdecCommon::NvdecRegisters current_frame_info = curr_frame.info; bitstream = std::move(curr_frame.bit_stream); } - // The uncompressed header routine sets PrevProb parameters needed for the compressed header auto uncomp_writer = ComposeUncompressedHeader(); std::vector compressed_header = ComposeCompressedHeader(); diff --git a/src/video_core/command_classes/codecs/vp9.h b/src/video_core/command_classes/codecs/vp9.h index 8396c8105..bb4d0d972 100644 --- a/src/video_core/command_classes/codecs/vp9.h +++ b/src/video_core/command_classes/codecs/vp9.h @@ -184,7 +184,7 @@ private: std::array frame_ctxs{}; Vp9FrameContainer next_frame{}; Vp9FrameContainer next_next_frame{}; - bool swap_next_golden{}; + bool swap_ref_indices{}; Vp9PictureInfo current_frame_info{}; Vp9EntropyProbs prev_frame_probs{}; From e3688f0627aa460bffb9b79c7a45e4ed615a1fa7 Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Mon, 2 Aug 2021 00:08:41 -0400 Subject: [PATCH 3/4] vp9: Cleanup unused variables With reference frames refreshes fix, we no longer need to buffer two frames in advance. We can also remove other unused or otherwise unneeded variables. --- src/video_core/command_classes/codecs/vp9.cpp | 57 +++++-------------- src/video_core/command_classes/codecs/vp9.h | 12 +--- .../command_classes/codecs/vp9_types.h | 6 -- 3 files changed, 17 insertions(+), 58 deletions(-) diff --git a/src/video_core/command_classes/codecs/vp9.cpp b/src/video_core/command_classes/codecs/vp9.cpp index b3e3462d6..7eecb3991 100644 --- a/src/video_core/command_classes/codecs/vp9.cpp +++ b/src/video_core/command_classes/codecs/vp9.cpp @@ -11,6 +11,9 @@ namespace Tegra::Decoder { namespace { +constexpr u32 diff_update_probability = 252; +constexpr u32 frame_sync_code = 0x498342; + // Default compressed header probabilities once frame context resets constexpr Vp9EntropyProbs default_probs{ .y_mode_prob{ @@ -361,8 +364,7 @@ Vp9PictureInfo VP9::GetVp9PictureInfo(const NvdecCommon::NvdecRegisters& state) InsertEntropy(state.vp9_entropy_probs_offset, vp9_info.entropy); // surface_luma_offset[0:3] contains the address of the reference frame offsets in the following - // order: last, golden, altref, current. It may be worthwhile to track the updates done here - // to avoid buffering frame data needed for reference frame updating in the header composition. + // order: last, golden, altref, current. std::copy(state.surface_luma_offset.begin(), state.surface_luma_offset.begin() + 4, vp9_info.frame_offsets.begin()); @@ -384,33 +386,18 @@ Vp9FrameContainer VP9::GetCurrentFrame(const NvdecCommon::NvdecRegisters& state) gpu.MemoryManager().ReadBlock(state.frame_bitstream_offset, current_frame.bit_stream.data(), current_frame.info.bitstream_size); } - // Buffer two frames, saving the last show frame info - if (!next_next_frame.bit_stream.empty()) { + if (!next_frame.bit_stream.empty()) { Vp9FrameContainer temp{ .info = current_frame.info, .bit_stream = std::move(current_frame.bit_stream), }; - next_next_frame.info.show_frame = current_frame.info.last_frame_shown; - current_frame.info = next_next_frame.info; - current_frame.bit_stream = std::move(next_next_frame.bit_stream); - next_next_frame = std::move(temp); - - if (!next_frame.bit_stream.empty()) { - Vp9FrameContainer temp2{ - .info = current_frame.info, - .bit_stream = std::move(current_frame.bit_stream), - }; - next_frame.info.show_frame = current_frame.info.last_frame_shown; - current_frame.info = next_frame.info; - current_frame.bit_stream = std::move(next_frame.bit_stream); - next_frame = std::move(temp2); - } else { - next_frame.info = current_frame.info; - next_frame.bit_stream = std::move(current_frame.bit_stream); - } + next_frame.info.show_frame = current_frame.info.last_frame_shown; + current_frame.info = next_frame.info; + current_frame.bit_stream = std::move(next_frame.bit_stream); + next_frame = std::move(temp); } else { - next_next_frame.info = current_frame.info; - next_next_frame.bit_stream = std::move(current_frame.bit_stream); + next_frame.info = current_frame.info; + next_frame.bit_stream = std::move(current_frame.bit_stream); } return current_frame; } @@ -616,13 +603,7 @@ VpxBitStreamWriter VP9::ComposeUncompressedHeader() { swap_ref_indices = false; loop_filter_ref_deltas.fill(0); loop_filter_mode_deltas.fill(0); - - // allow frames offsets to stabilize before checking for golden frames - grace_period = 4; - - // On key frames, all frame slots are set to the current frame, - // so the value of the selected slot doesn't really matter. - frame_ctxs.fill({current_frame_number, false, default_probs}); + frame_ctxs.fill(default_probs); // intra only, meaning the frame can be recreated with no other references current_frame_info.intra_only = true; @@ -698,10 +679,9 @@ VpxBitStreamWriter VP9::ComposeUncompressedHeader() { frame_ctx_idx = 1; } - uncomp_writer.WriteU(frame_ctx_idx, 2); // Frame context index. - prev_frame_probs = - frame_ctxs[frame_ctx_idx].probs; // reference probabilities for compressed header - frame_ctxs[frame_ctx_idx] = {current_frame_number, false, current_frame_info.entropy}; + uncomp_writer.WriteU(frame_ctx_idx, 2); // Frame context index. + prev_frame_probs = frame_ctxs[frame_ctx_idx]; // reference probabilities for compressed header + frame_ctxs[frame_ctx_idx] = current_frame_info.entropy; uncomp_writer.WriteU(current_frame_info.first_level, 6); uncomp_writer.WriteU(current_frame_info.sharpness_level, 3); @@ -811,13 +791,6 @@ const std::vector& VP9::ComposeFrameHeader(const NvdecCommon::NvdecRegisters frame.begin() + uncompressed_header.size()); std::copy(bitstream.begin(), bitstream.end(), frame.begin() + uncompressed_header.size() + compressed_header.size()); - - // keep track of frame number - current_frame_number++; - grace_period--; - - // don't display hidden frames - hidden = !current_frame_info.show_frame; return frame; } diff --git a/src/video_core/command_classes/codecs/vp9.h b/src/video_core/command_classes/codecs/vp9.h index bb4d0d972..e6e9fc17e 100644 --- a/src/video_core/command_classes/codecs/vp9.h +++ b/src/video_core/command_classes/codecs/vp9.h @@ -14,7 +14,6 @@ namespace Tegra { class GPU; -enum class FrameType { KeyFrame = 0, InterFrame = 1 }; namespace Decoder { /// The VpxRangeEncoder, and VpxBitStreamWriter classes are used to compose the @@ -124,7 +123,7 @@ public: /// Returns true if the most recent frame was a hidden frame. [[nodiscard]] bool WasFrameHidden() const { - return hidden; + return !current_frame_info.show_frame; } private: @@ -178,19 +177,12 @@ private: std::array loop_filter_ref_deltas{}; std::array loop_filter_mode_deltas{}; - bool hidden = false; - s64 current_frame_number = -2; // since we buffer 2 frames - s32 grace_period = 6; // frame offsets need to stabilize - std::array frame_ctxs{}; Vp9FrameContainer next_frame{}; - Vp9FrameContainer next_next_frame{}; + std::array frame_ctxs{}; bool swap_ref_indices{}; Vp9PictureInfo current_frame_info{}; Vp9EntropyProbs prev_frame_probs{}; - - s32 diff_update_probability = 252; - s32 frame_sync_code = 0x498342; }; } // namespace Decoder diff --git a/src/video_core/command_classes/codecs/vp9_types.h b/src/video_core/command_classes/codecs/vp9_types.h index 2da14f3ca..6820afa26 100644 --- a/src/video_core/command_classes/codecs/vp9_types.h +++ b/src/video_core/command_classes/codecs/vp9_types.h @@ -296,12 +296,6 @@ struct RefPoolElement { bool refresh{}; }; -struct FrameContexts { - s64 from; - bool adapted; - Vp9EntropyProbs probs; -}; - #define ASSERT_POSITION(field_name, position) \ static_assert(offsetof(Vp9EntropyProbs, field_name) == position, \ "Field " #field_name " has invalid position") From 156ea746a3459cb8db676946b532d0c6363f0b52 Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Fri, 6 Aug 2021 20:11:12 -0400 Subject: [PATCH 4/4] nvhost_nvdec_common: Remove BufferMap This was mainly used to keep track of mapped buffers for later unmapping. Since unmap is no longer implemented, this no longer seves a valuable purpose. --- .../nvdrv/devices/nvhost_nvdec_common.cpp | 31 ------------- .../nvdrv/devices/nvhost_nvdec_common.h | 45 ------------------- 2 files changed, 76 deletions(-) diff --git a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp index 207a5b90d..845de724d 100644 --- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp +++ b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp @@ -166,8 +166,6 @@ NvResult nvhost_nvdec_common::MapBuffer(const std::vector& input, std::vecto LOG_ERROR(Service_NVDRV, "failed to map size={}", object->size); } else { cmd_buffer.map_address = object->dma_map_addr; - AddBufferMap(object->dma_map_addr, object->size, object->addr, - object->status == nvmap::Object::Status::Allocated); } } std::memcpy(output.data(), ¶ms, sizeof(IoctlMapBuffer)); @@ -193,33 +191,4 @@ NvResult nvhost_nvdec_common::SetSubmitTimeout(const std::vector& input, return NvResult::Success; } -std::optional nvhost_nvdec_common::FindBufferMap( - GPUVAddr gpu_addr) const { - const auto it = std::find_if( - buffer_mappings.begin(), buffer_mappings.upper_bound(gpu_addr), [&](const auto& entry) { - return (gpu_addr >= entry.second.StartAddr() && gpu_addr < entry.second.EndAddr()); - }); - - ASSERT(it != buffer_mappings.end()); - return it->second; -} - -void nvhost_nvdec_common::AddBufferMap(GPUVAddr gpu_addr, std::size_t size, VAddr cpu_addr, - bool is_allocated) { - buffer_mappings.insert_or_assign(gpu_addr, BufferMap{gpu_addr, size, cpu_addr, is_allocated}); -} - -std::optional nvhost_nvdec_common::RemoveBufferMap(GPUVAddr gpu_addr) { - const auto iter{buffer_mappings.find(gpu_addr)}; - if (iter == buffer_mappings.end()) { - return std::nullopt; - } - std::size_t size = 0; - if (iter->second.IsAllocated()) { - size = iter->second.Size(); - } - buffer_mappings.erase(iter); - return size; -} - } // namespace Service::Nvidia::Devices diff --git a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h index da10f5f41..af59f00d2 100644 --- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h +++ b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h @@ -23,45 +23,6 @@ public: ~nvhost_nvdec_common() override; protected: - class BufferMap final { - public: - constexpr BufferMap() = default; - - constexpr BufferMap(GPUVAddr start_addr_, std::size_t size_) - : start_addr{start_addr_}, end_addr{start_addr_ + size_} {} - - constexpr BufferMap(GPUVAddr start_addr_, std::size_t size_, VAddr cpu_addr_, - bool is_allocated_) - : start_addr{start_addr_}, end_addr{start_addr_ + size_}, cpu_addr{cpu_addr_}, - is_allocated{is_allocated_} {} - - constexpr VAddr StartAddr() const { - return start_addr; - } - - constexpr VAddr EndAddr() const { - return end_addr; - } - - constexpr std::size_t Size() const { - return end_addr - start_addr; - } - - constexpr VAddr CpuAddr() const { - return cpu_addr; - } - - constexpr bool IsAllocated() const { - return is_allocated; - } - - private: - GPUVAddr start_addr{}; - GPUVAddr end_addr{}; - VAddr cpu_addr{}; - bool is_allocated{}; - }; - struct IoctlSetNvmapFD { s32_le nvmap_fd{}; }; @@ -154,17 +115,11 @@ protected: NvResult UnmapBuffer(const std::vector& input, std::vector& output); NvResult SetSubmitTimeout(const std::vector& input, std::vector& output); - std::optional FindBufferMap(GPUVAddr gpu_addr) const; - void AddBufferMap(GPUVAddr gpu_addr, std::size_t size, VAddr cpu_addr, bool is_allocated); - std::optional RemoveBufferMap(GPUVAddr gpu_addr); - s32_le nvmap_fd{}; u32_le submit_timeout{}; std::shared_ptr nvmap_dev; SyncpointManager& syncpoint_manager; std::array device_syncpoints{}; - // This is expected to be ordered, therefore we must use a map, not unordered_map - std::map buffer_mappings; }; }; // namespace Devices } // namespace Service::Nvidia