From b8c906f9d1acbd59c264bd3ad335c6d75f92da5c Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Sat, 24 Jun 2023 19:58:45 -0400
Subject: [PATCH 1/5] scratch_buffer: Add member types to ScratchBuffer
Allows for implicit conversion to std::span<T>.
---
src/common/scratch_buffer.h | 46 +++++++++++++++++++++----------------
1 file changed, 26 insertions(+), 20 deletions(-)
diff --git a/src/common/scratch_buffer.h b/src/common/scratch_buffer.h
index 6fe9079531..d5961b0208 100644
--- a/src/common/scratch_buffer.h
+++ b/src/common/scratch_buffer.h
@@ -5,7 +5,6 @@
#include <iterator>
-#include "common/concepts.h"
#include "common/make_unique_for_overwrite.h"
namespace Common {
@@ -19,15 +18,22 @@ namespace Common {
template <typename T>
class ScratchBuffer {
public:
- using iterator = T*;
- using const_iterator = const T*;
- using value_type = T;
using element_type = T;
- using iterator_category = std::contiguous_iterator_tag;
+ using value_type = T;
+ using size_type = size_t;
+ using difference_type = std::ptrdiff_t;
+ using pointer = T*;
+ using const_pointer = const T*;
+ using reference = T&;
+ using const_reference = const T&;
+ using iterator = pointer;
+ using const_iterator = const_pointer;
+ using iterator_category = std::random_access_iterator_tag;
+ using iterator_concept = std::contiguous_iterator_tag;
ScratchBuffer() = default;
- explicit ScratchBuffer(size_t initial_capacity)
+ explicit ScratchBuffer(size_type initial_capacity)
: last_requested_size{initial_capacity}, buffer_capacity{initial_capacity},
buffer{Common::make_unique_for_overwrite<T[]>(initial_capacity)} {}
@@ -39,7 +45,7 @@ public:
/// This will only grow the buffer's capacity if size is greater than the current capacity.
/// The previously held data will remain intact.
- void resize(size_t size) {
+ void resize(size_type size) {
if (size > buffer_capacity) {
auto new_buffer = Common::make_unique_for_overwrite<T[]>(size);
std::move(buffer.get(), buffer.get() + buffer_capacity, new_buffer.get());
@@ -51,7 +57,7 @@ public:
/// This will only grow the buffer's capacity if size is greater than the current capacity.
/// The previously held data will be destroyed if a reallocation occurs.
- void resize_destructive(size_t size) {
+ void resize_destructive(size_type size) {
if (size > buffer_capacity) {
buffer_capacity = size;
buffer = Common::make_unique_for_overwrite<T[]>(buffer_capacity);
@@ -59,43 +65,43 @@ public:
last_requested_size = size;
}
- [[nodiscard]] T* data() noexcept {
+ [[nodiscard]] pointer data() noexcept {
return buffer.get();
}
- [[nodiscard]] const T* data() const noexcept {
+ [[nodiscard]] const_pointer data() const noexcept {
return buffer.get();
}
- [[nodiscard]] T* begin() noexcept {
+ [[nodiscard]] iterator begin() noexcept {
return data();
}
- [[nodiscard]] const T* begin() const noexcept {
+ [[nodiscard]] const_iterator begin() const noexcept {
return data();
}
- [[nodiscard]] T* end() noexcept {
+ [[nodiscard]] iterator end() noexcept {
return data() + last_requested_size;
}
- [[nodiscard]] const T* end() const noexcept {
+ [[nodiscard]] const_iterator end() const noexcept {
return data() + last_requested_size;
}
- [[nodiscard]] T& operator[](size_t i) {
+ [[nodiscard]] reference operator[](size_type i) {
return buffer[i];
}
- [[nodiscard]] const T& operator[](size_t i) const {
+ [[nodiscard]] const_reference operator[](size_type i) const {
return buffer[i];
}
- [[nodiscard]] size_t size() const noexcept {
+ [[nodiscard]] size_type size() const noexcept {
return last_requested_size;
}
- [[nodiscard]] size_t capacity() const noexcept {
+ [[nodiscard]] size_type capacity() const noexcept {
return buffer_capacity;
}
@@ -106,8 +112,8 @@ public:
}
private:
- size_t last_requested_size{};
- size_t buffer_capacity{};
+ size_type last_requested_size{};
+ size_type buffer_capacity{};
std::unique_ptr<T[]> buffer{};
};
From fbd85417ffdfe23dd4b4d3d13518244bd00be361 Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Sat, 24 Jun 2023 20:31:20 -0400
Subject: [PATCH 2/5] ring_buffer: Fix const usage on std::span
---
src/common/ring_buffer.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/src/common/ring_buffer.h b/src/common/ring_buffer.h
index 416680d445..5c961b202f 100644
--- a/src/common/ring_buffer.h
+++ b/src/common/ring_buffer.h
@@ -54,7 +54,7 @@ public:
return push_count;
}
- std::size_t Push(const std::span<T> input) {
+ std::size_t Push(std::span<const T> input) {
return Push(input.data(), input.size());
}
From 310b6cf4af940fa07666400426bbcca815c5375c Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Sat, 24 Jun 2023 21:58:23 -0400
Subject: [PATCH 3/5] general: Use ScratchBuffer where possible
---
src/core/hle/service/audio/audin_u.cpp | 16 ++++++------
src/core/hle/service/audio/audout_u.cpp | 20 ++++++++-------
src/core/hle/service/audio/audren_u.cpp | 23 +++++++++--------
src/core/hle/service/audio/hwopus.cpp | 9 ++++---
.../hle/service/nvdrv/nvdrv_interface.cpp | 25 ++++++++++---------
src/core/hle/service/nvdrv/nvdrv_interface.h | 5 ++--
src/video_core/host1x/codecs/codec.cpp | 2 +-
src/video_core/host1x/codecs/h264.cpp | 14 +++++------
src/video_core/host1x/codecs/h264.h | 12 ++++++---
src/video_core/host1x/codecs/vp8.cpp | 2 +-
src/video_core/host1x/codecs/vp8.h | 7 +++---
src/video_core/host1x/codecs/vp9.cpp | 1 +
src/video_core/host1x/codecs/vp9.h | 8 +++---
src/video_core/host1x/codecs/vp9_types.h | 1 +
14 files changed, 81 insertions(+), 64 deletions(-)
diff --git a/src/core/hle/service/audio/audin_u.cpp b/src/core/hle/service/audio/audin_u.cpp
index c8d574993b..526a391307 100644
--- a/src/core/hle/service/audio/audin_u.cpp
+++ b/src/core/hle/service/audio/audin_u.cpp
@@ -5,7 +5,7 @@
#include "audio_core/renderer/audio_device.h"
#include "common/common_funcs.h"
#include "common/logging/log.h"
-#include "common/settings.h"
+#include "common/scratch_buffer.h"
#include "common/string_util.h"
#include "core/core.h"
#include "core/hle/kernel/k_event.h"
@@ -124,12 +124,15 @@ private:
void GetReleasedAudioInBuffer(HLERequestContext& ctx) {
const auto write_buffer_size = ctx.GetWriteBufferNumElements<u64>();
- tmp_buffer.resize_destructive(write_buffer_size);
- tmp_buffer[0] = 0;
+ released_buffer.resize_destructive(write_buffer_size);
+ released_buffer[0] = 0;
- const auto count = impl->GetReleasedBuffers(tmp_buffer);
+ const auto count = impl->GetReleasedBuffers(released_buffer);
- ctx.WriteBuffer(tmp_buffer);
+ LOG_TRACE(Service_Audio, "called. Session {} released {} buffers",
+ impl->GetSystem().GetSessionId(), count);
+
+ ctx.WriteBuffer(released_buffer);
IPC::ResponseBuilder rb{ctx, 3};
rb.Push(ResultSuccess);
@@ -155,7 +158,6 @@ private:
LOG_DEBUG(Service_Audio, "called. Buffer count={}", buffer_count);
IPC::ResponseBuilder rb{ctx, 3};
-
rb.Push(ResultSuccess);
rb.Push(buffer_count);
}
@@ -195,7 +197,7 @@ private:
KernelHelpers::ServiceContext service_context;
Kernel::KEvent* event;
std::shared_ptr<AudioCore::AudioIn::In> impl;
- Common::ScratchBuffer<u64> tmp_buffer;
+ Common::ScratchBuffer<u64> released_buffer;
};
AudInU::AudInU(Core::System& system_)
diff --git a/src/core/hle/service/audio/audout_u.cpp b/src/core/hle/service/audio/audout_u.cpp
index 032c8c11f5..23f84a29f7 100644
--- a/src/core/hle/service/audio/audout_u.cpp
+++ b/src/core/hle/service/audio/audout_u.cpp
@@ -9,6 +9,7 @@
#include "audio_core/renderer/audio_device.h"
#include "common/common_funcs.h"
#include "common/logging/log.h"
+#include "common/scratch_buffer.h"
#include "common/string_util.h"
#include "common/swap.h"
#include "core/core.h"
@@ -102,8 +103,8 @@ private:
AudioOutBuffer buffer{};
std::memcpy(&buffer, in_buffer.data(), sizeof(AudioOutBuffer));
- [[maybe_unused]] auto sessionid{impl->GetSystem().GetSessionId()};
- LOG_TRACE(Service_Audio, "called. Session {} Appending buffer {:08X}", sessionid, tag);
+ LOG_TRACE(Service_Audio, "called. Session {} Appending buffer {:08X}",
+ impl->GetSystem().GetSessionId(), tag);
auto result = impl->AppendBuffer(buffer, tag);
@@ -123,12 +124,15 @@ private:
void GetReleasedAudioOutBuffers(HLERequestContext& ctx) {
const auto write_buffer_size = ctx.GetWriteBufferNumElements<u64>();
- tmp_buffer.resize_destructive(write_buffer_size);
- tmp_buffer[0] = 0;
+ released_buffer.resize_destructive(write_buffer_size);
+ released_buffer[0] = 0;
- const auto count = impl->GetReleasedBuffers(tmp_buffer);
+ const auto count = impl->GetReleasedBuffers(released_buffer);
- ctx.WriteBuffer(tmp_buffer);
+ ctx.WriteBuffer(released_buffer);
+
+ LOG_TRACE(Service_Audio, "called. Session {} released {} buffers",
+ impl->GetSystem().GetSessionId(), count);
IPC::ResponseBuilder rb{ctx, 3};
rb.Push(ResultSuccess);
@@ -154,7 +158,6 @@ private:
LOG_DEBUG(Service_Audio, "called. Buffer count={}", buffer_count);
IPC::ResponseBuilder rb{ctx, 3};
-
rb.Push(ResultSuccess);
rb.Push(buffer_count);
}
@@ -165,7 +168,6 @@ private:
LOG_DEBUG(Service_Audio, "called. Played samples={}", samples_played);
IPC::ResponseBuilder rb{ctx, 4};
-
rb.Push(ResultSuccess);
rb.Push(samples_played);
}
@@ -205,7 +207,7 @@ private:
KernelHelpers::ServiceContext service_context;
Kernel::KEvent* event;
std::shared_ptr<AudioCore::AudioOut::Out> impl;
- Common::ScratchBuffer<u64> tmp_buffer;
+ Common::ScratchBuffer<u64> released_buffer;
};
AudOutU::AudOutU(Core::System& system_)
diff --git a/src/core/hle/service/audio/audren_u.cpp b/src/core/hle/service/audio/audren_u.cpp
index 12845c23a3..003870176b 100644
--- a/src/core/hle/service/audio/audren_u.cpp
+++ b/src/core/hle/service/audio/audren_u.cpp
@@ -15,6 +15,7 @@
#include "common/common_funcs.h"
#include "common/logging/log.h"
#include "common/polyfill_ranges.h"
+#include "common/scratch_buffer.h"
#include "common/string_util.h"
#include "core/core.h"
#include "core/hle/kernel/k_event.h"
@@ -119,23 +120,23 @@ private:
auto is_buffer_b{ctx.BufferDescriptorB()[0].Size() != 0};
if (is_buffer_b) {
const auto buffersB{ctx.BufferDescriptorB()};
- tmp_output.resize_destructive(buffersB[0].Size());
- tmp_performance.resize_destructive(buffersB[1].Size());
+ output_buffer.resize_destructive(buffersB[0].Size());
+ performance_buffer.resize_destructive(buffersB[1].Size());
} else {
const auto buffersC{ctx.BufferDescriptorC()};
- tmp_output.resize_destructive(buffersC[0].Size());
- tmp_performance.resize_destructive(buffersC[1].Size());
+ output_buffer.resize_destructive(buffersC[0].Size());
+ performance_buffer.resize_destructive(buffersC[1].Size());
}
- auto result = impl->RequestUpdate(input, tmp_performance, tmp_output);
+ auto result = impl->RequestUpdate(input, performance_buffer, output_buffer);
if (result.IsSuccess()) {
if (is_buffer_b) {
- ctx.WriteBufferB(tmp_output.data(), tmp_output.size(), 0);
- ctx.WriteBufferB(tmp_performance.data(), tmp_performance.size(), 1);
+ ctx.WriteBufferB(output_buffer.data(), output_buffer.size(), 0);
+ ctx.WriteBufferB(performance_buffer.data(), performance_buffer.size(), 1);
} else {
- ctx.WriteBufferC(tmp_output.data(), tmp_output.size(), 0);
- ctx.WriteBufferC(tmp_performance.data(), tmp_performance.size(), 1);
+ ctx.WriteBufferC(output_buffer.data(), output_buffer.size(), 0);
+ ctx.WriteBufferC(performance_buffer.data(), performance_buffer.size(), 1);
}
} else {
LOG_ERROR(Service_Audio, "RequestUpdate failed error 0x{:02X}!", result.description);
@@ -233,8 +234,8 @@ private:
Kernel::KEvent* rendered_event;
Manager& manager;
std::unique_ptr<Renderer> impl;
- Common::ScratchBuffer<u8> tmp_output;
- Common::ScratchBuffer<u8> tmp_performance;
+ Common::ScratchBuffer<u8> output_buffer;
+ Common::ScratchBuffer<u8> performance_buffer;
};
class IAudioDevice final : public ServiceFramework<IAudioDevice> {
diff --git a/src/core/hle/service/audio/hwopus.cpp b/src/core/hle/service/audio/hwopus.cpp
index c835f6cb76..fa77007f37 100644
--- a/src/core/hle/service/audio/hwopus.cpp
+++ b/src/core/hle/service/audio/hwopus.cpp
@@ -11,6 +11,7 @@
#include "common/assert.h"
#include "common/logging/log.h"
+#include "common/scratch_buffer.h"
#include "core/hle/service/audio/hwopus.h"
#include "core/hle/service/ipc_helpers.h"
@@ -68,13 +69,13 @@ private:
ExtraBehavior extra_behavior) {
u32 consumed = 0;
u32 sample_count = 0;
- tmp_samples.resize_destructive(ctx.GetWriteBufferNumElements<opus_int16>());
+ samples.resize_destructive(ctx.GetWriteBufferNumElements<opus_int16>());
if (extra_behavior == ExtraBehavior::ResetContext) {
ResetDecoderContext();
}
- if (!DecodeOpusData(consumed, sample_count, ctx.ReadBuffer(), tmp_samples, performance)) {
+ if (!DecodeOpusData(consumed, sample_count, ctx.ReadBuffer(), samples, performance)) {
LOG_ERROR(Audio, "Failed to decode opus data");
IPC::ResponseBuilder rb{ctx, 2};
// TODO(ogniK): Use correct error code
@@ -90,7 +91,7 @@ private:
if (performance) {
rb.Push<u64>(*performance);
}
- ctx.WriteBuffer(tmp_samples);
+ ctx.WriteBuffer(samples);
}
bool DecodeOpusData(u32& consumed, u32& sample_count, std::span<const u8> input,
@@ -154,7 +155,7 @@ private:
OpusDecoderPtr decoder;
u32 sample_rate;
u32 channel_count;
- Common::ScratchBuffer<opus_int16> tmp_samples;
+ Common::ScratchBuffer<opus_int16> samples;
};
class IHardwareOpusDecoderManager final : public ServiceFramework<IHardwareOpusDecoderManager> {
diff --git a/src/core/hle/service/nvdrv/nvdrv_interface.cpp b/src/core/hle/service/nvdrv/nvdrv_interface.cpp
index 348207e250..c8a880e841 100644
--- a/src/core/hle/service/nvdrv/nvdrv_interface.cpp
+++ b/src/core/hle/service/nvdrv/nvdrv_interface.cpp
@@ -2,7 +2,6 @@
// SPDX-FileCopyrightText: 2021 Skyline Team and Contributors
// SPDX-License-Identifier: GPL-3.0-or-later
-#include <cinttypes>
#include "common/logging/log.h"
#include "core/core.h"
#include "core/hle/kernel/k_event.h"
@@ -63,12 +62,12 @@ void NVDRV::Ioctl1(HLERequestContext& ctx) {
}
// Check device
- tmp_output.resize_destructive(ctx.GetWriteBufferSize(0));
+ output_buffer.resize_destructive(ctx.GetWriteBufferSize(0));
const auto input_buffer = ctx.ReadBuffer(0);
- const auto nv_result = nvdrv->Ioctl1(fd, command, input_buffer, tmp_output);
+ const auto nv_result = nvdrv->Ioctl1(fd, command, input_buffer, output_buffer);
if (command.is_out != 0) {
- ctx.WriteBuffer(tmp_output);
+ ctx.WriteBuffer(output_buffer);
}
IPC::ResponseBuilder rb{ctx, 3};
@@ -90,12 +89,12 @@ void NVDRV::Ioctl2(HLERequestContext& ctx) {
const auto input_buffer = ctx.ReadBuffer(0);
const auto input_inlined_buffer = ctx.ReadBuffer(1);
- tmp_output.resize_destructive(ctx.GetWriteBufferSize(0));
+ output_buffer.resize_destructive(ctx.GetWriteBufferSize(0));
const auto nv_result =
- nvdrv->Ioctl2(fd, command, input_buffer, input_inlined_buffer, tmp_output);
+ nvdrv->Ioctl2(fd, command, input_buffer, input_inlined_buffer, output_buffer);
if (command.is_out != 0) {
- ctx.WriteBuffer(tmp_output);
+ ctx.WriteBuffer(output_buffer);
}
IPC::ResponseBuilder rb{ctx, 3};
@@ -116,12 +115,14 @@ void NVDRV::Ioctl3(HLERequestContext& ctx) {
}
const auto input_buffer = ctx.ReadBuffer(0);
- tmp_output.resize_destructive(ctx.GetWriteBufferSize(0));
- tmp_output_inline.resize_destructive(ctx.GetWriteBufferSize(1));
- const auto nv_result = nvdrv->Ioctl3(fd, command, input_buffer, tmp_output, tmp_output_inline);
+ output_buffer.resize_destructive(ctx.GetWriteBufferSize(0));
+ inline_output_buffer.resize_destructive(ctx.GetWriteBufferSize(1));
+
+ const auto nv_result =
+ nvdrv->Ioctl3(fd, command, input_buffer, output_buffer, inline_output_buffer);
if (command.is_out != 0) {
- ctx.WriteBuffer(tmp_output, 0);
- ctx.WriteBuffer(tmp_output_inline, 1);
+ ctx.WriteBuffer(output_buffer, 0);
+ ctx.WriteBuffer(inline_output_buffer, 1);
}
IPC::ResponseBuilder rb{ctx, 3};
diff --git a/src/core/hle/service/nvdrv/nvdrv_interface.h b/src/core/hle/service/nvdrv/nvdrv_interface.h
index 4b593ff90d..6e98115dcd 100644
--- a/src/core/hle/service/nvdrv/nvdrv_interface.h
+++ b/src/core/hle/service/nvdrv/nvdrv_interface.h
@@ -4,6 +4,7 @@
#pragma once
#include <memory>
+
#include "common/scratch_buffer.h"
#include "core/hle/service/nvdrv/nvdrv.h"
#include "core/hle/service/service.h"
@@ -34,8 +35,8 @@ private:
u64 pid{};
bool is_initialized{};
- Common::ScratchBuffer<u8> tmp_output;
- Common::ScratchBuffer<u8> tmp_output_inline;
+ Common::ScratchBuffer<u8> output_buffer;
+ Common::ScratchBuffer<u8> inline_output_buffer;
};
} // namespace Service::Nvidia
diff --git a/src/video_core/host1x/codecs/codec.cpp b/src/video_core/host1x/codecs/codec.cpp
index cd6a3a9b83..da07a556f4 100644
--- a/src/video_core/host1x/codecs/codec.cpp
+++ b/src/video_core/host1x/codecs/codec.cpp
@@ -290,7 +290,7 @@ void Codec::Decode() {
return vp9_decoder->GetFrameBytes();
default:
ASSERT(false);
- return std::vector<u8>{};
+ return std::span<const u8>{};
}
}();
AVPacketPtr packet{av_packet_alloc(), AVPacketDeleter};
diff --git a/src/video_core/host1x/codecs/h264.cpp b/src/video_core/host1x/codecs/h264.cpp
index ce827eb6c2..862904e397 100644
--- a/src/video_core/host1x/codecs/h264.cpp
+++ b/src/video_core/host1x/codecs/h264.cpp
@@ -29,15 +29,15 @@ H264::H264(Host1x::Host1x& host1x_) : host1x{host1x_} {}
H264::~H264() = default;
-const std::vector<u8>& H264::ComposeFrame(const Host1x::NvdecCommon::NvdecRegisters& state,
- bool is_first_frame) {
+std::span<const u8> H264::ComposeFrame(const Host1x::NvdecCommon::NvdecRegisters& state,
+ bool is_first_frame) {
H264DecoderContext context;
host1x.MemoryManager().ReadBlock(state.picture_info_offset, &context,
sizeof(H264DecoderContext));
const s64 frame_number = context.h264_parameter_set.frame_number.Value();
if (!is_first_frame && frame_number != 0) {
- frame.resize(context.stream_len);
+ frame.resize_destructive(context.stream_len);
host1x.MemoryManager().ReadBlock(state.frame_bitstream_offset, frame.data(), frame.size());
return frame;
}
@@ -135,14 +135,14 @@ const std::vector<u8>& H264::ComposeFrame(const Host1x::NvdecCommon::NvdecRegist
for (s32 index = 0; index < 6; index++) {
writer.WriteBit(true);
std::span<const u8> matrix{context.weight_scale};
- writer.WriteScalingList(matrix, index * 16, 16);
+ writer.WriteScalingList(scan, matrix, index * 16, 16);
}
if (context.h264_parameter_set.transform_8x8_mode_flag) {
for (s32 index = 0; index < 2; index++) {
writer.WriteBit(true);
std::span<const u8> matrix{context.weight_scale_8x8};
- writer.WriteScalingList(matrix, index * 64, 64);
+ writer.WriteScalingList(scan, matrix, index * 64, 64);
}
}
@@ -188,8 +188,8 @@ void H264BitWriter::WriteBit(bool state) {
WriteBits(state ? 1 : 0, 1);
}
-void H264BitWriter::WriteScalingList(std::span<const u8> list, s32 start, s32 count) {
- static Common::ScratchBuffer<u8> scan{};
+void H264BitWriter::WriteScalingList(Common::ScratchBuffer<u8>& scan, std::span<const u8> list,
+ s32 start, s32 count) {
scan.resize_destructive(count);
if (count == 16) {
std::memcpy(scan.data(), zig_zag_scan.data(), scan.size());
diff --git a/src/video_core/host1x/codecs/h264.h b/src/video_core/host1x/codecs/h264.h
index 5cc86454e4..d6b5563223 100644
--- a/src/video_core/host1x/codecs/h264.h
+++ b/src/video_core/host1x/codecs/h264.h
@@ -5,9 +5,11 @@
#include <span>
#include <vector>
+
#include "common/bit_field.h"
#include "common/common_funcs.h"
#include "common/common_types.h"
+#include "common/scratch_buffer.h"
#include "video_core/host1x/nvdec_common.h"
namespace Tegra {
@@ -37,7 +39,8 @@ public:
/// Based on section 7.3.2.1.1.1 and Table 7-4 in the H.264 specification
/// Writes the scaling matrices of the sream
- void WriteScalingList(std::span<const u8> list, s32 start, s32 count);
+ void WriteScalingList(Common::ScratchBuffer<u8>& scan, std::span<const u8> list, s32 start,
+ s32 count);
/// Return the bitstream as a vector.
[[nodiscard]] std::vector<u8>& GetByteArray();
@@ -63,11 +66,12 @@ public:
~H264();
/// Compose the H264 frame for FFmpeg decoding
- [[nodiscard]] const std::vector<u8>& ComposeFrame(
- const Host1x::NvdecCommon::NvdecRegisters& state, bool is_first_frame = false);
+ [[nodiscard]] std::span<const u8> ComposeFrame(const Host1x::NvdecCommon::NvdecRegisters& state,
+ bool is_first_frame = false);
private:
- std::vector<u8> frame;
+ Common::ScratchBuffer<u8> frame;
+ Common::ScratchBuffer<u8> scan;
Host1x::Host1x& host1x;
struct H264ParameterSet {
diff --git a/src/video_core/host1x/codecs/vp8.cpp b/src/video_core/host1x/codecs/vp8.cpp
index 28fb12cb8e..ee6392ff9c 100644
--- a/src/video_core/host1x/codecs/vp8.cpp
+++ b/src/video_core/host1x/codecs/vp8.cpp
@@ -12,7 +12,7 @@ VP8::VP8(Host1x::Host1x& host1x_) : host1x{host1x_} {}
VP8::~VP8() = default;
-const std::vector<u8>& VP8::ComposeFrame(const Host1x::NvdecCommon::NvdecRegisters& state) {
+std::span<const u8> VP8::ComposeFrame(const Host1x::NvdecCommon::NvdecRegisters& state) {
VP8PictureInfo info;
host1x.MemoryManager().ReadBlock(state.picture_info_offset, &info, sizeof(VP8PictureInfo));
diff --git a/src/video_core/host1x/codecs/vp8.h b/src/video_core/host1x/codecs/vp8.h
index 5bf07ecab1..7926b73f31 100644
--- a/src/video_core/host1x/codecs/vp8.h
+++ b/src/video_core/host1x/codecs/vp8.h
@@ -4,10 +4,11 @@
#pragma once
#include <array>
-#include <vector>
+#include <span>
#include "common/common_funcs.h"
#include "common/common_types.h"
+#include "common/scratch_buffer.h"
#include "video_core/host1x/nvdec_common.h"
namespace Tegra {
@@ -24,11 +25,11 @@ public:
~VP8();
/// Compose the VP8 frame for FFmpeg decoding
- [[nodiscard]] const std::vector<u8>& ComposeFrame(
+ [[nodiscard]] std::span<const u8> ComposeFrame(
const Host1x::NvdecCommon::NvdecRegisters& state);
private:
- std::vector<u8> frame;
+ Common::ScratchBuffer<u8> frame;
Host1x::Host1x& host1x;
struct VP8PictureInfo {
diff --git a/src/video_core/host1x/codecs/vp9.cpp b/src/video_core/host1x/codecs/vp9.cpp
index cf40c90121..306c3d0e8c 100644
--- a/src/video_core/host1x/codecs/vp9.cpp
+++ b/src/video_core/host1x/codecs/vp9.cpp
@@ -3,6 +3,7 @@
#include <algorithm> // for std::copy
#include <numeric>
+
#include "common/assert.h"
#include "video_core/host1x/codecs/vp9.h"
#include "video_core/host1x/host1x.h"
diff --git a/src/video_core/host1x/codecs/vp9.h b/src/video_core/host1x/codecs/vp9.h
index d4083e8d3e..f1ed195081 100644
--- a/src/video_core/host1x/codecs/vp9.h
+++ b/src/video_core/host1x/codecs/vp9.h
@@ -4,9 +4,11 @@
#pragma once
#include <array>
+#include <span>
#include <vector>
#include "common/common_types.h"
+#include "common/scratch_buffer.h"
#include "common/stream.h"
#include "video_core/host1x/codecs/vp9_types.h"
#include "video_core/host1x/nvdec_common.h"
@@ -128,8 +130,8 @@ public:
return !current_frame_info.show_frame;
}
- /// Returns a const reference to the composed frame data.
- [[nodiscard]] const std::vector<u8>& GetFrameBytes() const {
+ /// Returns a const span to the composed frame data.
+ [[nodiscard]] std::span<const u8> GetFrameBytes() const {
return frame;
}
@@ -181,7 +183,7 @@ private:
[[nodiscard]] VpxBitStreamWriter ComposeUncompressedHeader();
Host1x::Host1x& host1x;
- std::vector<u8> frame;
+ Common::ScratchBuffer<u8> frame;
std::array<s8, 4> loop_filter_ref_deltas{};
std::array<s8, 2> loop_filter_mode_deltas{};
diff --git a/src/video_core/host1x/codecs/vp9_types.h b/src/video_core/host1x/codecs/vp9_types.h
index adad8ed7ed..cc9b256907 100644
--- a/src/video_core/host1x/codecs/vp9_types.h
+++ b/src/video_core/host1x/codecs/vp9_types.h
@@ -5,6 +5,7 @@
#include <array>
#include <vector>
+
#include "common/common_funcs.h"
#include "common/common_types.h"
From 5a09fa50122af7c56ca3a05b18a1d2ab1e6b0e8b Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Sat, 24 Jun 2023 22:25:35 -0400
Subject: [PATCH 4/5] maxwell_dma: Specify dst_operand.pitch instead of a temp
var
---
src/video_core/engines/maxwell_dma.cpp | 7 +++----
1 file changed, 3 insertions(+), 4 deletions(-)
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index a290d6ea7e..f8598fd980 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -174,8 +174,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
src_operand.address = regs.offset_in;
DMA::BufferOperand dst_operand;
- u32 abs_pitch_out = std::abs(static_cast<s32>(regs.pitch_out));
- dst_operand.pitch = abs_pitch_out;
+ dst_operand.pitch = static_cast<u32>(std::abs(regs.pitch_out));
dst_operand.width = regs.line_length_in;
dst_operand.height = regs.line_count;
dst_operand.address = regs.offset_out;
@@ -222,7 +221,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
const size_t src_size =
CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
- const size_t dst_size = static_cast<size_t>(abs_pitch_out) * regs.line_count;
+ const size_t dst_size = dst_operand.pitch * regs.line_count;
read_buffer.resize_destructive(src_size);
write_buffer.resize_destructive(dst_size);
@@ -231,7 +230,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
src_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
- abs_pitch_out);
+ dst_operand.pitch);
memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
}
From 1a46823ec59b0ad4556e494c0488d5a8160aa19b Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Fri, 30 Jun 2023 21:50:44 -0400
Subject: [PATCH 5/5] parcel: Optimize small_vector sizes
---
src/core/hle/service/nvnflinger/parcel.h | 24 +++++++++++++-----------
1 file changed, 13 insertions(+), 11 deletions(-)
diff --git a/src/core/hle/service/nvnflinger/parcel.h b/src/core/hle/service/nvnflinger/parcel.h
index 23ba315a05..e2c9bbd50f 100644
--- a/src/core/hle/service/nvnflinger/parcel.h
+++ b/src/core/hle/service/nvnflinger/parcel.h
@@ -6,6 +6,7 @@
#include <memory>
#include <span>
#include <vector>
+
#include <boost/container/small_vector.hpp>
#include "common/alignment.h"
@@ -148,9 +149,9 @@ public:
this->WriteImpl(0U, m_object_buffer);
}
- std::vector<u8> Serialize() const {
- std::vector<u8> output_buffer(sizeof(ParcelHeader) + m_data_buffer.size() +
- m_object_buffer.size());
+ std::span<u8> Serialize() {
+ m_output_buffer.resize(sizeof(ParcelHeader) + m_data_buffer.size() +
+ m_object_buffer.size());
ParcelHeader header{};
header.data_size = static_cast<u32>(m_data_buffer.size());
@@ -158,17 +159,17 @@ public:
header.objects_size = static_cast<u32>(m_object_buffer.size());
header.objects_offset = header.data_offset + header.data_size;
- std::memcpy(output_buffer.data(), &header, sizeof(header));
- std::ranges::copy(m_data_buffer, output_buffer.data() + header.data_offset);
- std::ranges::copy(m_object_buffer, output_buffer.data() + header.objects_offset);
+ std::memcpy(m_output_buffer.data(), &header, sizeof(ParcelHeader));
+ std::ranges::copy(m_data_buffer, m_output_buffer.data() + header.data_offset);
+ std::ranges::copy(m_object_buffer, m_output_buffer.data() + header.objects_offset);
- return output_buffer;
+ return m_output_buffer;
}
private:
- template <typename T>
+ template <typename T, size_t BufferSize>
requires(std::is_trivially_copyable_v<T>)
- void WriteImpl(const T& val, boost::container::small_vector<u8, 0x200>& buffer) {
+ void WriteImpl(const T& val, boost::container::small_vector<u8, BufferSize>& buffer) {
const size_t aligned_size = Common::AlignUp(sizeof(T), 4);
const size_t old_size = buffer.size();
buffer.resize(old_size + aligned_size);
@@ -177,8 +178,9 @@ private:
}
private:
- boost::container::small_vector<u8, 0x200> m_data_buffer;
- boost::container::small_vector<u8, 0x200> m_object_buffer;
+ boost::container::small_vector<u8, 0x1B0> m_data_buffer;
+ boost::container::small_vector<u8, 0x40> m_object_buffer;
+ boost::container::small_vector<u8, 0x200> m_output_buffer;
};
} // namespace Service::android