ShuriZma
/
xenia-canary
mirror of https://github.com/xenia-canary/xenia-canary.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[APU] XMA: Vectorize 2-channel ConvertFrame

This commit is contained in:
Triang3l 2021-06-05 16:06:16 +03:00
parent 866a29e153
commit 60b24b2d3a
2 changed files with 62 additions and 41 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

99
src/xenia/apu/xma_context.cc
View File

@ -352,14 +352,12 @@ void XmaContext::Decode(XMA_CONTEXT_DATA* data) {
output_rb.set_read_offset(output_read_offset);
output_rb.set_write_offset(output_write_offset);
int num_channels = data->is_stereo ? 2 : 1;
// We can only decode an entire frame and write it out at a time, so
// don't save any samples.
// TODO(JoelLinn): subframes when looping
size_t output_remaining_bytes = output_rb.write_count();
output_remaining_bytes -=
output_remaining_bytes % (kBytesPerFrameChannel * num_channels);
output_remaining_bytes % (kBytesPerFrameChannel << data->is_stereo);
// is_dirty_ = true; // TODO
// is_dirty_ = false; // TODO
@ -487,7 +485,7 @@ void XmaContext::Decode(XMA_CONTEXT_DATA* data) {
std::tie(frame_count, frame_last_split) = GetPacketFrameCount(packet);
assert_true(frame_count >= 0); // TODO end
PrepareDecoder(packet, data->sample_rate, num_channels);
PrepareDecoder(packet, data->sample_rate, bool(data->is_stereo));
// Current frame is split to next packet:
bool frame_is_split = frame_last_split && (frame_idx >= frame_count - 1);
@ -581,11 +579,11 @@ void XmaContext::Decode(XMA_CONTEXT_DATA* data) {
// assert_true(frame_is_split == (frame_idx == -1));
// dump_raw(av_frame_, id());
ConvertFrame((const uint8_t**)av_frame_->data, num_channels,
ConvertFrame((const uint8_t**)av_frame_->data, bool(data->is_stereo),
raw_frame_.data());
// decoded_consumed_samples_ += kSamplesPerFrame;
auto byte_count = kBytesPerFrameChannel * num_channels;
auto byte_count = kBytesPerFrameChannel << data->is_stereo;
assert_true(output_remaining_bytes >= byte_count);
output_rb.Write(raw_frame_.data(), byte_count);
output_remaining_bytes -= byte_count;
@ -781,13 +779,15 @@ std::tuple<int, bool> XmaContext::GetPacketFrameCount(uint8_t* packet) {
}
}
int XmaContext::PrepareDecoder(uint8_t* packet, int sample_rate, int channels) {
int XmaContext::PrepareDecoder(uint8_t* packet, int sample_rate,
bool is_two_channel) {
// Sanity check: Packet metadata is always 1 for XMA2/0 for XMA
assert_true((packet[2] & 0x7) == 1 || (packet[2] & 0x7) == 0);
sample_rate = GetSampleRate(sample_rate);
// Re-initialize the context with new sample rate and channels.
uint32_t channels = is_two_channel ? 2 : 1;
if (av_context_->sample_rate != sample_rate ||
av_context_->channels != channels) {
// We have to reopen the codec so it'll realloc whatever data it needs.
@ -806,7 +806,7 @@ int XmaContext::PrepareDecoder(uint8_t* packet, int sample_rate, int channels) {
return 0;
}
bool XmaContext::ConvertFrame(const uint8_t** samples, int num_channels,
bool XmaContext::ConvertFrame(const uint8_t** samples, bool is_two_channel,
uint8_t* output_buffer) {
// Loop through every sample, convert and drop it into the output array.
// If more than one channel, we need to interleave the samples from each
@ -815,50 +815,71 @@ bool XmaContext::ConvertFrame(const uint8_t** samples, int num_channels,
constexpr float scale = (1 << 15) - 1;
auto out = reinterpret_cast<int16_t*>(output_buffer);
// For testing of vectorized versions, stereo audio is common in Halo 3, since
// the first menu frame; the intro cutscene also has more than 2 channels.
#if XE_ARCH_AMD64
static_assert(kSamplesPerFrame % 8 == 0);
// Most audio is single channel, no need to optimize for the game music
if (num_channels == 1) {
const auto in = reinterpret_cast<const float*>(samples[0]);
const __m128 scale_mm = _mm_set1_ps(scale);
const auto in_channel_0 = reinterpret_cast<const float*>(samples[0]);
const __m128 scale_mm = _mm_set1_ps(scale);
if (is_two_channel) {
const auto in_channel_1 = reinterpret_cast<const float*>(samples[1]);
const __m128i shufmask =
_mm_set_epi8(14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1);
for (int i = 0; i < kSamplesPerFrame; i += 8) {
// load 8 samples
__m128 in_mm0 = _mm_loadu_ps(&in[i]);
__m128 in_mm1 = _mm_loadu_ps(&in[i + 4]);
// rescale
_mm_set_epi8(14, 15, 6, 7, 12, 13, 4, 5, 10, 11, 2, 3, 8, 9, 0, 1);
for (uint32_t i = 0; i < kSamplesPerFrame; i += 4) {
// Load 8 samples, 4 for each channel.
__m128 in_mm0 = _mm_loadu_ps(&in_channel_0[i]);
__m128 in_mm1 = _mm_loadu_ps(&in_channel_1[i]);
// Rescale.
in_mm0 = _mm_mul_ps(in_mm0, scale_mm);
in_mm1 = _mm_mul_ps(in_mm1, scale_mm);
// cast to int32
// Cast to int32.
__m128i out_mm0 = _mm_cvtps_epi32(in_mm0);
__m128i out_mm1 = _mm_cvtps_epi32(in_mm1);
// saturated cast and pack to int16
// Saturated cast and pack to int16.
__m128i out_mm = _mm_packs_epi32(out_mm0, out_mm1);
// byte swap
// Interleave channels and byte swap.
out_mm = _mm_shuffle_epi8(out_mm, shufmask);
// store
_mm_storeu_si128(reinterpret_cast<__m128i*>(&out[i]), out_mm);
// Store, as [out + i * 4] movqdu.
_mm_storeu_si128(reinterpret_cast<__m128i*>(&out[i * 2]), out_mm);
}
} else {
#else
{
#endif
uint32_t o = 0;
for (int i = 0; i < kSamplesPerFrame; i++) {
for (int j = 0; j < num_channels; j++) {
// Select the appropriate array based on the current channel.
auto in = reinterpret_cast<const float*>(samples[j]);
// Raw samples sometimes aren't within [-1, 1]
float scaled_sample = xe::saturate_signed(in[i]) * scale;
// Convert the sample and output it in big endian.
auto sample = static_cast<int16_t>(scaled_sample);
out[o++] = xe::byte_swap(sample);
}
const __m128i shufmask =
_mm_set_epi8(14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1);
for (uint32_t i = 0; i < kSamplesPerFrame; i += 8) {
// Load 8 samples, as [in_channel_0 + i * 4] and
// [in_channel_0 + i * 4 + 16] movups.
__m128 in_mm0 = _mm_loadu_ps(&in_channel_0[i]);
__m128 in_mm1 = _mm_loadu_ps(&in_channel_0[i + 4]);
// Rescale.
in_mm0 = _mm_mul_ps(in_mm0, scale_mm);
in_mm1 = _mm_mul_ps(in_mm1, scale_mm);
// Cast to int32.
__m128i out_mm0 = _mm_cvtps_epi32(in_mm0);
__m128i out_mm1 = _mm_cvtps_epi32(in_mm1);
// Saturated cast and pack to int16.
__m128i out_mm = _mm_packs_epi32(out_mm0, out_mm1);
// Byte swap.
out_mm = _mm_shuffle_epi8(out_mm, shufmask);
// Store, as [out + i * 2] movqdu.
_mm_storeu_si128(reinterpret_cast<__m128i*>(&out[i]), out_mm);
}
}
#else
uint32_t o = 0;
for (uint32_t i = 0; i < kSamplesPerFrame; i++) {
for (uint32_t j = 0; j <= uint32_t(is_two_channel); j++) {
// Select the appropriate array based on the current channel.
auto in = reinterpret_cast<const float*>(samples[j]);
// Raw samples sometimes aren't within [-1, 1]
float scaled_sample = xe::saturate_signed(in[i]) * scale;
// Convert the sample and output it in big endian.
auto sample = static_cast<int16_t>(scaled_sample);
out[o++] = xe::byte_swap(sample);
}
}
#endif
return true;
}

4
src/xenia/apu/xma_context.h
View File

@ -186,13 +186,13 @@ class XmaContext {
static std::tuple<int, bool> GetPacketFrameCount(uint8_t* packet);
// Convert sample format and swap bytes
static bool ConvertFrame(const uint8_t** samples, int num_channels,
static bool ConvertFrame(const uint8_t** samples, bool is_two_channel,
uint8_t* output_buffer);
bool ValidFrameOffset(uint8_t* block, size_t size_bytes,
size_t frame_offset_bits);
void Decode(XMA_CONTEXT_DATA* data);
int PrepareDecoder(uint8_t* packet, int sample_rate, int channels);
int PrepareDecoder(uint8_t* packet, int sample_rate, bool is_two_channel);
Memory* memory_ = nullptr;