Skip to content

cactus-kernels: bounds & robustness in conv/DSP/WAV/thread-pool #700

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ncylich wants to merge 1 commit into
base: main
Choose a base branch
from
+65 −84
Open

cactus-kernels: bounds & robustness in conv/DSP/WAV/thread-pool #700

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
67 changes: 4 additions & 63 deletions cactus-kernels/src/conv.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,71 +18,12 @@ constexpr size_t T_TILE_F16 = 2;
constexpr size_t ACCELERATE_K_THRESHOLD = 32;
constexpr size_t ACCELERATE_L_THRESHOLD = 128;

#ifdef __APPLE__
static void conv1d_causal_depthwise_f16_accelerate(
const __fp16* input,
const __fp16* weight,
__fp16* output,
size_t N, size_t L, size_t C, size_t K, size_t dilation)
{
const size_t in_bs = L * C;
const size_t out_bs = L * C;
CactusThreading::parallel_for_2d(N, C, CactusThreading::Thresholds::ATTENTION, [&](size_t n, size_t c) {
const __fp16* Xb = input + n * in_bs;
__fp16* Yb = output + n * out_bs;
const __fp16* Wc = weight + c * K;

if (dilation == 1) {

size_t i = 0;
for(; i + 7 < L; i += 8){
float16x8_t acc = vdupq_n_f16(0.0f);

for (size_t k = 0; k < K; ++k) {
float16x8_t input_vec = vld1q_f16(&Xb[(i + k) * C + c]);
float16x8_t weight_vec = vdupq_n_f16(Wc[k]);
acc = vfmaq_f16(acc, input_vec, weight_vec);
}
vst1q_f16(&Yb[i * C + c], acc);
}

for (; i < L; ++i) {
float acc = 0.0f;
for (size_t k = 0; k < K; ++k) {
acc += float(Xb[(i + k) * C + c]) * float(Wc[k]);
}
Yb[i * C + c] = (__fp16)acc;
}
} else {

for (size_t i = 0; i < L; ++i) {
float acc = 0.0f;

for (size_t k = 0; k < K; ++k) {
size_t idx = i + k * dilation;
acc += float(Xb[idx * C + c]) * float(Wc[k]);
}

Yb[i * C + c] = (__fp16)acc;
}
}
});
}
#endif

void cactus_conv1d_causal_depthwise_f16(
const __fp16* input,
const __fp16* weight,
__fp16* output,
size_t N, size_t L, size_t C, size_t K, size_t dilation)
{
#ifdef __APPLE__
if (K >= ACCELERATE_K_THRESHOLD && L >= ACCELERATE_L_THRESHOLD) {
conv1d_causal_depthwise_f16_accelerate(input, weight, output, N, L, C, K, dilation);
return;
}
#endif

const size_t in_bs = L * C;
const size_t out_bs = L * C;

Expand Down Expand Up @@ -316,7 +257,7 @@ static void conv1d_f16_accelerate(
size_t K,
size_t stride
){
const size_t out_len = ((L - K) / stride) + 1;
const size_t out_len = (L < K) ? 0 : ((L - K) / stride) + 1;
const size_t in_bs = C_in * L;
const size_t out_bs = C_out * out_len;

Expand Down Expand Up @@ -378,7 +319,7 @@ static void conv1d_f16_neon(
size_t K,
size_t stride
){
const size_t out_len = ((L - K) / stride) + 1;
const size_t out_len = (L < K) ? 0 : ((L - K) / stride) + 1;
const size_t in_bs = C_in * L;
const size_t out_bs = C_out * out_len;

Expand Down Expand Up @@ -448,7 +389,7 @@ static void conv1d_f16_gemm(
size_t K,
size_t stride
) {
const size_t out_len = (L - K) / stride + 1;
const size_t out_len = (L < K) ? 0 : (L - K) / stride + 1;
const size_t col_K = C_in * K;

std::vector<float> W_f32(C_out * col_K);
Expand Down Expand Up @@ -791,7 +732,7 @@ void cactus_stft_f16(
size_t K, size_t stride,
size_t num_fft_bins
) {
const size_t out_len = ((L - K) / stride) + 1;
const size_t out_len = (L < K) ? 0 : ((L - K) / stride) + 1;
const size_t in_bs = C_in * L;
const size_t out_bs = 2 * num_fft_bins * out_len;

Expand Down
7 changes: 7 additions & 0 deletions cactus-kernels/src/dsp.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -452,6 +452,9 @@ void cactus_compute_spectrogram_f32(
padded_waveform.assign(padded_length, 0.0f);

if (std::strcmp(pad_mode, "reflect") == 0) {
if (waveform_length <= pad_length) {
throw std::invalid_argument("reflect padding requires waveform_length > frame_length/2");
}
for (size_t i = 0; i < pad_length; i++) padded_waveform[i] = waveform[pad_length - i];
std::copy(waveform, waveform + waveform_length, padded_waveform.data() + pad_length);
for (size_t i = 0; i < pad_length; i++) padded_waveform[pad_length + waveform_length + i] = waveform[waveform_length - 2 - i];
Expand All @@ -465,6 +468,10 @@ void cactus_compute_spectrogram_f32(
input_length = padded_length;
}

if (input_length < frame_length) {
throw std::invalid_argument("waveform shorter than frame_length");
}

const size_t num_frames = 1 + (input_length - frame_length) / hop_length;
const size_t num_frequency_bins = (actual_fft_length / 2) + 1;

Expand Down
56 changes: 36 additions & 20 deletions cactus-kernels/src/threading.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -386,15 +386,25 @@ namespace CactusThreading {
private:
static constexpr size_t MAX_WORKERS = 16;

struct Batch {
std::mutex mutex;
std::condition_variable cv;
size_t remaining;
explicit Batch(size_t count) : remaining(count) {}
};

static std::shared_ptr<Batch>& current_batch() {
static thread_local std::shared_ptr<Batch> batch;
return batch;
}

std::vector<std::thread> workers;
std::deque<std::function<void()>> tasks;

std::mutex mutex;
std::condition_variable work_available;
std::condition_variable work_done;

bool stop{false};
std::atomic<size_t> pending_tasks{0};
size_t num_workers_;

void worker_thread() {
Expand All @@ -415,17 +425,21 @@ namespace CactusThreading {
}

task();
}
}

if (pending_tasks.fetch_sub(1, std::memory_order_acq_rel) == 1) {
std::lock_guard<std::mutex> lock(mutex);
work_done.notify_one();
}
static void finish_batch(const std::shared_ptr<Batch>& batch) {
size_t left;
{
std::lock_guard<std::mutex> lock(batch->mutex);
left = --batch->remaining;
}
if (left == 0) batch->cv.notify_all();
}

public:
explicit ThreadPool(size_t num_threads = std::thread::hardware_concurrency())
: stop(false), pending_tasks(0) {
: stop(false) {
num_workers_ = std::min(num_threads, MAX_WORKERS);
if (num_workers_ == 0) num_workers_ = 1;

Expand Down Expand Up @@ -477,7 +491,6 @@ namespace CactusThreading {

{
std::lock_guard<std::mutex> lock(mutex);
pending_tasks.fetch_add(1, std::memory_order_relaxed);
tasks.emplace_back([task](){ (*task)(); });
}
work_available.notify_one();
Expand All @@ -487,35 +500,37 @@ namespace CactusThreading {

template<typename F>
void enqueue_batch(size_t total_work, F task_func) {
if (total_work == 0) return;
if (total_work == 0) { current_batch().reset(); return; }

const size_t num_tasks = std::min(num_workers_, total_work);
const size_t per_worker = total_work / num_tasks;
const size_t remainder = total_work % num_tasks;

auto batch = std::make_shared<Batch>(num_tasks);
current_batch() = batch;

{
std::lock_guard<std::mutex> lock(mutex);
pending_tasks.fetch_add(num_tasks, std::memory_order_relaxed);

for (size_t w = 0; w < num_tasks; ++w) {
size_t start = w * per_worker + std::min(w, remainder);
size_t end = start + per_worker + (w < remainder ? 1 : 0);
tasks.emplace_back([=]() { task_func(start, end); });
tasks.emplace_back([=]() { task_func(start, end); finish_batch(batch); });
}
}
work_available.notify_all();
}

void wait_all() {
std::unique_lock<std::mutex> lock(mutex);
work_done.wait(lock, [this] {
return pending_tasks.load(std::memory_order_acquire) == 0;
});
auto batch = current_batch();
if (!batch) return;
current_batch().reset();
std::unique_lock<std::mutex> lock(batch->mutex);
batch->cv.wait(lock, [&] { return batch->remaining == 0; });
}

template<typename F>
void enqueue_n_threads(size_t total_work, size_t num_threads, F task_func) {
if (total_work == 0 || num_threads == 0) return;
if (total_work == 0 || num_threads == 0) { current_batch().reset(); return; }

num_threads = std::min(num_threads, std::min(num_workers_, total_work));
size_t num_tasks = num_threads;
Expand All @@ -527,14 +542,15 @@ namespace CactusThreading {
const size_t per_task = total_work / num_tasks;
const size_t remainder = total_work % num_tasks;

auto batch = std::make_shared<Batch>(num_tasks);
current_batch() = batch;

{
std::lock_guard<std::mutex> lock(mutex);
pending_tasks.fetch_add(num_tasks, std::memory_order_relaxed);

for (size_t t = 0; t < num_tasks; ++t) {
size_t start = t * per_task + std::min(t, remainder);
size_t end = start + per_task + (t < remainder ? 1 : 0);
tasks.emplace_back([=]() { task_func(start, end); });
tasks.emplace_back([=]() { task_func(start, end); finish_batch(batch); });
}
}
work_available.notify_all();
Expand Down
19 changes: 18 additions & 1 deletion cactus-kernels/src/wav.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,10 @@ inline AudioFP32 load_wav_fp32(const std::string& path) {
if (!file)
throw std::runtime_error("Could not open WAV file: " + path);

file.seekg(0, std::ios::end);
const std::streamoff file_size = file.tellg();
file.seekg(0, std::ios::beg);

char riff[4];
file.read(riff, 4);
if (std::string(riff, 4) != "RIFF") throw std::runtime_error("Not RIFF");
Expand Down Expand Up @@ -66,12 +70,25 @@ inline AudioFP32 load_wav_fp32(const std::string& path) {
if (std::string(data_id, 4) == "data")
break;

const std::streamoff after_header = file.tellg();
if (data_size > (uint64_t)(file_size - after_header))
throw std::runtime_error("Malformed WAV: chunk exceeds file");
file.seekg(data_size, std::ios::cur);
}

const std::streamoff data_offset = file.tellg();
if (data_size > (uint64_t)(file_size - data_offset))
throw std::runtime_error("Malformed WAV: data chunk exceeds file");

const size_t bytes_per_frame = (size_t)num_channels * 2;
if (num_channels == 0 || data_size % bytes_per_frame != 0)
throw std::runtime_error("Malformed WAV: data size not frame-aligned");

size_t num_samples = data_size / 2;
std::vector<int16_t> raw(num_samples);
file.read(reinterpret_cast<char*>(raw.data()), data_size);
if ((uint64_t)file.gcount() != data_size)
throw std::runtime_error("Malformed WAV: truncated data chunk");

std::vector<float> tmp(num_samples);
constexpr float scale = 1.0f / 32768.0f;
Expand All @@ -85,7 +102,7 @@ inline AudioFP32 load_wav_fp32(const std::string& path) {
mono = std::move(tmp);
} else if (num_channels == 2) {
mono.reserve(num_samples / 2);
for (size_t i = 0; i < num_samples; i += 2)
for (size_t i = 0; i + 1 < num_samples; i += 2)
mono.push_back(0.5f * (tmp[i] + tmp[i + 1]));
} else {
throw std::runtime_error("Unsupported channel count");
Expand Down
Loading