Taconite

https://docs.rs/taconite/latest/taconite/
Admin
Taconite is a Rust crate for replaying IRON kernels on AMD XDNA NPUs through XRT or the amdxdna...
Tokens:90,118
Snippets:660
Trust Score:9.1
Update:3 hours ago
Show doc for...
Tokens:
Raw
### Example: eltwise_mul.rs using KernelSource::new

Source: https://docs.rs/taconite/latest/taconite/compile/struct.KernelSource.html

Example from the repository (examples/eltwise_mul.rs) showing how to create a KernelSource with new and pass it to Toolchain::compile_kernel. It also demonstrates compiling a design, loading a kernel, and running it on the NPU.

```rust
fn main() -> Result<(), Box<dyn std::error::Error>> {
    let args: Vec<String> = std::env::args().skip(1).collect();
    let [mlir, source, out] = &args[..] else {
        return Err("usage: eltwise_mul <design.mlir> <mul.cc> <out dir>".into());
    };
    let (mlir, out) = (PathBuf::from(mlir), PathBuf::from(out));
    let text = std::fs::read_to_string(&mlir)?;
    let arch = Arch::of_mlir(&text).ok_or("the MLIR names no npu1/npu2 device")?;
    let n = elements(&text).ok_or("no memref<Nxbf16> in the runtime sequence")?;

    let mut tc = Toolchain::from_env()?;
    if let Some(launcher) = std::env::var_os("IRON_TOOL_LAUNCHER") {
        tc = tc.with_launcher([launcher]);
    }
    let obj = out.join(arch.as_str()).join("mul.o");
    let (xclbin, insts) = (out.join("eltwise_mul.xclbin"), out.join("eltwise_mul.insts.bin"));

    let t = Instant::now();
    tc.compile_kernel(&KernelSource::new(source, arch), &obj)?;
    println!("kernel  {} ({:.1} s)", obj.display(), t.elapsed().as_secs_f64());
    let t = Instant::now();
    tc.compile_design(&Design::new(&mlir, out.join("work")).link(&obj), &xclbin, &insts)?;
    println!("design  {} + {} ({:.1} s)", xclbin.display(), insts.display(), t.elapsed().as_secs_f64());

    let session = Session::open(0)?;
    let kernel = session.load_kernel(&xclbin, &insts, None, n as u64)?;
    let (mut a, mut b, c) = (session.alloc_of::<u16>(n)?, session.alloc_of::<u16>(n)?, session.alloc_of::<u16>(n)?);
    // A fixed LCG over [-4, 4): reproducible without a rand dependency.
    let mut seed = 0x2545_f491u32;
    let mut next = || {
        seed = seed.wrapping_mul(1_664_525).wrapping_add(1_013_904_223);
        f32_to_bf16((seed >> 8) as f32 / (1u32 << 24) as f32 * 8.0 - 4.0)
    };
    a.as_mut_slice::<u16>().iter_mut().for_each(|x| *x = next());
    b.as_mut_slice::<u16>().iter_mut().for_each(|x| *x = next());
    a.sync_to_device()?;
    b.sync_to_device()?;
    let took = kernel.run(&[&a, &b, &c])?;
    c.sync_from_device()?;

    let (a, b, c) = (a.as_slice::<u16>(), b.as_slice::<u16>(), c.as_slice::<u16>());
    let mut bad = 0;
    for i in 0..n {
        let want = bf16_to_f32(f32_to_bf16(bf16_to_f32(a[i]) * bf16_to_f32(b[i])));
        let got = bf16_to_f32(c[i]);
        if (got - want).abs() > 0.04 * want.abs() + 1e-6 {
            if bad < 5 {
                eprintln!("[{i}] {} * {} = {got}, want {want}", bf16_to_f32(a[i]), bf16_to_f32(b[i]));
            }
            bad += 1;
        }
    }
    println!("npu     {n} elements in {:.3} ms, {bad} mismatches", took.as_secs_f64() * 1e3);
    if bad > 0 {
        return Err(format!("{bad} of {n} elements differ from the CPU").into());
    }
    Ok(())
}
```

--------------------------------

### Run XRT kernel and read result with as_slice (direct_eltwise_mul.rs)

Source: https://docs.rs/taconite/latest/taconite/struct.Buffer.html

Example from examples/direct_eltwise_mul.rs (line 76) showing how to open a session, load a kernel, run it, sync from device, and use `as_slice::<u16>()` to get the result buffer as a slice.

```rust
48fn run_xrt(
49    xclbin: &Path,
50    insts: &Path,
51    a: &[u16],
52    b: &[u16],
53) -> Result<(Vec<u16>, Vec<Duration>), Box<dyn std::error::Error>> {
54    let n = a.len();
55    let t = Instant::now();
56    let session = Session::open(0)?;
57    let opened = t.elapsed();
58    let t = Instant::now();
59    let kernel = session.load_kernel(xclbin, insts, None, n as u64)?;
60    let loaded = t.elapsed();
61    let t = Instant::now();
62    let (mut ab, mut bb, cb) = (session.alloc_of::<u16>(n)?, session.alloc_of::<u16>(n)?, session.alloc_of::<u16>(n)?);
63    println!(
64        "xrt: open {:.1} ms, load {:.1} ms, 3 buffers {:.1} ms",
65        opened.as_secs_f64() * 1e3,
66        loaded.as_secs_f64() * 1e3,
67        t.elapsed().as_secs_f64() * 1e3
68    );
69    ab.write(a)?;
70    bb.write(b)?;
71    let mut times = Vec::new();
72    for _ in 0..RUNS {
73        times.push(kernel.run(&[&ab, &bb, &cb])?);
74    }
75    cb.sync_from_device()?;
76    Ok((cb.as_slice::<u16>().to_vec(), times))
77}
```

--------------------------------

### sync_from_device example from examples/direct_eltwise_mul.rs

Source: https://docs.rs/taconite/latest/taconite/direct/struct.Buffer.html

This example shows a complete run_direct function that opens a direct session, loads a kernel, allocates buffers, runs the kernel multiple times, and finally calls cb.sync_from_device() to synchronize the buffer from the device before reading its contents. It prints timing and context information. The sync_from_device call is at line 125.

```rust
fn run_direct(
    xclbin: &Path,
    insts: &Path,
    a: &[u16],
    b: &[u16],
) -> Result<(Vec<u16>, Vec<Duration>), Box<dyn std::error::Error>> {
    use taconite::direct::Session;
    let n = a.len();
    let t = Instant::now();
    let device = Session::open(0)?;
    let opened = t.elapsed();
    let (cols, rows) = device.array();
    println!(
        "direct: {cols}x{rows} array, AIE {:?}, firmware {:?}, power {:?}",
        device.aie_version()?,
        device.firmware_version()?,
        device.power_mode()?
    );
    let t = Instant::now();
    let kernel = device.load_kernel(xclbin, insts, None, n as u64)?;
    let loaded = t.elapsed();
    let t = Instant::now();
    let (mut ab, mut bb, cb) = (device.alloc_of::<u16>(n)?, device.alloc_of::<u16>(n)?, device.alloc_of::<u16>(n)?);
    println!(
        "direct: open {:.1} ms, load {:.1} ms, 3 buffers {:.1} ms",
        opened.as_secs_f64() * 1e3,
        loaded.as_secs_f64() * 1e3,
        t.elapsed().as_secs_f64() * 1e3
    );
    ab.write(a)?;
    bb.write(b)?;
    let mut times = Vec::new();
    for _ in 0..RUNS {
        match kernel.run(&[&ab, &bb, &cb]) {
            Ok(t) => times.push(t),
            Err(e) => {
                for c in device.contexts().unwrap_or_default() {
                    eprintln!("  {c:?}");
                }
                return Err(e.into());
            }
        }
    }
    for c in device.contexts()? {
        println!("direct: context {c:?}");
    }
    cb.sync_from_device()?;
    Ok((cb.as_slice::<u16>().to_vec(), times))
}
```

--------------------------------

### Example: eltwise_mul.rs using Toolchain::from_env()

Source: https://docs.rs/taconite/latest/taconite/compile/struct.Toolchain.html

This example from the repository demonstrates using Toolchain::from_env() to compile a kernel and design, then run it on an NPU. It shows how to optionally set a launcher via IRON_TOOL_LAUNCHER, compile with compile_kernel and compile_design, and verify results against a CPU computation. The code includes a fixed LCG for reproducible random data and checks for mismatches.

```rust
25fn main() -> Result<(), Box<dyn std::error::Error>> {
26    let args: Vec<String> = std::env::args().skip(1).collect();
27    let [mlir, source, out] = &args[..] else {
28        return Err("usage: eltwise_mul <design.mlir> <mul.cc> <out dir>".into());
29    };
30    let (mlir, out) = (PathBuf::from(mlir), PathBuf::from(out));
31    let text = std::fs::read_to_string(&mlir)?;
32    let arch = Arch::of_mlir(&text).ok_or("the MLIR names no npu1/npu2 device")?;
33    let n = elements(&text).ok_or("no memref<Nxbf16> in the runtime sequence")?;
34
35    let mut tc = Toolchain::from_env()?;
36    if let Some(launcher) = std::env::var_os("IRON_TOOL_LAUNCHER") {
37        tc = tc.with_launcher([launcher]);
38    }
39    let obj = out.join(arch.as_str()).join("mul.o");
40    let (xclbin, insts) = (out.join("eltwise_mul.xclbin"), out.join("eltwise_mul.insts.bin"));
41
42    let t = Instant::now();
43    tc.compile_kernel(&KernelSource::new(source, arch), &obj)?;
44    println!("kernel  {} ({:.1} s)", obj.display(), t.elapsed().as_secs_f64());
45    let t = Instant::now();
46    tc.compile_design(&Design::new(&mlir, out.join("work")).link(&obj), &xclbin, &insts)?;
47    println!("design  {} + {} ({:.1} s)", xclbin.display(), insts.display(), t.elapsed().as_secs_f64());
48
49    let session = Session::open(0)?;
50    let kernel = session.load_kernel(&xclbin, &insts, None, n as u64)?;
51    let (mut a, mut b, c) = (session.alloc_of::<u16>(n)?, session.alloc_of::<u16>(n)?, session.alloc_of::<u16>(n)?);
52    // A fixed LCG over [-4, 4): reproducible without a rand dependency.
53    let mut seed = 0x2545_f491u32;
54    let mut next = || {
55        seed = seed.wrapping_mul(1_664_525).wrapping_add(1_013_904_223);
56        f32_to_bf16((seed >> 8) as f32 / (1u32 << 24) as f32 * 8.0 - 4.0)
57    };
58    a.as_mut_slice::<u16>().iter_mut().for_each(|x| *x = next());
59    b.as_mut_slice::<u16>().iter_mut().for_each(|x| *x = next());
60    a.sync_to_device()?;
61    b.sync_to_device()?;
62    let took = kernel.run(&[&a, &b, &c])?;
63    c.sync_from_device()?;
64
65    let (a, b, c) = (a.as_slice::<u16>(), b.as_slice::<u16>(), c.as_slice::<u16>());
66    let mut bad = 0;
67    for i in 0..n {
68        let want = bf16_to_f32(f32_to_bf16(bf16_to_f32(a[i]) * bf16_to_f32(b[i])));
69        let got = bf16_to_f32(c[i]);
70        if (got - want).abs() > 0.04 * want.abs() + 1e-6 {
71            if bad < 5 {
72                eprintln!("[{i}] {} * {} = {got}, want {want}", bf16_to_f32(a[i]), bf16_to_f32(b[i]));
73            }
74            bad += 1;
75        }
76    }
77    println!("npu     {n} elements in {:.3} ms, {bad} mismatches", took.as_secs_f64() * 1e3);
78    if bad > 0 {
79        return Err(format!("{bad} of {n} elements differ from the CPU").into());
80    }
81    Ok(())
82}
```

--------------------------------

### Toolchain::with_peano

Source: https://docs.rs/taconite/latest/taconite/compile/struct.Toolchain.html

Sets the Peano install directory to use. The directory should contain the `bin/clang++` executable.

```APIDOC
## Toolchain::with_peano

### Description
Sets the Peano install directory to use. The directory should contain the `bin/clang++` executable.

### Method
`pub fn with_peano(self, peano: impl Into<PathBuf>) -> Self`

### Parameters
- **peano** (impl Into<PathBuf>) - Required - Path to the Peano install directory.

### Returns
- `Self` - The Toolchain with the updated Peano path.

### Example
```rust
let tc = Toolchain::from_env()?.with_peano("/path/to/peano");
```
```

--------------------------------

### Vk::record

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.Vk.html

Starts recording a GPU plan, returning a Recorder that can be used to build the plan.

```APIDOC
## Vk::record

### Description
Starts recording a GPU plan, returning a Recorder that can be used to build the plan.

### Method
pub fn record(&mut self) -> Result<Recorder<'_>, Error>

### Parameters
None

### Returns
- **Result<Recorder<'_>, Error>** - A Recorder for building the plan or an error.
```

--------------------------------

### Run direct eltwise mul and read buffer as slice

Source: https://docs.rs/taconite/latest/taconite/direct/struct.Buffer.html

Example from examples/direct_eltwise_mul.rs showing how to open a device, load a kernel, allocate buffers, run the kernel, sync from device, and read the result as a slice using as_slice::<u16>().

```rust
79fn run_direct(
80    xclbin: &Path,
81    insts: &Path,
82    a: &[u16],
83    b: &[u16],
84) -> Result<(Vec<u16>, Vec<Duration>), Box<dyn std::error::Error>> {
85    use taconite::direct::Session;
86    let n = a.len();
87    let t = Instant::now();
88    let device = Session::open(0)?;
89    let opened = t.elapsed();
90    let (cols, rows) = device.array();
91    println!(
92        "direct: {cols}x{rows} array, AIE {:?}, firmware {:?}, power {:?}",
93        device.aie_version()?,
94        device.firmware_version()?,
95        device.power_mode()?
96    );
97    let t = Instant::now();
98    let kernel = device.load_kernel(xclbin, insts, None, n as u64)?;
99    let loaded = t.elapsed();
100    let t = Instant::now();
101    let (mut ab, mut bb, cb) = (device.alloc_of::<u16>(n)?, device.alloc_of::<u16>(n)?, device.alloc_of::<u16>(n)?);
102    println!(
103        "direct: open {:.1} ms, load {:.1} ms, 3 buffers {:.1} ms",
104        opened.as_secs_f64() * 1e3,
105        loaded.as_secs_f64() * 1e3,
106        t.elapsed().as_secs_f64() * 1e3
107    );
108    ab.write(a)?;
109    bb.write(b)?;
110    let mut times = Vec::new();
111    for _ in 0..RUNS {
112        match kernel.run(&[&ab, &bb, &cb]) {
113            Ok(t) => times.push(t),
114            Err(e) => {
115                for c in device.contexts().unwrap_or_default() {
116                    eprintln!("  {c:?}");
117                }
118                return Err(e.into());
119            }
120        }
121    }
122    for c in device.contexts()? {
123        println!("direct: context {c:?}");
124    }
125    cb.sync_from_device()?;
126    Ok((cb.as_slice::<u16>().to_vec(), times))
127}
```

--------------------------------

### View::write

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.View.html

Writes the given data at the start of the view. Only works with host-visible memory.

```APIDOC
## View::write

### Description
Writes the given data at the start of the view. Only works with host-visible memory.

### Method
`pub fn write<T: Copy>(&self, data: &[T])`

### Parameters
- `data` (&[T]) - Required - The data to write. `T` must implement `Copy`.

### Returns
- `()` - No return value.
```

--------------------------------

### sync_from_device in direct_eltwise_mul.rs

Source: https://docs.rs/taconite/latest/taconite/struct.Buffer.html

Example from `examples/direct_eltwise_mul.rs` (line 75) showing `cb.sync_from_device()?` after a kernel run, then reading the result via `as_slice::<u16>()`.

```rust
fn run_xrt(
    xclbin: &Path,
    insts: &Path,
    a: &[u16],
    b: &[u16],
) -> Result<(Vec<u16>, Vec<Duration>), Box<dyn std::error::Error>> {
    let n = a.len();
    let t = Instant::now();
    let session = Session::open(0)?;
    let opened = t.elapsed();
    let t = Instant::now();
    let kernel = session.load_kernel(xclbin, insts, None, n as u64)?;
    let loaded = t.elapsed();
    let t = Instant::now();
    let (mut ab, mut bb, cb) = (session.alloc_of::<u16>(n)?, session.alloc_of::<u16>(n)?, session.alloc_of::<u16>(n)?);
    println!(
        "xrt: open {:.1} ms, load {:.1} ms, 3 buffers {:.1} ms",
        opened.as_secs_f64() * 1e3,
        loaded.as_secs_f64() * 1e3,
        t.elapsed().as_secs_f64() * 1e3
    );
    ab.write(a)?;
    bb.write(b)?;
    let mut times = Vec::new();
    for _ in 0..RUNS {
        times.push(kernel.run(&[&ab, &bb, &cb])?);
    }
    cb.sync_from_device()?;
    Ok((cb.as_slice::<u16>().to_vec(), times))
}
```

--------------------------------

### Toolchain::new

Source: https://docs.rs/taconite/latest/taconite/compile/struct.Toolchain.html

Creates a new Toolchain from the path to an mlir-aie install. The Peano compiler is located automatically via environment variables or standard locations.

```APIDOC
## Toolchain::new

### Description
Creates a new Toolchain from the path to an mlir-aie install. The Peano compiler is located automatically via environment variables or standard locations.

### Method
new

### Signature
pub fn new(mlir_aie: impl Into<PathBuf>) -> Self

### Parameters
- **mlir_aie** (impl Into<PathBuf>) - Required - The path to the mlir-aie install directory (the `mlir_aie` directory of the pip wheel, or `$MLIR_AIE_INSTALL_DIR`).

### Behavior
Peano is looked for where mlir-aie itself looks: `$PEANO_INSTALL_DIR`, then `<mlir_aie>/peano`, then the `llvm-aie` wheel beside it.

### Returns
A new Toolchain instance.
```

--------------------------------

### pub fn threads() -> usize

Source: https://docs.rs/taconite/latest/taconite/cpu/fn.threads.html

Returns the number of host threads to use. The default is the machine's thread count, capped at 16. The value is fixed once the pool has started.

```APIDOC
## pub fn threads() -> usize

### Description
Returns the number of host threads to use. The default is the machine's thread count, capped at 16. The value is fixed once the pool has started.

### Function Signature
```rust
pub fn threads() -> usize
```

### Parameters
None.

### Return Value
- **usize** - The number of host threads to use.

### Notes
- Defaults to the machine's thread count, up to 16.
- The value is fixed once the pool has started.
```

--------------------------------

### Recorder::copy

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.Recorder.html

Copies the contents of a source view to the start of a destination view.

```APIDOC
## Recorder::copy

### Description
Copies `src` over the start of `dst`.

### Method
`pub fn copy(&mut self, src: View, dst: View)`

### Parameters
- `src` (View) - Required - The source view to copy from.
- `dst` (View) - Required - The destination view to copy to.

### Returns
`()`
```

--------------------------------

### Session::open_path

Source: https://docs.rs/taconite/latest/taconite/direct/struct.Session.html

Opens the accel node at the given path and sets up the device heap. Returns a Result with the Session or an Error.

```APIDOC
## Session::open_path

### Description
Opens the accel node at the given path and sets up the device heap. Returns a `Result` with the `Session` or an `Error`.

### Method
`pub fn open_path(path: &Path) -> Result<Self, Error>`

### Parameters
- `path` (&Path) - Required - The path to the accel node.

### Returns
- `Result<Self, Error>` - The opened `Session` or an `Error`.
```

--------------------------------

### Regex::is_match_at_start

Source: https://docs.rs/taconite/latest/taconite/pyre/struct.Regex.html

Checks if the text matches the pattern from the start, equivalent to Python's re.match.

```APIDOC
## Regex::is_match_at_start

### Description
Checks if the text matches the pattern from the start, equivalent to Python's re.match.

### Method
`pub fn is_match_at_start(&self, text: &str) -> bool`

### Parameters
- `text` (&str) - Required - The text to match from the start.

### Returns
- `bool` - True if the text matches from the start, false otherwise.

### Example
```rust
let regex = Regex::new(r"\d+", Flags::default())?;
let is_match = regex.is_match_at_start("123abc");
```
```

--------------------------------

### Kernel::start

Source: https://docs.rs/taconite/latest/taconite/direct/struct.Kernel.html

Launches the kernel without waiting. The returned Run borrows the argument buffers until Run::wait is called.

```APIDOC
## Kernel::start

### Description
Launches the kernel without waiting. The returned `Run` borrows the argument buffers until `Run::wait` is called.

### Method
`pub fn start<'a>(&'a self, args: &[&'a Buffer]) -> Result<Run<'a>, Error>`

### Parameters
- **args** (`&[&Buffer]`) - Required - The argument buffers for the kernel.

### Returns
- `Result<Run<'a>, Error>` - A `Run` handle that borrows the argument buffers, or an `Error` if the launch fails.

### Example
```rust
let kernel = Kernel::new(...);
let buffers = vec![buffer1, buffer2];
let run = kernel.start(&buffers)?;
```
```

--------------------------------

### View::read

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.View.html

Reads `n` elements of type `T` from the start of the view. Only works with host-visible memory.

```APIDOC
## View::read

### Description
Reads `n` elements of type `T` from the start of the view. Only works with host-visible memory.

### Method
`pub fn read<T: Copy + Default>(&self, n: usize) -> Vec<T>`

### Parameters
- `n` (usize) - Required - The number of elements to read.

### Returns
- `Vec<T>` - A vector containing the read elements. `T` must implement `Copy` and `Default`.
```

--------------------------------

### Session::open

Source: https://docs.rs/taconite/latest/taconite/direct/struct.Session.html

Opens the NPU device at /dev/accel/accel<device_index> (0 on a laptop). Returns a Result with the Session or an Error.

```APIDOC
## Session::open

### Description
Opens the NPU device at `/dev/accel/accel<device_index>` (0 on a laptop). Returns a `Result` with the `Session` or an `Error`.

### Method
`pub fn open(device_index: u32) -> Result<Self, Error>`

### Parameters
- `device_index` (u32) - Required - The device index to open.

### Returns
- `Result<Self, Error>` - The opened `Session` or an `Error`.

### Example
```rust
let device = Session::open(0)?;
```
```

--------------------------------

### run

Source: https://docs.rs/taconite/latest/taconite/direct/struct.Kernel.html

Launches the kernel over the given buffer arguments and waits for it, returning the time from submission to completion. Inputs must have been synced to device, outputs need sync from device before reading.

```APIDOC
## run

### Description
Launches the kernel over `args` (its buffer arguments in order, at most `MAX_ARGS`) and waits for it; returns the time from submission to completion. Inputs must have been `sync_to_device`d, outputs need `sync_from_device` before reading.

### Method
`pub fn run(&self, args: &[&Buffer]) -> Result<Duration, Error>`

### Parameters
- **args** (`&[&Buffer]`) - Required - A slice of references to `Buffer` objects representing the kernel's arguments in order. At most `MAX_ARGS` arguments are allowed.

### Return Value
- `Result<Duration, Error>` - On success, the duration from submission to completion. On failure, an `Error`.
```

--------------------------------

### View::at

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.View.html

Returns a new View representing `bytes` bytes starting at offset `off` (which must be a multiple of 256).

```APIDOC
## View::at

### Description
Returns a new View representing `bytes` bytes starting at offset `off` (which must be a multiple of 256).

### Method
`pub fn at(&self, off: u64, bytes: u64) -> View`

### Parameters
- `off` (u64) - Required - The offset in bytes, must be a multiple of 256.
- `bytes` (u64) - Required - The number of bytes to include in the new view.

### Returns
- `View` - A new View representing the specified range.
```

--------------------------------

### pub fn npu_total(&self) -> Duration

Source: https://docs.rs/taconite/latest/taconite/struct.Timing.html

Returns the total NPU dispatch time, summing all entries under keys starting with 'npu:'.

```APIDOC
## pub fn npu_total(&self) -> Duration

### Description
Returns the total NPU dispatch time, summing all entries under keys starting with 'npu:'.

### Method
npu_total

### Parameters
None

### Returns
- **Duration** - The total NPU dispatch time.
```

--------------------------------

### pub fn load_kernel(&self, xclbin: &[Path], insts: &[Path], kernel_name: Option<&str>, ops_per_run: u64) -> Result<Kernel, Error>

Source: https://docs.rs/taconite/latest/taconite/struct.Session.html

Loads an xclbin and instruction stream, returning a Kernel. Kernels sharing the same xclbin and kernel name share a hardware context. The ops_per_run parameter declares the work per Kernel::run for QoS reporting.

```APIDOC
## pub fn load_kernel(&self, xclbin: &[Path], insts: &[Path], kernel_name: Option<&str>, ops_per_run: u64) -> Result<Kernel, Error>

### Description
Loads an xclbin + instruction stream. Kernels naming the same xclbin (and kernel name) share one hardware context, resident until the last of them is dropped – dropping them frees the context for another (NPU2 holds 16, across every process). `kernel_name` is the kernel inside the xclbin (`None`: its first — IRON’s is `MLIR_AIE`).

`ops_per_run` is the work one `Kernel::run` performs — `2·M·K·N` for a GEMM, a multiply-add counting two — declared to the driver as the context’s QoS `gops` (rounded to whole GOP, the field’s unit; a kernel under half a GOP declares nothing). It is how `xrt-smi` and the desktop’s NPU ACTIVITY card turn the context’s completion count into operations per second; see `iron_xrt_shim.h` for why it can’t change how the context is scheduled.

### Method
load_kernel

### Signature
pub fn load_kernel(&self, xclbin: &[Path], insts: &[Path], kernel_name: Option<&str>, ops_per_run: u64) -> Result<Kernel, Error>

### Parameters
- **xclbin** (&[Path]) - Required - Path to the xclbin file.
- **insts** (&[Path]) - Required - Path to the instruction stream.
- **kernel_name** (Option<&str>) - Optional - Name of the kernel inside the xclbin. `None` selects the first kernel.
- **ops_per_run** (u64) - Required - Work performed per `Kernel::run`, used for QoS reporting.

### Returns
- **Result<Kernel, Error>** - A `Kernel` handle on success, or an `Error` on failure.
```

--------------------------------

### pub fn get(&self, key: &str) -> Duration

Source: https://docs.rs/taconite/latest/taconite/struct.Timing.html

Retrieves the duration associated with the given key. Returns a core::time::Duration.

```APIDOC
## pub fn get(&self, key: &str) -> Duration

### Description
Retrieves the duration associated with the given key.

### Method
get

### Parameters
- **key** (&str) - Required - The key identifying the timing stage.

### Returns
- **Duration** - The duration associated with the key.
```

--------------------------------

### Toolchain::from_env

Source: https://docs.rs/taconite/latest/taconite/compile/struct.Toolchain.html

Creates a new Toolchain from environment variables. It uses the $MLIR_AIE_INSTALL_DIR environment variable (exported by mlir-aie's env_setup.sh) and optionally $AIECC_PATH to name the aiecc binary.

```APIDOC
## Toolchain::from_env

### Description
Creates a new Toolchain from environment variables. It uses the `$MLIR_AIE_INSTALL_DIR` environment variable (exported by mlir-aie's `env_setup.sh`) and optionally `$AIECC_PATH` to name the aiecc binary.

### Method
`pub fn from_env() -> Result<Self, Error>`

### Parameters
None.

### Returns
- `Result<Self, Error>` - A Result containing the new Toolchain or an Error.

### Example
```rust
let mut tc = Toolchain::from_env()?;
```
```

--------------------------------

### Recorder::cut

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.Recorder.html

Ends the plan recorded so far and starts a new one, allowing the new plan to see everything written by the previous plan.

```APIDOC
## Recorder::cut

### Description
Ends the plan recorded so far and goes on recording a new one: submitted after it, the new one sees everything it wrote.

### Method
`pub fn cut(&mut self) -> Result<Plan, Error>`

### Parameters
None.

### Returns
`Result<Plan, Error>`
```

--------------------------------

### Plan::submit

Source: https://docs.rs/taconite/latest/taconite/gpu/index.html

Submits the recorded command buffer for execution without waiting for completion.

```APIDOC
## Plan::submit

### Description
Submits the recorded command buffer for execution without waiting for completion.

### Method
submit

### Parameters
- None explicitly documented.

### Returns
- **Result** (type not specified) - Indicates success or failure.

```

--------------------------------

### Plan::run

Source: https://docs.rs/taconite/latest/taconite/gpu/index.html

Replays a recorded chain of dispatches by submitting the command buffer and waiting for completion.

```APIDOC
## Plan::run

### Description
Replays a recorded chain of dispatches by submitting the command buffer and waiting for completion.

### Method
run

### Parameters
- None explicitly documented.

### Returns
- **Result** (type not specified) - Indicates success or failure.

```

--------------------------------

### direct_eltwise_mul main function

Source: https://docs.rs/taconite/latest/taconite/fn.f32_to_bf16.html

Parses command-line arguments (mode, xclbin, insts.bin, elements), generates inputs, runs the kernel via run_xrt or run_direct, and reports mismatches against CPU results.

```rust
fn main() -> Result<(), Box<dyn std::error::Error>> {
    let args: Vec<String> = std::env::args().skip(1).collect();
    let [mode, xclbin, insts, n] = &args[..] else {
        return Err("usage: direct_eltwise_mul <xrt|direct> <xclbin> <insts.bin> <elements>".into());
    };
    let (xclbin, insts, n) = (Path::new(xclbin), Path::new(insts), n.parse::<usize>()?);
    let (a, b) = (inputs(n, 0x2545_f491), inputs(n, 0x9e37_79b9));

    let t = Instant::now();
    let (c, times) = match mode.as_str() {
        "xrt" => run_xrt(xclbin, insts, &a, &b)?,
        "direct" => run_direct(xclbin, insts, &a, &b)?,
        other => return Err(format!("unknown mode {other}").into()),
    };
    println!("{mode}: setup+runs {:.1} ms", t.elapsed().as_secs_f64() * 1e3);
    report(&times);

    let bad = (0..n)
        .filter(|&i| {
            let want = bf16_to_f32(f32_to_bf16(bf16_to_f32(a[i]) * bf16_to_f32(b[i])));
            (bf16_to_f32(c[i]) - want).abs() > 0.04 * want.abs() + 1e-6
        })
        .count();
    println!("{n} elements, {bad} mismatches");
    if bad > 0 {
        return Err(format!("{bad} of {n} elements differ from the CPU").into());
    }
    Ok(())
}
```

--------------------------------

### lower_simple(c: char) -> char

Source: https://docs.rs/taconite/latest/taconite/unicode/fn.lower_simple.html

Returns the simple (one-character) lowercase mapping of the given character, as used by Python's `re` module for case folding. For example, U+0130 (İ) maps to `i`.

```APIDOC
## lower_simple(c: char) -> char

### Description
Returns the simple (one-character) lowercase mapping of the given character, as used by Python's `re` module for case folding. For example, U+0130 (İ) maps to `i`.

### Function Signature
`pub fn lower_simple(c: char) -> char`

### Parameters
- **c** (char) - The input character to map to its lowercase equivalent.

### Return Value
- **char** - The simple lowercase mapping of the input character.

### Example
```rust
let result = lower_simple('İ');
assert_eq!(result, 'i');
```
```

--------------------------------

### Plan::run

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.Plan.html

Submits the plan and waits for it to complete. Inputs must be written before calling, and outputs are readable after it returns.

```APIDOC
## Plan::run

### Description
Submits the plan and waits for it to complete. Inputs must be written before calling, and outputs are readable after it returns.

### Method
run

### Signature
pub fn run(&self, vk: &Vk) -> Result<Duration, Error>

### Parameters
- **vk** (&Vk) - Required - Reference to the Vulkan context.

### Returns
- **Result<Duration, Error>** - Duration of the run on success, or an Error on failure.

```

--------------------------------

### Session::load_kernel

Source: https://docs.rs/taconite/latest/taconite/direct/struct.Session.html

Loads an xclbin and instruction stream, creating a Kernel that shares a hardware context with other kernels using the same xclbin and kernel name. The context runs the xclbin's PDI, and ops_per_run is declared as QoS gops.

```APIDOC
## Session::load_kernel

### Description
Loads an xclbin + instruction stream, as `crate::Session::load_kernel` does: kernels naming the same xclbin (and kernel name) share one hardware context, resident until the last of them is dropped. The context runs the xclbin's one PDI whatever `kernel_name` says; the name only tells contexts apart. `ops_per_run` is declared to the driver as the context's QoS `gops`.

### Method
`pub fn load_kernel(&self, xclbin: &Path, insts: &Path, kernel_name: Option<&str>, ops_per_run: u64) -> Result<Kernel, Error>`

### Parameters
- `xclbin` (`&Path`) - Required - Path to the xclbin file.
- `insts` (`&Path`) - Required - Path to the instruction stream file.
- `kernel_name` (`Option<&str>`) - Optional - Name to distinguish contexts; the actual kernel is determined by the xclbin's PDI.
- `ops_per_run` (`u64`) - Required - Declared QoS gops for the context.

### Returns
- `Result<Kernel, Error>` - A `Kernel` handle on success, or an `Error` on failure.

### Example
```rust
let kernel = device.load_kernel(xclbin, insts, None, n as u64)?;
```
```

--------------------------------

### Buffer::sub

Source: https://docs.rs/taconite/latest/taconite/direct/struct.Buffer.html

Creates a view of a portion of the buffer, starting at an offset and spanning a specified number of bytes. The returned Buffer is a kernel argument with syncs covering only its bytes, and it keeps the allocation alive.

```APIDOC
## Buffer::sub

### Description
Creates a view of `bytes` bytes of this buffer from `offset`, a `Buffer` in its own right: a kernel argument (the NPU is handed its address) with syncs that cover its bytes only. It keeps the allocation alive itself, as `crate::Buffer::sub`’s does.

### Method
sub

### Signature
pub fn sub(&self, offset: usize, bytes: usize) -> Result<Buffer, Error>

### Parameters
- **offset** (usize) - Required - The byte offset into the buffer where the sub-buffer starts.
- **bytes** (usize) - Required - The number of bytes the sub-buffer spans.

### Returns
- **Result<Buffer, Error>** - A `Buffer` representing the sub-buffer, or an `Error` if the parameters are invalid.

```

--------------------------------

### category(c: char) -> &'static str

Source: https://docs.rs/taconite/latest/taconite/unicode/fn.category.html

Returns the Unicode general category of the given character as a static string. For example, 'Lu' for uppercase letter, 'Nd' for decimal digit, and 'Cn' if the character is unassigned.

```APIDOC
## category(c: char) -> &'static str

### Description
Returns the Unicode general category of the given character as a static string. For example, 'Lu' for uppercase letter, 'Nd' for decimal digit, and 'Cn' if the character is unassigned.

### Function Signature
pub fn category(c: char) -> &'static str

### Parameters
- **c** (char) - Required - The character to query.

### Return Value
- **&'static str** - The Unicode general category of the character (e.g., "Lu", "Nd", "Cn").

### Example
```rust
let cat = taconite::unicode::category('A');
assert_eq!(cat, "Lu");
```
```

--------------------------------

### Vk::new

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.Vk.html

Creates a new Vk instance using the first non-CPU GPU device (e.g., not llvmpipe) if a Vulkan loader and such a device are available.

```APIDOC
## Vk::new

### Description
Creates a new Vk instance using the first non-CPU GPU device (e.g., not llvmpipe) if a Vulkan loader and such a device are available.

### Method
pub fn new() -> Result<Self, Error>

### Parameters
None

### Returns
- **Result<Self, Error>** - The new Vk instance or an error if no suitable GPU is found.
```

--------------------------------

### CloneToUninit::clone_to_uninit

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.View.html

Performs copy-assignment from self to dest. This is a nightly-only experimental API.

```APIDOC
## CloneToUninit::clone_to_uninit

### Description
Performs copy-assignment from `self` to `dest`. This is a nightly-only experimental API.

### Method
unsafe fn clone_to_uninit(&self, dest: *mut u8)

### Parameters
- `dest` (*mut u8) - Required - Destination pointer for the copy.

### Notes
- This method is part of the `CloneToUninit` trait, which is nightly-only experimental.

```

--------------------------------

### pub fn take(&mut self, bytes: u64) -> Result<View, Error>

Source: https://docs.rs/taconite/latest/taconite/gpu/struct.Arena.html

Takes a view of the specified number of bytes from the arena. Returns a Result containing a View on success or an Error on failure.

```APIDOC
## pub fn take(&mut self, bytes: u64) -> Result<View, Error>

### Description
Takes a view of the specified number of bytes from the arena. Returns a Result containing a View on success or an Error on failure.

### Method
pub fn take

### Parameters
- **bytes** (u64) - Required - The number of bytes to allocate for the view.

### Return
- **Result<View, Error>** - A View on success, or an Error on failure.

### Example
```rust
let view = arena.take(1024)?;
```
```

--------------------------------

### Toolchain::with_launcher

Source: https://docs.rs/taconite/latest/taconite/compile/struct.Toolchain.html

Sets a launcher for all tools invoked by the Toolchain. The launcher is prepended to the tool command, allowing execution through a script or container.

```APIDOC
## Toolchain::with_launcher

### Description
Runs every tool as `launcher… tool args…`, e.g. through a script that execs its arguments inside a container. This is useful on hosts that cannot run the binaries directly, such as NixOS.

### Method
`pub fn with_launcher<I, S>(self, launcher: I) -> Self`

### Parameters
- **launcher** (I) - Required - An iterator of items that can be converted into `OsString`, representing the launcher command and its arguments.

### Returns
- **Self** - The updated `Toolchain` instance.

### Example
```rust
let mut tc = Toolchain::from_env()?;
if let Some(launcher) = std::env::var_os("IRON_TOOL_LAUNCHER") {
    tc = tc.with_launcher([launcher]);
}
```
```

--------------------------------

### Toolchain::peano

Source: https://docs.rs/taconite/latest/taconite/compile/struct.Toolchain.html

Returns the path to the peano tool.

```APIDOC
## Toolchain::peano

### Description
Returns the path to the peano tool.

### Method
`pub fn peano(&self) -> &Path`

### Returns
- **&Path** - The path to the peano executable.
```

--------------------------------

### KernelSource::new

Source: https://docs.rs/taconite/latest/taconite/compile/struct.KernelSource.html

Creates a new KernelSource instance with the given source path and architecture. This is used to specify the kernel source for compilation.

```APIDOC
## KernelSource::new

### Description
Creates a new `KernelSource` instance with the given source path and architecture. This is used to specify the kernel source for compilation.

### Method
`pub fn new(source: impl Into<PathBuf>, arch: Arch) -> Self`

### Parameters
- **source** (impl Into<PathBuf>) - Required - The path to the kernel source file.
- **arch** (Arch) - Required - The target architecture for the kernel.

### Returns
- **Self** - A new `KernelSource` instance.

### Example
```rust
let kernel_source = KernelSource::new("kernel.cc", Arch::Npu1);
```
```

--------------------------------

### include

Source: https://docs.rs/taconite/latest/taconite/compile/struct.KernelSource.html

Adds an include directory to the compiler invocation using the -I flag.

```APIDOC
## include

### Description
Adds an include directory to the compiler invocation using the `-I` flag.

### Method
`pub fn include(self, dir: impl Into<PathBuf>) -> Self`

### Parameters
- **dir** (PathBuf) - Required - The directory to add as an include path.

### Returns
- **Self** - The updated `KernelSource` instance for chaining.
```
Tokens:
Raw