mirror of
https://github.com/huggingface/candle.git
synced 2025-06-20 04:00:28 +00:00
Follow crate conventions
This commit is contained in:
@ -22,7 +22,11 @@ fn criterion_benchmark(c: &mut Criterion) {
|
||||
bencher.iter_custom(|iters| {
|
||||
let start = Instant::now();
|
||||
for _i in 0..iters {
|
||||
run(black_box((b, rows, columns)), black_box(DType::U8), black_box(&device1));
|
||||
run(
|
||||
black_box((b, rows, columns)),
|
||||
black_box(DType::U8),
|
||||
black_box(&device1),
|
||||
);
|
||||
}
|
||||
if let Device::Metal(device) = &device1 {
|
||||
device.wait_until_completed().unwrap();
|
||||
@ -35,12 +39,18 @@ fn criterion_benchmark(c: &mut Criterion) {
|
||||
group.finish();
|
||||
|
||||
let mut group = c.benchmark_group("fill_metal_f32");
|
||||
group.throughput(Throughput::Bytes((flops * DType::F32.size_in_bytes()) as u64));
|
||||
group.throughput(Throughput::Bytes(
|
||||
(flops * DType::F32.size_in_bytes()) as u64,
|
||||
));
|
||||
group.bench_function("iter", move |bencher| {
|
||||
bencher.iter_custom(|iters| {
|
||||
let start = Instant::now();
|
||||
for _i in 0..iters {
|
||||
run(black_box((b, rows, columns)), black_box(DType::F32), black_box(&device2));
|
||||
run(
|
||||
black_box((b, rows, columns)),
|
||||
black_box(DType::F32),
|
||||
black_box(&device2),
|
||||
);
|
||||
}
|
||||
if let Device::Metal(device) = &device2 {
|
||||
device.wait_until_completed().unwrap();
|
||||
|
@ -3,7 +3,7 @@ use crate::conv::{ParamsConv1D, ParamsConv2D, ParamsConvTranspose1D, ParamsConvT
|
||||
use crate::op::{BinaryOpT, CmpOp, ReduceOp, UnaryOpT};
|
||||
use crate::{CpuStorage, DType, Layout, Result, Shape};
|
||||
use candle_metal_kernels;
|
||||
use candle_metal_kernels::{FillOp, Unary, Kernels};
|
||||
use candle_metal_kernels::Kernels;
|
||||
use half::{bf16, f16};
|
||||
use metal;
|
||||
use metal::{Buffer, CommandBuffer, CommandQueue, MTLResourceOptions, NSUInteger};
|
||||
@ -1405,15 +1405,14 @@ impl BackendDevice for MetalDevice {
|
||||
let command_buffer = self.command_buffer()?;
|
||||
command_buffer.set_label("zeros");
|
||||
|
||||
// This assumes the zero value of this DType is equal to 0x00u8
|
||||
// This kernel assumes the zero value of this DType is equal to 0x00u8
|
||||
// (which is true for all current types)
|
||||
Unary::fill(
|
||||
&self.device,
|
||||
candle_metal_kernels::call_fill_u8(
|
||||
&command_buffer,
|
||||
&self.kernels,
|
||||
shape.elem_count(),
|
||||
&buffer,
|
||||
0u8,
|
||||
0,
|
||||
)
|
||||
.map_err(MetalError::from)?;
|
||||
|
||||
@ -1427,7 +1426,7 @@ impl BackendDevice for MetalDevice {
|
||||
|
||||
macro_rules! fill {
|
||||
($value:expr) => {
|
||||
Unary::fill(
|
||||
candle_metal_kernels::call_fill(
|
||||
&self.device,
|
||||
&command_buffer,
|
||||
&self.kernels,
|
||||
@ -1439,7 +1438,14 @@ impl BackendDevice for MetalDevice {
|
||||
};
|
||||
}
|
||||
match dtype {
|
||||
DType::U8 => fill!(1u8),
|
||||
DType::U8 => candle_metal_kernels::call_fill_u8(
|
||||
&command_buffer,
|
||||
&self.kernels,
|
||||
shape.elem_count(),
|
||||
&buffer,
|
||||
1u8,
|
||||
)
|
||||
.map_err(MetalError::from)?,
|
||||
DType::U32 => fill!(1u32),
|
||||
DType::I64 => fill!(1i64),
|
||||
DType::BF16 => fill!(bf16::ONE),
|
||||
|
Reference in New Issue
Block a user