sigmoid

linalg/arm64/arm64simd/arm64simd_act_f32_32n.tmpl

@@ -339,8 +339,41 @@
  fmax v7.4s, v7.4s, v24.4s
  b .inner_loop
 
-.fma:
- b .unsupported
+.fma: 
+ // a <- a * b + k
+ // vfma a,b,c does a <- a + b * c
+ // mov d,a ; mov a,#k ; vfma a, b, d
+
+ and v24.16b, v0.16b, v0.16b
+ and v25.16b, v1.16b, v1.16b
+ and v26.16b, v2.16b, v2.16b
+ and v27.16b, v3.16b, v3.16b
+ and v28.16b, v4.16b, v4.16b
+ and v29.16b, v5.16b, v5.16b
+ and v30.16b, v6.16b, v6.16b
+ and v31.16b, v7.16b, v7.16b
+
+ ins v0.s[0], w3
+ add x5, x5, 4
+ dup v0.4s, v0.s[0]
+ dup v1.4s, v0.s[0]
+ dup v2.4s, v0.s[0]
+ dup v3.4s, v0.s[0]
+ dup v4.4s, v0.s[0]
+ dup v5.4s, v0.s[0]
+ dup v6.4s, v0.s[0]
+ dup v7.4s, v0.s[0]
+
+ fmla v0.4s, v24.4s, v8.4s 
+ fmla v1.4s, v25.4s, v9.4s 
+ fmla v2.4s, v26.4s, v10.4s
+ fmla v3.4s, v27.4s, v11.4s
+ fmla v4.4s, v28.4s, v12.4s
+ fmla v5.4s, v29.4s, v13.4s
+ fmla v6.4s, v30.4s, v14.4s
+ fmla v7.4s, v31.4s, v15.4s
+
+ b .inner_loop
 
 .if_pos_then_else:
  fcmge v0.4s, v0.4s, #0.0
@@ -362,7 +395,34 @@
  b .inner_loop
 
 .swap_b_c:
- b .unsupported
+// move d <- b
+ and v24.16b, v8.16b , v8.16b
+ and v25.16b, v9.16b , v9.16b
+ and v26.16b, v10.16b, v10.16b
+ and v27.16b, v11.16b, v11.16b
+ and v28.16b, v12.16b, v12.16b
+ and v29.16b, v13.16b, v13.16b
+ and v30.16b, v14.16b, v14.16b
+ and v31.16b, v15.16b, v15.16b
+// move b <- c
+ and v8.16b , v16.16b, v16.16b
+ and v9.16b , v17.16b, v17.16b
+ and v10.16b, v18.16b, v18.16b
+ and v11.16b, v19.16b, v19.16b
+ and v12.16b, v20.16b, v20.16b
+ and v13.16b, v21.16b, v21.16b
+ and v14.16b, v22.16b, v22.16b
+ and v15.16b, v23.16b, v23.16b
+// move c <- d
+ and v16.16b, v24.16b, v24.16b
+ and v17.16b, v25.16b, v25.16b
+ and v18.16b, v26.16b, v26.16b
+ and v19.16b, v27.16b, v27.16b
+ and v20.16b, v28.16b, v28.16b
+ and v21.16b, v29.16b, v29.16b
+ and v22.16b, v30.16b, v30.16b
+ and v23.16b, v31.16b, v31.16b
+ b .inner_loop
 
 .floor:
  b .unsupported

linalg/benches/activations.rs

@@ -1,9 +1,11 @@
 use criterion::{black_box, criterion_group, criterion_main, BatchSize, BenchmarkId, Criterion};
 use tract_linalg::frame::activations::{definitions, reference, ActivationKer, Program};
 
+const SIZES:&[i32] = &[32, 256, 1024, 8192];
+
 fn crit(c: &mut Criterion, name: &str, r: impl Fn(f32) -> f32, prog: &Program<f32>) {
  let mut group = c.benchmark_group(name);
- for size in [1i32, 32, 256, 1024, 8192].iter() {
+ for size in SIZES {
  group.throughput(criterion::Throughput::Elements(*size as u64));
  group.bench_with_input(BenchmarkId::new("Reference", size), size, |b, size| {
  b.iter_batched(
@@ -14,7 +16,7 @@ fn crit(c: &mut Criterion, name: &str, r: impl Fn(f32) -> f32, prog: &Program<f3
  }
  },
  BatchSize::LargeInput,
- )
+  )
  });
  #[allow(unused_mut)]
  let mut vms = vec!(tract_linalg::generic::activations::SActivations::act());
@@ -29,7 +31,17 @@ fn crit(c: &mut Criterion, name: &str, r: impl Fn(f32) -> f32, prog: &Program<f3
  || vec![1.0f32; *size as usize],
  |mut v| vm.run(prog, &mut v),
  BatchSize::LargeInput,
- )
+ )
+ });
+ }
+ if name == "sigmoid" {
+ let sigmoid = (tract_linalg::ops().sigmoid_f32)();
+ group.bench_with_input(BenchmarkId::new("handcrafted", size), size, |b, size| {
+ b.iter_batched(
+ || vec![1.0f32; *size as usize],
+ |mut v| sigmoid.run(&mut v),
+ BatchSize::LargeInput,
+ )
  });
  }
  }
@@ -38,11 +50,12 @@ fn crit(c: &mut Criterion, name: &str, r: impl Fn(f32) -> f32, prog: &Program<f3
 fn criterion_benchmark(c: &mut Criterion) {
  crit(c, "relu", reference::relu, &definitions::relu());
  crit(c, "hardswish", reference::hardswish, &definitions::hard_swish());
- /*
- crit(c, "exp2f", reference::exp2f, &definitions::exp2f());
  crit(c, "sigmoid", reference::sigmoid, &definitions::sigmoid());
- */
-}
 
-criterion_group!(benches, criterion_benchmark);
-criterion_main!(benches);
+ /*
+ crit(c, "exp2f", reference::exp2f, &definitions::exp2f());
+ */
+ }
+
+ criterion_group!(benches, criterion_benchmark);
+ criterion_main!(benches);