bicg.rs

#![allow(non_snake_case)]

fn init_array<const M: usize, const N: usize>(
    A: &mut [[f32; N]; M],
    r: &mut [f32; M],
    p: &mut [f32; N],
) {
    for i in 0..N {
        p[i] = (i % N) as f32 / N as f32;
    }
    for i in 0..M {
        r[i] = (i % M) as f32 / M as f32;
        for j in 0..N {
            A[i][j] = (i * (j + 1) % M) as f32 / M as f32;
        }
    }
}

fn kernel_bicg<const M: usize, const N: usize>(
    A: &[[f32; N]; M],
    s: &mut [f32; N],
    q: &mut [f32; M],
    p: &[f32; N],
    r: &[f32; M],
) {
    for i in 0..N {
        s[i] = 0.0;
    }
    for i in 0..M {
        q[i] = 0.0;
        for j in 0..N {
            s[j] = s[j] + r[i] * A[i][j];
            q[i] = q[i] + A[i][j] * p[j];
        }
    }
}

#[no_mangle]
pub extern "C" fn bench() {
    const M: usize = 1;
    const N: usize = 1;

    let mut A = [[0_f32; N]; M];
    let mut s = [0_f32; N];
    let mut q = [0_f32; M];
    let mut p = [0_f32; N];
    let mut r = [0_f32; M];

    init_array(&mut A, &mut r, &mut p);
    kernel_bicg(&A, &mut s, &mut q, &p, &r);
}

bicg.wat

(module
  (type $t0 (func))
  (func $bench (type $t0)
    (local $l0 i32) (local $l1 i32) (local $l2 i32) (local $l3 f32)
    global.get $g0
    i32.const 32
    i32.sub
    local.set $l0
    i32.const 0
    local.set $l1
    block $B0
      loop $L1
        local.get $l1
        i32.const 4
        i32.eq
        br_if $B0
        local.get $l0
        i32.const 28
        i32.add
        local.get $l1
        i32.add
        i32.const 0
        i32.store
        local.get $l1
        i32.const 4
        i32.add
        local.set $l1
        br $L1
      end
    end
    i32.const 0
    local.set $l1
    local.get $l0
    i32.load offset=28
    local.set $l2
    block $B2
      loop $L3
        local.get $l1
        i32.const 4
        i32.eq
        br_if $B2
        local.get $l0
        i32.const 12
        i32.add
        local.get $l1
        i32.add
        local.get $l2
        i32.store
        local.get $l1
        i32.const 4
        i32.add
        local.set $l1
        br $L3
      end
    end
    i32.const 0
    local.set $l1
    block $B4
      loop $L5
        local.get $l1
        i32.const 4
        i32.eq
        br_if $B4
        local.get $l0
        i32.const 16
        i32.add
        local.get $l1
        i32.add
        i32.const 0
        i32.store
        local.get $l1
        i32.const 4
        i32.add
        local.set $l1
        br $L5
      end
    end
    i32.const 0
    local.set $l1
    block $B6
      loop $L7
        local.get $l1
        i32.const 4
        i32.eq
        br_if $B6
        local.get $l0
        i32.const 20
        i32.add
        local.get $l1
        i32.add
        i32.const 0
        i32.store
        local.get $l1
        i32.const 4
        i32.add
        local.set $l1
        br $L7
      end
    end
    i32.const 0
    local.set $l1
    block $B8
      loop $L9
        local.get $l1
        i32.const 4
        i32.eq
        br_if $B8
        local.get $l0
        i32.const 24
        i32.add
        local.get $l1
        i32.add
        i32.const 0
        i32.store
        local.get $l1
        i32.const 4
        i32.add
        local.set $l1
        br $L9
      end
    end
    i32.const 0
    local.set $l1
    block $B10
      loop $L11
        local.get $l1
        i32.const 4
        i32.eq
        br_if $B10
        local.get $l0
        i32.const 28
        i32.add
        local.get $l1
        i32.add
        i32.const 0
        i32.store
        local.get $l1
        i32.const 4
        i32.add
        local.set $l1
        br $L11
      end
    end
    local.get $l0
    i32.const 0
    i32.store offset=28
    local.get $l0
    i32.const 0
    i32.store offset=24
    local.get $l0
    i32.const 0
    i32.store offset=12
    local.get $l0
    i32.const 0
    i32.store offset=16
    f32.const 0x0p+0 (;=0;)
    local.set $l3
    i32.const 1
    local.set $l1
    block $B12
      loop $L13
        local.get $l1
        i32.const 1
        i32.and
        i32.eqz
        br_if $B12
        local.get $l0
        f32.const 0x0p+0 (;=0;)
        local.get $l3
        f32.add
        local.tee $l3
        f32.store offset=20
        local.get $l0
        f32.const 0x0p+0 (;=0;)
        local.get $l0
        f32.load offset=16
        f32.add
        f32.store offset=16
        i32.const 0
        local.set $l1
        br $L13
      end
    end)
  (memory $memory 1)
  (global $g0 (mut i32) (i32.const 16))
  (global $__data_end i32 (i32.const 16))
  (global $__heap_base i32 (i32.const 16))
  (export "memory" (memory 0))
  (export "bench" (func $bench))
  (export "__data_end" (global 1))
  (export "__heap_base" (global 2)))