# Fortran MAXLOC SVP64 demo

<https://bugs.libre-soc.org/show_bug.cgi?id=676>

MAXLOC is a notoriously difficult function for SIMD to cope with.
Typical approaches are to perform leaf-node (depth-first) parallel
operations, merging the results mapreduce-style to guage a final
index.

SVP64 however has similar capabilities to Z80 CPIR and LDIR and
therefore hardware may transparently implement back-end parallel
operations whilst the developer programs in a simple sequential
algorithm.

A clear reference implementation of MAXLOC is as follows:

```
int maxloc(int * const restrict a, int n) {
    int m, nm = INT_MIN, 0;
    for (int i=0; i<n; i++) {
        if (a[i] > m) {
            m = a[i];
            nm = i;
        }
    }
    return nm; 
}
```

**AVX-512**

An informative article by Vamsi Sripathi of Intel shows the extent of
the problem faced by SIMD ISAs (in the case below, AVX-512). Significant
loop-unrolling is performed which leaves blocks that need to be merged:
this is carried out with "blending" instructions.

Article:
<https://www.intel.com/content/www/us/en/developer/articles/technical/optimizing-maxloc-operation-using-avx-512-vector-instructions.html#gs.12t5y0>

<img src="https://www.intel.com/content/dam/developer/articles/technical/optimizing-maxloc-operation-using-avx-512-vector-instructions/optimizing-maxloc-operation-using-avx-512-vector-instructions-code2.png
" alt="NLnet foundation logo" width="100%" />

**ARM NEON**

From stackexchange in ARM NEON intrinsics, one developer (Pavel P) wrote the
subroutine below, explaining that it finds the index of a minimum value within
a group of eight unsigned bytes. It is necessary to use a second outer loop
to perform many of these searches in parallel, followed by conditionally
offsetting each of the block-results. 

<https://stackoverflow.com/questions/49683866/find-min-and-position-of-the-min-element-in-uint8x8-t-neon-register>

```
#define VMIN8(x, index, value)                               \
do {                                                         \
    uint8x8_t m = vpmin_u8(x, x);                            \
    m = vpmin_u8(m, m);                                      \
    m = vpmin_u8(m, m);                                      \
    uint8x8_t r = vceq_u8(x, m);                             \
                                                             \
    uint8x8_t z = vand_u8(vmask, r);                         \
                                                             \
    z = vpadd_u8(z, z);                                      \
    z = vpadd_u8(z, z);                                      \
    z = vpadd_u8(z, z);                                      \
                                                             \
    unsigned u32 = vget_lane_u32(vreinterpret_u32_u8(z), 0); \
    index = __lzcnt(u32);                                    \
    value = vget_lane_u8(m, 0);                              \
} while (0)


uint8_t v[8] = { ... };

static const uint8_t mask[] = { 0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01 };
uint8x8_t vmask = vld1_u8(mask);

uint8x8_t v8 = vld1_u8(v);
int ret;
int ret_pos;
VMIN8(v8, ret_pos, ret);
```

**Rust Assembler Intrinsics**

An approach by jvdd shows that the two stage approach of "blending"
arrays of results in a type of parallelised "leaf node depth first"
search seems to be a common technique.

<https://github.com/jvdd/argminmax/blob/main/src/simd/simd_u64.rs>

**Python implementation**

A variant on the c reference implementation allows for skipping
of updating m (the largest value found so far), followed by
always updating m whilst a batch of numbers is found that are
(in their order of sequence) always continuously greater than
all previous numbers.  The algorithm therefore flips back and
forth between "skip whilst smaller" and "update whilst bigger",
only updating the index during "bigger" sequences. This is key
later when doing SVP64 assembler.

```
def m2(a): # array a
    m, nm, i, n = 0, 0, 0, len(a)
    while i<n:
        while i<n and a[i]<=m: i += 1 # skip whilst smaller
        while i<n and a[i]> m: m, nm, i = a[i], i, i+1
    return nm;
```

# Implementation in SVP64 Assembler

The core algorithm (inner part, in-register) is below: 11 instructions.
Loading of data, and offsetting the "core" below is relatively
straightforward: estimated another 6 instructions and needing one
more branch (outer loop).

```
# while (i<n)
setvl 2,0,4,0,1,1            # set MVL=4, VL=MIN(MVL,CTR)
#    while (i<n and a[i]<=m) : i += 1
sv.cmp/ff=gt/m=ge *0,0,*10,4 # truncates VL to min
sv.creqv *16,*16,*16         # set mask on already-tested
setvl 2,0,4,0,1,1            # set MVL=4, VL=MIN(MVL,CTR)
mtcrf 128, 0                 # clear CR0 (in case VL=0?)
#    while (i<n and a[i]>m):
sv.minmax./ff=le/m=ge 4,*10,4,1 # uses r4 as accumulator
crternlogi 0,1,2,127         # test greater/equal or VL=0
sv.crand *19,*16,0           # clear if CR0.eq=0
#      nm = i (count masked bits. could use crweirds here TODO)
sv.svstep/mr/m=so 1, 0, 6, 1 # svstep: get vector dststep
sv.creqv *16,*16,*16         # set mask on already-tested
bc 12,0, -0x40               # CR0 lt bit clear, branch back
```

`sv.cmp` can be used in the first while loop because m (r4, the current
largest-found value so far) does not change.
However `sv.minmax.` has to be used in the key while loop
because r4 is sequentially replaced each time, and mapreduce (`/mr`)
is used to request this rolling-chain (accumulator) mode.

Also note that `i` (the `"while i<n"` part) is represented as an
unary bitmask (in CR bits 16,20,24,28) which turns out to be easier to
use than a binary index, as it can be used directly as a Predicate Mask
(`/m=ge`).

The algorithm works by excluding previous operations using `i-in-unary`,
combined with VL being truncated due to use of Data-Dependent Fail-First.
What therefore happens for example on the `sv.com/ff=gt/m=ge` operation
is that it is *VL* (the Vector Length) that gets truncated to only
contain those elements that are smaller than the current largest value
found (`m` aka `r4`). Calling `sv.creqv` then sets **only** the
CR field bits up to length `VL`, which on the next loop will exclude
them because the Predicate Mask is `m=ge` (ok if the CR field bit is
**zero**).

[[!tag svp64_cookbook ]]