Work on porting SIMD vt arser to Go for the kittens

This commit is contained in:
Kovid Goyal 2024-01-31 19:02:19 +05:30
parent 131716da00
commit de8c1e0206
No known key found for this signature in database
GPG key ID: 06BC317B515ACE7C
7 changed files with 1660 additions and 0 deletions

3
tools/simdstring/.gitignore vendored Normal file
View file

@ -0,0 +1,3 @@
tags
simdstring.test
asm_*

1252
tools/simdstring/generate.go Normal file

File diff suppressed because it is too large Load diff

8
tools/simdstring/generate.sh Executable file
View file

@ -0,0 +1,8 @@
#!/bin/sh
#
# generate.sh
# Copyright (C) 2024 Kovid Goyal <kovid at kovidgoyal.net>
#
# Distributed under terms of the MIT license.
#
go run generate.go && GOARCH=amd64 go vet ./... && GOARCH=arm64 go vet ./...

View file

@ -0,0 +1,106 @@
// License: GPLv3 Copyright: 2024, Kovid Goyal, <kovid at kovidgoyal.net>
package simdstring
import (
"runtime"
"unsafe"
"golang.org/x/sys/cpu"
)
var Have128bit = false
var Have256bit = false
var VectorSize = 1
// Return the index at which either a or b first occurs in data. If neither is
// found -1 is returned. Can read upto VectorSize-1 bytes beyond the end of data.
var UnsafeIndexByte2 func(data []byte, a, b byte) int = index_byte2_scalar
// Return the index at which either a or b first occurs in text. If neither is
// found -1 is returned. Can read upto VectorSize-1 bytes beyond the end of text.
var UnsafeIndexByte2String func(text string, a, b byte) int = index_byte2_string_scalar
func get_safe_slice(addr uintptr, mask, datalen int) (align_len, vecsafe_len int) {
vecsize := mask + 1
if datalen < vecsize {
return datalen, 0
}
align_len = vecsize - int(addr&uintptr(mask))
datalen -= align_len
vecsafe_len = datalen - datalen&mask
return
}
func get_safe_slice_with_cap(addr uintptr, mask, datalen, datacap int) (align_len, vecsafe_len int) {
if datacap > datalen+mask {
return 0, datalen
}
return get_safe_slice(addr, mask, datalen)
}
func IndexByte2(data []byte, a, b byte) int {
align_len, vecsafe_len := get_safe_slice_with_cap(uintptr(unsafe.Pointer(&data)), VectorSize-1, len(data), cap(data))
if align_len > 0 {
if ans := index_byte2_scalar(data[:align_len], a, b); ans > -1 {
return ans
}
data = data[align_len:]
}
if vecsafe_len > 0 {
if ans := UnsafeIndexByte2(data[:vecsafe_len], a, b); ans > -1 {
return align_len + ans
}
data = data[vecsafe_len:]
}
if len(data) > 0 {
if ans := index_byte2_scalar(data, a, b); ans > -1 {
return ans + align_len + vecsafe_len
}
}
return -1
}
func IndexByte2String(data string, a, b byte) int {
align_len, vecsafe_len := get_safe_slice(uintptr(unsafe.Pointer(&data)), VectorSize-1, len(data))
if align_len > 0 {
if ans := index_byte2_string_scalar(data[:align_len], a, b); ans > -1 {
return ans
}
data = data[align_len:]
}
if vecsafe_len > 0 {
if ans := UnsafeIndexByte2String(data[:vecsafe_len], a, b); ans > -1 {
return align_len + ans
}
data = data[vecsafe_len:]
}
if len(data) > 0 {
if ans := index_byte2_string_scalar(data, a, b); ans > -1 {
return ans + align_len + vecsafe_len
}
}
return -1
}
func init() {
switch runtime.GOARCH {
case "amd64":
if cpu.Initialized {
Have128bit = cpu.X86.HasSSE42 && HasSIMD128Code
Have256bit = cpu.X86.HasAVX2 && HasSIMD256Code
}
case "arm64":
Have128bit = HasSIMD128Code
Have128bit = HasSIMD256Code
}
if Have256bit {
UnsafeIndexByte2 = index_byte2_asm_256
UnsafeIndexByte2String = index_byte2_string_asm_256
VectorSize = 32
} else if Have128bit {
UnsafeIndexByte2 = index_byte2_asm_128
UnsafeIndexByte2String = index_byte2_string_asm_128
VectorSize = 16
}
}

View file

@ -0,0 +1,256 @@
// License: GPLv3 Copyright: 2024, Kovid Goyal, <kovid at kovidgoyal.net>
package simdstring
import (
"bytes"
"fmt"
"runtime"
"strings"
"testing"
"github.com/google/go-cmp/cmp"
)
var _ = fmt.Print
func test_load(src []byte) []byte {
ans := make([]byte, len(src))
if len(src) == 16 {
test_load_asm_128(src, ans)
} else {
test_load_asm_256(src, ans)
}
return ans
}
func test_set1_epi8(b byte, sz int) []byte {
ans := make([]byte, sz)
if sz == 16 {
test_set1_epi8_asm_128(b, ans)
} else {
test_set1_epi8_asm_256(b, ans)
}
return ans
}
func test_cmpeq_epi8(a, b []byte) []byte {
ans := make([]byte, len(a))
if len(ans) == 16 {
test_cmpeq_epi8_asm_128(a, b, ans)
} else {
test_cmpeq_epi8_asm_256(a, b, ans)
}
return ans
}
func test_or(a, b []byte) []byte {
ans := make([]byte, len(a))
if len(ans) == 16 {
test_or_asm_128(a, b, ans)
} else {
test_or_asm_256(a, b, ans)
}
return ans
}
func test_jump_if_zero(a []byte) int {
if len(a) == 16 {
return test_jump_if_zero_asm_128(a)
}
return test_jump_if_zero_asm_256(a)
}
func test_count_to_match(a []byte, b byte) int {
if len(a) == 16 {
return test_count_to_match_asm_128(a, b)
}
return test_count_to_match_asm_256(a, b)
}
func ordered_bytes(size int) []byte {
ans := make([]byte, size)
for i := range ans {
ans[i] = byte(i)
}
return ans
}
func broadcast_byte(b byte, size int) []byte {
return bytes.Repeat([]byte{b}, size)
}
func TestSIMDStringOps(t *testing.T) {
test := func(haystack []byte, a, b byte) {
sizes := []int{}
if Have128bit {
sizes = append(sizes, 16)
}
if Have256bit {
sizes = append(sizes, 32)
}
var actual int
safe_haystack := append(bytes.Repeat([]byte{'<'}, 64), haystack...)
safe_haystack = append(safe_haystack, bytes.Repeat([]byte{'>'}, 64)...)
haystack = safe_haystack[64 : 64+len(haystack)]
expected := index_byte2_scalar(haystack, a, b)
for _, sz := range sizes {
switch sz {
case 16:
actual = index_byte2_asm_128(haystack, a, b)
case 32:
actual = index_byte2_asm_256(haystack, a, b)
}
if actual != expected {
t.Fatalf("Failed to find '%c' or '%c' in: %#v (%d != %d) at size: %d", a, b, string(haystack), expected, actual, sz)
}
}
}
tests := func(h string, a, b byte) {
test([]byte(h), a, b)
for _, sz := range []int{16, 32, 64, 79} {
q := strings.Repeat(" ", sz) + h
test([]byte(q), a, b)
}
}
test(nil, '1', '2')
test([]byte{}, '1', '2')
tests("", '<', '>')
tests("a", 0, 0)
tests("a", '<', '>')
tests("dsdfsfa", '1', 'a')
tests("xa", 'a', 'a')
tests("bbb", 'a', '1')
tests("bba", 'a', '<')
tests("baa", '>', 'a')
tbs := func(addr, datalen int) {
align_len, vecsafelen := get_safe_slice(uintptr(addr), 15, datalen)
if vecsafelen < 0 || align_len+vecsafelen > datalen || datalen-vecsafelen-align_len > 15 || align_len < 0 {
t.Fatalf("Invalid bounds for addr=%d datalen=%d (align_len=%d vecsafelen=%d)", addr, datalen, align_len, vecsafelen)
}
if vecsafelen > 0 {
pos := addr + align_len
if pos&15 != 0 {
t.Fatalf("Non-aligned vector read for addr=%d datalen=%d (align_len=%d vecsafelen=%d)", addr, datalen, align_len, vecsafelen)
}
limit := pos + vecsafelen
read := func() {
if pos+16 > addr+datalen {
t.Fatalf("Read past limit for addr=%d datalen=%d (align_len=%d vecsafelen=%d pos=%d)", addr, datalen, align_len, vecsafelen, pos)
}
}
read()
for ; pos < limit; pos += 16 {
read()
}
}
}
for datalen := 0; datalen < 33; datalen++ {
for addr := 0; addr < 32; addr++ {
tbs(addr, datalen)
}
}
}
func TestIntrinsics(t *testing.T) {
switch runtime.GOARCH {
case "amd64":
if !HasSIMD128Code {
t.Fatal("SIMD 128bit code not built")
}
if !HasSIMD256Code {
t.Fatal("SIMD 256bit code not built")
}
case "arm64":
if !HasSIMD128Code {
t.Fatal("SIMD 128bit code not built")
}
}
ae := func(sz int, func_name string, a, b any) {
if s := cmp.Diff(a, b); s != "" {
t.Fatalf("%s failed with size: %d\n%s", func_name, sz, s)
}
}
tests := []func(int){}
tests = append(tests, func(sz int) {
a := ordered_bytes(sz)
ae(sz, `load_test`, a, test_load(a))
})
tests = append(tests, func(sz int) {
for _, b := range []byte{1, 0b110111, 0xff, 0} {
ae(sz, `set1_epi8_test`, broadcast_byte(b, sz), test_set1_epi8(b, sz))
}
})
tests = append(tests, func(sz int) {
a := ordered_bytes(sz)
b := ordered_bytes(sz)
ans := test_cmpeq_epi8(a, b)
ae(sz, `cmpeq_epi8_test`, broadcast_byte(0xff, sz), ans)
})
tests = append(tests, func(sz int) {
a := make([]byte, sz)
b := make([]byte, sz)
c := make([]byte, sz)
a[0] = 0xff
b[0] = 0xff
b[1] = 0xff
a[sz-1] = 1
b[sz-1] = 2
for i := range c {
c[i] = a[i] | b[i]
}
ans := test_or(a, b)
ae(sz, `or_test`, c, ans)
})
tests = append(tests, func(sz int) {
a := make([]byte, sz)
if e := test_jump_if_zero(a); e != 0 {
t.Fatalf("Did not detect zero register")
}
for i := 0; i < sz; i++ {
a = make([]byte, sz)
a[i] = 1
if e := test_jump_if_zero(a); e != 1 {
t.Fatalf("Did not detect non-zero register")
}
}
})
tests = append(tests, func(sz int) {
a := ordered_bytes(sz)
if e := test_count_to_match(a, 77); e != -1 {
t.Fatalf("Unexpectedly found byte at: %d", e)
}
for i := 0; i < sz; i++ {
if e := test_count_to_match(a, byte(i)); e != i {
t.Fatalf("Failed to find the byte: %d (%d != %d)", i, i, e)
}
}
a[7] = 0x34
if e := test_count_to_match(a, 0x34); e != 7 {
t.Fatalf("Failed to find the byte: %d (%d != %d)", 0x34, 7, e)
}
})
sizes := []int{}
if Have128bit {
sizes = append(sizes, 16)
}
if Have256bit {
sizes = append(sizes, 32)
}
for _, sz := range sizes {
for _, test := range tests {
test(sz)
}
}
}

View file

@ -0,0 +1,23 @@
// License: GPLv3 Copyright: 2024, Kovid Goyal, <kovid at kovidgoyal.net>
package simdstring
func index_byte2_scalar(data []byte, a, b byte) int {
for i, ch := range data {
switch ch {
case a, b:
return i
}
}
return -1
}
func index_byte2_string_scalar(data string, a, b byte) int {
for i := 0; i < len(data); i++ {
switch data[i] {
case a, b:
return i
}
}
return -1
}

12
tools/simdstring/test.sh Executable file
View file

@ -0,0 +1,12 @@
#!/bin/sh
#
# test.sh
# Copyright (C) 2024 Kovid Goyal <kovid at kovidgoyal.net>
#
# Distributed under terms of the MIT license.
#
echo -e "\x1b[32mtesting amd64\x1b[m" && go test -v &&
echo -e "\x1b[32mtesting arm64\x1b[m" && GOARCH=arm64 go test -c && qemu-aarch64-static simdstring.test -test.v