ed/Odin
1
0
Fork 0
mirror of https://github.com/Ed94/Odin.git synced 2026-08-07 08:08:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

ZLIB: Moar faster.

Browse Source
This commit is contained in:
Jeroen van Rijn
2021-06-26 20:40:39 +02:00
parent 8ba1c9a6cd
commit 30a5808460
3 changed files with 109 additions and 159 deletions
Download Patch File Download Diff File Expand all files Collapse all files
core/compress/zlib/zlib.odin
+56 -77
View File
@@ -30,8 +30,7 @@ import "core:bytes"
`Context.rolling_hash` if not inlining it is still faster.
*/
Context :: compress.Context;
Code_Buffer :: compress.Code_Buffer;
Context :: compress.Context;
Compression_Method :: enum u8 {
DEFLATE = 8,
@@ -166,7 +165,7 @@ grow_buffer :: proc(buf: ^[dynamic]u8) -> (err: compress.Error) {
*/
@(optimization_mode="speed")
write_byte :: #force_inline proc(z: ^Context, cb: ^Code_Buffer, c: u8) -> (err: io.Error) #no_bounds_check {
write_byte :: #force_inline proc(z: ^Context, c: u8) -> (err: io.Error) #no_bounds_check {
/*
Resize if needed.
*/
@@ -179,14 +178,13 @@ write_byte :: #force_inline proc(z: ^Context, cb: ^Code_Buffer, c: u8) -> (err:
#no_bounds_check {
z.output.buf[z.bytes_written] = c;
cb.last[z.bytes_written & cb.window_mask] = c;
}
z.bytes_written += 1;
return .None;
}
@(optimization_mode="speed")
repl_byte :: proc(z: ^Context, cb: ^Code_Buffer, count: u16, c: u8) -> (err: io.Error) #no_bounds_check {
repl_byte :: proc(z: ^Context, count: u16, c: u8) -> (err: io.Error) #no_bounds_check {
/*
TODO(Jeroen): Once we have a magic ring buffer, we can just peek/write into it
without having to worry about wrapping, so no need for a temp allocation to give to
@@ -206,7 +204,6 @@ repl_byte :: proc(z: ^Context, cb: ^Code_Buffer, count: u16, c: u8) -> (err: io.
#no_bounds_check {
for _ in 0..<count {
z.output.buf[z.bytes_written] = c;
cb.last[z.bytes_written & cb.window_mask] = c;
z.bytes_written += 1;
}
}
@@ -215,14 +212,14 @@ repl_byte :: proc(z: ^Context, cb: ^Code_Buffer, count: u16, c: u8) -> (err: io.
}
@(optimization_mode="speed")
repl_bytes :: proc(z: ^Context, cb: ^Code_Buffer, count: u16, distance: u16) -> (err: io.Error) {
repl_bytes :: proc(z: ^Context, count: u16, distance: u16) -> (err: io.Error) {
/*
TODO(Jeroen): Once we have a magic ring buffer, we can just peek/write into it
without having to worry about wrapping, so no need for a temp allocation to give to
the output stream, just give it _that_ slice.
*/
offset := z.bytes_written - i64(distance);
offset := i64(distance);
if int(z.bytes_written) + int(count) >= len(z.output.buf) {
e := grow_buffer(&z.output.buf);
@@ -233,10 +230,9 @@ repl_bytes :: proc(z: ^Context, cb: ^Code_Buffer, count: u16, distance: u16) ->
#no_bounds_check {
for _ in 0..<count {
c := cb.last[offset & cb.window_mask];
c := z.output.buf[z.bytes_written - offset];
z.output.buf[z.bytes_written] = c;
cb.last[z.bytes_written & cb.window_mask] = c;
z.bytes_written += 1; offset += 1;
z.bytes_written += 1;
}
}
@@ -308,8 +304,8 @@ build_huffman :: proc(z: ^Huffman_Table, code_lengths: []u8) -> (err: Error) {
}
@(optimization_mode="speed")
decode_huffman_slowpath :: proc(z: ^Context, cb: ^Code_Buffer, t: ^Huffman_Table) -> (r: u16, err: Error) #no_bounds_check {
code := u16(compress.peek_bits_lsb(z, cb, 16));
decode_huffman_slowpath :: proc(z: ^Context, t: ^Huffman_Table) -> (r: u16, err: Error) #no_bounds_check {
code := u16(compress.peek_bits_lsb(z,16));
k := int(z_bit_reverse(code, 16));
s: u8;
@@ -332,41 +328,41 @@ decode_huffman_slowpath :: proc(z: ^Context, cb: ^Code_Buffer, t: ^Huffman_Table
return 0, E_Deflate.Bad_Huffman_Code;
}
compress.consume_bits_lsb(cb, s);
compress.consume_bits_lsb(z, s);
r = t.value[b];
return r, nil;
}
@(optimization_mode="speed")
decode_huffman :: proc(z: ^Context, cb: ^Code_Buffer, t: ^Huffman_Table) -> (r: u16, err: Error) #no_bounds_check {
if cb.num_bits < 16 {
if cb.num_bits > 63 {
decode_huffman :: proc(z: ^Context, t: ^Huffman_Table) -> (r: u16, err: Error) #no_bounds_check {
if z.num_bits < 16 {
if z.num_bits > 63 {
return 0, E_ZLIB.Code_Buffer_Malformed;
}
compress.refill_lsb(z, cb);
if cb.num_bits > 63 {
compress.refill_lsb(z);
if z.num_bits > 63 {
return 0, E_General.Stream_Too_Short;
}
}
#no_bounds_check b := t.fast[cb.code_buffer & ZFAST_MASK];
#no_bounds_check b := t.fast[z.code_buffer & ZFAST_MASK];
if b != 0 {
s := u8(b >> ZFAST_BITS);
compress.consume_bits_lsb(cb, s);
compress.consume_bits_lsb(z, s);
return b & 511, nil;
}
return decode_huffman_slowpath(z, cb, t);
return decode_huffman_slowpath(z, t);
}
@(optimization_mode="speed")
parse_huffman_block :: proc(z: ^Context, cb: ^Code_Buffer, z_repeat, z_offset: ^Huffman_Table) -> (err: Error) #no_bounds_check {
parse_huffman_block :: proc(z: ^Context, z_repeat, z_offset: ^Huffman_Table) -> (err: Error) #no_bounds_check {
#no_bounds_check for {
value, e := decode_huffman(z, cb, z_repeat);
value, e := decode_huffman(z, z_repeat);
if e != nil {
return err;
}
if value < 256 {
e := write_byte(z, cb, u8(value));
e := write_byte(z, u8(value));
if e != .None {
return E_General.Output_Too_Short;
}
@@ -379,17 +375,17 @@ parse_huffman_block :: proc(z: ^Context, cb: ^Code_Buffer, z_repeat, z_offset: ^
value -= 257;
length := Z_LENGTH_BASE[value];
if Z_LENGTH_EXTRA[value] > 0 {
length += u16(compress.read_bits_lsb(z, cb, Z_LENGTH_EXTRA[value]));
length += u16(compress.read_bits_lsb(z, Z_LENGTH_EXTRA[value]));
}
value, e = decode_huffman(z, cb, z_offset);
value, e = decode_huffman(z, z_offset);
if e != nil {
return E_Deflate.Bad_Huffman_Code;
}
distance := Z_DIST_BASE[value];
if Z_DIST_EXTRA[value] > 0 {
distance += u16(compress.read_bits_lsb(z, cb, Z_DIST_EXTRA[value]));
distance += u16(compress.read_bits_lsb(z, Z_DIST_EXTRA[value]));
}
if z.bytes_written < i64(distance) {
@@ -397,7 +393,6 @@ parse_huffman_block :: proc(z: ^Context, cb: ^Code_Buffer, z_repeat, z_offset: ^
return E_Deflate.Bad_Distance;
}
offset := i64(z.bytes_written - i64(distance));
/*
These might be sped up with a repl_byte call that copies
from the already written output more directly, and that
@@ -410,15 +405,15 @@ parse_huffman_block :: proc(z: ^Context, cb: ^Code_Buffer, z_repeat, z_offset: ^
Replicate the last outputted byte, length times.
*/
if length > 0 {
c := cb.last[offset & cb.window_mask];
e := repl_byte(z, cb, length, c);
c := z.output.buf[z.bytes_written - i64(distance)];
e := repl_byte(z, length, c);
if e != .None {
return E_General.Output_Too_Short;
}
}
} else {
if length > 0 {
e := repl_bytes(z, cb, length, distance);
e := repl_bytes(z, length, distance);
if e != .None {
return E_General.Output_Too_Short;
}
@@ -442,9 +437,6 @@ inflate_from_stream :: proc(using ctx: ^Context, raw := false, expected_output_s
DEFLATE stream.
*/
code_buffer := Code_Buffer{};
cb := &code_buffer;
if !raw {
data_size := io.size(ctx.input);
if data_size < 6 {
@@ -462,8 +454,6 @@ inflate_from_stream :: proc(using ctx: ^Context, raw := false, expected_output_s
if cinfo > 7 {
return E_ZLIB.Unsupported_Window_Size;
}
cb.window_mask = i64((1 << (cinfo + 8) - 1));
flg, _ := compress.read_u8(ctx);
fcheck := flg & 0x1f;
@@ -488,23 +478,21 @@ inflate_from_stream :: proc(using ctx: ^Context, raw := false, expected_output_s
at the end to compare checksums.
*/
// Seed the Adler32 rolling checksum.
ctx.rolling_hash = 1;
}
// Parse ZLIB stream without header.
err = inflate_raw(z=ctx, cb=cb, expected_output_size=expected_output_size);
err = inflate_raw(z=ctx, expected_output_size=expected_output_size);
if err != nil {
return err;
}
if !raw {
compress.discard_to_next_byte_lsb(cb);
adler32 := compress.read_bits_lsb(ctx, cb, 8) << 24 | compress.read_bits_lsb(ctx, cb, 8) << 16 | compress.read_bits_lsb(ctx, cb, 8) << 8 | compress.read_bits_lsb(ctx, cb, 8);
compress.discard_to_next_byte_lsb(ctx);
adler32 := compress.read_bits_lsb(ctx, 8) << 24 | compress.read_bits_lsb(ctx, 8) << 16 | compress.read_bits_lsb(ctx, 8) << 8 | compress.read_bits_lsb(ctx, 8);
ctx.rolling_hash = hash.adler32(ctx.output.buf[:]);
output_hash := hash.adler32(ctx.output.buf[:]);
if ctx.rolling_hash != u32(adler32) {
if output_hash != u32(adler32) {
return E_General.Checksum_Failed;
}
}
@@ -512,7 +500,7 @@ inflate_from_stream :: proc(using ctx: ^Context, raw := false, expected_output_s
}
@(optimization_mode="speed")
inflate_from_stream_raw :: proc(z: ^Context, cb: ^Code_Buffer, expected_output_size := -1, allocator := context.allocator) -> (err: Error) #no_bounds_check {
inflate_from_stream_raw :: proc(z: ^Context, expected_output_size := -1, allocator := context.allocator) -> (err: Error) #no_bounds_check {
expected_output_size := expected_output_size;
if expected_output_size <= 0 {
@@ -536,8 +524,8 @@ inflate_from_stream_raw :: proc(z: ^Context, cb: ^Code_Buffer, expected_output_s
return .Resize_Failed;
}
cb.num_bits = 0;
cb.code_buffer = 0;
z.num_bits = 0;
z.code_buffer = 0;
z_repeat: ^Huffman_Table;
z_offset: ^Huffman_Table;
@@ -559,21 +547,12 @@ inflate_from_stream_raw :: proc(z: ^Context, cb: ^Code_Buffer, expected_output_s
defer free(z_offset);
defer free(codelength_ht);
if cb.window_mask == 0 {
cb.window_mask = DEFLATE_MAX_DISTANCE - 1;
}
// Allocate rolling window buffer.
cb.last = mem.make_dynamic_array_len_cap([dynamic]u8, cb.window_mask + 1, cb.window_mask + 1, allocator);
defer delete(cb.last);
final := u32(0);
type := u32(0);
for {
final = compress.read_bits_lsb(z, cb, 1);
type = compress.read_bits_lsb(z, cb, 2);
final = compress.read_bits_lsb(z, 1);
type = compress.read_bits_lsb(z, 2);
// fmt.printf("Final: %v | Type: %v\n", final, type);
@@ -582,10 +561,10 @@ inflate_from_stream_raw :: proc(z: ^Context, cb: ^Code_Buffer, expected_output_s
// Uncompressed block
// Discard bits until next byte boundary
compress.discard_to_next_byte_lsb(cb);
compress.discard_to_next_byte_lsb(z);
uncompressed_len := i16(compress.read_bits_lsb(z, cb, 16));
length_check := i16(compress.read_bits_lsb(z, cb, 16));
uncompressed_len := i16(compress.read_bits_lsb(z, 16));
length_check := i16(compress.read_bits_lsb(z, 16));
// fmt.printf("LEN: %v, ~LEN: %v, NLEN: %v, ~NLEN: %v\n", uncompressed_len, ~uncompressed_len, length_check, ~length_check);
@@ -599,9 +578,9 @@ inflate_from_stream_raw :: proc(z: ^Context, cb: ^Code_Buffer, expected_output_s
and a single Adler32 update after.
*/
#no_bounds_check for uncompressed_len > 0 {
compress.refill_lsb(z, cb);
lit := compress.read_bits_lsb(z, cb, 8);
write_byte(z, cb, u8(lit));
compress.refill_lsb(z);
lit := compress.read_bits_lsb(z, 8);
write_byte(z, u8(lit));
uncompressed_len -= 1;
}
case 3:
@@ -625,14 +604,14 @@ inflate_from_stream_raw :: proc(z: ^Context, cb: ^Code_Buffer, expected_output_s
//i: u32;
n: u32;
compress.refill_lsb(z, cb, 14);
hlit := compress.read_bits_no_refill_lsb(z, cb, 5) + 257;
hdist := compress.read_bits_no_refill_lsb(z, cb, 5) + 1;
hclen := compress.read_bits_no_refill_lsb(z, cb, 4) + 4;
compress.refill_lsb(z, 14);
hlit := compress.read_bits_no_refill_lsb(z, 5) + 257;
hdist := compress.read_bits_no_refill_lsb(z, 5) + 1;
hclen := compress.read_bits_no_refill_lsb(z, 4) + 4;
ntot := hlit + hdist;
#no_bounds_check for i in 0..<hclen {
s := compress.read_bits_lsb(z, cb, 3);
s := compress.read_bits_lsb(z, 3);
codelength_sizes[Z_LENGTH_DEZIGZAG[i]] = u8(s);
}
err = build_huffman(codelength_ht, codelength_sizes[:]);
@@ -644,7 +623,7 @@ inflate_from_stream_raw :: proc(z: ^Context, cb: ^Code_Buffer, expected_output_s
c: u16;
for n < ntot {
c, err = decode_huffman(z, cb, codelength_ht);
c, err = decode_huffman(z, codelength_ht);
if err != nil {
return err;
}
@@ -657,18 +636,18 @@ inflate_from_stream_raw :: proc(z: ^Context, cb: ^Code_Buffer, expected_output_s
n += 1;
} else {
fill := u8(0);
compress.refill_lsb(z, cb, 7);
compress.refill_lsb(z, 7);
switch c {
case 16:
c = u16(compress.read_bits_no_refill_lsb(z, cb, 2) + 3);
c = u16(compress.read_bits_no_refill_lsb(z, 2) + 3);
if n == 0 {
return E_Deflate.Huffman_Bad_Code_Lengths;
}
fill = lencodes[n - 1];
case 17:
c = u16(compress.read_bits_no_refill_lsb(z, cb, 3) + 3);
c = u16(compress.read_bits_no_refill_lsb(z, 3) + 3);
case 18:
c = u16(compress.read_bits_no_refill_lsb(z, cb, 7) + 11);
c = u16(compress.read_bits_no_refill_lsb(z, 7) + 11);
case:
return E_Deflate.Huffman_Bad_Code_Lengths;
}
@@ -698,7 +677,7 @@ inflate_from_stream_raw :: proc(z: ^Context, cb: ^Code_Buffer, expected_output_s
return err;
}
}
err = parse_huffman_block(z, cb, z_repeat, z_offset);
err = parse_huffman_block(z, z_repeat, z_offset);
// log.debugf("Err: %v | Final: %v | Type: %v\n", err, final, type);
if err != nil {
return err;
@@ -734,7 +713,7 @@ inflate_from_byte_array :: proc(input: []u8, buf: ^bytes.Buffer, raw := false, e
return err;
}
inflate_from_byte_array_raw :: proc(input: []u8, buf: ^bytes.Buffer, cb: ^Code_Buffer, raw := false, expected_output_size := -1) -> (err: Error) {
inflate_from_byte_array_raw :: proc(input: []u8, buf: ^bytes.Buffer, raw := false, expected_output_size := -1) -> (err: Error) {
ctx := Context{};
r := bytes.Reader{};
@@ -746,7 +725,7 @@ inflate_from_byte_array_raw :: proc(input: []u8, buf: ^bytes.Buffer, cb: ^Code_B
ctx.output = buf;
return inflate_from_stream_raw(z=&ctx, cb=cb, expected_output_size=expected_output_size);
return inflate_from_stream_raw(z=&ctx, expected_output_size=expected_output_size);
}
inflate :: proc{inflate_from_stream, inflate_from_byte_array};