mirror of
https://github.com/ddidderr/wuffs.git
synced 2026-11-18 00:32:11 +01:00
std/jpeg: comments for "read Huffman LUT earlier"
This commit is the second of two. It just adds comments to the code added by the previous commit.
This commit is contained in:
1 parent
d9c7740ea6
commit
255516bc21
5 files changed
+74
-6
No files matched your search
@@ -75,7 +75,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
|
||||
{
|
||||
this.mcu_blocks[.. 1].bulk_memset!(byte_value: 0)
|
||||
|
||||
// Load at least 31 bits, by keeping 56 .. 64 loaded.
|
||||
// Load at least 47 bits, by keeping 56 .. 64 loaded.
|
||||
if r.length() < 8 {
|
||||
ret = 1 // Request another fill_bitstream call.
|
||||
break.goto_done
|
||||
@@ -149,8 +149,16 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
|
||||
break.goto_done
|
||||
}
|
||||
|
||||
// Ensure the informal "the lower bound on the number of loaded
|
||||
// bits is at least 16" loop pre-condition below.
|
||||
if n_bits < 16 {
|
||||
bits |= r.peek_u64be() >> (n_bits & 63)
|
||||
// Skip these lines that are normally part of "load more bits".
|
||||
// It's unnecessary, just for the informal (NB > 16) below, and
|
||||
// skipping it avoids a small but negative performance impact.
|
||||
//
|
||||
// r.skip_u32_fast!( etc )
|
||||
// n_bits |= 56
|
||||
}
|
||||
|
||||
z = 1
|
||||
@@ -161,9 +169,32 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
|
||||
inv this.mcu_current_block < 10,
|
||||
inv ac_huff_table_fast <> nullptr,
|
||||
{
|
||||
// Use the high bits of the bits variable to look up the
|
||||
// Huffman table. Conceptually, we should do this *after*
|
||||
// "Load at least 47 bits" below (at "LOOK-UP" further below),
|
||||
// but it is noticably faster to do it beforehand:
|
||||
// https://github.com/google/wuffs/commit/d9c7740ea6c4a450401c393787838436895b87bb
|
||||
//
|
||||
// Earlier look-up is OK because there is a informal loop
|
||||
// pre-condition (informal means not enforced by the Wuffs
|
||||
// compiler) that (NB >= 16) here, where NB is "the number of
|
||||
// loaded bits in the bits variable". It is literally n_bits
|
||||
// for almost all loop iterations, but it can be higher on the
|
||||
// first iteration, set up by "Ensure the informal" above.
|
||||
//
|
||||
// For subsequent iterations, the "Load at least 47 bits" below
|
||||
// does what it says. The Huffman code consumes up to 16 bits
|
||||
// and processing the Huffman symbol's ssss nibble consumes up
|
||||
// to 15 bits. (47 - 16 - 15) = 16 so at the end of the loop
|
||||
// body, (n_bits >= 16).
|
||||
//
|
||||
// This 16 bound has some slack, as (NB >= (64 - 56)) is all we
|
||||
// need, but a bound of 16 (the longest JPEG Huffman code bit
|
||||
// length) won't need to change even if, in the future, we
|
||||
// enlarge the look-up table size.
|
||||
ac_ht_fast = ac_huff_table_fast[bits >> 56] as base.u32
|
||||
|
||||
// Load at least 31 bits, by keeping 56 .. 64 loaded.
|
||||
// Load at least 47 bits, by keeping 56 .. 64 loaded.
|
||||
//
|
||||
// TODO: eliminate this bounds check entirely.
|
||||
if r.length() < 8 {
|
||||
@@ -175,6 +206,9 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
|
||||
n_bits |= 56
|
||||
|
||||
// Read the Huffman-encoded ac_symbol, up to 16 bits long.
|
||||
//
|
||||
// LOOK-UP: "ac_ht_fast = etc" conceptually happens here, just
|
||||
// before deriving ac_bl from ac_ht_fast.
|
||||
ac_bl = ac_ht_fast >> 8
|
||||
if n_bits >= ac_bl {
|
||||
ac_symbol = 0xFF & ac_ht_fast
|
||||
|
||||
@@ -65,8 +65,16 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
|
||||
break.goto_done
|
||||
}
|
||||
|
||||
// Ensure the informal "the lower bound on the number of loaded
|
||||
// bits is at least 16" loop pre-condition below.
|
||||
if n_bits < 16 {
|
||||
bits |= r.peek_u64be() >> (n_bits & 63)
|
||||
// Skip these lines that are normally part of "load more bits".
|
||||
// It's unnecessary, just for the informal (NB > 16) below, and
|
||||
// skipping it avoids a small but negative performance impact.
|
||||
//
|
||||
// r.skip_u32_fast!( etc )
|
||||
// n_bits |= 56
|
||||
}
|
||||
|
||||
z = this.mcu_zig_index
|
||||
@@ -78,9 +86,32 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
|
||||
{
|
||||
assert z < 64 via "a < b: a <= c; c < b"(c: (this.scan_se as base.u32))
|
||||
|
||||
// Use the high bits of the bits variable to look up the
|
||||
// Huffman table. Conceptually, we should do this *after*
|
||||
// "Load at least 47 bits" below (at "LOOK-UP" further below),
|
||||
// but it is noticably faster to do it beforehand:
|
||||
// https://github.com/google/wuffs/commit/d9c7740ea6c4a450401c393787838436895b87bb
|
||||
//
|
||||
// Earlier look-up is OK because there is a informal loop
|
||||
// pre-condition (informal means not enforced by the Wuffs
|
||||
// compiler) that (NB >= 16) here, where NB is "the number of
|
||||
// loaded bits in the bits variable". It is literally n_bits
|
||||
// for almost all loop iterations, but it can be higher on the
|
||||
// first iteration, set up by "Ensure the informal" above.
|
||||
//
|
||||
// For subsequent iterations, the "Load at least 47 bits" below
|
||||
// does what it says. The Huffman code consumes up to 16 bits
|
||||
// and processing the Huffman symbol's ssss nibble consumes up
|
||||
// to 15 bits. (47 - 16 - 15) = 16 so at the end of the loop
|
||||
// body, (n_bits >= 16).
|
||||
//
|
||||
// This 16 bound has some slack, as (NB >= (64 - 56)) is all we
|
||||
// need, but a bound of 16 (the longest JPEG Huffman code bit
|
||||
// length) won't need to change even if, in the future, we
|
||||
// enlarge the look-up table size.
|
||||
ac_ht_fast = ac_huff_table_fast[bits >> 56] as base.u32
|
||||
|
||||
// Load at least 31 bits, by keeping 56 .. 64 loaded.
|
||||
// Load at least 47 bits, by keeping 56 .. 64 loaded.
|
||||
//
|
||||
// TODO: eliminate this bounds check entirely.
|
||||
if r.length() < 8 {
|
||||
@@ -92,6 +123,9 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
|
||||
n_bits |= 56
|
||||
|
||||
// Read the Huffman-encoded ac_symbol, up to 16 bits long.
|
||||
//
|
||||
// LOOK-UP: "ac_ht_fast = etc" conceptually happens here, just
|
||||
// before deriving ac_bl from ac_ht_fast.
|
||||
ac_bl = ac_ht_fast >> 8
|
||||
if n_bits >= ac_bl {
|
||||
ac_symbol = 0xFF & ac_ht_fast
|
||||
|
||||
@@ -62,7 +62,7 @@ pri func decoder.decode_mcu_progressive_ac_low_bit!(workbuf: slice base.u8, mx:
|
||||
while.ac_components true,
|
||||
inv ac_huff_table_fast <> nullptr,
|
||||
{
|
||||
// Load at least 31 bits, by keeping 56 .. 64 loaded (call it 8
|
||||
// Load at least 47 bits, by keeping 56 .. 64 loaded (call it 8
|
||||
// bytes). Just like the other decode_mcu_etc functions, this
|
||||
// is enough for the Huffman-encoded ac_symbol (up to 16 bits)
|
||||
// and processing the ac_value (up to 15 bits).
|
||||
|
||||
@@ -47,7 +47,7 @@ pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx
|
||||
while.dc_component true,
|
||||
inv this.mcu_current_block < 10,
|
||||
{
|
||||
// Load at least 31 bits, by keeping 56 .. 64 loaded.
|
||||
// Load at least 47 bits, by keeping 56 .. 64 loaded.
|
||||
if r.length() < 8 {
|
||||
ret = 1 // Request another fill_bitstream call.
|
||||
break.goto_done
|
||||
|
||||
@@ -39,7 +39,7 @@ pri func decoder.decode_mcu_progressive_dc_low_bit!(workbuf: slice base.u8, mx:
|
||||
while.dc_component true,
|
||||
inv this.mcu_current_block < 10,
|
||||
{
|
||||
// Load at least 31 bits, by keeping 56 .. 64 loaded.
|
||||
// Load at least 47 bits, by keeping 56 .. 64 loaded.
|
||||
if r.length() < 8 {
|
||||
ret = 1 // Request another fill_bitstream call.
|
||||
break.goto_done
|
||||
|
||||
Reference in new issue
Block a user