add remaining encoding config options + presets for x86 and z80

Cargo.lock

@@ -117,7 +117,7 @@ dependencies = [
 
 [[package]]
 name = "upkr"
-version = "0.1.0"
+version = "0.2.0-pre1"
 dependencies = [
  "anyhow",
  "cdivsufsort",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "upkr"
-version = "0.2.0"
+version = "0.2.0-pre1"
 edition = "2021"
 
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html

asm_unpackers/Makefile

@@ -6,10 +6,25 @@ test_riscv64: build/unpack_riscv64
 	qemu-riscv64 $< test_data.upk /tmp/out.bin
 	cmp test_data.bin /tmp/out.bin
 
-build/unpack_riscv64.bin: unpack_riscv.S
+build/unpack_riscv64.o: unpack_riscv.S
 	mkdir -p build
-	riscv64-linux-gnu-gcc -c -o build/unpack_riscv64.o $?
+	riscv64-linux-gnu-gcc -c -o $@ $?
-	riscv64-linux-gnu-objcopy -O binary --only-section=.text build/unpack_riscv64.o $@
+
+build/unpack_riscv64.bin: build/unpack_riscv64.o
+	riscv64-linux-gnu-objcopy -O binary --only-section=.text $? $@
+
+disas-riscv64: build/unpack_riscv64.o
+	riscv64-linux-gnu-objdump -d $?
+
+build/unpack_riscv32.o: unpack_riscv.S
+	mkdir -p build
+	riscv64-linux-gnu-gcc -march=rv32imc -mabi=ilp32 -c -o $@ $?
+	
+build/unpack_riscv32.bin: build/unpack_riscv32.o
+	riscv64-linux-gnu-objcopy -O binary --only-section=.text $? $@
+
+disas-riscv32: build/unpack_riscv32.o
+	riscv64-linux-gnu-objdump -d $?
 
 build/unpack_armv6m: ../c_unpacker/main.c unpack_armv6m.S
 	mkdir -p build
@@ -32,5 +47,5 @@ test_c: build/unpack_c
 	$< test_data.upk /tmp/out.bin
 	cmp test_data.bin /tmp/out.bin
 
-sizes: build/unpack_armv6m.bin build/unpack_riscv64.bin
+sizes: build/unpack_armv6m.bin build/unpack_riscv64.bin build/unpack_riscv32.bin
 	ls -l build/*.bin

asm_unpackers/unpack_riscv.S

@@ -1,6 +1,6 @@
 .section .text
 
-#define FRAME_SIZE (256+64*4+4)
+#define FRAME_SIZE (256+32*4+4)
 
 // x8 prob array ptr
 // x9 prev was literal
@@ -28,39 +28,39 @@ upkr_unpack:
 	jal upkr_decode_bit
 	beqz x15, .Lliteral
 	
-	li x14, 256
+	slli x14, x14, 8
-	beqz x9, .Lread_offset
+	beqz x9, .Lread_offset_inc_x14
 	jal upkr_decode_bit
-	beqz x15, .Lskip_offset
+	bnez x15, .Lread_offset
-.Lread_offset:
-	jal t3, upkr_decode_number
-	addi x12, x9, -1
-	beqz x12, .Ldone
 
-.Lskip_offset:
+.Lfinished_offset:
-	li x14, 256+64
+	addi x14, x14, 64
 	jal t3, upkr_decode_number
 1:
-	sub x15, x10, x12
+	add x14, x10, t0
-	lbu x15, (x15)
+	lbu x14, (x14)
-	sb x15, (x10)
+.Lstore_byte:
+	sb x14, (x10)
 	addi x10, x10, 1
-	addi x9, x9, -1
+	addi x9, x9, 1
-	bnez x9, 1b
+	blt x9, x0, 1b
 	j .Lmainloop
 	
 .Lliteral:
-	li x14, 1
-1:
 	jal upkr_decode_bit
+	addi x14, x14, -1
 	slli x14, x14, 1
 	add x14, x14, x15
 	srli x9, x14, 8
-	beqz x9, 1b
+	beqz x9, .Lliteral
-	sb x14, 0(x10)
+	j .Lstore_byte
-	addi x10, x10, 1
-	j .Lmainloop
 
+.Lread_offset_inc_x14:
+	addi x14, x14, 1
+.Lread_offset:
+	jal t3, upkr_decode_number
+	addi t0, x9, 1
+	bnez t0, .Lfinished_offset
 .Ldone:
 	addi sp, sp, FRAME_SIZE
 	mv x8, x17
@@ -68,16 +68,14 @@ upkr_unpack:
 	jr t4
 
 // x14 context index
-// return: x9 decoded number 
+// return: x9 negtive decoded number 
 upkr_decode_number:
 	mv t5, x14
 	li x9, 0
-	li x8, 1
+	li x8, -1
 1:
-	addi x14, x14, 1
 	jal upkr_decode_bit
 	beqz x15, 1f
-	addi x14, x14, 1
 	jal upkr_decode_bit
 	beqz x15, 2f
 	add x9, x9, x8
@@ -99,46 +97,46 @@ upkr_load_byte:
 // x11 in ptr
 // x13 state
 // x14 context index
-// return: x15 decoded bit
+// return:
+//   x14 context index + 1
+//   x15 decoded bit
 upkr_decode_bit:
 	srli x15, x13, 12
 	beqz x15, upkr_load_byte
 
-	mv t0, x9
 	mv t1, x14
 	mv t2, x10
 	
 	add x14, x14, sp
-	lbu x9, 0(x14)
+	lbu x12, 0(x14)
 	
 	andi x10, x13, 255
-	sltu x15, x10, x9
+	sltu x15, x10, x12
 	srli x13, x13, 8
 	beqz x15, .Lelse
 	
-	mul x13, x13, x9
+	mul x13, x13, x12
 	add x13, x13, x10
 	li x10, 256 + 8
-	sub x10, x10, x9
+	sub x10, x10, x12
 	srli x10, x10, 4
-	add x9, x9, x10
+	add x12, x12, x10
 	j .Lendif
 
 .Lelse:
 	li x16, 256
-	sub x16, x16, x9
+	sub x16, x16, x12
 	mul x13, x13, x16
 	add x13, x13, x10
-	sub x13, x13, x9
+	sub x13, x13, x12
-	addi x10, x9, 8
+	addi x10, x12, 8
 	srli x10, x10, 4
-	sub x9, x9, x10
+	sub x12, x12, x10
 
 .Lendif:
 
-	sb x9, 0(x14)
+	sb x12, 0(x14)
 	
-	mv x9, t0
+	addi x14, t1, 1
-	mv x14, t1
 	mv x10, t2
 	ret

src/context_state.rs

@@ -1,4 +1,7 @@
-use crate::rans::{PROB_BITS, ONE_PROB};
+use crate::{
+    rans::{ONE_PROB, PROB_BITS},
+    Config,
+};
 
 const INIT_PROB: u16 = 1 << (PROB_BITS - 1);
 const UPDATE_RATE: u32 = 4;
@@ -7,6 +10,8 @@ const UPDATE_ADD: u32 = 8;
 #[derive(Clone)]
 pub struct ContextState {
     contexts: Vec<u8>,
+    invert_bit_encoding: bool,
+    simplified_prob_update: bool,
 }
 
 pub struct Context<'a> {
@@ -15,9 +20,11 @@ pub struct Context<'a> {
 }
 
 impl ContextState {
-    pub fn new(size: usize) -> ContextState {
+    pub fn new(size: usize, config: &Config) -> ContextState {
         ContextState {
             contexts: vec![INIT_PROB as u8; size],
+            invert_bit_encoding: config.invert_bit_encoding,
+            simplified_prob_update: config.simplified_prob_update,
         }
     }
 
@@ -33,10 +40,21 @@ impl<'a> Context<'a> {
 
     pub fn update(&mut self, bit: bool) {
         let old = self.state.contexts[self.index];
-        self.state.contexts[self.index] = if bit {
+
+        self.state.contexts[self.index] = if self.state.simplified_prob_update {
+            let offset = if bit ^ self.state.invert_bit_encoding {
+                ONE_PROB as i32 >> UPDATE_RATE
+            } else {
+                0
+            };
+
+            (offset + old as i32 - ((old as i32 + UPDATE_ADD as i32) >> UPDATE_RATE)) as u8
+        } else {
+            if bit ^ self.state.invert_bit_encoding {
                 old + ((ONE_PROB - old as u32 + UPDATE_ADD) >> UPDATE_RATE) as u8
             } else {
                 old - ((old as u32 + UPDATE_ADD) >> UPDATE_RATE) as u8
+            }
         };
     }
 }

src/greedy_packer.rs

@@ -1,17 +1,16 @@
-use crate::lz;
 use crate::match_finder::MatchFinder;
 use crate::rans::RansCoder;
 use crate::ProgressCallback;
+use crate::{lz, Config};
 
 pub fn pack(
     data: &[u8],
-    use_bitstream: bool,
+    config: &Config,
-    parity_contexts: usize,
     mut progress_callback: Option<ProgressCallback>,
 ) -> Vec<u8> {
     let mut match_finder = MatchFinder::new(data);
-    let mut rans_coder = RansCoder::new(use_bitstream);
+    let mut rans_coder = RansCoder::new(config);
-    let mut state = lz::CoderState::new(parity_contexts);
+    let mut state = lz::CoderState::new(config);
 
     let mut pos = 0;
     while pos < data.len() {
@@ -27,7 +26,7 @@ pub fn pack(
                     offset: offset as u32,
                     len: m.length as u32,
                 }
-                .encode(&mut rans_coder, &mut state);
+                .encode(&mut rans_coder, &mut state, config);
                 pos += m.length;
                 encoded_match = true;
             }
@@ -46,7 +45,7 @@ pub fn pack(
                         offset: offset as u32,
                         len: length as u32,
                     }
-                    .encode(&mut rans_coder, &mut state);
+                    .encode(&mut rans_coder, &mut state, config);
                     pos += length;
                     encoded_match = true;
                 }
@@ -54,11 +53,11 @@ pub fn pack(
         }
 
         if !encoded_match {
-            lz::Op::Literal(data[pos]).encode(&mut rans_coder, &mut state);
+            lz::Op::Literal(data[pos]).encode(&mut rans_coder, &mut state, config);
             pos += 1;
         }
     }
 
-    lz::encode_eof(&mut rans_coder, &mut state);
+    lz::encode_eof(&mut rans_coder, &mut state, config);
     rans_coder.finish()
 }

src/lib.rs

@@ -12,6 +12,14 @@ pub type ProgressCallback<'a> = &'a mut dyn FnMut(usize);
 pub struct Config {
     pub use_bitstream: bool,
     pub parity_contexts: usize,
+
+    pub invert_bit_encoding: bool,
+    pub is_match_bit: bool,
+    pub new_offset_bit: bool,
+    pub continue_value_bit: bool,
+
+    pub bitstream_is_big_endian: bool,
+    pub simplified_prob_update: bool,
 }
 
 impl Default for Config {
@@ -19,6 +27,14 @@ impl Default for Config {
         Config {
             use_bitstream: false,
             parity_contexts: 1,
+
+            invert_bit_encoding: false,
+            is_match_bit: true,
+            new_offset_bit: true,
+            continue_value_bit: true,
+
+            bitstream_is_big_endian: false,
+            simplified_prob_update: false,
         }
     }
 }
@@ -30,20 +46,9 @@ pub fn pack(
     progress_callback: Option<ProgressCallback>,
 ) -> Vec<u8> {
     if level == 0 {
-        greedy_packer::pack(
+        greedy_packer::pack(data, &config, progress_callback)
-            data,
-            config.use_bitstream,
-            config.parity_contexts,
-            progress_callback,
-        )
     } else {
-        parsing_packer::pack(
+        parsing_packer::pack(data, level, &config, progress_callback)
-            data,
-            level,
-            config.use_bitstream,
-            config.parity_contexts,
-            progress_callback,
-        )
     }
 }
 

src/lz.rs

@@ -1,5 +1,6 @@
 use crate::context_state::ContextState;
 use crate::rans::{EntropyCoder, RansDecoder};
+use crate::Config;
 
 #[derive(Copy, Clone, Debug)]
 pub enum Op {
@@ -8,11 +9,11 @@ pub enum Op {
 }
 
 impl Op {
-    pub fn encode(&self, coder: &mut dyn EntropyCoder, state: &mut CoderState) {
+    pub fn encode(&self, coder: &mut dyn EntropyCoder, state: &mut CoderState, config: &Config) {
         let literal_base = state.pos % state.parity_contexts * 256;
         match self {
             &Op::Literal(lit) => {
-                encode_bit(coder, state, literal_base, false);
+                encode_bit(coder, state, literal_base, !config.is_match_bit);
                 let mut context_index = 1;
                 for i in (0..8).rev() {
                     let bit = (lit >> i) & 1 != 0;
@@ -23,22 +24,28 @@ impl Op {
                 state.pos += 1;
             }
             &Op::Match { offset, len } => {
-                encode_bit(coder, state, literal_base, true);
+                encode_bit(coder, state, literal_base, config.is_match_bit);
                 if !state.prev_was_match {
                     encode_bit(
                         coder,
                         state,
                         256 * state.parity_contexts,
-                        offset != state.last_offset,
+                        (offset != state.last_offset) == config.new_offset_bit,
                     );
                 } else {
                     assert!(offset != state.last_offset);
                 }
                 if offset != state.last_offset {
-                    encode_length(coder, state, 256 * state.parity_contexts + 1, offset + 1);
+                    encode_length(
+                        coder,
+                        state,
+                        256 * state.parity_contexts + 1,
+                        offset + 1,
+                        config,
+                    );
                     state.last_offset = offset;
                 }
-                encode_length(coder, state, 256 * state.parity_contexts + 65, len);
+                encode_length(coder, state, 256 * state.parity_contexts + 65, len, config);
                 state.prev_was_match = true;
                 state.pos += len as usize;
             }
@@ -46,12 +53,22 @@ impl Op {
     }
 }
 
-pub fn encode_eof(coder: &mut dyn EntropyCoder, state: &mut CoderState) {
+pub fn encode_eof(coder: &mut dyn EntropyCoder, state: &mut CoderState, config: &Config) {
-    encode_bit(coder, state, state.pos % state.parity_contexts * 256, true);
+    encode_bit(
+        coder,
+        state,
+        state.pos % state.parity_contexts * 256,
+        config.is_match_bit,
+    );
     if !state.prev_was_match {
-        encode_bit(coder, state, 256 * state.parity_contexts, true);
+        encode_bit(
+            coder,
+            state,
+            256 * state.parity_contexts,
+            config.new_offset_bit,
+        );
     }
-    encode_length(coder, state, 256 * state.parity_contexts + 1, 1);
+    encode_length(coder, state, 256 * state.parity_contexts + 1, 1, config);
 }
 
 fn encode_bit(
@@ -68,36 +85,37 @@ fn encode_length(
     state: &mut CoderState,
     context_start: usize,
     mut value: u32,
+    config: &Config,
 ) {
     assert!(value >= 1);
 
     let mut context_index = context_start;
     while value >= 2 {
-        encode_bit(coder, state, context_index, true);
+        encode_bit(coder, state, context_index, config.continue_value_bit);
         encode_bit(coder, state, context_index + 1, value & 1 != 0);
         context_index += 2;
         value >>= 1;
     }
-    encode_bit(coder, state, context_index, false);
+    encode_bit(coder, state, context_index, !config.continue_value_bit);
 }
 
 #[derive(Clone)]
 pub struct CoderState {
     contexts: ContextState,
-    parity_contexts: usize,
     last_offset: u32,
     prev_was_match: bool,
     pos: usize,
+    parity_contexts: usize,
 }
 
 impl CoderState {
-    pub fn new(parity_contexts: usize) -> CoderState {
+    pub fn new(config: &Config) -> CoderState {
         CoderState {
-            contexts: ContextState::new((1 + 255) * parity_contexts + 1 + 64 + 64),
+            contexts: ContextState::new((1 + 255) * config.parity_contexts + 1 + 64 + 64, config),
             last_offset: 0,
-            parity_contexts,
             prev_was_match: false,
             pos: 0,
+            parity_contexts: config.parity_contexts,
         }
     }
 
@@ -106,9 +124,9 @@ impl CoderState {
     }
 }
 
-pub fn unpack(packed_data: &[u8], config: crate::Config) -> Vec<u8> {
+pub fn unpack(packed_data: &[u8], config: Config) -> Vec<u8> {
-    let mut decoder = RansDecoder::new(packed_data, config.use_bitstream);
+    let mut decoder = RansDecoder::new(packed_data, &config);
-    let mut contexts = ContextState::new((1 + 255) * config.parity_contexts + 1 + 64 + 64);
+    let mut contexts = ContextState::new((1 + 255) * config.parity_contexts + 1 + 64 + 64, &config);
     let mut result = vec![];
     let mut offset = 0;
     let mut prev_was_match = false;
@@ -117,10 +135,13 @@ pub fn unpack(packed_data: &[u8], config: crate::Config) -> Vec<u8> {
         decoder: &mut RansDecoder,
         contexts: &mut ContextState,
         mut context_index: usize,
+        config: &Config,
     ) -> usize {
         let mut length = 0;
         let mut bit_pos = 0;
-        while decoder.decode_with_context(&mut contexts.context_mut(context_index)) {
+        while decoder.decode_with_context(&mut contexts.context_mut(context_index))
+            == config.continue_value_bit
+        {
             length |= (decoder.decode_with_context(&mut contexts.context_mut(context_index + 1))
                 as usize)
                 << bit_pos;
@@ -132,15 +153,19 @@ pub fn unpack(packed_data: &[u8], config: crate::Config) -> Vec<u8> {
 
     loop {
         let literal_base = result.len() % config.parity_contexts * 256;
-        if decoder.decode_with_context(&mut contexts.context_mut(literal_base)) {
+        if decoder.decode_with_context(&mut contexts.context_mut(literal_base))
+            == config.is_match_bit
+        {
             if prev_was_match
                 || decoder
                     .decode_with_context(&mut contexts.context_mut(256 * config.parity_contexts))
+                    == config.new_offset_bit
             {
                 offset = decode_length(
                     &mut decoder,
                     &mut contexts,
                     256 * config.parity_contexts + 1,
+                    &config,
                 ) - 1;
                 if offset == 0 {
                     break;
@@ -150,6 +175,7 @@ pub fn unpack(packed_data: &[u8], config: crate::Config) -> Vec<u8> {
                 &mut decoder,
                 &mut contexts,
                 256 * config.parity_contexts + 65,
+                &config,
             );
             for _ in 0..length {
                 result.push(result[result.len() - offset]);

src/main.rs

@@ -19,6 +19,28 @@ fn main() -> Result<()> {
             Short('b') | Long("bitstream") => config.use_bitstream = true,
             Short('p') | Long("parity") => config.parity_contexts = parser.value()?.parse()?,
             Short('r') | Long("reverse") => reverse = true,
+            Long("invert-is-match-bit") => config.is_match_bit = false,
+            Long("invert-new-offset-bit") => config.new_offset_bit = false,
+            Long("invert-continue-value-bit") => config.continue_value_bit = false,
+            Long("invert-bit-encoding") => config.invert_bit_encoding = true,
+            Long("simplified-prob-update") => config.simplified_prob_update = true,
+            Long("big-endian-bitstream") => {
+                config.use_bitstream = true;
+                config.bitstream_is_big_endian = true;
+            }
+
+            Long("z80") => {
+                config.use_bitstream = true;
+                config.bitstream_is_big_endian = true;
+                config.invert_bit_encoding = true;
+                config.simplified_prob_update = true;
+            }
+            Long("x86") => {
+                config.use_bitstream = true;
+                config.continue_value_bit = false;
+                config.is_match_bit = false;
+            }
+
             Short('u') | Long("unpack") => unpack = true,
             Short('l') | Long("level") => level = parser.value()?.parse()?,
             Short('h') | Long("help") => print_help(0),
@@ -107,9 +129,21 @@ fn print_help(exit_code: i32) -> ! {
     eprintln!(" -l, --level N       compression level 0-9");
     eprintln!(" -u, --unpack        unpack infile");
     eprintln!();
+    eprintln!("Config presets for specific unpackers:");
+    eprintln!(" --z80               --big-endian-bitstream --invert-bit-encoding --simplified-prob-update");
+    eprintln!(" --x86               --bitstream --invert-is-match-bit --invert-continue-value-bit");
+    eprintln!();
     eprintln!("Config options (need to match when packing/unpacking):");
     eprintln!(" -b, --bitstream     bitstream mode");
     eprintln!(" -p, --parity N      use N (2/4) parity contexts");
     eprintln!(" -r, --reverse       reverse input & output");
+    eprintln!();
+    eprintln!("Config options to tailor output to specific optimized unpackers:");
+    eprintln!(" --invert-is-match-bit");
+    eprintln!(" --invert-new-offset-bit");
+    eprintln!(" --invert-continue-value-bit");
+    eprintln!(" --invert-bit-encoding");
+    eprintln!(" --simplified-prob-update");
+    eprintln!(" --big-endian-bitstream   (implies --bitstream)");
     process::exit(exit_code);
 }

src/parsing_packer.rs

@@ -9,27 +9,21 @@ use crate::{lz, ProgressCallback};
 pub fn pack(
     data: &[u8],
     level: u8,
-    use_bitstream: bool,
+    config: &crate::Config,
-    parity_contexts: usize,
     progress_cb: Option<ProgressCallback>,
 ) -> Vec<u8> {
-    let mut parse = parse(
+    let mut parse = parse(data, Config::from_level(level), config, progress_cb);
-        data,
-        Config::from_level(level),
-        parity_contexts,
-        progress_cb,
-    );
     let mut ops = vec![];
     while let Some(link) = parse {
         ops.push(link.op);
         parse = link.prev.clone();
     }
-    let mut state = lz::CoderState::new(parity_contexts);
+    let mut state = lz::CoderState::new(config);
-    let mut coder = RansCoder::new(use_bitstream);
+    let mut coder = RansCoder::new(config);
     for op in ops.into_iter().rev() {
-        op.encode(&mut coder, &mut state);
+        op.encode(&mut coder, &mut state, config);
     }
-    lz::encode_eof(&mut coder, &mut state);
+    lz::encode_eof(&mut coder, &mut state, config);
     coder.finish()
 }
 
@@ -49,7 +43,7 @@ type Arrivals = HashMap<usize, Vec<Arrival>>;
 fn parse(
     data: &[u8],
     config: Config,
-    parity_contexts: usize,
+    encoding_config: &crate::Config,
     mut progress_cb: Option<ProgressCallback>,
 ) -> Option<Rc<Parse>> {
     let mut match_finder = MatchFinder::new(data)
@@ -114,6 +108,7 @@ fn parse(
         length: usize,
         arrival: &Arrival,
         max_arrivals: usize,
+        config: &crate::Config,
     ) {
         cost_counter.reset();
         let mut state = arrival.state.clone();
@@ -121,7 +116,7 @@ fn parse(
             offset: offset as u32,
             len: length as u32,
         };
-        op.encode(cost_counter, &mut state);
+        op.encode(cost_counter, &mut state, config);
         add_arrival(
             arrivals,
             pos + length,
@@ -141,13 +136,13 @@ fn parse(
         0,
         Arrival {
             parse: None,
-            state: lz::CoderState::new(parity_contexts),
+            state: lz::CoderState::new(encoding_config),
             cost: 0.0,
         },
         max_arrivals,
     );
 
-    let cost_counter = &mut CostCounter::new();
+    let cost_counter = &mut CostCounter::new(encoding_config);
     let mut best_per_offset = HashMap::new();
     for pos in 0..data.len() {
         let match_length = |offset: usize| {
@@ -197,6 +192,7 @@ fn parse(
                     m.length,
                     &arrival,
                     max_arrivals,
+                    encoding_config,
                 );
                 if m.length >= config.greedy_size {
                     break 'arrival_loop;
@@ -220,6 +216,7 @@ fn parse(
                     length,
                     &arrival,
                     max_arrivals,
+                    encoding_config,
                 );
                 found_last_offset |= offset as u32 == arrival.state.last_offset();
                 if offset < near_matches.len() {
@@ -240,6 +237,7 @@ fn parse(
                         length,
                         &arrival,
                         max_arrivals,
+                        encoding_config,
                     );
                 }
             }
@@ -247,7 +245,7 @@ fn parse(
             cost_counter.reset();
             let mut state = arrival.state;
             let op = lz::Op::Literal(data[pos]);
-            op.encode(cost_counter, &mut state);
+            op.encode(cost_counter, &mut state, encoding_config);
             add_arrival(
                 &mut arrivals,
                 pos + 1,

src/rans.rs

@@ -1,4 +1,4 @@
-use crate::context_state::Context;
+use crate::{context_state::Context, Config};
 
 pub const PROB_BITS: u32 = 8;
 pub const ONE_PROB: u32 = 1 << PROB_BITS;
@@ -15,20 +15,25 @@ pub trait EntropyCoder {
 pub struct RansCoder {
     bits: Vec<u16>,
     use_bitstream: bool,
+    bitstream_is_big_endian: bool,
+    invert_bit_encoding: bool,
 }
 
 impl EntropyCoder for RansCoder {
     fn encode_bit(&mut self, bit: bool, prob: u16) {
         assert!(prob < 32768);
-        self.bits.push(prob | ((bit as u16) << 15));
+        self.bits
+            .push(prob | (((bit ^ self.invert_bit_encoding) as u16) << 15));
     }
 }
 
 impl RansCoder {
-    pub fn new(use_bitstream: bool) -> RansCoder {
+    pub fn new(config: &Config) -> RansCoder {
         RansCoder {
             bits: Vec::new(),
-            use_bitstream,
+            use_bitstream: config.use_bitstream,
+            bitstream_is_big_endian: config.bitstream_is_big_endian,
+            invert_bit_encoding: config.invert_bit_encoding,
         }
     }
 
@@ -38,8 +43,20 @@ impl RansCoder {
         let mut state = 1 << l_bits;
 
         let mut byte = 0u8;
-        let mut bit = 8;
+        let mut bit = if self.bitstream_is_big_endian { 0 } else { 8 };
         let mut flush_state: Box<dyn FnMut(&mut u32)> = if self.use_bitstream {
+            if self.bitstream_is_big_endian {
+                Box::new(|state: &mut u32| {
+                    byte |= ((*state & 1) as u8) << bit;
+                    bit += 1;
+                    if bit == 8 {
+                        buffer.push(byte);
+                        byte = 0;
+                        bit = 0;
+                    }
+                    *state >>= 1;
+                })
+            } else {
                 Box::new(|state: &mut u32| {
                     bit -= 1;
                     byte |= ((*state & 1) as u8) << bit;
@@ -50,6 +67,7 @@ impl RansCoder {
                     }
                     *state >>= 1;
                 })
+            }
         } else {
             Box::new(|state: &mut u32| {
                 buffer.push(*state as u8);
@@ -91,10 +109,11 @@ impl RansCoder {
 pub struct CostCounter {
     cost: f64,
     log2_table: Vec<f64>,
+    invert_bit_encoding: bool,
 }
 
 impl CostCounter {
-    pub fn new() -> CostCounter {
+    pub fn new(config: &Config) -> CostCounter {
         let log2_table = (0..ONE_PROB)
             .map(|prob| {
                 let inv_prob = ONE_PROB as f64 / prob as f64;
@@ -104,6 +123,7 @@ impl CostCounter {
         CostCounter {
             cost: 0.0,
             log2_table,
+            invert_bit_encoding: config.invert_bit_encoding,
         }
     }
 
@@ -118,7 +138,7 @@ impl CostCounter {
 
 impl EntropyCoder for CostCounter {
     fn encode_bit(&mut self, bit: bool, prob: u16) {
-        let prob = if bit {
+        let prob = if bit ^ self.invert_bit_encoding {
             prob as u32
         } else {
             ONE_PROB - prob as u32
@@ -133,18 +153,22 @@ pub struct RansDecoder<'a> {
     use_bitstream: bool,
     byte: u8,
     bits_left: u8,
+    invert_bit_encoding: bool,
+    bitstream_is_big_endian: bool,
 }
 
 const PROB_MASK: u32 = ONE_PROB - 1;
 
 impl<'a> RansDecoder<'a> {
-    pub fn new(data: &'a [u8], use_bitstream: bool) -> RansDecoder<'a> {
+    pub fn new(data: &'a [u8], config: &Config) -> RansDecoder<'a> {
         RansDecoder {
             data,
             state: 0,
-            use_bitstream,
+            use_bitstream: config.use_bitstream,
             byte: 0,
             bits_left: 0,
+            invert_bit_encoding: config.invert_bit_encoding,
+            bitstream_is_big_endian: config.bitstream_is_big_endian,
         }
     }
 
@@ -163,8 +187,13 @@ impl<'a> RansDecoder<'a> {
                     self.data = &self.data[1..];
                     self.bits_left = 8;
                 }
+                if self.bitstream_is_big_endian {
+                    self.state = (self.state << 1) | (self.byte >> 7) as u32;
+                    self.byte <<= 1;
+                } else {
                     self.state = (self.state << 1) | (self.byte & 1) as u32;
                     self.byte >>= 1;
+                }
                 self.bits_left -= 1;
             }
         } else {
@@ -183,6 +212,6 @@ impl<'a> RansDecoder<'a> {
         };
         self.state = prob * (self.state >> PROB_BITS) + (self.state & PROB_MASK) - start;
 
-        bit
+        bit ^ self.invert_bit_encoding
     }
 }