/* apply Q-permutation to x */
void Q(hashState *ctx, u8 x[ROWS][COLS1024]) {
u8 i;
Variant v = ctx->columns==8?Q512:Q1024;
#ifndef TURN_OFF_PRINTING
printf(":: BEGIN Q\n");
printf("Input:\n");
PrintState(ctx, x);
#endif
for (i = 0; i < ctx->rounds; i++) {
AddRoundConstant(x, ctx->columns, i, v);
#ifndef TURN_OFF_PRINTING
printf("t=%d (AddRoundConstant):\n", i);
PrintState(ctx, x);
#endif
SubBytes(x, ctx->columns);
#ifndef TURN_OFF_PRINTING
printf("t=%d (SubBytes):\n", i);
PrintState(ctx, x);
#endif
ShiftBytes(x, ctx->columns, v);
#ifndef TURN_OFF_PRINTING
printf("t=%d (ShiftBytes):\n", i);
PrintState(ctx, x);
#endif
MixBytes(x, ctx->columns);
#ifndef TURN_OFF_PRINTING
printf("t=%d (MixBytes):\n", i);
PrintState(ctx, x);
#endif
}
#ifndef TURN_OFF_PRINTING
printf(":: END Q\n\n");
#endif
}
void VP8LFillBitWindow(VP8LBitReader* const br) {
if (br->bit_pos_ >= WBITS) {
#if (defined(__x86_64__) || defined(_M_X64))
if (br->pos_ + sizeof(br->val_) < br->len_) {
br->val_ >>= WBITS;
br->bit_pos_ -= WBITS;
// The expression below needs a little-endian arch to work correctly.
// This gives a large speedup for decoding speed.
br->val_ |= *(const vp8l_val_t*)(br->buf_ + br->pos_) << (LBITS - WBITS);
br->pos_ += LOG8_WBITS;
return;
}
#endif
ShiftBytes(br); // Slow path.
br->eos_ = IsEndOfStreamSpecial(br);
}
uint32_t VP8LReadBits(VP8LBitReader* const br, int n_bits) {
assert(n_bits >= 0);
// Flag an error if end_of_stream or n_bits is more than allowed limit.
if (!br->eos_ && n_bits < MAX_NUM_BIT_READ) {
const uint32_t val =
(uint32_t)(br->val_ >> br->bit_pos_) & kBitMask[n_bits];
const int new_bits = br->bit_pos_ + n_bits;
br->bit_pos_ = new_bits;
// If this read is going to cross the read buffer, set the eos flag.
if (br->pos_ == br->len_) {
if (new_bits >= LBITS) {
br->eos_ = 1;
}
}
ShiftBytes(br);
return val;
} else {
