jessikitty/mikunyan
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add ASTC/HDR support and improve ImageDecoder

Browse Source
This commit is contained in:
Ishotihadus
2019-12-10 04:25:58 +09:00
parent e1ac05e540
commit bd4b10f871
11 changed files with 1745 additions and 622 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ext/decoders/native/astc.c
+791 -345
View File
File diff suppressed because it is too large Load Diff
ext/decoders/native/common.h
+25
View File
@@ -0,0 +1,25 @@
#include <ruby.h>
/* https://github.com/ruby/ruby/blob/master/siphash.c */
#ifdef _WIN32
#define BYTE_ORDER __LITTLE_ENDIAN
#elif !defined BYTE_ORDER
#include <endian.h>
#endif
#ifndef LITTLE_ENDIAN
#define LITTLE_ENDIAN __LITTLE_ENDIAN
#endif
#ifndef BIG_ENDIAN
#define BIG_ENDIAN __BIG_ENDIAN
#endif
#if BYTE_ORDER == LITTLE_ENDIAN
#define IS_LITTLE_ENDIAN 1
#define IS_BIG_ENDIAN 0
#elif BYTE_ORDER == BIG_ENDIAN
#define IS_LITTLE_ENDIAN 0
#define IS_BIG_ENDIAN 1
#else
#error "Only strictly little or big endian supported"
#endif
ext/decoders/native/dxtc.c
+19 -13
View File
@@ -1,12 +1,14 @@
#include "dxtc.h"
#include <stdint.h>
#include <string.h>
#include "dxtc.h"
static inline uint_fast32_t color(uint_fast32_t r, uint_fast32_t g, uint_fast32_t b, uint_fast32_t a) {
static inline uint_fast32_t color(uint_fast32_t r, uint_fast32_t g, uint_fast32_t b, uint_fast32_t a)
{
return r | g << 8 | b << 16 | a << 24;
}
static inline void rgb565(const uint_fast16_t c, int *r, int *g, int *b) {
static inline void rgb565(const uint_fast16_t c, int* r, int* g, int* b)
{
*r = (c & 0xf800) >> 8;
*g = (c & 0x07e0) >> 3;
*b = (c & 0x001f) << 3;
@@ -15,7 +17,8 @@ static inline void rgb565(const uint_fast16_t c, int *r, int *g, int *b) {
*b |= *b >> 5;
}
static inline void decode_dxt1_block(const uint64_t *data, uint32_t *outbuf) {
static inline void decode_dxt1_block(const uint64_t* data, uint32_t* outbuf)
{
int r0, g0, b0, r1, g1, b1;
int q0 = ((uint16_t*)data)[0];
int q1 = ((uint16_t*)data)[1];
@@ -33,12 +36,13 @@ static inline void decode_dxt1_block(const uint64_t *data, uint32_t *outbuf) {
outbuf[i] = c[d & 3];
}
void decode_dxt1(const uint64_t *data, const int w, const int h, uint32_t *image) {
void decode_dxt1(const uint64_t* data, const int w, const int h, uint32_t* image)
{
int bcw = (w + 3) / 4;
int bch = (h + 3) / 4;
int clen_last = (w + 3) % 4 + 1;
uint32_t buf[16];
const uint64_t *d = data;
const uint64_t* d = data;
for (int t = 0; t < bch; t++) {
for (int s = 0; s < bcw; s++, d++) {
decode_dxt1_block(d, buf);
@@ -49,20 +53,21 @@ void decode_dxt1(const uint64_t *data, const int w, const int h, uint32_t *image
}
}
static inline void decode_dxt5_block(const uint64_t *data, uint32_t *outbuf) {
static inline void decode_dxt5_block(const uint64_t* data, uint32_t* outbuf)
{
uint_fast32_t a[8] = { ((uint8_t*)data)[0], ((uint8_t*)data)[1] };
if (a[0] > a[1]) {
a[2] = (a[0] * 6 + a[1] ) / 7;
a[2] = (a[0] * 6 + a[1]) / 7;
a[3] = (a[0] * 5 + a[1] * 2) / 7;
a[4] = (a[0] * 4 + a[1] * 3) / 7;
a[5] = (a[0] * 3 + a[1] * 4) / 7;
a[6] = (a[0] * 2 + a[1] * 5) / 7;
a[7] = (a[0] + a[1] * 6) / 7;
a[7] = (a[0] + a[1] * 6) / 7;
} else {
a[2] = (a[0] * 4 + a[1] ) / 5;
a[2] = (a[0] * 4 + a[1]) / 5;
a[3] = (a[0] * 3 + a[1] * 2) / 5;
a[4] = (a[0] * 2 + a[1] * 3) / 5;
a[5] = (a[0] + a[1] * 4) / 5;
a[5] = (a[0] + a[1] * 4) / 5;
a[7] = 255;
}
for (int i = 0; i < 8; i++)
@@ -87,12 +92,13 @@ static inline void decode_dxt5_block(const uint64_t *data, uint32_t *outbuf) {
outbuf[i] = a[da & 7] | c[dc & 3];
}
void decode_dxt5(const uint64_t *data, const int w, const int h, uint32_t *image) {
void decode_dxt5(const uint64_t* data, const int w, const int h, uint32_t* image)
{
int bcw = (w + 3) / 4;
int bch = (h + 3) / 4;
int clen_last = (w + 3) % 4 + 1;
uint32_t buf[16];
const uint64_t *d = data;
const uint64_t* d = data;
for (int t = 0; t < bch; t++) {
for (int s = 0; s < bcw; s++, d += 2) {
decode_dxt5_block(d, buf);
ext/decoders/native/etc.c
+157 -139
View File
@@ -1,52 +1,63 @@
#include "etc.h"
#include "common.h"
#include <stdint.h>
#include <string.h>
#include "etc.h"
uint_fast8_t WriteOrderTable[16] = { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 };
uint_fast8_t WriteOrderTableRev[16] = { 15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0 };
uint_fast8_t Etc1ModifierTable[8][2] = {{2, 8}, {5, 17}, {9, 29}, {13, 42}, {18, 60}, {24, 80}, {33, 106}, {47, 183}};
uint_fast8_t Etc1SubblockTable[2][16] = {{0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1}, {0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1}};
uint_fast8_t Etc2DistanceTable[8] = {3, 6, 11, 16, 23, 32, 41, 64};
int_fast8_t Etc2AlphaModTable[16][8] = {
{-3, -6, -9, -15, 2, 5, 8, 14},
{-3, -7, -10, -13, 2, 6, 9, 12},
{-2, -5, -8, -13, 1, 4, 7, 12},
{-2, -4, -6, -13, 1, 3, 5, 12},
{-3, -6, -8, -12, 2, 5, 7, 11},
{-3, -7, -9, -11, 2, 6, 8, 10},
{-4, -7, -8, -11, 3, 6, 7, 10},
{-3, -5, -8, -11, 2, 4, 7, 10},
{-2, -6, -8, -10, 1, 5, 7, 9},
{-2, -5, -8, -10, 1, 4, 7, 9},
{-2, -4, -8, -10, 1, 3, 7, 9},
{-2, -5, -7, -10, 1, 4, 6, 9},
{-3, -4, -7, -10, 2, 3, 6, 9},
{-1, -2, -3, -10, 0, 1, 2, 9},
{-4, -6, -8, -9, 3, 5, 7, 8},
{-3, -5, -7, -9, 2, 4, 6, 8}
const uint_fast8_t WriteOrderTable[16] = { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 };
const uint_fast8_t WriteOrderTableRev[16] = { 15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0 };
const uint_fast8_t Etc1ModifierTable[8][2] = { { 2, 8 }, { 5, 17 }, { 9, 29 }, { 13, 42 }, { 18, 60 }, { 24, 80 }, { 33, 106 }, { 47, 183 } };
const uint_fast8_t Etc1SubblockTable[2][16] = { { 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1 }, { 0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1 } };
const uint_fast8_t Etc2DistanceTable[8] = { 3, 6, 11, 16, 23, 32, 41, 64 };
const int_fast8_t Etc2AlphaModTable[16][8] = {
{ -3, -6, -9, -15, 2, 5, 8, 14 },
{ -3, -7, -10, -13, 2, 6, 9, 12 },
{ -2, -5, -8, -13, 1, 4, 7, 12 },
{ -2, -4, -6, -13, 1, 3, 5, 12 },
{ -3, -6, -8, -12, 2, 5, 7, 11 },
{ -3, -7, -9, -11, 2, 6, 8, 10 },
{ -4, -7, -8, -11, 3, 6, 7, 10 },
{ -3, -5, -8, -11, 2, 4, 7, 10 },
{ -2, -6, -8, -10, 1, 5, 7, 9 },
{ -2, -5, -8, -10, 1, 4, 7, 9 },
{ -2, -4, -8, -10, 1, 3, 7, 9 },
{ -2, -5, -7, -10, 1, 4, 6, 9 },
{ -3, -4, -7, -10, 2, 3, 6, 9 },
{ -1, -2, -3, -10, 0, 1, 2, 9 },
{ -4, -6, -8, -9, 3, 5, 7, 8 },
{ -3, -5, -7, -9, 2, 4, 6, 8 }
};
static inline uint_fast32_t color(uint_fast32_t r, uint_fast32_t g, uint_fast32_t b, uint_fast32_t a) {
static inline uint_fast32_t color(uint_fast8_t r, uint_fast8_t g, uint_fast8_t b, uint_fast8_t a)
{
#if BYTE_ORDER == LITTLE_ENDIAN
return r | g << 8 | b << 16 | a << 24;
#else
return a | b << 8 | g << 16 | r << 24;
#endif
}
static inline uint_fast8_t clamp(const int n) {
static inline uint_fast8_t clamp(const int n)
{
return n < 0 ? 0 : n > 255 ? 255 : n;
}
static inline uint32_t applicate_color(uint_fast8_t c[3], int_fast16_t m) {
static inline uint32_t applicate_color(uint_fast8_t c[3], int_fast16_t m)
{
return color(clamp(c[0] + m), clamp(c[1] + m), clamp(c[2] + m), 255);
}
static inline uint32_t applicate_color_raw(uint_fast8_t c[3]) {
static inline uint32_t applicate_color_raw(uint_fast8_t c[3])
{
return color(c[0], c[1], c[2], 255);
}
static inline void decode_etc1_block(const uint8_t *data, uint32_t *outbuf) {
uint_fast8_t code[2] = { data[3] >> 5, data[3] >> 2 & 7 };
uint_fast8_t *table = Etc1SubblockTable[data[3] & 1];
static inline void decode_etc1_block(const uint8_t* data, uint32_t* outbuf)
{
const uint_fast8_t code[2] = { data[3] >> 5, data[3] >> 2 & 7 }; // Table codewords
const uint_fast8_t* table = Etc1SubblockTable[data[3] & 1];
uint_fast8_t c[2][3];
if (data[3] & 2) {
// diff bit == 1
c[0][0] = data[0] & 0xf8;
c[0][1] = data[1] & 0xf8;
c[0][2] = data[2] & 0xf8;
@@ -60,16 +71,17 @@ static inline void decode_etc1_block(const uint8_t *data, uint32_t *outbuf) {
c[1][1] |= c[1][1] >> 5;
c[1][2] |= c[1][2] >> 5;
} else {
c[0][0] = data[0] & 0xf0 | data[0] >> 4;
c[1][0] = data[0] & 0x0f | data[0] << 4;
c[0][1] = data[1] & 0xf0 | data[1] >> 4;
c[1][1] = data[1] & 0x0f | data[1] << 4;
c[0][2] = data[2] & 0xf0 | data[2] >> 4;
c[1][2] = data[2] & 0x0f | data[2] << 4;
// diff bit == 0
c[0][0] = (data[0] & 0xf0) | data[0] >> 4;
c[1][0] = (data[0] & 0x0f) | data[0] << 4;
c[0][1] = (data[1] & 0xf0) | data[1] >> 4;
c[1][1] = (data[1] & 0x0f) | data[1] << 4;
c[0][2] = (data[2] & 0xf0) | data[2] >> 4;
c[1][2] = (data[2] & 0x0f) | data[2] << 4;
}
uint_fast16_t j = data[6] << 8 | data[7];
uint_fast16_t k = data[4] << 8 | data[5];
uint_fast16_t j = data[6] << 8 | data[7]; // less significant pixel index bits
uint_fast16_t k = data[4] << 8 | data[5]; // more significant pixel index bits
for (int i = 0; i < 16; i++, j >>= 1, k >>= 1) {
uint_fast8_t s = table[i];
uint_fast8_t m = Etc1ModifierTable[code[s]][j & 1];
@@ -77,28 +89,33 @@ static inline void decode_etc1_block(const uint8_t *data, uint32_t *outbuf) {
}
}
void decode_etc1(const void *data, const int w, const int h, uint32_t *image) {
int bcw = (w + 3) / 4;
int bch = (h + 3) / 4;
int clen_last = (w + 3) % 4 + 1;
void decode_etc1(const void* data, const int w, const int h, uint32_t* image)
{
int num_blocks_x = (w + 3) / 4;
int num_blocks_y = (h + 3) / 4;
int copy_length_last = (w + 3) % 4 + 1;
uint32_t buf[16];
const uint8_t *d = (uint8_t*)data;
for (int t = 0; t < bch; t++) {
for (int s = 0; s < bcw; s++, d += 8) {
uint32_t* buf_end = buf + 16;
const uint8_t* d = (uint8_t*)data;
for (int by = 0; by < num_blocks_y; by++) {
for (int bx = 0, x = 0; bx < num_blocks_x; bx++, d += 8, x += 4) {
decode_etc1_block(d, buf);
int clen = (s < bcw - 1 ? 4 : clen_last) * 4;
for (int i = 0, y = h - t * 4 - 1; i < 4 && y >= 0; i++, y--)
memcpy(image + y * w + s * 4, buf + i * 4, clen);
int copy_length = (bx < num_blocks_x - 1 ? 4 : copy_length_last) * 4;
uint32_t* b = buf;
for (int y = h - 1 - by * 4; b < buf_end && y >= 0; y--, b += 4)
memcpy(image + y * w + x, b, copy_length);
}
}
}
static inline void decode_etc2_block(const uint8_t *data, uint32_t *outbuf) {
uint_fast16_t j = data[6] << 8 | data[7];
uint_fast16_t k = data[4] << 8 | data[5];
static inline void decode_etc2_block(const uint8_t* data, uint32_t* outbuf)
{
uint_fast16_t j = data[6] << 8 | data[7]; // 15 -> 0
uint_fast32_t k = data[4] << 8 | data[5]; // 31 -> 16
uint_fast8_t c[3][3] = {};
if (data[3] & 2) {
// diff bit == 1
uint_fast8_t r = data[0] & 0xf8;
int_fast16_t dr = (data[0] << 3 & 0x18) - (data[0] << 3 & 0x20);
uint_fast8_t g = data[1] & 0xf8;
@@ -107,57 +124,49 @@ static inline void decode_etc2_block(const uint8_t *data, uint32_t *outbuf) {
int_fast16_t db = (data[2] << 3 & 0x18) - (data[2] << 3 & 0x20);
if (r + dr < 0 || r + dr > 255) {
// T
c[0][0] = data[0] << 3 & 0xc0 | data[0] << 4 & 0x30 | data[0] >> 1 & 0xc | data[0] & 3;
c[0][1] = data[1] & 0xf0 | data[1] >> 4;
c[0][2] = data[1] & 0x0f | data[1] << 4;
c[1][0] = data[2] & 0xf0 | data[2] >> 4;
c[1][1] = data[2] & 0x0f | data[2] << 4;
c[1][2] = data[3] & 0xf0 | data[3] >> 4;
uint_fast8_t d = Etc2DistanceTable[data[3] >> 1 & 6 | data[3] & 1];
uint_fast32_t color_set[4] = {
applicate_color_raw(c[0]),
applicate_color(c[1], d),
applicate_color_raw(c[1]),
applicate_color(c[1], -d)
};
c[0][0] = (data[0] << 3 & 0xc0) | (data[0] << 4 & 0x30) | (data[0] >> 1 & 0xc) | (data[0] & 3);
c[0][1] = (data[1] & 0xf0) | data[1] >> 4;
c[0][2] = (data[1] & 0x0f) | data[1] << 4;
c[1][0] = (data[2] & 0xf0) | data[2] >> 4;
c[1][1] = (data[2] & 0x0f) | data[2] << 4;
c[1][2] = (data[3] & 0xf0) | data[3] >> 4;
const uint_fast8_t d = Etc2DistanceTable[(data[3] >> 1 & 6) | (data[3] & 1)];
uint_fast32_t color_set[4] = { applicate_color_raw(c[0]), applicate_color(c[1], d), applicate_color_raw(c[1]), applicate_color(c[1], -d) };
k <<= 1;
for (int i = 0; i < 16; i++, j >>= 1, k >>= 1)
outbuf[WriteOrderTable[i]] = color_set[k << 1 & 2 | j & 1];
outbuf[WriteOrderTable[i]] = color_set[(k & 2) | (j & 1)];
} else if (g + dg < 0 || g + dg > 255) {
// H
c[0][0] = data[0] << 1 & 0xf0 | data[0] >> 3 & 0xf;
c[0][1] = data[0] << 5 & 0xe0 | data[1] & 0x10;
c[0][0] = (data[0] << 1 & 0xf0) | (data[0] >> 3 & 0xf);
c[0][1] = (data[0] << 5 & 0xe0) | (data[1] & 0x10);
c[0][1] |= c[0][1] >> 4;
c[0][2] = data[1] & 8 | data[1] << 1 & 6 | data[2] >> 7;
c[0][2] = (data[1] & 8) | (data[1] << 1 & 6) | data[2] >> 7;
c[0][2] |= c[0][2] << 4;
c[1][0] = data[2] << 1 & 0xf0 | data[2] >> 3 & 0xf;
c[1][1] = data[2] << 5 & 0xe0 | data[3] >> 3 & 0x10;
c[1][0] = (data[2] << 1 & 0xf0) | (data[2] >> 3 & 0xf);
c[1][1] = (data[2] << 5 & 0xe0) | (data[3] >> 3 & 0x10);
c[1][1] |= c[1][1] >> 4;
c[1][2] = data[3] << 1 & 0xf0 | data[3] >> 3 & 0xf;
uint_fast8_t d = data[3] & 4 | data[3] << 1 & 2;
c[1][2] = (data[3] << 1 & 0xf0) | (data[3] >> 3 & 0xf);
uint_fast8_t d = (data[3] & 4) | (data[3] << 1 & 2);
if (c[0][0] > c[1][0] || (c[0][0] == c[1][0] && (c[0][1] > c[1][1] || (c[0][1] == c[1][1] && c[0][2] >= c[1][2]))))
++d;
d = Etc2DistanceTable[d];
uint_fast32_t color_set[4] = {
applicate_color(c[0], d),
applicate_color(c[0], -d),
applicate_color(c[1], d),
applicate_color(c[1], -d)
};
uint_fast32_t color_set[4] = { applicate_color(c[0], d), applicate_color(c[0], -d), applicate_color(c[1], d), applicate_color(c[1], -d) };
k <<= 1;
for (int i = 0; i < 16; i++, j >>= 1, k >>= 1)
outbuf[WriteOrderTable[i]] = color_set[k << 1 & 2 | j & 1];
outbuf[WriteOrderTable[i]] = color_set[(k & 2) | (j & 1)];
} else if (b + db < 0 || b + db > 255) {
// planar
c[0][0] = data[0] << 1 & 0xfc | data[0] >> 5 & 3;
c[0][1] = data[0] << 7 & 0x80 | data[1] & 0x7e | data[0] & 1;
c[0][2] = data[1] << 7 & 0x80 | data[2] << 2 & 0x60 | data[2] << 3 & 0x18 | data[3] >> 5 & 4;
c[0][0] = (data[0] << 1 & 0xfc) | (data[0] >> 5 & 3);
c[0][1] = (data[0] << 7 & 0x80) | (data[1] & 0x7e) | (data[0] & 1);
c[0][2] = (data[1] << 7 & 0x80) | (data[2] << 2 & 0x60) | (data[2] << 3 & 0x18) | (data[3] >> 5 & 4);
c[0][2] |= c[0][2] >> 6;
c[1][0] = data[3] << 1 & 0xf8 | data[3] << 2 & 4 | data[3] >> 5 & 3;
c[1][1] = data[4] & 0xfe | data[4] >> 7;
c[1][2] = data[4] << 7 & 0x80 | data[5] >> 1 & 0x7c;
c[1][0] = (data[3] << 1 & 0xf8) | (data[3] << 2 & 4) | (data[3] >> 5 & 3);
c[1][1] = (data[4] & 0xfe) | data[4] >> 7;
c[1][2] = (data[4] << 7 & 0x80) | (data[5] >> 1 & 0x7c);
c[1][2] |= c[1][2] >> 6;
c[2][0] = data[5] << 5 & 0xe0 | data[6] >> 3 & 0x1c | data[5] >> 1 & 3;
c[2][1] = data[6] << 3 & 0xf8 | data[7] >> 5 & 0x6 | data[6] >> 4 & 1;
c[2][2] = data[7] << 2 | data[7] >> 4 & 3;
c[2][0] = (data[5] << 5 & 0xe0) | (data[6] >> 3 & 0x1c) | (data[5] >> 1 & 3);
c[2][1] = (data[6] << 3 & 0xf8) | (data[7] >> 5 & 0x6) | (data[6] >> 4 & 1);
c[2][2] = data[7] << 2 | (data[7] >> 4 & 3);
for (int y = 0, i = 0; y < 4; y++) {
for (int x = 0; x < 4; x++, i++) {
uint8_t r = clamp((x * (c[1][0] - c[0][0]) + y * (c[2][0] - c[0][0]) + 4 * c[0][0] + 2) >> 2);
@@ -168,8 +177,8 @@ static inline void decode_etc2_block(const uint8_t *data, uint32_t *outbuf) {
}
} else {
// differential
uint_fast8_t code[2] = { data[3] >> 5, data[3] >> 2 & 7 };
uint_fast8_t *table = Etc1SubblockTable[data[3] & 1];
const uint_fast8_t code[2] = { data[3] >> 5, data[3] >> 2 & 7 };
const uint_fast8_t* table = Etc1SubblockTable[data[3] & 1];
c[0][0] = r | r >> 5;
c[0][1] = g | g >> 5;
c[0][2] = b | b >> 5;
@@ -186,15 +195,15 @@ static inline void decode_etc2_block(const uint8_t *data, uint32_t *outbuf) {
}
}
} else {
// individual
uint_fast8_t code[2] = { data[3] >> 5, data[3] >> 2 & 7 };
uint_fast8_t *table = Etc1SubblockTable[data[3] & 1];
c[0][0] = data[0] & 0xf0 | data[0] >> 4;
c[1][0] = data[0] & 0x0f | data[0] << 4;
c[0][1] = data[1] & 0xf0 | data[1] >> 4;
c[1][1] = data[1] & 0x0f | data[1] << 4;
c[0][2] = data[2] & 0xf0 | data[2] >> 4;
c[1][2] = data[2] & 0x0f | data[2] << 4;
// individual (diff bit == 0)
const uint_fast8_t code[2] = { data[3] >> 5, data[3] >> 2 & 7 };
const uint_fast8_t* table = Etc1SubblockTable[data[3] & 1];
c[0][0] = (data[0] & 0xf0) | data[0] >> 4;
c[1][0] = (data[0] & 0x0f) | data[0] << 4;
c[0][1] = (data[1] & 0xf0) | data[1] >> 4;
c[1][1] = (data[1] & 0x0f) | data[1] << 4;
c[0][2] = (data[2] & 0xf0) | data[2] >> 4;
c[1][2] = (data[2] & 0x0f) | data[2] << 4;
for (int i = 0; i < 16; i++, j >>= 1, k >>= 1) {
uint_fast8_t s = table[i];
uint_fast8_t m = Etc1ModifierTable[code[s]][j & 1];
@@ -203,69 +212,78 @@ static inline void decode_etc2_block(const uint8_t *data, uint32_t *outbuf) {
}
}
static inline void decode_etc2a8_block(const uint8_t *data, uint32_t *outbuf) {
static inline void decode_etc2a8_block(const uint8_t* data, uint32_t* outbuf)
{
if (data[1] & 0xf0) {
uint_fast8_t mult = data[1] >> 4;
int_fast8_t *table = Etc2AlphaModTable[data[1] & 0xf];
uint_fast64_t l =
data[7] | (uint_fast16_t)data[6] << 8 |
(uint_fast32_t)data[5] << 16 | (uint_fast32_t)data[4] << 24 |
(uint_fast64_t)data[3] << 32 | (uint_fast64_t)data[2] << 40;
// multiplier != 0
const uint_fast8_t multiplier = data[1] >> 4;
const int_fast8_t* table = Etc2AlphaModTable[data[1] & 0xf];
uint_fast64_t l = data[7] | (uint_fast16_t)data[6] << 8 | (uint_fast32_t)data[5] << 16 | (uint_fast32_t)data[4] << 24 | (uint_fast64_t)data[3] << 32 | (uint_fast64_t)data[2] << 40;
for (int i = 0; i < 16; i++, l >>= 3)
((uint8_t*)(outbuf + WriteOrderTableRev[i]))[3] = clamp(data[0] + mult * table[l & 7]);
((uint8_t*)(outbuf + WriteOrderTableRev[i]))[3] = clamp(data[0] + multiplier * table[l & 7]);
} else {
for (int i = 0; i < 16; i++)
((uint8_t*)(outbuf + i))[3] = data[0];
// multiplier == 0 (always same as base codeword)
for (int i = 0; i < 16; i++, outbuf++)
((uint8_t*)outbuf)[3] = data[0];
}
}
void decode_etc2(const void *data, const int w, const int h, uint32_t *image) {
int bcw = (w + 3) / 4;
int bch = (h + 3) / 4;
int clen_last = (w + 3) % 4 + 1;
void decode_etc2(const void* data, const int w, const int h, uint32_t* image)
{
int num_blocks_x = (w + 3) / 4;
int num_blocks_y = (h + 3) / 4;
int copy_length_last = (w + 3) % 4 + 1;
uint32_t buf[16];
const uint8_t *d = (uint8_t*)data;
for (int t = 0; t < bch; t++) {
for (int s = 0; s < bcw; s++, d += 8) {
uint32_t* buf_end = buf + 16;
const uint8_t* d = (uint8_t*)data;
for (int by = 0; by < num_blocks_y; by++) {
for (int bx = 0, x = 0; bx < num_blocks_x; bx++, d += 8, x += 4) {
decode_etc2_block(d, buf);
int clen = (s < bcw - 1 ? 4 : clen_last) * 4;
for (int i = 0, y = h - t * 4 - 1; i < 4 && y >= 0; i++, y--)
memcpy(image + y * w + s * 4, buf + i * 4, clen);
int copy_length = (bx < num_blocks_x - 1 ? 4 : copy_length_last) * 4;
uint32_t* b = buf;
for (int y = h - by * 4 - 1; b < buf_end && y >= 0; y--, b += 4)
memcpy(image + y * w + x, b, copy_length);
}
}
}
void decode_etc2a1(const void *data, const int w, const int h, uint32_t *image) {
int bcw = (w + 3) / 4;
int bch = (h + 3) / 4;
int clen_last = (w + 3) % 4 + 1;
void decode_etc2a1(const void* data, const int w, const int h, uint32_t* image)
{
int num_blocks_x = (w + 3) / 4;
int num_blocks_y = (h + 3) / 4;
int copy_length_last = (w + 3) % 4 + 1;
uint32_t buf[16];
const uint8_t *d = (uint8_t*)data;
for (int t = 0; t < bch; t++) {
for (int s = 0; s < bcw; s++, d += 9) {
uint32_t* buf_end = buf + 16;
const uint8_t* d = (uint8_t*)data;
for (int by = 0; by < num_blocks_y; by++) {
for (int bx = 0, x = 0; bx < num_blocks_x; bx++, d += 9, x += 4) {
decode_etc2_block(d + 1, buf);
for (int i = 0; i < 16; i++)
((uint8_t*)(buf + i))[3] = d[0];
int clen = (s < bcw - 1 ? 4 : clen_last) * 4;
for (int i = 0, y = h - t * 4 - 1; i < 4 && y >= 0; i++, y--)
memcpy(image + y * w + s * 4, buf + i * 4, clen);
int copy_length = (bx < num_blocks_x - 1 ? 4 : copy_length_last) * 4;
uint32_t* b = buf;
for (int y = h - by * 4 - 1; b < buf_end && y >= 0; y--, b += 4)
memcpy(image + y * w + x, b, copy_length);
}
}
}
void decode_etc2a8(const void *data, const int w, const int h, uint32_t *image) {
int bcw = (w + 3) / 4;
int bch = (h + 3) / 4;
int clen_last = (w + 3) % 4 + 1;
void decode_etc2a8(const void* data, const int w, const int h, uint32_t* image)
{
int num_blocks_x = (w + 3) / 4;
int num_blocks_y = (h + 3) / 4;
int copy_length_last = (w + 3) % 4 + 1;
uint32_t buf[16];
const uint8_t *d = (uint8_t*)data;
for (int t = 0; t < bch; t++) {
for (int s = 0; s < bcw; s++, d += 16) {
uint32_t* buf_end = buf + 16;
const uint8_t* d = (uint8_t*)data;
for (int by = 0; by < num_blocks_y; by++) {
for (int bx = 0, x = 0; bx < num_blocks_x; bx++, d += 16, x += 4) {
decode_etc2_block(d + 8, buf);
decode_etc2a8_block(d, buf);
int clen = (s < bcw - 1 ? 4 : clen_last) * 4;
for (int i = 0, y = h - t * 4 - 1; i < 4 && y >= 0; i++, y--)
memcpy(image + y * w + s * 4, buf + i * 4, clen);
int copy_length = (bx < num_blocks_x - 1 ? 4 : copy_length_last) * 4;
uint32_t* b = buf;
for (int y = h - by * 4 - 1; b < buf_end && y >= 0; y--, b += 4)
memcpy(image + y * w + x, b, copy_length);
}
}
}
ext/decoders/native/fp16.h
+40
View File
@@ -0,0 +1,40 @@
#pragma once
#ifndef FP16_H
#define FP16_H
#include "fp16/fp16.h"
#endif /* FP16_H */
/**
*
* License Information
*
* FP16 library is derived from https://github.com/Maratyszcza/FP16.
* The library is licensed under the MIT License shown below.
The MIT License (MIT)
Copyright (c) 2017 Facebook Inc.
Copyright (c) 2017 Georgia Institute of Technology
Copyright 2019 Google LLC
Permission is hereby granted, free of charge, to any person obtaining a copy of
this software and associated documentation files (the "Software"), to deal in
the Software without restriction, including without limitation the rights to
use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
of the Software, and to permit persons to whom the Software is furnished to do
so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
**/
ext/decoders/native/fp16/bitcasts.h
+76
View File
@@ -0,0 +1,76 @@
#pragma once
#ifndef FP16_BITCASTS_H
#define FP16_BITCASTS_H
#if defined(__cplusplus) && (__cplusplus >= 201103L)
#include <cstdint>
#elif !defined(__OPENCL_VERSION__)
#include <stdint.h>
#endif
static inline float fp32_from_bits(uint32_t w) {
#if defined(__OPENCL_VERSION__)
return as_float(w);
#elif defined(__CUDA_ARCH__)
return __uint_as_float((unsigned int) w);
#elif defined(__INTEL_COMPILER)
return _castu32_f32(w);
#else
union {
uint32_t as_bits;
float as_value;
} fp32 = { w };
return fp32.as_value;
#endif
}
static inline uint32_t fp32_to_bits(float f) {
#if defined(__OPENCL_VERSION__)
return as_uint(f);
#elif defined(__CUDA_ARCH__)
return (uint32_t) __float_as_uint(f);
#elif defined(__INTEL_COMPILER)
return _castf32_u32(f);
#else
union {
float as_value;
uint32_t as_bits;
} fp32 = { f };
return fp32.as_bits;
#endif
}
static inline double fp64_from_bits(uint64_t w) {
#if defined(__OPENCL_VERSION__)
return as_double(w);
#elif defined(__CUDA_ARCH__)
return __longlong_as_double((long long) w);
#elif defined(__INTEL_COMPILER)
return _castu64_f64(w);
#else
union {
uint64_t as_bits;
double as_value;
} fp64 = { w };
return fp64.as_value;
#endif
}
static inline uint64_t fp64_to_bits(double f) {
#if defined(__OPENCL_VERSION__)
return as_ulong(f);
#elif defined(__CUDA_ARCH__)
return (uint64_t) __double_as_longlong(f);
#elif defined(__INTEL_COMPILER)
return _castf64_u64(f);
#else
union {
double as_value;
uint64_t as_bits;
} fp64 = { f };
return fp64.as_bits;
#endif
}
#endif /* FP16_BITCASTS_H */
ext/decoders/native/fp16/fp16.h
+451
View File
@@ -0,0 +1,451 @@
#pragma once
#ifndef FP16_FP16_H
#define FP16_FP16_H
#if defined(__cplusplus) && (__cplusplus >= 201103L)
#include <cstdint>
#include <cmath>
#elif !defined(__OPENCL_VERSION__)
#include <stdint.h>
#include <math.h>
#endif
#ifdef _MSC_VER
#include <intrin.h>
#endif
#include "fp16/bitcasts.h"
/*
* Convert a 16-bit floating-point number in IEEE half-precision format, in bit representation, to
* a 32-bit floating-point number in IEEE single-precision format, in bit representation.
*
* @note The implementation doesn't use any floating-point operations.
*/
static inline uint32_t fp16_ieee_to_fp32_bits(uint16_t h) {
/*
* Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
* +---+-----+------------+-------------------+
* | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 31 26-30 16-25 0-15
*
* S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.
*/
const uint32_t w = (uint32_t) h << 16;
/*
* Extract the sign of the input number into the high bit of the 32-bit word:
*
* +---+----------------------------------+
* | S |0000000 00000000 00000000 00000000|
* +---+----------------------------------+
* Bits 31 0-31
*/
const uint32_t sign = w & UINT32_C(0x80000000);
/*
* Extract mantissa and biased exponent of the input number into the bits 0-30 of the 32-bit word:
*
* +---+-----+------------+-------------------+
* | 0 |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 30 27-31 17-26 0-16
*/
const uint32_t nonsign = w & UINT32_C(0x7FFFFFFF);
/*
* Renorm shift is the number of bits to shift mantissa left to make the half-precision number normalized.
* If the initial number is normalized, some of its high 6 bits (sign == 0 and 5-bit exponent) equals one.
* In this case renorm_shift == 0. If the number is denormalize, renorm_shift > 0. Note that if we shift
* denormalized nonsign by renorm_shift, the unit bit of mantissa will shift into exponent, turning the
* biased exponent into 1, and making mantissa normalized (i.e. without leading 1).
*/
#ifdef _MSC_VER
unsigned long nonsign_bsr;
_BitScanReverse(&nonsign_bsr, (unsigned long) nonsign);
uint32_t renorm_shift = (uint32_t) nonsign_bsr ^ 31;
#else
uint32_t renorm_shift = __builtin_clz(nonsign);
#endif
renorm_shift = renorm_shift > 5 ? renorm_shift - 5 : 0;
/*
* Iff half-precision number has exponent of 15, the addition overflows it into bit 31,
* and the subsequent shift turns the high 9 bits into 1. Thus
* inf_nan_mask ==
* 0x7F800000 if the half-precision number had exponent of 15 (i.e. was NaN or infinity)
* 0x00000000 otherwise
*/
const int32_t inf_nan_mask = ((int32_t) (nonsign + 0x04000000) >> 8) & INT32_C(0x7F800000);
/*
* Iff nonsign is 0, it overflows into 0xFFFFFFFF, turning bit 31 into 1. Otherwise, bit 31 remains 0.
* The signed shift right by 31 broadcasts bit 31 into all bits of the zero_mask. Thus
* zero_mask ==
* 0xFFFFFFFF if the half-precision number was zero (+0.0h or -0.0h)
* 0x00000000 otherwise
*/
const int32_t zero_mask = (int32_t) (nonsign - 1) >> 31;
/*
* 1. Shift nonsign left by renorm_shift to normalize it (if the input was denormal)
* 2. Shift nonsign right by 3 so the exponent (5 bits originally) becomes an 8-bit field and 10-bit mantissa
* shifts into the 10 high bits of the 23-bit mantissa of IEEE single-precision number.
* 3. Add 0x70 to the exponent (starting at bit 23) to compensate the different in exponent bias
* (0x7F for single-precision number less 0xF for half-precision number).
* 4. Subtract renorm_shift from the exponent (starting at bit 23) to account for renormalization. As renorm_shift
* is less than 0x70, this can be combined with step 3.
* 5. Binary OR with inf_nan_mask to turn the exponent into 0xFF if the input was NaN or infinity.
* 6. Binary ANDNOT with zero_mask to turn the mantissa and exponent into zero if the input was zero.
* 7. Combine with the sign of the input number.
*/
return sign | ((((nonsign << renorm_shift >> 3) + ((0x70 - renorm_shift) << 23)) | inf_nan_mask) & ~zero_mask);
}
/*
* Convert a 16-bit floating-point number in IEEE half-precision format, in bit representation, to
* a 32-bit floating-point number in IEEE single-precision format.
*
* @note The implementation relies on IEEE-like (no assumption about rounding mode and no operations on denormals)
* floating-point operations and bitcasts between integer and floating-point variables.
*/
static inline float fp16_ieee_to_fp32_value(uint16_t h) {
/*
* Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
* +---+-----+------------+-------------------+
* | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 31 26-30 16-25 0-15
*
* S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.
*/
const uint32_t w = (uint32_t) h << 16;
/*
* Extract the sign of the input number into the high bit of the 32-bit word:
*
* +---+----------------------------------+
* | S |0000000 00000000 00000000 00000000|
* +---+----------------------------------+
* Bits 31 0-31
*/
const uint32_t sign = w & UINT32_C(0x80000000);
/*
* Extract mantissa and biased exponent of the input number into the high bits of the 32-bit word:
*
* +-----+------------+---------------------+
* |EEEEE|MM MMMM MMMM|0 0000 0000 0000 0000|
* +-----+------------+---------------------+
* Bits 27-31 17-26 0-16
*/
const uint32_t two_w = w + w;
/*
* Shift mantissa and exponent into bits 23-28 and bits 13-22 so they become mantissa and exponent
* of a single-precision floating-point number:
*
* S|Exponent | Mantissa
* +-+---+-----+------------+----------------+
* |0|000|EEEEE|MM MMMM MMMM|0 0000 0000 0000|
* +-+---+-----+------------+----------------+
* Bits | 23-31 | 0-22
*
* Next, there are some adjustments to the exponent:
* - The exponent needs to be corrected by the difference in exponent bias between single-precision and half-precision
* formats (0x7F - 0xF = 0x70)
* - Inf and NaN values in the inputs should become Inf and NaN values after conversion to the single-precision number.
* Therefore, if the biased exponent of the half-precision input was 0x1F (max possible value), the biased exponent
* of the single-precision output must be 0xFF (max possible value). We do this correction in two steps:
* - First, we adjust the exponent by (0xFF - 0x1F) = 0xE0 (see exp_offset below) rather than by 0x70 suggested
* by the difference in the exponent bias (see above).
* - Then we multiply the single-precision result of exponent adjustment by 2**(-112) to reverse the effect of
* exponent adjustment by 0xE0 less the necessary exponent adjustment by 0x70 due to difference in exponent bias.
* The floating-point multiplication hardware would ensure than Inf and NaN would retain their value on at least
* partially IEEE754-compliant implementations.
*
* Note that the above operations do not handle denormal inputs (where biased exponent == 0). However, they also do not
* operate on denormal inputs, and do not produce denormal results.
*/
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
/*
* Convert denormalized half-precision inputs into single-precision results (always normalized).
* Zero inputs are also handled here.
*
* In a denormalized number the biased exponent is zero, and mantissa has on-zero bits.
* First, we shift mantissa into bits 0-9 of the 32-bit word.
*
* zeros | mantissa
* +---------------------------+------------+
* |0000 0000 0000 0000 0000 00|MM MMMM MMMM|
* +---------------------------+------------+
* Bits 10-31 0-9
*
* Now, remember that denormalized half-precision numbers are represented as:
* FP16 = mantissa * 2**(-24).
* The trick is to construct a normalized single-precision number with the same mantissa and thehalf-precision input
* and with an exponent which would scale the corresponding mantissa bits to 2**(-24).
* A normalized single-precision floating-point number is represented as:
* FP32 = (1 + mantissa * 2**(-23)) * 2**(exponent - 127)
* Therefore, when the biased exponent is 126, a unit change in the mantissa of the input denormalized half-precision
* number causes a change of the constructud single-precision number by 2**(-24), i.e. the same ammount.
*
* The last step is to adjust the bias of the constructed single-precision number. When the input half-precision number
* is zero, the constructed single-precision number has the value of
* FP32 = 1 * 2**(126 - 127) = 2**(-1) = 0.5
* Therefore, we need to subtract 0.5 from the constructed single-precision number to get the numerical equivalent of
* the input half-precision number.
*/
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
/*
* - Choose either results of conversion of input as a normalized number, or as a denormalized number, depending on the
* input exponent. The variable two_w contains input exponent in bits 27-31, therefore if its smaller than 2**27, the
* input is either a denormal number, or zero.
* - Combine the result of conversion of exponent and mantissa with the sign of the input number.
*/
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
/*
* Convert a 32-bit floating-point number in IEEE single-precision format to a 16-bit floating-point number in
* IEEE half-precision format, in bit representation.
*
* @note The implementation relies on IEEE-like (no assumption about rounding mode and no operations on denormals)
* floating-point operations and bitcasts between integer and floating-point variables.
*/
static inline uint16_t fp16_ieee_from_fp32_value(float f) {
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
/*
* Convert a 16-bit floating-point number in ARM alternative half-precision format, in bit representation, to
* a 32-bit floating-point number in IEEE single-precision format, in bit representation.
*
* @note The implementation doesn't use any floating-point operations.
*/
static inline uint32_t fp16_alt_to_fp32_bits(uint16_t h) {
/*
* Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
* +---+-----+------------+-------------------+
* | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 31 26-30 16-25 0-15
*
* S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.
*/
const uint32_t w = (uint32_t) h << 16;
/*
* Extract the sign of the input number into the high bit of the 32-bit word:
*
* +---+----------------------------------+
* | S |0000000 00000000 00000000 00000000|
* +---+----------------------------------+
* Bits 31 0-31
*/
const uint32_t sign = w & UINT32_C(0x80000000);
/*
* Extract mantissa and biased exponent of the input number into the bits 0-30 of the 32-bit word:
*
* +---+-----+------------+-------------------+
* | 0 |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 30 27-31 17-26 0-16
*/
const uint32_t nonsign = w & UINT32_C(0x7FFFFFFF);
/*
* Renorm shift is the number of bits to shift mantissa left to make the half-precision number normalized.
* If the initial number is normalized, some of its high 6 bits (sign == 0 and 5-bit exponent) equals one.
* In this case renorm_shift == 0. If the number is denormalize, renorm_shift > 0. Note that if we shift
* denormalized nonsign by renorm_shift, the unit bit of mantissa will shift into exponent, turning the
* biased exponent into 1, and making mantissa normalized (i.e. without leading 1).
*/
#ifdef _MSC_VER
unsigned long nonsign_bsr;
_BitScanReverse(&nonsign_bsr, (unsigned long) nonsign);
uint32_t renorm_shift = (uint32_t) nonsign_bsr ^ 31;
#else
uint32_t renorm_shift = __builtin_clz(nonsign);
#endif
renorm_shift = renorm_shift > 5 ? renorm_shift - 5 : 0;
/*
* Iff nonsign is 0, it overflows into 0xFFFFFFFF, turning bit 31 into 1. Otherwise, bit 31 remains 0.
* The signed shift right by 31 broadcasts bit 31 into all bits of the zero_mask. Thus
* zero_mask ==
* 0xFFFFFFFF if the half-precision number was zero (+0.0h or -0.0h)
* 0x00000000 otherwise
*/
const int32_t zero_mask = (int32_t) (nonsign - 1) >> 31;
/*
* 1. Shift nonsign left by renorm_shift to normalize it (if the input was denormal)
* 2. Shift nonsign right by 3 so the exponent (5 bits originally) becomes an 8-bit field and 10-bit mantissa
* shifts into the 10 high bits of the 23-bit mantissa of IEEE single-precision number.
* 3. Add 0x70 to the exponent (starting at bit 23) to compensate the different in exponent bias
* (0x7F for single-precision number less 0xF for half-precision number).
* 4. Subtract renorm_shift from the exponent (starting at bit 23) to account for renormalization. As renorm_shift
* is less than 0x70, this can be combined with step 3.
* 5. Binary ANDNOT with zero_mask to turn the mantissa and exponent into zero if the input was zero.
* 6. Combine with the sign of the input number.
*/
return sign | (((nonsign << renorm_shift >> 3) + ((0x70 - renorm_shift) << 23)) & ~zero_mask);
}
/*
* Convert a 16-bit floating-point number in ARM alternative half-precision format, in bit representation, to
* a 32-bit floating-point number in IEEE single-precision format.
*
* @note The implementation relies on IEEE-like (no assumption about rounding mode and no operations on denormals)
* floating-point operations and bitcasts between integer and floating-point variables.
*/
static inline float fp16_alt_to_fp32_value(uint16_t h) {
/*
* Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
* +---+-----+------------+-------------------+
* | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 31 26-30 16-25 0-15
*
* S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.
*/
const uint32_t w = (uint32_t) h << 16;
/*
* Extract the sign of the input number into the high bit of the 32-bit word:
*
* +---+----------------------------------+
* | S |0000000 00000000 00000000 00000000|
* +---+----------------------------------+
* Bits 31 0-31
*/
const uint32_t sign = w & UINT32_C(0x80000000);
/*
* Extract mantissa and biased exponent of the input number into the high bits of the 32-bit word:
*
* +-----+------------+---------------------+
* |EEEEE|MM MMMM MMMM|0 0000 0000 0000 0000|
* +-----+------------+---------------------+
* Bits 27-31 17-26 0-16
*/
const uint32_t two_w = w + w;
/*
* Shift mantissa and exponent into bits 23-28 and bits 13-22 so they become mantissa and exponent
* of a single-precision floating-point number:
*
* S|Exponent | Mantissa
* +-+---+-----+------------+----------------+
* |0|000|EEEEE|MM MMMM MMMM|0 0000 0000 0000|
* +-+---+-----+------------+----------------+
* Bits | 23-31 | 0-22
*
* Next, the exponent is adjusted for the difference in exponent bias between single-precision and half-precision
* formats (0x7F - 0xF = 0x70). This operation never overflows or generates non-finite values, as the largest
* half-precision exponent is 0x1F and after the adjustment is can not exceed 0x8F < 0xFE (largest single-precision
* exponent for non-finite values).
*
* Note that this operation does not handle denormal inputs (where biased exponent == 0). However, they also do not
* operate on denormal inputs, and do not produce denormal results.
*/
const float exp_offset = UINT32_C(0x70) << 23;
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset);
/*
* Convert denormalized half-precision inputs into single-precision results (always normalized).
* Zero inputs are also handled here.
*
* In a denormalized number the biased exponent is zero, and mantissa has on-zero bits.
* First, we shift mantissa into bits 0-9 of the 32-bit word.
*
* zeros | mantissa
* +---------------------------+------------+
* |0000 0000 0000 0000 0000 00|MM MMMM MMMM|
* +---------------------------+------------+
* Bits 10-31 0-9
*
* Now, remember that denormalized half-precision numbers are represented as:
* FP16 = mantissa * 2**(-24).
* The trick is to construct a normalized single-precision number with the same mantissa and thehalf-precision input
* and with an exponent which would scale the corresponding mantissa bits to 2**(-24).
* A normalized single-precision floating-point number is represented as:
* FP32 = (1 + mantissa * 2**(-23)) * 2**(exponent - 127)
* Therefore, when the biased exponent is 126, a unit change in the mantissa of the input denormalized half-precision
* number causes a change of the constructud single-precision number by 2**(-24), i.e. the same ammount.
*
* The last step is to adjust the bias of the constructed single-precision number. When the input half-precision number
* is zero, the constructed single-precision number has the value of
* FP32 = 1 * 2**(126 - 127) = 2**(-1) = 0.5
* Therefore, we need to subtract 0.5 from the constructed single-precision number to get the numerical equivalent of
* the input half-precision number.
*/
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
/*
* - Choose either results of conversion of input as a normalized number, or as a denormalized number, depending on the
* input exponent. The variable two_w contains input exponent in bits 27-31, therefore if its smaller than 2**27, the
* input is either a denormal number, or zero.
* - Combine the result of conversion of exponent and mantissa with the sign of the input number.
*/
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
/*
* Convert a 32-bit floating-point number in IEEE single-precision format to a 16-bit floating-point number in
* ARM alternative half-precision format, in bit representation.
*
* @note The implementation relies on IEEE-like (no assumption about rounding mode and no operations on denormals)
* floating-point operations and bitcasts between integer and floating-point variables.
*/
static inline uint16_t fp16_alt_from_fp32_value(float f) {
const uint32_t w = fp32_to_bits(f);
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t shl1_w = w + w;
const uint32_t shl1_max_fp16_fp32 = UINT32_C(0x8FFFC000);
const uint32_t shl1_base = shl1_w > shl1_max_fp16_fp32 ? shl1_max_fp16_fp32 : shl1_w;
uint32_t shl1_bias = shl1_base & UINT32_C(0xFF000000);
const uint32_t exp_difference = 23 - 10;
const uint32_t shl1_bias_min = (127 - 1 - exp_difference) << 24;
if (shl1_bias < shl1_bias_min) {
shl1_bias = shl1_bias_min;
}
const float bias = fp32_from_bits((shl1_bias >> 1) + ((exp_difference + 2) << 23));
const float base = fp32_from_bits((shl1_base >> 1) + (2 << 23)) + bias;
const uint32_t exp_f = fp32_to_bits(base) >> 13;
return (sign >> 16) | ((exp_f & UINT32_C(0x00007C00)) + (fp32_to_bits(base) & UINT32_C(0x00000FFF)));
}
#endif /* FP16_FP16_H */
ext/decoders/native/main.c
+78 -30
View File
@@ -1,10 +1,45 @@
#include <stdlib.h>
#include <stdint.h>
#include <ruby.h>
#include "rgb.h"
#include "etc.h"
#include "astc.h"
#include "dxtc.h"
#include "etc.h"
#include "rgb.h"
#include <ruby.h>
#include <stdint.h>
#include <stdlib.h>
/*
* Decode image from A8 binary
*
* @param [String] rb_data binary to decode
* @param [Integer] size width * height
* @return [String] decoded rgb binary
*/
static VALUE rb_decode_a8(VALUE self, VALUE rb_data, VALUE size)
{
if (RSTRING_LEN(rb_data) < FIX2LONG(size))
rb_raise(rb_eStandardError, "Data size is not enough.");
VALUE ret = rb_str_buf_new(FIX2LONG(size) * 3);
decode_a8((uint8_t*)RSTRING_PTR(rb_data), FIX2INT(size), (uint8_t*)RSTRING_PTR(ret));
rb_str_set_len(ret, FIX2LONG(size) * 3);
return ret;
}
/*
* Decode image from R16 binary
*
* @param [String] rb_data binary to decode
* @param [Integer] size width * height
* @param [Boolean] big whether input data are big endian
* @return [String] decoded rgb binary
*/
static VALUE rb_decode_r16(VALUE self, VALUE rb_data, VALUE size, VALUE big)
{
if (RSTRING_LEN(rb_data) < FIX2LONG(size) * 2)
rb_raise(rb_eStandardError, "Data size is not enough.");
VALUE ret = rb_str_buf_new(FIX2LONG(size) * 3);
decode_r16((uint16_t*)RSTRING_PTR(rb_data), FIX2INT(size), RTEST(big), (uint8_t*)RSTRING_PTR(ret));
rb_str_set_len(ret, FIX2LONG(size) * 3);
return ret;
}
/*
* Decode image from RGB565 binary
@@ -12,15 +47,15 @@
* @param [String] rb_data binary to decode
* @param [Integer] size width * height
* @param [Boolean] big whether input data are big endian
* @return [String] decoded rgba binary
* @return [String] decoded rgb binary
*/
static VALUE rb_decode_rgb565(VALUE self, VALUE rb_data, VALUE size, VALUE big) {
static VALUE rb_decode_rgb565(VALUE self, VALUE rb_data, VALUE size, VALUE big)
{
if (RSTRING_LEN(rb_data) < FIX2LONG(size) * 2)
rb_raise(rb_eStandardError, "Data size is not enough.");
uint8_t *image = (uint8_t*)malloc(FIX2LONG(size) * 4);
decode_rgb565((uint16_t*)RSTRING_PTR(rb_data), FIX2INT(size), RTEST(big), image);
VALUE ret = rb_str_new((char*)image, FIX2LONG(size) * 4);
free(image);
VALUE ret = rb_str_buf_new(FIX2LONG(size) * 3);
decode_rgb565((uint16_t*)RSTRING_PTR(rb_data), FIX2INT(size), RTEST(big), (uint8_t*)RSTRING_PTR(ret));
rb_str_set_len(ret, FIX2LONG(size) * 3);
return ret;
}
@@ -32,10 +67,11 @@ static VALUE rb_decode_rgb565(VALUE self, VALUE rb_data, VALUE size, VALUE big)
* @param [Integer] h image height
* @return [String] decoded rgba binary
*/
static VALUE rb_decode_etc1(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
static VALUE rb_decode_etc1(VALUE self, VALUE rb_data, VALUE w, VALUE h)
{
if (RSTRING_LEN(rb_data) < ((FIX2LONG(w) + 3) / 4) * ((FIX2LONG(h) + 3) / 4) * 8)
rb_raise(rb_eStandardError, "Data size is not enough.");
uint32_t *image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
uint32_t* image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
decode_etc1((uint64_t*)RSTRING_PTR(rb_data), FIX2INT(w), FIX2INT(h), image);
VALUE ret = rb_str_new((char*)image, FIX2LONG(w) * FIX2LONG(h) * sizeof(uint32_t));
free(image);
@@ -50,10 +86,11 @@ static VALUE rb_decode_etc1(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
* @param [Integer] h image height
* @return [String] decoded rgba binary
*/
static VALUE rb_decode_etc2(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
static VALUE rb_decode_etc2(VALUE self, VALUE rb_data, VALUE w, VALUE h)
{
if (RSTRING_LEN(rb_data) < ((FIX2LONG(w) + 3) / 4) * ((FIX2LONG(h) + 3) / 4) * 8)
rb_raise(rb_eStandardError, "Data size is not enough.");
uint32_t *image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
uint32_t* image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
decode_etc2((uint64_t*)RSTRING_PTR(rb_data), FIX2INT(w), FIX2INT(h), image);
VALUE ret = rb_str_new((char*)image, FIX2LONG(w) * FIX2LONG(h) * sizeof(uint32_t));
free(image);
@@ -68,11 +105,13 @@ static VALUE rb_decode_etc2(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
* @param [Integer] h image height
* @return [String] decoded rgba binary
*/
static VALUE rb_decode_etc2a1(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
static VALUE rb_decode_etc2a1(VALUE self, VALUE rb_data, VALUE w, VALUE h)
{
if (RSTRING_LEN(rb_data) < ((FIX2LONG(w) + 3) / 4) * ((FIX2LONG(h) + 3) / 4) * 9)
rb_raise(rb_eStandardError, "Data size is not enough.");
uint32_t *image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
decode_etc2a8((uint64_t*)RSTRING_PTR(rb_data), FIX2INT(w), FIX2INT(h), image);
uint32_t* image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
decode_etc2a8((uint64_t*)RSTRING_PTR(rb_data), FIX2INT(w), FIX2INT(h),
image);
VALUE ret = rb_str_new((char*)image, FIX2LONG(w) * FIX2LONG(h) * sizeof(uint32_t));
free(image);
return ret;
@@ -86,11 +125,13 @@ static VALUE rb_decode_etc2a1(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
* @param [Integer] h image height
* @return [String] decoded rgba binary
*/
static VALUE rb_decode_etc2a8(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
static VALUE rb_decode_etc2a8(VALUE self, VALUE rb_data, VALUE w, VALUE h)
{
if (RSTRING_LEN(rb_data) < ((FIX2LONG(w) + 3) / 4) * ((FIX2LONG(h) + 3) / 4) * 16)
rb_raise(rb_eStandardError, "Data size is not enough.");
uint32_t *image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
decode_etc2a8((uint64_t*)RSTRING_PTR(rb_data), FIX2INT(w), FIX2INT(h), image);
uint32_t* image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
decode_etc2a8((uint64_t*)RSTRING_PTR(rb_data), FIX2INT(w), FIX2INT(h),
image);
VALUE ret = rb_str_new((char*)image, FIX2LONG(w) * FIX2LONG(h) * sizeof(uint32_t));
free(image);
return ret;
@@ -106,11 +147,13 @@ static VALUE rb_decode_etc2a8(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
* @param [Integer] bh block height
* @return [String] decoded rgba binary
*/
static VALUE rb_decode_astc(VALUE self, VALUE rb_data, VALUE w, VALUE h, VALUE bw, VALUE bh) {
static VALUE rb_decode_astc(VALUE self, VALUE rb_data, VALUE w, VALUE h,
VALUE bw, VALUE bh)
{
if (RSTRING_LEN(rb_data) < ((FIX2LONG(w) + FIX2LONG(bw) - 1) / FIX2LONG(bw)) * ((FIX2LONG(h) + FIX2LONG(bh) - 1) / FIX2LONG(bh)) * 16)
rb_raise(rb_eStandardError, "Data size is not enough.");
const uint8_t *data = (uint8_t*)RSTRING_PTR(rb_data);
uint32_t *image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
const uint8_t* data = (uint8_t*)RSTRING_PTR(rb_data);
uint32_t* image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
decode_astc(data, FIX2INT(w), FIX2INT(h), FIX2INT(bw), FIX2INT(bh), image);
VALUE ret = rb_str_new((char*)image, FIX2LONG(w) * FIX2LONG(h) * sizeof(uint32_t));
free(image);
@@ -125,10 +168,11 @@ static VALUE rb_decode_astc(VALUE self, VALUE rb_data, VALUE w, VALUE h, VALUE b
* @param [Integer] h image height
* @return [String] decoded rgba binary
*/
static VALUE rb_decode_dxt1(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
static VALUE rb_decode_dxt1(VALUE self, VALUE rb_data, VALUE w, VALUE h)
{
if (RSTRING_LEN(rb_data) < ((FIX2LONG(w) + 3) / 4) * ((FIX2LONG(h) + 3) / 4) * 8)
rb_raise(rb_eStandardError, "Data size is not enough.");
uint32_t *image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
uint32_t* image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
decode_dxt1((uint64_t*)RSTRING_PTR(rb_data), FIX2INT(w), FIX2INT(h), image);
VALUE ret = rb_str_new((char*)image, FIX2LONG(w) * FIX2LONG(h) * sizeof(uint32_t));
free(image);
@@ -143,25 +187,29 @@ static VALUE rb_decode_dxt1(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
* @param [Integer] h image height
* @return [String] decoded rgba binary
*/
static VALUE rb_decode_dxt5(VALUE self, VALUE rb_data, VALUE w, VALUE h) {
static VALUE rb_decode_dxt5(VALUE self, VALUE rb_data, VALUE w, VALUE h)
{
if (RSTRING_LEN(rb_data) < ((FIX2LONG(w) + 3) / 4) * ((FIX2LONG(h) + 3) / 4) * 16)
rb_raise(rb_eStandardError, "Data size is not enough.");
uint32_t *image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
uint32_t* image = (uint32_t*)calloc(FIX2LONG(w) * FIX2LONG(h), sizeof(uint32_t));
decode_dxt5((uint64_t*)RSTRING_PTR(rb_data), FIX2INT(w), FIX2INT(h), image);
VALUE ret = rb_str_new((char*)image, FIX2LONG(w) * FIX2LONG(h) * sizeof(uint32_t));
free(image);
return ret;
}
void Init_native() {
void Init_native()
{
VALUE mMikunyan = rb_define_module("Mikunyan");
VALUE mDecodeHelper = rb_define_module_under(mMikunyan, "DecodeHelper");
rb_define_module_function(mDecodeHelper, "decode_a8", rb_decode_a8, 2);
rb_define_module_function(mDecodeHelper, "decode_r16", rb_decode_r16, 3);
rb_define_module_function(mDecodeHelper, "decode_rgb565", rb_decode_rgb565, 3);
rb_define_module_function(mDecodeHelper, "decode_etc1", rb_decode_etc1, 3);
rb_define_module_function(mDecodeHelper, "decode_etc2", rb_decode_etc2, 3);
rb_define_module_function(mDecodeHelper, "decode_etc2a1", rb_decode_etc2a1, 3);
rb_define_module_function(mDecodeHelper, "decode_etc2a8", rb_decode_etc2a8, 3);
rb_define_module_function(mDecodeHelper, "decode_astc", rb_decode_astc, 5);
rb_define_module_function(mDecodeHelper, "decode_astc", rb_decode_astc, 5);
rb_define_module_function(mDecodeHelper, "decode_dxt1", rb_decode_dxt1, 3);
rb_define_module_function(mDecodeHelper, "decode_dxt5", rb_decode_dxt5, 3);
}
ext/decoders/native/rgb.c
+53 -22
View File
@@ -1,26 +1,57 @@
#include "rgb.h"
#include "common.h"
#include <stdint.h>
static inline int is_system_little() {
int x = 1;
return *(char*)&x == 1;
}
void decode_rgb565(const uint16_t* data, const int size, const int is_big_endian, uint8_t* image) {
const uint16_t *d = data;
if (is_big_endian == is_system_little()) {
uint8_t *p = image;
for (int i = 0; i < size; i++, d++, p += 4) {
uint_fast8_t r = *d & 0x00f8;
uint_fast8_t g = (*d & 0x0007) << 5 | (*d & 0xe000) >> 11;
uint_fast8_t b = (*d & 0x1f00) >> 5;
p[0] = r | r >> 5;
p[1] = g | g >> 6;
p[2] = b | b >> 5;
p[3] = 255;
}
} else {
uint32_t *p = (uint32_t*)image;
for (int i = 0; i < size; i++, d++, p++)
*p = (*d & 0xf800) >> 8 | *d >> 13 | (*d & 0x7e0) << 5 | (*d & 0x60) << 3 | *d << 19 | (*d & 0x1c) << 14 | 0xff000000;
void decode_a8(const uint8_t* data, const int size, uint8_t* image)
{
const uint8_t *d = data, *d_end = data + size;
for (int i = 0; d < d_end; d++) {
image[i++] = *d;
image[i++] = *d;
image[i++] = *d;
}
}
void decode_r16(const uint16_t* data, const int size, const int endian_big, uint8_t* image)
{
const uint16_t *d = data, *d_end = data + size;
if (IS_LITTLE_ENDIAN == !endian_big) {
// Same endian
for (int i = 0; d < d_end; d++) {
uint8_t c = *d >> 8;
image[i++] = c;
image[i++] = c;
image[i++] = c;
}
} else {
// Different endian
for (int i = 0; d < d_end; d++) {
uint8_t c = *d;
image[i++] = c;
image[i++] = c;
image[i++] = c;
}
}
}
void decode_rgb565(const uint16_t* data, const int size, const int endian_big, uint8_t* image)
{
const uint16_t *d = data, *d_end = data + size;
if (IS_LITTLE_ENDIAN == !endian_big) {
// Same endian
// RRRRR GGG | GGG BBBBB
for (int i = 0; d < d_end; d++) {
image[i++] = (*d >> 8 & 0xf8) | (*d >> 13);
image[i++] = (*d >> 3 & 0xfc) | (*d >> 9 & 3);
image[i++] = (*d << 3) | (*d >> 2 & 7);
}
} else {
// Different endian
// GGG BBBBB | RRRRR GGG
for (int i = 0; d < d_end; d++) {
image[i++] = (*d & 0xf8) | (*d >> 5 & 7);
image[i++] = (*d << 5 & 0xe0) | (*d >> 11 & 0x1c) | (*d >> 1 & 3);
image[i++] = (*d >> 5 & 0xf8) | (*d >> 10 & 0x7);
}
}
}
ext/decoders/native/rgb.h
+2
View File
@@ -3,6 +3,8 @@
#include <stdint.h>
void decode_a8(const uint8_t*, const int, uint8_t*);
void decode_r16(const uint16_t*, const int, const int, uint8_t*);
void decode_rgb565(const uint16_t*, const int, const int, uint8_t*);
#endif /* end of include guard: RGB_H */