Use a faster BinDCT.

* A slightly faster BinDCT.
* JPEG time: GS:22ms RGB:62ms (tested on color-bar images).
This commit is contained in:
iabdalkader 2016-02-13 04:28:26 +02:00
parent 2e9f25515b
commit ee6192aeba

View File

@ -13,12 +13,8 @@
#include "ff.h"
#include "xalloc.h"
#include "imlib.h"
// BinDCT Constants
#define C0 (1567) // 0.382683433f * 4096
#define C2 (5351) // 1.306562965f * 4096
#define C3 (2896) // 0.707106781f * 4096
#include "stm32f4xx_hal.h"
#define TIME_JPEG (0)
typedef struct {
int idx;
@ -219,17 +215,16 @@ static void jpeg_calcBits(int val, uint16_t bits[2]) {
bits[0] = val & ((1<<bits[1])-1);
}
static int jpeg_processDU(jpeg_buf_t *jpeg_buf, int16_t *CDU, int *fdtbl, int DC, const uint16_t (*HTDC)[2], const uint16_t (*HTAC)[2])
static int jpeg_processDU(jpeg_buf_t *jpeg_buf, int *CDU, int *fdtbl, int DC, const uint16_t (*HTDC)[2], const uint16_t (*HTAC)[2])
{
const uint16_t EOB[2] = { HTAC[0x00][0], HTAC[0x00][1] };
const uint16_t M16zeroes[2] = { HTAC[0xF0][0], HTAC[0xF0][1] };
int z1, z2, z3, z4, z5, z11, z13;
int t0, t1, t2, t3, t4, t5, t6, t7, t10, t11, t12, t13;
// BinDCT
// BinDCT-a1
// DCT rows
for (int16_t i=8, *p=CDU; i>0; i--, p+=8) {
for (int i=8, *p=CDU; i>0; i--, p+=8) {
t0 = p[0] + p[7];
t1 = p[1] + p[6];
t2 = p[2] + p[5];
@ -240,78 +235,85 @@ static int jpeg_processDU(jpeg_buf_t *jpeg_buf, int16_t *CDU, int *fdtbl, int DC
t5 = p[2] - p[5];
t4 = p[3] - p[4];
// Even part
t10 = t0 + t3;
/* Even part */
t10 = t0 + t3 ; /* phase 2 */
t13 = t0 - t3;
t11 = t1 + t2;
t11 = t1 + t2 ;
t12 = t1 - t2;
z1 = (t12 + t13) * C3>>12; // c4
p[0] = t10 + t11;
p[4] = t10 - t11;
p[2] = t13 + z1;
p[6] = t13 - z1;
p[0] = (t10 + t11); /* phase 3 */
p[4] = ((p[0] ) >> 1) - t11; /* Jie 05/18/00 */
// Odd part
t10 = t4 + t5;// phase 2
t11 = t5 + t6;
t12 = t6 + t7;
/*1/2, -1/2: alter the sign to get positive scaling factor */
p[6] = (( t13 ) >> 1) - t12;
p[2] = t13 - ((p[6] ) >> 1);
// The rotator is modified from fig 4-8 to avoid extra negations.
z5 = (t10 - t12) * C0>>12; // c6
z2 = (t10 >> 1) + z5; // c2-c6
z4 = (t12 * C2>>12) + z5; // c2+c6
z3 = (t11 * C3>>12); // c4
z11 = t7 + z3; // phase 5
z13 = t7 - z3;
/* Odd part */
/* pi/4 = -1/2u 3/4d -1/2u*/
t10 = t5 - (( t6 ) >> 1);
t6 = t6 + t10 - ((t10 ) >> 2);
t5 = ((t6 ) >> 1) - t10;
p[5] = z13 + z2;// phase 6
p[3] = z13 - z2;
p[1] = z11 + z4;
p[7] = z11 - z4;
t10 = t4 + t5;
t11 = t4 - t5;
t12 = t7 - t6;
t13 = t7 + t6;
/* 7pi/16 = 1/4u -1/4d: alter the sign to get positive scaling factor */
p[7] = ((t13 ) >> 2) - t10;
p[1] = t13 - ((p[7] ) >> 2);
/* 3pi/16 = */
/* new version: 1, -1/2 */
p[5] = t11 + t12;
p[3] = t12 - ((p[5] ) >> 1);
}
// DCT columns
for (int16_t i=8, *p=CDU; i>0; i--, p++) {
t0 = p[0] + p[56];
t1 = p[8] + p[48];
t2 = p[16] + p[40];
t3 = p[24] + p[32];
for (int i=8, *p=CDU; i>0; i--, p++) {
t0 = (p[8*0] + p[8*7]);
t1 = (p[8*1] + p[8*6]);
t2 = (p[8*2] + p[8*5]);
t3 = (p[8*3] + p[8*4]);
t7 = p[0] - p[56];
t6 = p[8] - p[48];
t5 = p[16] - p[40];
t4 = p[24] - p[32];
t7 = (p[8*0] - p[8*7]);
t6 = (p[8*1] - p[8*6]);
t5 = (p[8*2] - p[8*5]);
t4 = (p[8*3] - p[8*4]);
// Even part
t10 = t0 + t3; // phase 2
/* Even part */
t10 = t0 + t3; /* phase 2 */
t13 = t0 - t3;
t11 = t1 + t2;
t12 = t1 - t2;
z1 = (t12 + t13) * C3>>12; // c4
p[0] = t10 + t11; // phase 3
p[32] = t10 - t11;
p[16] = t13 + z1; // phase 5
p[48] = t13 - z1;
p[8*0] = (t10 + t11); /* phase 3 */
p[8*4] = ((p[8*0] ) >> 1) - t11; /* Jie 05/18/00 */
// Odd part
t10 = t4 + t5; // phase 2
t11 = t5 + t6;
t12 = t6 + t7;
// 1/2, 1/2
p[8*6] = ((t13 ) >> 1) - t12;
p[8*2] = t13 - ((p[8*6] ) >> 1);
// The rotator is modified from fig 4-8 to avoid extra negations.
z5 = (t10 - t12) * C0>>12; // c6
z2 = (t10 >> 1) + z5; // c2-c6
z4 = (t12 * C2>>12) + z5; // c2+c6
z3 = (t11 * C3>>12); // c4
z11 = t7 + z3; // phase 5
z13 = t7 - z3;
/* Odd part */
/* pi/4 = -1/2u 3/4d -1/2u*/
t10 = t5 - ((t6 ) >> 1);
t6 = t6 + t10 - ((t10 ) >> 2);
t5 = ((t6 ) >> 1) - t10;
p[40] = z13 + z2;// phase 6
p[24] = z13 - z2;
p[8] = z11 + z4;
p[56] = z11 - z4;
t10 = t4 + t5;
t11 = t4 - t5;
t12 = t7 - t6;
t13 = t7 + t6;
/* 7pi/16 = 1/4u -1/4d: alter sign to get positive scaling factor */
p[8*7] = ((t13 ) >> 2) - t10;
p[8*1] = t13 - ((p[8*7] ) >> 2);
/* 3pi/16 = */
/* new : 1 and -1/2 */
p[8*5] = t11 + t12 ;
p[8*3] = t12 - ((p[8*5] ) >> 1);
}
// first non-zero element in reverse order
@ -342,6 +344,7 @@ static int jpeg_processDU(jpeg_buf_t *jpeg_buf, int16_t *CDU, int *fdtbl, int DC
jpeg_writeBits(jpeg_buf, EOB);
return DU[0];
}
for(int i = 1; i <= end0pos; ++i) {
int startpos = i;
for (; DU[i]==0 && i<=end0pos ; ++i) {
@ -382,8 +385,8 @@ void jpeg_init(int quality)
for(int r = 0, k = 0; r < 8; ++r) {
for(int c = 0; c < 8; ++c, ++k) {
fdtbl_Y[k] = (int)((1/(YTable [s_jpeg_ZigZag[k]] * aasf[r] * aasf[c] * 8.0f))*4096);
fdtbl_UV[k] = (int)((1/(UVTable[s_jpeg_ZigZag[k]] * aasf[r] * aasf[c] * 8.0f))*4096);
fdtbl_Y[k] = (int)((aasf[r] * aasf[c] /(YTable [s_jpeg_ZigZag[k]] * 8.0f))*4096);
fdtbl_UV[k] = (int)((aasf[r] * aasf[c] /(UVTable[s_jpeg_ZigZag[k]] * 8.0f))*4096);
}
}
}
@ -425,7 +428,11 @@ void jpeg_write_headers(jpeg_buf_t *jpeg_buf, int width, int height)
void jpeg_compress(image_t *src, image_t *dst, int quality)
{
int DCY=0, DCU=0, DCV=0;
int16_t YDU[64], UDU[64], VDU[64];
int YDU[64], UDU[64], VDU[64];
#if (TIME_JPEG==1)
uint32_t start = HAL_GetTick();
#endif
// JPEG buffer
jpeg_buf_t jpeg_buf = {
@ -528,4 +535,8 @@ void jpeg_compress(image_t *src, image_t *dst, int quality)
dst->bpp = jpeg_buf.idx;
dst->data = jpeg_buf.buf;
#if (TIME_JPEG==1)
printf("time: %lums\n", HAL_GetTick() - start);
#endif
}