ThePain/Swiftgram
1
0
Fork 0
mirror of https://github.com/Swiftgram/Telegram-iOS.git synced 2025-12-22 22:25:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

Input node improvements

Browse Source
This commit is contained in:
Ali
2022-06-28 22:54:05 +02:00
parent 575605f4d0
commit 131be5aaab
51 changed files with 2586 additions and 1363 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
submodules/TelegramUI/Components/AnimationCache/BUILD
View File

@@ -13,7 +13,7 @@ swift_library(
"//submodules/SSignalKit/SwiftSignalKit:SwiftSignalKit",
"//submodules/CryptoUtils:CryptoUtils",
"//submodules/ManagedFile:ManagedFile",
"//submodules/TelegramUI/Components/AnimationCache/DCT:DCT",
"//submodules/TelegramUI/Components/AnimationCache/ImageDCT:ImageDCT",
],
visibility = [
"//visibility:public",

14
submodules/TelegramUI/Components/AnimationCache/DCT/PublicHeaders/DCT/DCT.h
View File

@@ -1,14 +0,0 @@
#ifndef DctImageTransform_h
#define DctImageTransform_h
#import <Foundation/Foundation.h>
#import <DCT/YuvConversion.h>
NSData *generateForwardDctData(int quality);
void performForwardDct(uint8_t const *pixels, int16_t *coefficients, int width, int height, int bytesPerRow, NSData *dctData);
NSData *generateInverseDctData(int quality);
void performInverseDct(int16_t const *coefficients, uint8_t *pixels, int width, int height, int coefficientsPerRow, int bytesPerRow, NSData *idctData);
#endif /* DctImageTransform_h */

991
submodules/TelegramUI/Components/AnimationCache/DCT/Sources/DCT.m
View File

@@ -1,991 +0,0 @@
#import <DCT/DCT.h>
typedef long JLONG;
typedef unsigned char JSAMPLE;
#define GETJSAMPLE(value) ((int)(value))
#define MAXJSAMPLE 255
#define CENTERJSAMPLE 128
typedef short JCOEF;
typedef unsigned int JDIMENSION;
#define JPEG_MAX_DIMENSION 65500L /* a tad under 64K to prevent overflows */
#define MULTIPLIER short /* prefer 16-bit with SIMD for parellelism */
typedef MULTIPLIER IFAST_MULT_TYPE; /* 16 bits is OK, use short if faster */
#define IFAST_SCALE_BITS 2 /* fractional bits in scale factors */
/* Various constants determining the sizes of things.
* All of these are specified by the JPEG standard, so don't change them
* if you want to be compatible.
*/
#define DCTSIZE 8 /* The basic DCT block is 8x8 samples */
#define DCTSIZE2 64 /* DCTSIZE squared; # of elements in a block */
#define NUM_QUANT_TBLS 4 /* Quantization tables are numbered 0..3 */
#define NUM_HUFF_TBLS 4 /* Huffman tables are numbered 0..3 */
#define NUM_ARITH_TBLS 16 /* Arith-coding tables are numbered 0..15 */
#define MAX_COMPS_IN_SCAN 4 /* JPEG limit on # of components in one scan */
#define MAX_SAMP_FACTOR 4 /* JPEG limit on sampling factors */
/* Unfortunately, some bozo at Adobe saw no reason to be bound by the standard;
* the PostScript DCT filter can emit files with many more than 10 blocks/MCU.
* If you happen to run across such a file, you can up D_MAX_BLOCKS_IN_MCU
* to handle it. We even let you do this from the jconfig.h file. However,
* we strongly discourage changing C_MAX_BLOCKS_IN_MCU; just because Adobe
* sometimes emits noncompliant files doesn't mean you should too.
*/
#define C_MAX_BLOCKS_IN_MCU 10 /* compressor's limit on blocks per MCU */
#ifndef D_MAX_BLOCKS_IN_MCU
#define D_MAX_BLOCKS_IN_MCU 10 /* decompressor's limit on blocks per MCU */
#endif
/* Data structures for images (arrays of samples and of DCT coefficients).
*/
typedef JSAMPLE *JSAMPROW; /* ptr to one image row of pixel samples. */
typedef JSAMPROW *JSAMPARRAY; /* ptr to some rows (a 2-D sample array) */
typedef JSAMPARRAY *JSAMPIMAGE; /* a 3-D sample array: top index is color */
typedef JCOEF JBLOCK[DCTSIZE2]; /* one block of coefficients */
typedef JBLOCK *JBLOCKROW; /* pointer to one row of coefficient blocks */
typedef JBLOCKROW *JBLOCKARRAY; /* a 2-D array of coefficient blocks */
typedef JBLOCKARRAY *JBLOCKIMAGE; /* a 3-D array of coefficient blocks */
typedef JCOEF *JCOEFPTR; /* useful in a couple of places */
#include <arm_neon.h>
/* jsimd_idct_ifast_neon() performs dequantization and a fast, not so accurate
* inverse DCT (Discrete Cosine Transform) on one block of coefficients. It
* uses the same calculations and produces exactly the same output as IJG's
* original jpeg_idct_ifast() function, which can be found in jidctfst.c.
*
* Scaled integer constants are used to avoid floating-point arithmetic:
* 0.082392200 = 2688 * 2^-15
* 0.414213562 = 13568 * 2^-15
* 0.847759065 = 27776 * 2^-15
* 0.613125930 = 20096 * 2^-15
*
* See jidctfst.c for further details of the IDCT algorithm. Where possible,
* the variable names and comments here in jsimd_idct_ifast_neon() match up
* with those in jpeg_idct_ifast().
*/
#define PASS1_BITS 2
#define F_0_082 2688
#define F_0_414 13568
#define F_0_847 27776
#define F_0_613 20096
__attribute__((aligned(16))) static const int16_t jsimd_idct_ifast_neon_consts[] = {
F_0_082, F_0_414, F_0_847, F_0_613
};
#define F_0_382 12544
#define F_0_541 17792
#define F_0_707 23168
#define F_0_306 9984
__attribute__((aligned(16))) static const int16_t jsimd_fdct_ifast_neon_consts[] = {
F_0_382, F_0_541, F_0_707, F_0_306
};
typedef short DCTELEM; /* prefer 16 bit with SIMD for parellelism */
typedef unsigned short UDCTELEM;
typedef unsigned int UDCTELEM2;
static void jsimd_fdct_ifast_neon(DCTELEM *data) {
/* Load an 8x8 block of samples into Neon registers. De-interleaving loads
* are used, followed by vuzp to transpose the block such that we have a
* column of samples per vector - allowing all rows to be processed at once.
*/
int16x8x4_t data1 = vld4q_s16(data);
int16x8x4_t data2 = vld4q_s16(data + 4 * DCTSIZE);
int16x8x2_t cols_04 = vuzpq_s16(data1.val[0], data2.val[0]);
int16x8x2_t cols_15 = vuzpq_s16(data1.val[1], data2.val[1]);
int16x8x2_t cols_26 = vuzpq_s16(data1.val[2], data2.val[2]);
int16x8x2_t cols_37 = vuzpq_s16(data1.val[3], data2.val[3]);
int16x8_t col0 = cols_04.val[0];
int16x8_t col1 = cols_15.val[0];
int16x8_t col2 = cols_26.val[0];
int16x8_t col3 = cols_37.val[0];
int16x8_t col4 = cols_04.val[1];
int16x8_t col5 = cols_15.val[1];
int16x8_t col6 = cols_26.val[1];
int16x8_t col7 = cols_37.val[1];
/* Pass 1: process rows. */
/* Load DCT conversion constants. */
const int16x4_t consts = vld1_s16(jsimd_fdct_ifast_neon_consts);
int16x8_t tmp0 = vaddq_s16(col0, col7);
int16x8_t tmp7 = vsubq_s16(col0, col7);
int16x8_t tmp1 = vaddq_s16(col1, col6);
int16x8_t tmp6 = vsubq_s16(col1, col6);
int16x8_t tmp2 = vaddq_s16(col2, col5);
int16x8_t tmp5 = vsubq_s16(col2, col5);
int16x8_t tmp3 = vaddq_s16(col3, col4);
int16x8_t tmp4 = vsubq_s16(col3, col4);
/* Even part */
int16x8_t tmp10 = vaddq_s16(tmp0, tmp3); /* phase 2 */
int16x8_t tmp13 = vsubq_s16(tmp0, tmp3);
int16x8_t tmp11 = vaddq_s16(tmp1, tmp2);
int16x8_t tmp12 = vsubq_s16(tmp1, tmp2);
col0 = vaddq_s16(tmp10, tmp11); /* phase 3 */
col4 = vsubq_s16(tmp10, tmp11);
int16x8_t z1 = vqdmulhq_lane_s16(vaddq_s16(tmp12, tmp13), consts, 2);
col2 = vaddq_s16(tmp13, z1); /* phase 5 */
col6 = vsubq_s16(tmp13, z1);
/* Odd part */
tmp10 = vaddq_s16(tmp4, tmp5); /* phase 2 */
tmp11 = vaddq_s16(tmp5, tmp6);
tmp12 = vaddq_s16(tmp6, tmp7);
int16x8_t z5 = vqdmulhq_lane_s16(vsubq_s16(tmp10, tmp12), consts, 0);
int16x8_t z2 = vqdmulhq_lane_s16(tmp10, consts, 1);
z2 = vaddq_s16(z2, z5);
int16x8_t z4 = vqdmulhq_lane_s16(tmp12, consts, 3);
z5 = vaddq_s16(tmp12, z5);
z4 = vaddq_s16(z4, z5);
int16x8_t z3 = vqdmulhq_lane_s16(tmp11, consts, 2);
int16x8_t z11 = vaddq_s16(tmp7, z3); /* phase 5 */
int16x8_t z13 = vsubq_s16(tmp7, z3);
col5 = vaddq_s16(z13, z2); /* phase 6 */
col3 = vsubq_s16(z13, z2);
col1 = vaddq_s16(z11, z4);
col7 = vsubq_s16(z11, z4);
/* Transpose to work on columns in pass 2. */
int16x8x2_t cols_01 = vtrnq_s16(col0, col1);
int16x8x2_t cols_23 = vtrnq_s16(col2, col3);
int16x8x2_t cols_45 = vtrnq_s16(col4, col5);
int16x8x2_t cols_67 = vtrnq_s16(col6, col7);
int32x4x2_t cols_0145_l = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[0]),
vreinterpretq_s32_s16(cols_45.val[0]));
int32x4x2_t cols_0145_h = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[1]),
vreinterpretq_s32_s16(cols_45.val[1]));
int32x4x2_t cols_2367_l = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[0]),
vreinterpretq_s32_s16(cols_67.val[0]));
int32x4x2_t cols_2367_h = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[1]),
vreinterpretq_s32_s16(cols_67.val[1]));
int32x4x2_t rows_04 = vzipq_s32(cols_0145_l.val[0], cols_2367_l.val[0]);
int32x4x2_t rows_15 = vzipq_s32(cols_0145_h.val[0], cols_2367_h.val[0]);
int32x4x2_t rows_26 = vzipq_s32(cols_0145_l.val[1], cols_2367_l.val[1]);
int32x4x2_t rows_37 = vzipq_s32(cols_0145_h.val[1], cols_2367_h.val[1]);
int16x8_t row0 = vreinterpretq_s16_s32(rows_04.val[0]);
int16x8_t row1 = vreinterpretq_s16_s32(rows_15.val[0]);
int16x8_t row2 = vreinterpretq_s16_s32(rows_26.val[0]);
int16x8_t row3 = vreinterpretq_s16_s32(rows_37.val[0]);
int16x8_t row4 = vreinterpretq_s16_s32(rows_04.val[1]);
int16x8_t row5 = vreinterpretq_s16_s32(rows_15.val[1]);
int16x8_t row6 = vreinterpretq_s16_s32(rows_26.val[1]);
int16x8_t row7 = vreinterpretq_s16_s32(rows_37.val[1]);
/* Pass 2: process columns. */
tmp0 = vaddq_s16(row0, row7);
tmp7 = vsubq_s16(row0, row7);
tmp1 = vaddq_s16(row1, row6);
tmp6 = vsubq_s16(row1, row6);
tmp2 = vaddq_s16(row2, row5);
tmp5 = vsubq_s16(row2, row5);
tmp3 = vaddq_s16(row3, row4);
tmp4 = vsubq_s16(row3, row4);
/* Even part */
tmp10 = vaddq_s16(tmp0, tmp3); /* phase 2 */
tmp13 = vsubq_s16(tmp0, tmp3);
tmp11 = vaddq_s16(tmp1, tmp2);
tmp12 = vsubq_s16(tmp1, tmp2);
row0 = vaddq_s16(tmp10, tmp11); /* phase 3 */
row4 = vsubq_s16(tmp10, tmp11);
z1 = vqdmulhq_lane_s16(vaddq_s16(tmp12, tmp13), consts, 2);
row2 = vaddq_s16(tmp13, z1); /* phase 5 */
row6 = vsubq_s16(tmp13, z1);
/* Odd part */
tmp10 = vaddq_s16(tmp4, tmp5); /* phase 2 */
tmp11 = vaddq_s16(tmp5, tmp6);
tmp12 = vaddq_s16(tmp6, tmp7);
z5 = vqdmulhq_lane_s16(vsubq_s16(tmp10, tmp12), consts, 0);
z2 = vqdmulhq_lane_s16(tmp10, consts, 1);
z2 = vaddq_s16(z2, z5);
z4 = vqdmulhq_lane_s16(tmp12, consts, 3);
z5 = vaddq_s16(tmp12, z5);
z4 = vaddq_s16(z4, z5);
z3 = vqdmulhq_lane_s16(tmp11, consts, 2);
z11 = vaddq_s16(tmp7, z3); /* phase 5 */
z13 = vsubq_s16(tmp7, z3);
row5 = vaddq_s16(z13, z2); /* phase 6 */
row3 = vsubq_s16(z13, z2);
row1 = vaddq_s16(z11, z4);
row7 = vsubq_s16(z11, z4);
vst1q_s16(data + 0 * DCTSIZE, row0);
vst1q_s16(data + 1 * DCTSIZE, row1);
vst1q_s16(data + 2 * DCTSIZE, row2);
vst1q_s16(data + 3 * DCTSIZE, row3);
vst1q_s16(data + 4 * DCTSIZE, row4);
vst1q_s16(data + 5 * DCTSIZE, row5);
vst1q_s16(data + 6 * DCTSIZE, row6);
vst1q_s16(data + 7 * DCTSIZE, row7);
}
static void jsimd_idct_ifast_neon(void *dct_table, JCOEFPTR coef_block,
JSAMPROW output_buf)
{
IFAST_MULT_TYPE *quantptr = dct_table;
/* Load DCT coefficients. */
int16x8_t row0 = vld1q_s16(coef_block + 0 * DCTSIZE);
int16x8_t row1 = vld1q_s16(coef_block + 1 * DCTSIZE);
int16x8_t row2 = vld1q_s16(coef_block + 2 * DCTSIZE);
int16x8_t row3 = vld1q_s16(coef_block + 3 * DCTSIZE);
int16x8_t row4 = vld1q_s16(coef_block + 4 * DCTSIZE);
int16x8_t row5 = vld1q_s16(coef_block + 5 * DCTSIZE);
int16x8_t row6 = vld1q_s16(coef_block + 6 * DCTSIZE);
int16x8_t row7 = vld1q_s16(coef_block + 7 * DCTSIZE);
/* Load quantization table values for DC coefficients. */
int16x8_t quant_row0 = vld1q_s16(quantptr + 0 * DCTSIZE);
/* Dequantize DC coefficients. */
row0 = vmulq_s16(row0, quant_row0);
/* Construct bitmap to test if all AC coefficients are 0. */
int16x8_t bitmap = vorrq_s16(row1, row2);
bitmap = vorrq_s16(bitmap, row3);
bitmap = vorrq_s16(bitmap, row4);
bitmap = vorrq_s16(bitmap, row5);
bitmap = vorrq_s16(bitmap, row6);
bitmap = vorrq_s16(bitmap, row7);
int64_t left_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 0);
int64_t right_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 1);
/* Load IDCT conversion constants. */
const int16x4_t consts = vld1_s16(jsimd_idct_ifast_neon_consts);
if (left_ac_bitmap == 0 && right_ac_bitmap == 0) {
/* All AC coefficients are zero.
* Compute DC values and duplicate into vectors.
*/
int16x8_t dcval = row0;
row1 = dcval;
row2 = dcval;
row3 = dcval;
row4 = dcval;
row5 = dcval;
row6 = dcval;
row7 = dcval;
} else if (left_ac_bitmap == 0) {
/* AC coefficients are zero for columns 0, 1, 2, and 3.
* Use DC values for these columns.
*/
int16x4_t dcval = vget_low_s16(row0);
/* Commence regular fast IDCT computation for columns 4, 5, 6, and 7. */
/* Load quantization table. */
int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE + 4);
int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE + 4);
int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE + 4);
int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE + 4);
int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE + 4);
int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE + 4);
int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE + 4);
/* Even part: dequantize DCT coefficients. */
int16x4_t tmp0 = vget_high_s16(row0);
int16x4_t tmp1 = vmul_s16(vget_high_s16(row2), quant_row2);
int16x4_t tmp2 = vmul_s16(vget_high_s16(row4), quant_row4);
int16x4_t tmp3 = vmul_s16(vget_high_s16(row6), quant_row6);
int16x4_t tmp10 = vadd_s16(tmp0, tmp2); /* phase 3 */
int16x4_t tmp11 = vsub_s16(tmp0, tmp2);
int16x4_t tmp13 = vadd_s16(tmp1, tmp3); /* phases 5-3 */
int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
tmp12 = vsub_s16(tmp12, tmp13);
tmp0 = vadd_s16(tmp10, tmp13); /* phase 2 */
tmp3 = vsub_s16(tmp10, tmp13);
tmp1 = vadd_s16(tmp11, tmp12);
tmp2 = vsub_s16(tmp11, tmp12);
/* Odd part: dequantize DCT coefficients. */
int16x4_t tmp4 = vmul_s16(vget_high_s16(row1), quant_row1);
int16x4_t tmp5 = vmul_s16(vget_high_s16(row3), quant_row3);
int16x4_t tmp6 = vmul_s16(vget_high_s16(row5), quant_row5);
int16x4_t tmp7 = vmul_s16(vget_high_s16(row7), quant_row7);
int16x4_t z13 = vadd_s16(tmp6, tmp5); /* phase 6 */
int16x4_t neg_z10 = vsub_s16(tmp5, tmp6);
int16x4_t z11 = vadd_s16(tmp4, tmp7);
int16x4_t z12 = vsub_s16(tmp4, tmp7);
tmp7 = vadd_s16(z11, z13); /* phase 5 */
int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vadd_s16(tmp11, z11_sub_z13);
int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
z5 = vadd_s16(z5, z10_add_z12);
tmp10 = vqdmulh_lane_s16(z12, consts, 0);
tmp10 = vadd_s16(tmp10, z12);
tmp10 = vsub_s16(tmp10, z5);
tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
tmp12 = vadd_s16(tmp12, z5);
tmp6 = vsub_s16(tmp12, tmp7); /* phase 2 */
tmp5 = vsub_s16(tmp11, tmp6);
tmp4 = vadd_s16(tmp10, tmp5);
row0 = vcombine_s16(dcval, vadd_s16(tmp0, tmp7));
row7 = vcombine_s16(dcval, vsub_s16(tmp0, tmp7));
row1 = vcombine_s16(dcval, vadd_s16(tmp1, tmp6));
row6 = vcombine_s16(dcval, vsub_s16(tmp1, tmp6));
row2 = vcombine_s16(dcval, vadd_s16(tmp2, tmp5));
row5 = vcombine_s16(dcval, vsub_s16(tmp2, tmp5));
row4 = vcombine_s16(dcval, vadd_s16(tmp3, tmp4));
row3 = vcombine_s16(dcval, vsub_s16(tmp3, tmp4));
} else if (right_ac_bitmap == 0) {
/* AC coefficients are zero for columns 4, 5, 6, and 7.
* Use DC values for these columns.
*/
int16x4_t dcval = vget_high_s16(row0);
/* Commence regular fast IDCT computation for columns 0, 1, 2, and 3. */
/* Load quantization table. */
int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE);
int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE);
int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE);
int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE);
int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE);
int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE);
int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE);
/* Even part: dequantize DCT coefficients. */
int16x4_t tmp0 = vget_low_s16(row0);
int16x4_t tmp1 = vmul_s16(vget_low_s16(row2), quant_row2);
int16x4_t tmp2 = vmul_s16(vget_low_s16(row4), quant_row4);
int16x4_t tmp3 = vmul_s16(vget_low_s16(row6), quant_row6);
int16x4_t tmp10 = vadd_s16(tmp0, tmp2); /* phase 3 */
int16x4_t tmp11 = vsub_s16(tmp0, tmp2);
int16x4_t tmp13 = vadd_s16(tmp1, tmp3); /* phases 5-3 */
int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
tmp12 = vsub_s16(tmp12, tmp13);
tmp0 = vadd_s16(tmp10, tmp13); /* phase 2 */
tmp3 = vsub_s16(tmp10, tmp13);
tmp1 = vadd_s16(tmp11, tmp12);
tmp2 = vsub_s16(tmp11, tmp12);
/* Odd part: dequantize DCT coefficients. */
int16x4_t tmp4 = vmul_s16(vget_low_s16(row1), quant_row1);
int16x4_t tmp5 = vmul_s16(vget_low_s16(row3), quant_row3);
int16x4_t tmp6 = vmul_s16(vget_low_s16(row5), quant_row5);
int16x4_t tmp7 = vmul_s16(vget_low_s16(row7), quant_row7);
int16x4_t z13 = vadd_s16(tmp6, tmp5); /* phase 6 */
int16x4_t neg_z10 = vsub_s16(tmp5, tmp6);
int16x4_t z11 = vadd_s16(tmp4, tmp7);
int16x4_t z12 = vsub_s16(tmp4, tmp7);
tmp7 = vadd_s16(z11, z13); /* phase 5 */
int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vadd_s16(tmp11, z11_sub_z13);
int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
z5 = vadd_s16(z5, z10_add_z12);
tmp10 = vqdmulh_lane_s16(z12, consts, 0);
tmp10 = vadd_s16(tmp10, z12);
tmp10 = vsub_s16(tmp10, z5);
tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
tmp12 = vadd_s16(tmp12, z5);
tmp6 = vsub_s16(tmp12, tmp7); /* phase 2 */
tmp5 = vsub_s16(tmp11, tmp6);
tmp4 = vadd_s16(tmp10, tmp5);
row0 = vcombine_s16(vadd_s16(tmp0, tmp7), dcval);
row7 = vcombine_s16(vsub_s16(tmp0, tmp7), dcval);
row1 = vcombine_s16(vadd_s16(tmp1, tmp6), dcval);
row6 = vcombine_s16(vsub_s16(tmp1, tmp6), dcval);
row2 = vcombine_s16(vadd_s16(tmp2, tmp5), dcval);
row5 = vcombine_s16(vsub_s16(tmp2, tmp5), dcval);
row4 = vcombine_s16(vadd_s16(tmp3, tmp4), dcval);
row3 = vcombine_s16(vsub_s16(tmp3, tmp4), dcval);
} else {
/* Some AC coefficients are non-zero; full IDCT calculation required. */
/* Load quantization table. */
int16x8_t quant_row1 = vld1q_s16(quantptr + 1 * DCTSIZE);
int16x8_t quant_row2 = vld1q_s16(quantptr + 2 * DCTSIZE);
int16x8_t quant_row3 = vld1q_s16(quantptr + 3 * DCTSIZE);
int16x8_t quant_row4 = vld1q_s16(quantptr + 4 * DCTSIZE);
int16x8_t quant_row5 = vld1q_s16(quantptr + 5 * DCTSIZE);
int16x8_t quant_row6 = vld1q_s16(quantptr + 6 * DCTSIZE);
int16x8_t quant_row7 = vld1q_s16(quantptr + 7 * DCTSIZE);
/* Even part: dequantize DCT coefficients. */
int16x8_t tmp0 = row0;
int16x8_t tmp1 = vmulq_s16(row2, quant_row2);
int16x8_t tmp2 = vmulq_s16(row4, quant_row4);
int16x8_t tmp3 = vmulq_s16(row6, quant_row6);
int16x8_t tmp10 = vaddq_s16(tmp0, tmp2); /* phase 3 */
int16x8_t tmp11 = vsubq_s16(tmp0, tmp2);
int16x8_t tmp13 = vaddq_s16(tmp1, tmp3); /* phases 5-3 */
int16x8_t tmp1_sub_tmp3 = vsubq_s16(tmp1, tmp3);
int16x8_t tmp12 = vqdmulhq_lane_s16(tmp1_sub_tmp3, consts, 1);
tmp12 = vaddq_s16(tmp12, tmp1_sub_tmp3);
tmp12 = vsubq_s16(tmp12, tmp13);
tmp0 = vaddq_s16(tmp10, tmp13); /* phase 2 */
tmp3 = vsubq_s16(tmp10, tmp13);
tmp1 = vaddq_s16(tmp11, tmp12);
tmp2 = vsubq_s16(tmp11, tmp12);
/* Odd part: dequantize DCT coefficients. */
int16x8_t tmp4 = vmulq_s16(row1, quant_row1);
int16x8_t tmp5 = vmulq_s16(row3, quant_row3);
int16x8_t tmp6 = vmulq_s16(row5, quant_row5);
int16x8_t tmp7 = vmulq_s16(row7, quant_row7);
int16x8_t z13 = vaddq_s16(tmp6, tmp5); /* phase 6 */
int16x8_t neg_z10 = vsubq_s16(tmp5, tmp6);
int16x8_t z11 = vaddq_s16(tmp4, tmp7);
int16x8_t z12 = vsubq_s16(tmp4, tmp7);
tmp7 = vaddq_s16(z11, z13); /* phase 5 */
int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vaddq_s16(tmp11, z11_sub_z13);
int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
z5 = vaddq_s16(z5, z10_add_z12);
tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
tmp10 = vaddq_s16(tmp10, z12);
tmp10 = vsubq_s16(tmp10, z5);
tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
tmp12 = vaddq_s16(tmp12, z5);
tmp6 = vsubq_s16(tmp12, tmp7); /* phase 2 */
tmp5 = vsubq_s16(tmp11, tmp6);
tmp4 = vaddq_s16(tmp10, tmp5);
row0 = vaddq_s16(tmp0, tmp7);
row7 = vsubq_s16(tmp0, tmp7);
row1 = vaddq_s16(tmp1, tmp6);
row6 = vsubq_s16(tmp1, tmp6);
row2 = vaddq_s16(tmp2, tmp5);
row5 = vsubq_s16(tmp2, tmp5);
row4 = vaddq_s16(tmp3, tmp4);
row3 = vsubq_s16(tmp3, tmp4);
}
/* Transpose rows to work on columns in pass 2. */
int16x8x2_t rows_01 = vtrnq_s16(row0, row1);
int16x8x2_t rows_23 = vtrnq_s16(row2, row3);
int16x8x2_t rows_45 = vtrnq_s16(row4, row5);
int16x8x2_t rows_67 = vtrnq_s16(row6, row7);
int32x4x2_t rows_0145_l = vtrnq_s32(vreinterpretq_s32_s16(rows_01.val[0]),
vreinterpretq_s32_s16(rows_45.val[0]));
int32x4x2_t rows_0145_h = vtrnq_s32(vreinterpretq_s32_s16(rows_01.val[1]),
vreinterpretq_s32_s16(rows_45.val[1]));
int32x4x2_t rows_2367_l = vtrnq_s32(vreinterpretq_s32_s16(rows_23.val[0]),
vreinterpretq_s32_s16(rows_67.val[0]));
int32x4x2_t rows_2367_h = vtrnq_s32(vreinterpretq_s32_s16(rows_23.val[1]),
vreinterpretq_s32_s16(rows_67.val[1]));
int32x4x2_t cols_04 = vzipq_s32(rows_0145_l.val[0], rows_2367_l.val[0]);
int32x4x2_t cols_15 = vzipq_s32(rows_0145_h.val[0], rows_2367_h.val[0]);
int32x4x2_t cols_26 = vzipq_s32(rows_0145_l.val[1], rows_2367_l.val[1]);
int32x4x2_t cols_37 = vzipq_s32(rows_0145_h.val[1], rows_2367_h.val[1]);
int16x8_t col0 = vreinterpretq_s16_s32(cols_04.val[0]);
int16x8_t col1 = vreinterpretq_s16_s32(cols_15.val[0]);
int16x8_t col2 = vreinterpretq_s16_s32(cols_26.val[0]);
int16x8_t col3 = vreinterpretq_s16_s32(cols_37.val[0]);
int16x8_t col4 = vreinterpretq_s16_s32(cols_04.val[1]);
int16x8_t col5 = vreinterpretq_s16_s32(cols_15.val[1]);
int16x8_t col6 = vreinterpretq_s16_s32(cols_26.val[1]);
int16x8_t col7 = vreinterpretq_s16_s32(cols_37.val[1]);
/* 1-D IDCT, pass 2 */
/* Even part */
int16x8_t tmp10 = vaddq_s16(col0, col4);
int16x8_t tmp11 = vsubq_s16(col0, col4);
int16x8_t tmp13 = vaddq_s16(col2, col6);
int16x8_t col2_sub_col6 = vsubq_s16(col2, col6);
int16x8_t tmp12 = vqdmulhq_lane_s16(col2_sub_col6, consts, 1);
tmp12 = vaddq_s16(tmp12, col2_sub_col6);
tmp12 = vsubq_s16(tmp12, tmp13);
int16x8_t tmp0 = vaddq_s16(tmp10, tmp13);
int16x8_t tmp3 = vsubq_s16(tmp10, tmp13);
int16x8_t tmp1 = vaddq_s16(tmp11, tmp12);
int16x8_t tmp2 = vsubq_s16(tmp11, tmp12);
/* Odd part */
int16x8_t z13 = vaddq_s16(col5, col3);
int16x8_t neg_z10 = vsubq_s16(col3, col5);
int16x8_t z11 = vaddq_s16(col1, col7);
int16x8_t z12 = vsubq_s16(col1, col7);
int16x8_t tmp7 = vaddq_s16(z11, z13); /* phase 5 */
int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vaddq_s16(tmp11, z11_sub_z13);
int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
z5 = vaddq_s16(z5, z10_add_z12);
tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
tmp10 = vaddq_s16(tmp10, z12);
tmp10 = vsubq_s16(tmp10, z5);
tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
tmp12 = vaddq_s16(tmp12, z5);
int16x8_t tmp6 = vsubq_s16(tmp12, tmp7); /* phase 2 */
int16x8_t tmp5 = vsubq_s16(tmp11, tmp6);
int16x8_t tmp4 = vaddq_s16(tmp10, tmp5);
col0 = vaddq_s16(tmp0, tmp7);
col7 = vsubq_s16(tmp0, tmp7);
col1 = vaddq_s16(tmp1, tmp6);
col6 = vsubq_s16(tmp1, tmp6);
col2 = vaddq_s16(tmp2, tmp5);
col5 = vsubq_s16(tmp2, tmp5);
col4 = vaddq_s16(tmp3, tmp4);
col3 = vsubq_s16(tmp3, tmp4);
/* Scale down by a factor of 8, narrowing to 8-bit. */
int8x16_t cols_01_s8 = vcombine_s8(vqshrn_n_s16(col0, PASS1_BITS + 3),
vqshrn_n_s16(col1, PASS1_BITS + 3));
int8x16_t cols_45_s8 = vcombine_s8(vqshrn_n_s16(col4, PASS1_BITS + 3),
vqshrn_n_s16(col5, PASS1_BITS + 3));
int8x16_t cols_23_s8 = vcombine_s8(vqshrn_n_s16(col2, PASS1_BITS + 3),
vqshrn_n_s16(col3, PASS1_BITS + 3));
int8x16_t cols_67_s8 = vcombine_s8(vqshrn_n_s16(col6, PASS1_BITS + 3),
vqshrn_n_s16(col7, PASS1_BITS + 3));
/* Clamp to range [0-255]. */
uint8x16_t cols_01 =
vreinterpretq_u8_s8
(vaddq_s8(cols_01_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
uint8x16_t cols_45 =
vreinterpretq_u8_s8
(vaddq_s8(cols_45_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
uint8x16_t cols_23 =
vreinterpretq_u8_s8
(vaddq_s8(cols_23_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
uint8x16_t cols_67 =
vreinterpretq_u8_s8
(vaddq_s8(cols_67_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
/* Transpose block to prepare for store. */
uint32x4x2_t cols_0415 = vzipq_u32(vreinterpretq_u32_u8(cols_01),
vreinterpretq_u32_u8(cols_45));
uint32x4x2_t cols_2637 = vzipq_u32(vreinterpretq_u32_u8(cols_23),
vreinterpretq_u32_u8(cols_67));
uint8x16x2_t cols_0145 = vtrnq_u8(vreinterpretq_u8_u32(cols_0415.val[0]),
vreinterpretq_u8_u32(cols_0415.val[1]));
uint8x16x2_t cols_2367 = vtrnq_u8(vreinterpretq_u8_u32(cols_2637.val[0]),
vreinterpretq_u8_u32(cols_2637.val[1]));
uint16x8x2_t rows_0426 = vtrnq_u16(vreinterpretq_u16_u8(cols_0145.val[0]),
vreinterpretq_u16_u8(cols_2367.val[0]));
uint16x8x2_t rows_1537 = vtrnq_u16(vreinterpretq_u16_u8(cols_0145.val[1]),
vreinterpretq_u16_u8(cols_2367.val[1]));
uint8x16_t rows_04 = vreinterpretq_u8_u16(rows_0426.val[0]);
uint8x16_t rows_15 = vreinterpretq_u8_u16(rows_1537.val[0]);
uint8x16_t rows_26 = vreinterpretq_u8_u16(rows_0426.val[1]);
uint8x16_t rows_37 = vreinterpretq_u8_u16(rows_1537.val[1]);
JSAMPROW outptr0 = output_buf + DCTSIZE * 0;
JSAMPROW outptr1 = output_buf + DCTSIZE * 1;
JSAMPROW outptr2 = output_buf + DCTSIZE * 2;
JSAMPROW outptr3 = output_buf + DCTSIZE * 3;
JSAMPROW outptr4 = output_buf + DCTSIZE * 4;
JSAMPROW outptr5 = output_buf + DCTSIZE * 5;
JSAMPROW outptr6 = output_buf + DCTSIZE * 6;
JSAMPROW outptr7 = output_buf + DCTSIZE * 7;
/* Store DCT block to memory. */
vst1q_lane_u64((uint64_t *)outptr0, vreinterpretq_u64_u8(rows_04), 0);
vst1q_lane_u64((uint64_t *)outptr1, vreinterpretq_u64_u8(rows_15), 0);
vst1q_lane_u64((uint64_t *)outptr2, vreinterpretq_u64_u8(rows_26), 0);
vst1q_lane_u64((uint64_t *)outptr3, vreinterpretq_u64_u8(rows_37), 0);
vst1q_lane_u64((uint64_t *)outptr4, vreinterpretq_u64_u8(rows_04), 1);
vst1q_lane_u64((uint64_t *)outptr5, vreinterpretq_u64_u8(rows_15), 1);
vst1q_lane_u64((uint64_t *)outptr6, vreinterpretq_u64_u8(rows_26), 1);
vst1q_lane_u64((uint64_t *)outptr7, vreinterpretq_u64_u8(rows_37), 1);
}
static int flss(uint16_t val) {
int bit;
bit = 16;
if (!val)
return 0;
if (!(val & 0xff00)) {
bit -= 8;
val <<= 8;
}
if (!(val & 0xf000)) {
bit -= 4;
val <<= 4;
}
if (!(val & 0xc000)) {
bit -= 2;
val <<= 2;
}
if (!(val & 0x8000)) {
bit -= 1;
val <<= 1;
}
return bit;
}
static int compute_reciprocal(uint16_t divisor, DCTELEM *dtbl) {
UDCTELEM2 fq, fr;
UDCTELEM c;
int b, r;
if (divisor == 1) {
/* divisor == 1 means unquantized, so these reciprocal/correction/shift
* values will cause the C quantization algorithm to act like the
* identity function. Since only the C quantization algorithm is used in
* these cases, the scale value is irrelevant.
*/
dtbl[DCTSIZE2 * 0] = (DCTELEM)1; /* reciprocal */
dtbl[DCTSIZE2 * 1] = (DCTELEM)0; /* correction */
dtbl[DCTSIZE2 * 2] = (DCTELEM)1; /* scale */
dtbl[DCTSIZE2 * 3] = -(DCTELEM)(sizeof(DCTELEM) * 8); /* shift */
return 0;
}
b = flss(divisor) - 1;
r = sizeof(DCTELEM) * 8 + b;
fq = ((UDCTELEM2)1 << r) / divisor;
fr = ((UDCTELEM2)1 << r) % divisor;
c = divisor / 2; /* for rounding */
if (fr == 0) { /* divisor is power of two */
/* fq will be one bit too large to fit in DCTELEM, so adjust */
fq >>= 1;
r--;
} else if (fr <= (divisor / 2U)) { /* fractional part is < 0.5 */
c++;
} else { /* fractional part is > 0.5 */
fq++;
}
dtbl[DCTSIZE2 * 0] = (DCTELEM)fq; /* reciprocal */
dtbl[DCTSIZE2 * 1] = (DCTELEM)c; /* correction + roundfactor */
#ifdef WITH_SIMD
dtbl[DCTSIZE2 * 2] = (DCTELEM)(1 << (sizeof(DCTELEM) * 8 * 2 - r)); /* scale */
#else
dtbl[DCTSIZE2 * 2] = 1;
#endif
dtbl[DCTSIZE2 * 3] = (DCTELEM)r - sizeof(DCTELEM) * 8; /* shift */
if (r <= 16) return 0;
else return 1;
}
#define DESCALE(x, n) RIGHT_SHIFT(x, n)
/* Multiply a DCTELEM variable by an JLONG constant, and immediately
* descale to yield a DCTELEM result.
*/
#define MULTIPLY(var, const) ((DCTELEM)DESCALE((var) * (const), CONST_BITS))
#define MULTIPLY16V16(var1, var2) ((var1) * (var2))
static DCTELEM std_luminance_quant_tbl[DCTSIZE2] = {
16, 11, 10, 16, 24, 40, 51, 61,
12, 12, 14, 19, 26, 58, 60, 55,
14, 13, 16, 24, 40, 57, 69, 56,
14, 17, 22, 29, 51, 87, 80, 62,
18, 22, 37, 56, 68, 109, 103, 77,
24, 35, 55, 64, 81, 104, 113, 92,
49, 64, 78, 87, 103, 121, 120, 101,
72, 92, 95, 98, 112, 100, 103, 99
};
static int jpeg_quality_scaling(int quality)
/* Convert a user-specified quality rating to a percentage scaling factor
* for an underlying quantization table, using our recommended scaling curve.
* The input 'quality' factor should be 0 (terrible) to 100 (very good).
*/
{
/* Safety limit on quality factor. Convert 0 to 1 to avoid zero divide. */
if (quality <= 0) quality = 1;
if (quality > 100) quality = 100;
/* The basic table is used as-is (scaling 100) for a quality of 50.
* Qualities 50..100 are converted to scaling percentage 200 - 2*Q;
* note that at Q=100 the scaling is 0, which will cause jpeg_add_quant_table
* to make all the table entries 1 (hence, minimum quantization loss).
* Qualities 1..50 are converted to scaling percentage 5000/Q.
*/
if (quality < 50)
quality = 5000 / quality;
else
quality = 200 - quality * 2;
return quality;
}
static void jpeg_add_quant_table(DCTELEM *qtable, DCTELEM *basicTable, int scale_factor, bool forceBaseline)
/* Define a quantization table equal to the basic_table times
* a scale factor (given as a percentage).
* If force_baseline is TRUE, the computed quantization table entries
* are limited to 1..255 for JPEG baseline compatibility.
*/
{
int i;
long temp;
for (i = 0; i < DCTSIZE2; i++) {
temp = ((long)basicTable[i] * scale_factor + 50L) / 100L;
/* limit the values to the valid range */
if (temp <= 0L) temp = 1L;
if (temp > 32767L) temp = 32767L; /* max quantizer needed for 12 bits */
if (forceBaseline && temp > 255L)
temp = 255L; /* limit to baseline range if requested */
qtable[i] = (uint16_t)temp;
}
}
static void jpeg_set_quality(DCTELEM *qtable, int quality)
/* Set or change the 'quality' (quantization) setting, using default tables.
* This is the standard quality-adjusting entry point for typical user
* interfaces; only those who want detailed control over quantization tables
* would use the preceding three routines directly.
*/
{
/* Convert user 0-100 rating to percentage scaling */
quality = jpeg_quality_scaling(quality);
/* Set up standard quality tables */
jpeg_add_quant_table(qtable, std_luminance_quant_tbl, quality, false);
}
static void getDivisors(DCTELEM *dtbl, DCTELEM *qtable) {
#define CONST_BITS 14
#define RIGHT_SHIFT(x, shft) ((x) >> (shft))
static const int16_t aanscales[DCTSIZE2] = {
/* precomputed values scaled up by 14 bits */
16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
};
for (int i = 0; i < DCTSIZE2; i++) {
if (!compute_reciprocal(
DESCALE(MULTIPLY16V16((JLONG)qtable[i],
(JLONG)aanscales[i]),
CONST_BITS - 3), &dtbl[i])) {
//fdct->quantize = quantize;
printf("here\n");
}
}
}
static void quantize(JCOEFPTR coef_block, DCTELEM *divisors, DCTELEM *workspace)
{
int i;
DCTELEM temp;
JCOEFPTR output_ptr = coef_block;
UDCTELEM recip, corr;
int shift;
UDCTELEM2 product;
for (i = 0; i < DCTSIZE2; i++) {
temp = workspace[i];
recip = divisors[i + DCTSIZE2 * 0];
corr = divisors[i + DCTSIZE2 * 1];
shift = divisors[i + DCTSIZE2 * 3];
if (temp < 0) {
temp = -temp;
product = (UDCTELEM2)(temp + corr) * recip;
product >>= shift + sizeof(DCTELEM) * 8;
temp = (DCTELEM)product;
temp = -temp;
} else {
product = (UDCTELEM2)(temp + corr) * recip;
product >>= shift + sizeof(DCTELEM) * 8;
temp = (DCTELEM)product;
}
output_ptr[i] = (JCOEF)temp;
}
}
NSData *generateForwardDctData(int quality) {
NSMutableData *divisors = [[NSMutableData alloc] initWithLength:DCTSIZE2 * 4 * sizeof(DCTELEM)];
DCTELEM qtable[DCTSIZE2];
jpeg_set_quality(qtable, quality);
getDivisors((DCTELEM *)divisors.mutableBytes, qtable);
return divisors;
}
NSData *generateInverseDctData(int quality) {
NSMutableData *divisors = [[NSMutableData alloc] initWithLength:DCTSIZE2 * sizeof(IFAST_MULT_TYPE)];
IFAST_MULT_TYPE *ifmtbl = (IFAST_MULT_TYPE *)divisors.mutableBytes;
DCTELEM qtable[DCTSIZE2];
jpeg_set_quality(qtable, quality);
#define CONST_BITS 14
static const int16_t aanscales[DCTSIZE2] = {
/* precomputed values scaled up by 14 bits */
16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
};
for (int i = 0; i < DCTSIZE2; i++) {
ifmtbl[i] = (IFAST_MULT_TYPE)
DESCALE(MULTIPLY16V16((JLONG)qtable[i],
(JLONG)aanscales[i]),
CONST_BITS - IFAST_SCALE_BITS);
}
return divisors;
}
static const int zigZagInv[DCTSIZE2] = {
0,1,8,16,9,2,3,10,
17,24,32,25,18,11,4,5,
12,19,26,33,40,48,41,34,
27,20,13,6,7,14,21,28,
35,42,49,56,57,50,43,36,
29,22,15,23,30,37,44,51,
58,59,52,45,38,31,39,46,
53,60,61,54,47,55,62,63
};
static const int zigZag[DCTSIZE2] = {
0, 1, 8, 16, 9, 2, 3, 10, 17, 24, 32, 25, 18, 11, 4, 5, 12, 19, 26, 33, 40, 48, 41, 34, 27, 20, 13, 6, 7, 14, 21, 28, 35, 42, 49, 56, 57, 50, 43, 36, 29, 22, 15, 23, 30, 37, 44, 51, 58, 59, 52, 45, 38, 31, 39, 46, 53, 60, 61, 54, 47, 55, 62, 63
};
void performForwardDct(uint8_t const *pixels, int16_t *coefficients, int width, int height, int bytesPerRow, NSData *dctData) {
DCTELEM *divisors = (DCTELEM *)dctData.bytes;
DCTELEM block[DCTSIZE2];
JCOEF coefBlock[DCTSIZE2];
for (int y = 0; y < height; y += DCTSIZE) {
for (int x = 0; x < width; x += DCTSIZE) {
for (int blockY = 0; blockY < DCTSIZE; blockY++) {
for (int blockX = 0; blockX < DCTSIZE; blockX++) {
block[blockY * DCTSIZE + blockX] = ((DCTELEM)pixels[(y + blockY) * bytesPerRow + (x + blockX)]) - CENTERJSAMPLE;
}
}
jsimd_fdct_ifast_neon(block);
quantize(coefBlock, divisors, block);
for (int blockY = 0; blockY < DCTSIZE; blockY++) {
for (int blockX = 0; blockX < DCTSIZE; blockX++) {
coefficients[(y + blockY) * bytesPerRow + (x + blockX)] = coefBlock[zigZagInv[blockY * DCTSIZE + blockX]];
}
}
}
}
}
void performInverseDct(int16_t const *coefficients, uint8_t *pixels, int width, int height, int coefficientsPerRow, int bytesPerRow, NSData *idctData) {
IFAST_MULT_TYPE *ifmtbl = (IFAST_MULT_TYPE *)idctData.bytes;
DCTELEM coefficientBlock[DCTSIZE2];
JSAMPLE pixelBlock[DCTSIZE2];
for (int y = 0; y < height; y += DCTSIZE) {
for (int x = 0; x < width; x += DCTSIZE) {
for (int blockY = 0; blockY < DCTSIZE; blockY++) {
for (int blockX = 0; blockX < DCTSIZE; blockX++) {
coefficientBlock[zigZag[blockY * DCTSIZE + blockX]] = coefficients[(y + blockY) * coefficientsPerRow + (x + blockX)];
}
}
jsimd_idct_ifast_neon(ifmtbl, coefficientBlock, pixelBlock);
for (int blockY = 0; blockY < DCTSIZE; blockY++) {
for (int blockX = 0; blockX < DCTSIZE; blockX++) {
pixels[(y + blockY) * bytesPerRow + (x + blockX)] = pixelBlock[blockY * DCTSIZE + blockX];
}
}
}
}
}

7
submodules/TelegramUI/Components/AnimationCache/DCT/BUILD → submodules/TelegramUI/Components/AnimationCache/ImageDCT/BUILD
View File

@@ -1,10 +1,13 @@
objc_library(
name = "DCT",
name = "ImageDCT",
enable_modules = True,
module_name = "DCT",
module_name = "ImageDCT",
srcs = glob([
"Sources/**/*.m",
"Sources/**/*.mm",
"Sources/**/*.c",
"Sources/**/*.cpp",
"Sources/**/*.h",
]),
hdrs = glob([

17
submodules/TelegramUI/Components/AnimationCache/ImageDCT/PublicHeaders/ImageDCT/ImageDCT.h Normal file
View File

@@ -0,0 +1,17 @@
#ifndef DctImageTransform_h
#define DctImageTransform_h
#import <Foundation/Foundation.h>
#import <ImageDCT/YuvConversion.h>
@interface ImageDCT : NSObject
- (instancetype _Nonnull)initWithQuality:(NSInteger)quality;
- (void)forwardWithPixels:(uint8_t const * _Nonnull)pixels coefficients:(int16_t * _Nonnull)coefficients width:(NSInteger)width height:(NSInteger)height bytesPerRow:(NSInteger)bytesPerRow __attribute__((objc_direct));
- (void)inverseWithCoefficients:(int16_t const * _Nonnull)coefficients pixels:(uint8_t * _Nonnull)pixels width:(NSInteger)width height:(NSInteger)height coefficientsPerRow:(NSInteger)coefficientsPerRow bytesPerRow:(NSInteger)bytesPerRow __attribute__((objc_direct));
@end
#endif /* DctImageTransform_h */

0
submodules/TelegramUI/Components/AnimationCache/DCT/PublicHeaders/DCT/YuvConversion.h → submodules/TelegramUI/Components/AnimationCache/ImageDCT/PublicHeaders/ImageDCT/YuvConversion.h
View File

376
submodules/TelegramUI/Components/AnimationCache/ImageDCT/Sources/DCT.cpp Normal file
View File

@@ -0,0 +1,376 @@
#import "DCT.h"
#include "DCTCommon.h"
#include <vector>
#define DCTSIZE 8 /* The basic DCT block is 8x8 samples */
#define DCTSIZE2 64 /* DCTSIZE squared; # of elements in a block */
typedef unsigned short UDCTELEM;
typedef unsigned int UDCTELEM2;
typedef long JLONG;
#define MULTIPLIER short /* prefer 16-bit with SIMD for parellelism */
typedef MULTIPLIER IFAST_MULT_TYPE; /* 16 bits is OK, use short if faster */
#define IFAST_SCALE_BITS 2 /* fractional bits in scale factors */
#define CENTERJSAMPLE 128
namespace {
int flss(uint16_t val) {
int bit;
bit = 16;
if (!val)
return 0;
if (!(val & 0xff00)) {
bit -= 8;
val <<= 8;
}
if (!(val & 0xf000)) {
bit -= 4;
val <<= 4;
}
if (!(val & 0xc000)) {
bit -= 2;
val <<= 2;
}
if (!(val & 0x8000)) {
bit -= 1;
val <<= 1;
}
return bit;
}
int compute_reciprocal(uint16_t divisor, DCTELEM *dtbl) {
UDCTELEM2 fq, fr;
UDCTELEM c;
int b, r;
if (divisor == 1) {
/* divisor == 1 means unquantized, so these reciprocal/correction/shift
* values will cause the C quantization algorithm to act like the
* identity function. Since only the C quantization algorithm is used in
* these cases, the scale value is irrelevant.
*/
dtbl[DCTSIZE2 * 0] = (DCTELEM)1; /* reciprocal */
dtbl[DCTSIZE2 * 1] = (DCTELEM)0; /* correction */
dtbl[DCTSIZE2 * 2] = (DCTELEM)1; /* scale */
dtbl[DCTSIZE2 * 3] = -(DCTELEM)(sizeof(DCTELEM) * 8); /* shift */
return 0;
}
b = flss(divisor) - 1;
r = sizeof(DCTELEM) * 8 + b;
fq = ((UDCTELEM2)1 << r) / divisor;
fr = ((UDCTELEM2)1 << r) % divisor;
c = divisor / 2; /* for rounding */
if (fr == 0) { /* divisor is power of two */
/* fq will be one bit too large to fit in DCTELEM, so adjust */
fq >>= 1;
r--;
} else if (fr <= (divisor / 2U)) { /* fractional part is < 0.5 */
c++;
} else { /* fractional part is > 0.5 */
fq++;
}
dtbl[DCTSIZE2 * 0] = (DCTELEM)fq; /* reciprocal */
dtbl[DCTSIZE2 * 1] = (DCTELEM)c; /* correction + roundfactor */
#ifdef WITH_SIMD
dtbl[DCTSIZE2 * 2] = (DCTELEM)(1 << (sizeof(DCTELEM) * 8 * 2 - r)); /* scale */
#else
dtbl[DCTSIZE2 * 2] = 1;
#endif
dtbl[DCTSIZE2 * 3] = (DCTELEM)r - sizeof(DCTELEM) * 8; /* shift */
if (r <= 16) return 0;
else return 1;
}
#define DESCALE(x, n) RIGHT_SHIFT(x, n)
/* Multiply a DCTELEM variable by an JLONG constant, and immediately
* descale to yield a DCTELEM result.
*/
#define MULTIPLY(var, const) ((DCTELEM)DESCALE((var) * (const), CONST_BITS))
#define MULTIPLY16V16(var1, var2) ((var1) * (var2))
static DCTELEM std_luminance_quant_tbl[DCTSIZE2] = {
16, 11, 10, 16, 24, 40, 51, 61,
12, 12, 14, 19, 26, 58, 60, 55,
14, 13, 16, 24, 40, 57, 69, 56,
14, 17, 22, 29, 51, 87, 80, 62,
18, 22, 37, 56, 68, 109, 103, 77,
24, 35, 55, 64, 81, 104, 113, 92,
49, 64, 78, 87, 103, 121, 120, 101,
72, 92, 95, 98, 112, 100, 103, 99
};
int jpeg_quality_scaling(int quality)
/* Convert a user-specified quality rating to a percentage scaling factor
* for an underlying quantization table, using our recommended scaling curve.
* The input 'quality' factor should be 0 (terrible) to 100 (very good).
*/
{
/* Safety limit on quality factor. Convert 0 to 1 to avoid zero divide. */
if (quality <= 0) quality = 1;
if (quality > 100) quality = 100;
/* The basic table is used as-is (scaling 100) for a quality of 50.
* Qualities 50..100 are converted to scaling percentage 200 - 2*Q;
* note that at Q=100 the scaling is 0, which will cause jpeg_add_quant_table
* to make all the table entries 1 (hence, minimum quantization loss).
* Qualities 1..50 are converted to scaling percentage 5000/Q.
*/
if (quality < 50)
quality = 5000 / quality;
else
quality = 200 - quality * 2;
return quality;
}
void jpeg_add_quant_table(DCTELEM *qtable, DCTELEM *basicTable, int scale_factor, bool forceBaseline)
/* Define a quantization table equal to the basic_table times
* a scale factor (given as a percentage).
* If force_baseline is TRUE, the computed quantization table entries
* are limited to 1..255 for JPEG baseline compatibility.
*/
{
int i;
long temp;
for (i = 0; i < DCTSIZE2; i++) {
temp = ((long)basicTable[i] * scale_factor + 50L) / 100L;
/* limit the values to the valid range */
if (temp <= 0L) temp = 1L;
if (temp > 32767L) temp = 32767L; /* max quantizer needed for 12 bits */
if (forceBaseline && temp > 255L)
temp = 255L; /* limit to baseline range if requested */
qtable[i] = (uint16_t)temp;
}
}
void jpeg_set_quality(DCTELEM *qtable, int quality)
/* Set or change the 'quality' (quantization) setting, using default tables.
* This is the standard quality-adjusting entry point for typical user
* interfaces; only those who want detailed control over quantization tables
* would use the preceding three routines directly.
*/
{
/* Convert user 0-100 rating to percentage scaling */
quality = jpeg_quality_scaling(quality);
/* Set up standard quality tables */
jpeg_add_quant_table(qtable, std_luminance_quant_tbl, quality, false);
}
void getDivisors(DCTELEM *dtbl, DCTELEM *qtable) {
#define CONST_BITS 14
#define RIGHT_SHIFT(x, shft) ((x) >> (shft))
static const int16_t aanscales[DCTSIZE2] = {
/* precomputed values scaled up by 14 bits */
16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
};
for (int i = 0; i < DCTSIZE2; i++) {
if (!compute_reciprocal(
DESCALE(MULTIPLY16V16((JLONG)qtable[i],
(JLONG)aanscales[i]),
CONST_BITS - 3), &dtbl[i])) {
}
}
}
void quantize(JCOEFPTR coef_block, DCTELEM *divisors, DCTELEM *workspace)
{
int i;
DCTELEM temp;
JCOEFPTR output_ptr = coef_block;
UDCTELEM recip, corr;
int shift;
UDCTELEM2 product;
for (i = 0; i < DCTSIZE2; i++) {
temp = workspace[i];
recip = divisors[i + DCTSIZE2 * 0];
corr = divisors[i + DCTSIZE2 * 1];
shift = divisors[i + DCTSIZE2 * 3];
if (temp < 0) {
temp = -temp;
product = (UDCTELEM2)(temp + corr) * recip;
product >>= shift + sizeof(DCTELEM) * 8;
temp = (DCTELEM)product;
temp = -temp;
} else {
product = (UDCTELEM2)(temp + corr) * recip;
product >>= shift + sizeof(DCTELEM) * 8;
temp = (DCTELEM)product;
}
output_ptr[i] = (JCOEF)temp;
}
}
void generateForwardDctData(int quality, std::vector<uint8_t> &data) {
data.resize(DCTSIZE2 * 4 * sizeof(DCTELEM));
DCTELEM qtable[DCTSIZE2];
jpeg_set_quality(qtable, quality);
getDivisors((DCTELEM *)data.data(), qtable);
}
void generateInverseDctData(int quality, std::vector<uint8_t> &data) {
data.resize(DCTSIZE2 * sizeof(IFAST_MULT_TYPE));
IFAST_MULT_TYPE *ifmtbl = (IFAST_MULT_TYPE *)data.data();
DCTELEM qtable[DCTSIZE2];
jpeg_set_quality(qtable, quality);
#define CONST_BITS 14
static const int16_t aanscales[DCTSIZE2] = {
/* precomputed values scaled up by 14 bits */
16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270,
21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906,
19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315,
16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520,
12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552,
8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446,
4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247
};
for (int i = 0; i < DCTSIZE2; i++) {
ifmtbl[i] = (IFAST_MULT_TYPE)
DESCALE(MULTIPLY16V16((JLONG)qtable[i],
(JLONG)aanscales[i]),
CONST_BITS - IFAST_SCALE_BITS);
}
}
static const int zigZagInv[DCTSIZE2] = {
0,1,8,16,9,2,3,10,
17,24,32,25,18,11,4,5,
12,19,26,33,40,48,41,34,
27,20,13,6,7,14,21,28,
35,42,49,56,57,50,43,36,
29,22,15,23,30,37,44,51,
58,59,52,45,38,31,39,46,
53,60,61,54,47,55,62,63
};
static const int zigZag[DCTSIZE2] = {
0, 1, 8, 16, 9, 2, 3, 10, 17, 24, 32, 25, 18, 11, 4, 5, 12, 19, 26, 33, 40, 48, 41, 34, 27, 20, 13, 6, 7, 14, 21, 28, 35, 42, 49, 56, 57, 50, 43, 36, 29, 22, 15, 23, 30, 37, 44, 51, 58, 59, 52, 45, 38, 31, 39, 46, 53, 60, 61, 54, 47, 55, 62, 63
};
void performForwardDct(uint8_t const *pixels, int16_t *coefficients, int width, int height, int bytesPerRow, DCTELEM *divisors) {
DCTELEM block[DCTSIZE2];
JCOEF coefBlock[DCTSIZE2];
for (int y = 0; y < height; y += DCTSIZE) {
for (int x = 0; x < width; x += DCTSIZE) {
for (int blockY = 0; blockY < DCTSIZE; blockY++) {
for (int blockX = 0; blockX < DCTSIZE; blockX++) {
block[blockY * DCTSIZE + blockX] = ((DCTELEM)pixels[(y + blockY) * bytesPerRow + (x + blockX)]) - CENTERJSAMPLE;
}
}
dct_jpeg_fdct_ifast(block);
quantize(coefBlock, divisors, block);
for (int blockY = 0; blockY < DCTSIZE; blockY++) {
for (int blockX = 0; blockX < DCTSIZE; blockX++) {
coefficients[(y + blockY) * bytesPerRow + (x + blockX)] = coefBlock[zigZagInv[blockY * DCTSIZE + blockX]];
}
}
}
}
}
void performInverseDct(int16_t const * coefficients, uint8_t *pixels, int width, int height, int coefficientsPerRow, int bytesPerRow, DctAuxiliaryData *auxiliaryData, IFAST_MULT_TYPE *ifmtbl) {
DCTELEM coefficientBlock[DCTSIZE2];
JSAMPLE pixelBlock[DCTSIZE2];
for (int y = 0; y < height; y += DCTSIZE) {
for (int x = 0; x < width; x += DCTSIZE) {
for (int blockY = 0; blockY < DCTSIZE; blockY++) {
for (int blockX = 0; blockX < DCTSIZE; blockX++) {
coefficientBlock[zigZag[blockY * DCTSIZE + blockX]] = coefficients[(y + blockY) * coefficientsPerRow + (x + blockX)];
}
}
dct_jpeg_idct_ifast(auxiliaryData, ifmtbl, coefficientBlock, pixelBlock);
for (int blockY = 0; blockY < DCTSIZE; blockY++) {
for (int blockX = 0; blockX < DCTSIZE; blockX++) {
pixels[(y + blockY) * bytesPerRow + (x + blockX)] = pixelBlock[blockY * DCTSIZE + blockX];
}
}
}
}
}
}
namespace dct {
class DCTInternal {
public:
DCTInternal(int quality) {
auxiliaryData = createDctAuxiliaryData();
generateForwardDctData(quality, forwardDctData);
generateInverseDctData(quality, inverseDctData);
}
~DCTInternal() {
freeDctAuxiliaryData(auxiliaryData);
}
public:
struct DctAuxiliaryData *auxiliaryData = nullptr;
std::vector<uint8_t> forwardDctData;
std::vector<uint8_t> inverseDctData;
};
DCT::DCT(int quality) {
_internal = new DCTInternal(quality);
}
DCT::~DCT() {
delete _internal;
}
void DCT::forward(uint8_t const *pixels, int16_t *coefficients, int width, int height, int bytesPerRow) {
performForwardDct(pixels, coefficients, width, height, bytesPerRow, (DCTELEM *)_internal->forwardDctData.data());
}
void DCT::inverse(int16_t const *coefficients, uint8_t *pixels, int width, int height, int coefficientsPerRow, int bytesPerRow) {
performInverseDct(coefficients, pixels, width, height, coefficientsPerRow, bytesPerRow, _internal->auxiliaryData, (IFAST_MULT_TYPE *)_internal->inverseDctData.data());
}
}

26
submodules/TelegramUI/Components/AnimationCache/ImageDCT/Sources/DCT.h Normal file
View File

@@ -0,0 +1,26 @@
#ifndef DCT_H
#define DCT_H
#include "DCTCommon.h"
#include <stdint.h>
namespace dct {
class DCTInternal;
class DCT {
public:
DCT(int quality);
~DCT();
void forward(uint8_t const *pixels, int16_t *coefficients, int width, int height, int bytesPerRow);
void inverse(int16_t const *coefficients, uint8_t *pixels, int width, int height, int coefficientsPerRow, int bytesPerRow);
private:
DCTInternal *_internal;
};
}
#endif

27
submodules/TelegramUI/Components/AnimationCache/ImageDCT/Sources/DCTCommon.h Normal file
View File

@@ -0,0 +1,27 @@
#ifndef DCT_COMMON_H
#define DCT_COMMON_H
#ifdef __cplusplus
extern "C" {
#endif
typedef short DCTELEM;
typedef short JCOEF;
typedef JCOEF *JCOEFPTR;
typedef unsigned char JSAMPLE;
typedef JSAMPLE *JSAMPROW;
struct DctAuxiliaryData;
struct DctAuxiliaryData *createDctAuxiliaryData();
void freeDctAuxiliaryData(struct DctAuxiliaryData *data);
void dct_jpeg_idct_ifast(struct DctAuxiliaryData *auxiliaryData, void *dct_table, JCOEFPTR coef_block, JSAMPROW output_buf);
void dct_jpeg_fdct_ifast(DCTELEM *data);
#ifdef __cplusplus
}
#endif
#endif

399
submodules/TelegramUI/Components/AnimationCache/ImageDCT/Sources/DCT_C.c Normal file
View File

@@ -0,0 +1,399 @@
#import "DCTCommon.h"
#if !defined(__aarch64__)
#include <string.h>
#include <stdlib.h>
typedef long JLONG;
#define CONST_BITS 8
#define PASS1_BITS 2
#define DCTSIZE 8 /* The basic DCT block is 8x8 samples */
#define DCTSIZE2 64 /* DCTSIZE squared; # of elements in a block */
#define FIX_0_382683433 ((JLONG)98) /* FIX(0.382683433) */
#define FIX_0_541196100 ((JLONG)139) /* FIX(0.541196100) */
#define FIX_0_707106781 ((JLONG)181) /* FIX(0.707106781) */
#define FIX_1_306562965 ((JLONG)334) /* FIX(1.306562965) */
#define FIX_1_082392200 ((JLONG)277) /* FIX(1.082392200) */
#define FIX_1_414213562 ((JLONG)362) /* FIX(1.414213562) */
#define FIX_1_847759065 ((JLONG)473) /* FIX(1.847759065) */
#define FIX_2_613125930 ((JLONG)669) /* FIX(2.613125930) */
#define RIGHT_SHIFT(x, shft) ((x) >> (shft))
#define IRIGHT_SHIFT(x, shft) ((x) >> (shft))
#define DESCALE(x, n) RIGHT_SHIFT(x, n)
#define IDESCALE(x, n) ((int)IRIGHT_SHIFT(x, n))
#define MULTIPLY(var, const) ((DCTELEM)DESCALE((var) * (const), CONST_BITS))
#define MULTIPLIER short /* prefer 16-bit with SIMD for parellelism */
typedef MULTIPLIER IFAST_MULT_TYPE; /* 16 bits is OK, use short if faster */
#define DEQUANTIZE(coef, quantval) (((IFAST_MULT_TYPE)(coef)) * (quantval))
#define RANGE_MASK (MAXJSAMPLE * 4 + 3) /* 2 bits wider than legal samples */
#define MAXJSAMPLE 255
#define CENTERJSAMPLE 128
typedef JSAMPROW *JSAMPARRAY; /* ptr to some rows (a 2-D sample array) */
typedef JSAMPARRAY *JSAMPIMAGE; /* a 3-D sample array: top index is color */
#define IDCT_range_limit(cinfo) ((cinfo)->sample_range_limit + CENTERJSAMPLE)
void dct_jpeg_fdct_ifast(DCTELEM *data)
{
DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
DCTELEM tmp10, tmp11, tmp12, tmp13;
DCTELEM z1, z2, z3, z4, z5, z11, z13;
DCTELEM *dataptr;
int ctr;
/* Pass 1: process rows. */
dataptr = data;
for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
tmp0 = dataptr[0] + dataptr[7];
tmp7 = dataptr[0] - dataptr[7];
tmp1 = dataptr[1] + dataptr[6];
tmp6 = dataptr[1] - dataptr[6];
tmp2 = dataptr[2] + dataptr[5];
tmp5 = dataptr[2] - dataptr[5];
tmp3 = dataptr[3] + dataptr[4];
tmp4 = dataptr[3] - dataptr[4];
/* Even part */
tmp10 = tmp0 + tmp3; /* phase 2 */
tmp13 = tmp0 - tmp3;
tmp11 = tmp1 + tmp2;
tmp12 = tmp1 - tmp2;
dataptr[0] = tmp10 + tmp11; /* phase 3 */
dataptr[4] = tmp10 - tmp11;
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
dataptr[2] = tmp13 + z1; /* phase 5 */
dataptr[6] = tmp13 - z1;
/* Odd part */
tmp10 = tmp4 + tmp5; /* phase 2 */
tmp11 = tmp5 + tmp6;
tmp12 = tmp6 + tmp7;
/* The rotator is modified from fig 4-8 to avoid extra negations. */
z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
z11 = tmp7 + z3; /* phase 5 */
z13 = tmp7 - z3;
dataptr[5] = z13 + z2; /* phase 6 */
dataptr[3] = z13 - z2;
dataptr[1] = z11 + z4;
dataptr[7] = z11 - z4;
dataptr += DCTSIZE; /* advance pointer to next row */
}
/* Pass 2: process columns. */
dataptr = data;
for (ctr = DCTSIZE - 1; ctr >= 0; ctr--) {
tmp0 = dataptr[DCTSIZE * 0] + dataptr[DCTSIZE * 7];
tmp7 = dataptr[DCTSIZE * 0] - dataptr[DCTSIZE * 7];
tmp1 = dataptr[DCTSIZE * 1] + dataptr[DCTSIZE * 6];
tmp6 = dataptr[DCTSIZE * 1] - dataptr[DCTSIZE * 6];
tmp2 = dataptr[DCTSIZE * 2] + dataptr[DCTSIZE * 5];
tmp5 = dataptr[DCTSIZE * 2] - dataptr[DCTSIZE * 5];
tmp3 = dataptr[DCTSIZE * 3] + dataptr[DCTSIZE * 4];
tmp4 = dataptr[DCTSIZE * 3] - dataptr[DCTSIZE * 4];
/* Even part */
tmp10 = tmp0 + tmp3; /* phase 2 */
tmp13 = tmp0 - tmp3;
tmp11 = tmp1 + tmp2;
tmp12 = tmp1 - tmp2;
dataptr[DCTSIZE * 0] = tmp10 + tmp11; /* phase 3 */
dataptr[DCTSIZE * 4] = tmp10 - tmp11;
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
dataptr[DCTSIZE * 2] = tmp13 + z1; /* phase 5 */
dataptr[DCTSIZE * 6] = tmp13 - z1;
/* Odd part */
tmp10 = tmp4 + tmp5; /* phase 2 */
tmp11 = tmp5 + tmp6;
tmp12 = tmp6 + tmp7;
/* The rotator is modified from fig 4-8 to avoid extra negations. */
z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
z11 = tmp7 + z3; /* phase 5 */
z13 = tmp7 - z3;
dataptr[DCTSIZE * 5] = z13 + z2; /* phase 6 */
dataptr[DCTSIZE * 3] = z13 - z2;
dataptr[DCTSIZE * 1] = z11 + z4;
dataptr[DCTSIZE * 7] = z11 - z4;
dataptr++; /* advance pointer to next column */
}
}
struct DctAuxiliaryData {
JSAMPLE *allocated_sample_range_limit;
JSAMPLE *sample_range_limit;
};
static void prepare_range_limit_table(struct DctAuxiliaryData *data)
/* Allocate and fill in the sample_range_limit table */
{
JSAMPLE *table;
int i;
table = (JSAMPLE *)malloc((5 * (MAXJSAMPLE + 1) + CENTERJSAMPLE) * sizeof(JSAMPLE));
data->allocated_sample_range_limit = table;
table += (MAXJSAMPLE + 1); /* allow negative subscripts of simple table */
data->sample_range_limit = table;
/* First segment of "simple" table: limit[x] = 0 for x < 0 */
memset(table - (MAXJSAMPLE + 1), 0, (MAXJSAMPLE + 1) * sizeof(JSAMPLE));
/* Main part of "simple" table: limit[x] = x */
for (i = 0; i <= MAXJSAMPLE; i++)
table[i] = (JSAMPLE)i;
table += CENTERJSAMPLE; /* Point to where post-IDCT table starts */
/* End of simple table, rest of first half of post-IDCT table */
for (i = CENTERJSAMPLE; i < 2 * (MAXJSAMPLE + 1); i++)
table[i] = MAXJSAMPLE;
/* Second half of post-IDCT table */
memset(table + (2 * (MAXJSAMPLE + 1)), 0,
(2 * (MAXJSAMPLE + 1) - CENTERJSAMPLE) * sizeof(JSAMPLE));
memcpy(table + (4 * (MAXJSAMPLE + 1) - CENTERJSAMPLE),
data->sample_range_limit, CENTERJSAMPLE * sizeof(JSAMPLE));
}
struct DctAuxiliaryData *createDctAuxiliaryData() {
struct DctAuxiliaryData *result = malloc(sizeof(struct DctAuxiliaryData));
memset(result, 0, sizeof(struct DctAuxiliaryData));
prepare_range_limit_table(result);
return result;
}
void freeDctAuxiliaryData(struct DctAuxiliaryData *data) {
if (data) {
free(data->allocated_sample_range_limit);
free(data);
}
}
void dct_jpeg_idct_ifast(struct DctAuxiliaryData *auxiliaryData, void *dct_table, JCOEFPTR coef_block, JSAMPROW output_buf) {
DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
DCTELEM tmp10, tmp11, tmp12, tmp13;
DCTELEM z5, z10, z11, z12, z13;
JCOEFPTR inptr;
IFAST_MULT_TYPE *quantptr;
int *wsptr;
JSAMPROW outptr;
JSAMPLE *range_limit = IDCT_range_limit(auxiliaryData);
int ctr;
int workspace[DCTSIZE2]; /* buffers data between passes */
/* Pass 1: process columns from input, store into work array. */
inptr = coef_block;
quantptr = dct_table;
wsptr = workspace;
for (ctr = DCTSIZE; ctr > 0; ctr--) {
/* Due to quantization, we will usually find that many of the input
* coefficients are zero, especially the AC terms. We can exploit this
* by short-circuiting the IDCT calculation for any column in which all
* the AC terms are zero. In that case each output is equal to the
* DC coefficient (with scale factor as needed).
* With typical images and quantization tables, half or more of the
* column DCT calculations can be simplified this way.
*/
if (inptr[DCTSIZE * 1] == 0 && inptr[DCTSIZE * 2] == 0 &&
inptr[DCTSIZE * 3] == 0 && inptr[DCTSIZE * 4] == 0 &&
inptr[DCTSIZE * 5] == 0 && inptr[DCTSIZE * 6] == 0 &&
inptr[DCTSIZE * 7] == 0) {
/* AC terms all zero */
int dcval = (int)DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
wsptr[DCTSIZE * 0] = dcval;
wsptr[DCTSIZE * 1] = dcval;
wsptr[DCTSIZE * 2] = dcval;
wsptr[DCTSIZE * 3] = dcval;
wsptr[DCTSIZE * 4] = dcval;
wsptr[DCTSIZE * 5] = dcval;
wsptr[DCTSIZE * 6] = dcval;
wsptr[DCTSIZE * 7] = dcval;
inptr++; /* advance pointers to next column */
quantptr++;
wsptr++;
continue;
}
/* Even part */
tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
tmp1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2]);
tmp2 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
tmp3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6]);
tmp10 = tmp0 + tmp2; /* phase 3 */
tmp11 = tmp0 - tmp2;
tmp13 = tmp1 + tmp3; /* phases 5-3 */
tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */
tmp0 = tmp10 + tmp13; /* phase 2 */
tmp3 = tmp10 - tmp13;
tmp1 = tmp11 + tmp12;
tmp2 = tmp11 - tmp12;
/* Odd part */
tmp4 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1]);
tmp5 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3]);
tmp6 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5]);
tmp7 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7]);
z13 = tmp6 + tmp5; /* phase 6 */
z10 = tmp6 - tmp5;
z11 = tmp4 + tmp7;
z12 = tmp4 - tmp7;
tmp7 = z11 + z13; /* phase 5 */
tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
tmp12 = MULTIPLY(z10, -FIX_2_613125930) + z5; /* -2*(c2+c6) */
tmp6 = tmp12 - tmp7; /* phase 2 */
tmp5 = tmp11 - tmp6;
tmp4 = tmp10 + tmp5;
wsptr[DCTSIZE * 0] = (int)(tmp0 + tmp7);
wsptr[DCTSIZE * 7] = (int)(tmp0 - tmp7);
wsptr[DCTSIZE * 1] = (int)(tmp1 + tmp6);
wsptr[DCTSIZE * 6] = (int)(tmp1 - tmp6);
wsptr[DCTSIZE * 2] = (int)(tmp2 + tmp5);
wsptr[DCTSIZE * 5] = (int)(tmp2 - tmp5);
wsptr[DCTSIZE * 4] = (int)(tmp3 + tmp4);
wsptr[DCTSIZE * 3] = (int)(tmp3 - tmp4);
inptr++; /* advance pointers to next column */
quantptr++;
wsptr++;
}
/* Pass 2: process rows from work array, store into output array. */
/* Note that we must descale the results by a factor of 8 == 2**3, */
/* and also undo the PASS1_BITS scaling. */
wsptr = workspace;
for (ctr = 0; ctr < DCTSIZE; ctr++) {
outptr = output_buf + ctr * DCTSIZE;
/* Rows of zeroes can be exploited in the same way as we did with columns.
* However, the column calculation has created many nonzero AC terms, so
* the simplification applies less often (typically 5% to 10% of the time).
* On machines with very fast multiplication, it's possible that the
* test takes more time than it's worth. In that case this section
* may be commented out.
*/
#ifndef NO_ZERO_ROW_TEST
if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
/* AC terms all zero */
JSAMPLE dcval =
range_limit[IDESCALE(wsptr[0], PASS1_BITS + 3) & RANGE_MASK];
outptr[0] = dcval;
outptr[1] = dcval;
outptr[2] = dcval;
outptr[3] = dcval;
outptr[4] = dcval;
outptr[5] = dcval;
outptr[6] = dcval;
outptr[7] = dcval;
wsptr += DCTSIZE; /* advance pointer to next row */
continue;
}
#endif
/* Even part */
tmp10 = ((DCTELEM)wsptr[0] + (DCTELEM)wsptr[4]);
tmp11 = ((DCTELEM)wsptr[0] - (DCTELEM)wsptr[4]);
tmp13 = ((DCTELEM)wsptr[2] + (DCTELEM)wsptr[6]);
tmp12 =
MULTIPLY((DCTELEM)wsptr[2] - (DCTELEM)wsptr[6], FIX_1_414213562) - tmp13;
tmp0 = tmp10 + tmp13;
tmp3 = tmp10 - tmp13;
tmp1 = tmp11 + tmp12;
tmp2 = tmp11 - tmp12;
/* Odd part */
z13 = (DCTELEM)wsptr[5] + (DCTELEM)wsptr[3];
z10 = (DCTELEM)wsptr[5] - (DCTELEM)wsptr[3];
z11 = (DCTELEM)wsptr[1] + (DCTELEM)wsptr[7];
z12 = (DCTELEM)wsptr[1] - (DCTELEM)wsptr[7];
tmp7 = z11 + z13; /* phase 5 */
tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
tmp12 = MULTIPLY(z10, -FIX_2_613125930) + z5; /* -2*(c2+c6) */
tmp6 = tmp12 - tmp7; /* phase 2 */
tmp5 = tmp11 - tmp6;
tmp4 = tmp10 + tmp5;
/* Final output stage: scale down by a factor of 8 and range-limit */
outptr[0] =
range_limit[IDESCALE(tmp0 + tmp7, PASS1_BITS + 3) & RANGE_MASK];
outptr[7] =
range_limit[IDESCALE(tmp0 - tmp7, PASS1_BITS + 3) & RANGE_MASK];
outptr[1] =
range_limit[IDESCALE(tmp1 + tmp6, PASS1_BITS + 3) & RANGE_MASK];
outptr[6] =
range_limit[IDESCALE(tmp1 - tmp6, PASS1_BITS + 3) & RANGE_MASK];
outptr[2] =
range_limit[IDESCALE(tmp2 + tmp5, PASS1_BITS + 3) & RANGE_MASK];
outptr[5] =
range_limit[IDESCALE(tmp2 - tmp5, PASS1_BITS + 3) & RANGE_MASK];
outptr[4] =
range_limit[IDESCALE(tmp3 + tmp4, PASS1_BITS + 3) & RANGE_MASK];
outptr[3] =
range_limit[IDESCALE(tmp3 - tmp4, PASS1_BITS + 3) & RANGE_MASK];
wsptr += DCTSIZE; /* advance pointer to next row */
}
}
#endif

677
submodules/TelegramUI/Components/AnimationCache/ImageDCT/Sources/DCT_Neon.c Normal file
View File

@@ -0,0 +1,677 @@
#import "DCTCommon.h"
#include <stdlib.h>
#if defined(__aarch64__)
typedef long JLONG;
#define GETJSAMPLE(value) ((int)(value))
#define MAXJSAMPLE 255
#define CENTERJSAMPLE 128
typedef unsigned int JDIMENSION;
#define JPEG_MAX_DIMENSION 65500L /* a tad under 64K to prevent overflows */
#define MULTIPLIER short /* prefer 16-bit with SIMD for parellelism */
typedef MULTIPLIER IFAST_MULT_TYPE; /* 16 bits is OK, use short if faster */
#define IFAST_SCALE_BITS 2 /* fractional bits in scale factors */
/* Various constants determining the sizes of things.
* All of these are specified by the JPEG standard, so don't change them
* if you want to be compatible.
*/
#define DCTSIZE 8 /* The basic DCT block is 8x8 samples */
#define DCTSIZE2 64 /* DCTSIZE squared; # of elements in a block */
#define NUM_QUANT_TBLS 4 /* Quantization tables are numbered 0..3 */
#define NUM_HUFF_TBLS 4 /* Huffman tables are numbered 0..3 */
#define NUM_ARITH_TBLS 16 /* Arith-coding tables are numbered 0..15 */
#define MAX_COMPS_IN_SCAN 4 /* JPEG limit on # of components in one scan */
#define MAX_SAMP_FACTOR 4 /* JPEG limit on sampling factors */
/* Unfortunately, some bozo at Adobe saw no reason to be bound by the standard;
* the PostScript DCT filter can emit files with many more than 10 blocks/MCU.
* If you happen to run across such a file, you can up D_MAX_BLOCKS_IN_MCU
* to handle it. We even let you do this from the jconfig.h file. However,
* we strongly discourage changing C_MAX_BLOCKS_IN_MCU; just because Adobe
* sometimes emits noncompliant files doesn't mean you should too.
*/
#define C_MAX_BLOCKS_IN_MCU 10 /* compressor's limit on blocks per MCU */
#ifndef D_MAX_BLOCKS_IN_MCU
#define D_MAX_BLOCKS_IN_MCU 10 /* decompressor's limit on blocks per MCU */
#endif
/* Data structures for images (arrays of samples and of DCT coefficients).
*/
typedef JSAMPROW *JSAMPARRAY; /* ptr to some rows (a 2-D sample array) */
typedef JSAMPARRAY *JSAMPIMAGE; /* a 3-D sample array: top index is color */
typedef JCOEF JBLOCK[DCTSIZE2]; /* one block of coefficients */
typedef JBLOCK *JBLOCKROW; /* pointer to one row of coefficient blocks */
typedef JBLOCKROW *JBLOCKARRAY; /* a 2-D array of coefficient blocks */
typedef JBLOCKARRAY *JBLOCKIMAGE; /* a 3-D array of coefficient blocks */
#include <arm_neon.h>
/* jsimd_idct_ifast_neon() performs dequantization and a fast, not so accurate
* inverse DCT (Discrete Cosine Transform) on one block of coefficients. It
* uses the same calculations and produces exactly the same output as IJG's
* original jpeg_idct_ifast() function, which can be found in jidctfst.c.
*
* Scaled integer constants are used to avoid floating-point arithmetic:
* 0.082392200 = 2688 * 2^-15
* 0.414213562 = 13568 * 2^-15
* 0.847759065 = 27776 * 2^-15
* 0.613125930 = 20096 * 2^-15
*
* See jidctfst.c for further details of the IDCT algorithm. Where possible,
* the variable names and comments here in jsimd_idct_ifast_neon() match up
* with those in jpeg_idct_ifast().
*/
#define PASS1_BITS 2
#define F_0_082 2688
#define F_0_414 13568
#define F_0_847 27776
#define F_0_613 20096
__attribute__((aligned(16))) static const int16_t jsimd_idct_ifast_neon_consts[] = {
F_0_082, F_0_414, F_0_847, F_0_613
};
#define F_0_382 12544
#define F_0_541 17792
#define F_0_707 23168
#define F_0_306 9984
__attribute__((aligned(16))) static const int16_t jsimd_fdct_ifast_neon_consts[] = {
F_0_382, F_0_541, F_0_707, F_0_306
};
void dct_jpeg_fdct_ifast(DCTELEM *data) {
/* Load an 8x8 block of samples into Neon registers. De-interleaving loads
* are used, followed by vuzp to transpose the block such that we have a
* column of samples per vector - allowing all rows to be processed at once.
*/
int16x8x4_t data1 = vld4q_s16(data);
int16x8x4_t data2 = vld4q_s16(data + 4 * DCTSIZE);
int16x8x2_t cols_04 = vuzpq_s16(data1.val[0], data2.val[0]);
int16x8x2_t cols_15 = vuzpq_s16(data1.val[1], data2.val[1]);
int16x8x2_t cols_26 = vuzpq_s16(data1.val[2], data2.val[2]);
int16x8x2_t cols_37 = vuzpq_s16(data1.val[3], data2.val[3]);
int16x8_t col0 = cols_04.val[0];
int16x8_t col1 = cols_15.val[0];
int16x8_t col2 = cols_26.val[0];
int16x8_t col3 = cols_37.val[0];
int16x8_t col4 = cols_04.val[1];
int16x8_t col5 = cols_15.val[1];
int16x8_t col6 = cols_26.val[1];
int16x8_t col7 = cols_37.val[1];
/* Pass 1: process rows. */
/* Load DCT conversion constants. */
const int16x4_t consts = vld1_s16(jsimd_fdct_ifast_neon_consts);
int16x8_t tmp0 = vaddq_s16(col0, col7);
int16x8_t tmp7 = vsubq_s16(col0, col7);
int16x8_t tmp1 = vaddq_s16(col1, col6);
int16x8_t tmp6 = vsubq_s16(col1, col6);
int16x8_t tmp2 = vaddq_s16(col2, col5);
int16x8_t tmp5 = vsubq_s16(col2, col5);
int16x8_t tmp3 = vaddq_s16(col3, col4);
int16x8_t tmp4 = vsubq_s16(col3, col4);
/* Even part */
int16x8_t tmp10 = vaddq_s16(tmp0, tmp3); /* phase 2 */
int16x8_t tmp13 = vsubq_s16(tmp0, tmp3);
int16x8_t tmp11 = vaddq_s16(tmp1, tmp2);
int16x8_t tmp12 = vsubq_s16(tmp1, tmp2);
col0 = vaddq_s16(tmp10, tmp11); /* phase 3 */
col4 = vsubq_s16(tmp10, tmp11);
int16x8_t z1 = vqdmulhq_lane_s16(vaddq_s16(tmp12, tmp13), consts, 2);
col2 = vaddq_s16(tmp13, z1); /* phase 5 */
col6 = vsubq_s16(tmp13, z1);
/* Odd part */
tmp10 = vaddq_s16(tmp4, tmp5); /* phase 2 */
tmp11 = vaddq_s16(tmp5, tmp6);
tmp12 = vaddq_s16(tmp6, tmp7);
int16x8_t z5 = vqdmulhq_lane_s16(vsubq_s16(tmp10, tmp12), consts, 0);
int16x8_t z2 = vqdmulhq_lane_s16(tmp10, consts, 1);
z2 = vaddq_s16(z2, z5);
int16x8_t z4 = vqdmulhq_lane_s16(tmp12, consts, 3);
z5 = vaddq_s16(tmp12, z5);
z4 = vaddq_s16(z4, z5);
int16x8_t z3 = vqdmulhq_lane_s16(tmp11, consts, 2);
int16x8_t z11 = vaddq_s16(tmp7, z3); /* phase 5 */
int16x8_t z13 = vsubq_s16(tmp7, z3);
col5 = vaddq_s16(z13, z2); /* phase 6 */
col3 = vsubq_s16(z13, z2);
col1 = vaddq_s16(z11, z4);
col7 = vsubq_s16(z11, z4);
/* Transpose to work on columns in pass 2. */
int16x8x2_t cols_01 = vtrnq_s16(col0, col1);
int16x8x2_t cols_23 = vtrnq_s16(col2, col3);
int16x8x2_t cols_45 = vtrnq_s16(col4, col5);
int16x8x2_t cols_67 = vtrnq_s16(col6, col7);
int32x4x2_t cols_0145_l = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[0]),
vreinterpretq_s32_s16(cols_45.val[0]));
int32x4x2_t cols_0145_h = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[1]),
vreinterpretq_s32_s16(cols_45.val[1]));
int32x4x2_t cols_2367_l = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[0]),
vreinterpretq_s32_s16(cols_67.val[0]));
int32x4x2_t cols_2367_h = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[1]),
vreinterpretq_s32_s16(cols_67.val[1]));
int32x4x2_t rows_04 = vzipq_s32(cols_0145_l.val[0], cols_2367_l.val[0]);
int32x4x2_t rows_15 = vzipq_s32(cols_0145_h.val[0], cols_2367_h.val[0]);
int32x4x2_t rows_26 = vzipq_s32(cols_0145_l.val[1], cols_2367_l.val[1]);
int32x4x2_t rows_37 = vzipq_s32(cols_0145_h.val[1], cols_2367_h.val[1]);
int16x8_t row0 = vreinterpretq_s16_s32(rows_04.val[0]);
int16x8_t row1 = vreinterpretq_s16_s32(rows_15.val[0]);
int16x8_t row2 = vreinterpretq_s16_s32(rows_26.val[0]);
int16x8_t row3 = vreinterpretq_s16_s32(rows_37.val[0]);
int16x8_t row4 = vreinterpretq_s16_s32(rows_04.val[1]);
int16x8_t row5 = vreinterpretq_s16_s32(rows_15.val[1]);
int16x8_t row6 = vreinterpretq_s16_s32(rows_26.val[1]);
int16x8_t row7 = vreinterpretq_s16_s32(rows_37.val[1]);
/* Pass 2: process columns. */
tmp0 = vaddq_s16(row0, row7);
tmp7 = vsubq_s16(row0, row7);
tmp1 = vaddq_s16(row1, row6);
tmp6 = vsubq_s16(row1, row6);
tmp2 = vaddq_s16(row2, row5);
tmp5 = vsubq_s16(row2, row5);
tmp3 = vaddq_s16(row3, row4);
tmp4 = vsubq_s16(row3, row4);
/* Even part */
tmp10 = vaddq_s16(tmp0, tmp3); /* phase 2 */
tmp13 = vsubq_s16(tmp0, tmp3);
tmp11 = vaddq_s16(tmp1, tmp2);
tmp12 = vsubq_s16(tmp1, tmp2);
row0 = vaddq_s16(tmp10, tmp11); /* phase 3 */
row4 = vsubq_s16(tmp10, tmp11);
z1 = vqdmulhq_lane_s16(vaddq_s16(tmp12, tmp13), consts, 2);
row2 = vaddq_s16(tmp13, z1); /* phase 5 */
row6 = vsubq_s16(tmp13, z1);
/* Odd part */
tmp10 = vaddq_s16(tmp4, tmp5); /* phase 2 */
tmp11 = vaddq_s16(tmp5, tmp6);
tmp12 = vaddq_s16(tmp6, tmp7);
z5 = vqdmulhq_lane_s16(vsubq_s16(tmp10, tmp12), consts, 0);
z2 = vqdmulhq_lane_s16(tmp10, consts, 1);
z2 = vaddq_s16(z2, z5);
z4 = vqdmulhq_lane_s16(tmp12, consts, 3);
z5 = vaddq_s16(tmp12, z5);
z4 = vaddq_s16(z4, z5);
z3 = vqdmulhq_lane_s16(tmp11, consts, 2);
z11 = vaddq_s16(tmp7, z3); /* phase 5 */
z13 = vsubq_s16(tmp7, z3);
row5 = vaddq_s16(z13, z2); /* phase 6 */
row3 = vsubq_s16(z13, z2);
row1 = vaddq_s16(z11, z4);
row7 = vsubq_s16(z11, z4);
vst1q_s16(data + 0 * DCTSIZE, row0);
vst1q_s16(data + 1 * DCTSIZE, row1);
vst1q_s16(data + 2 * DCTSIZE, row2);
vst1q_s16(data + 3 * DCTSIZE, row3);
vst1q_s16(data + 4 * DCTSIZE, row4);
vst1q_s16(data + 5 * DCTSIZE, row5);
vst1q_s16(data + 6 * DCTSIZE, row6);
vst1q_s16(data + 7 * DCTSIZE, row7);
}
struct DctAuxiliaryData {
};
struct DctAuxiliaryData *createDctAuxiliaryData() {
struct DctAuxiliaryData *result = malloc(sizeof(struct DctAuxiliaryData));
return result;
}
void freeDctAuxiliaryData(struct DctAuxiliaryData *data) {
if (data) {
free(data);
}
}
void dct_jpeg_idct_ifast(struct DctAuxiliaryData *auxiliaryData, void *dct_table, JCOEFPTR coef_block, JSAMPROW output_buf)
{
IFAST_MULT_TYPE *quantptr = dct_table;
/* Load DCT coefficients. */
int16x8_t row0 = vld1q_s16(coef_block + 0 * DCTSIZE);
int16x8_t row1 = vld1q_s16(coef_block + 1 * DCTSIZE);
int16x8_t row2 = vld1q_s16(coef_block + 2 * DCTSIZE);
int16x8_t row3 = vld1q_s16(coef_block + 3 * DCTSIZE);
int16x8_t row4 = vld1q_s16(coef_block + 4 * DCTSIZE);
int16x8_t row5 = vld1q_s16(coef_block + 5 * DCTSIZE);
int16x8_t row6 = vld1q_s16(coef_block + 6 * DCTSIZE);
int16x8_t row7 = vld1q_s16(coef_block + 7 * DCTSIZE);
/* Load quantization table values for DC coefficients. */
int16x8_t quant_row0 = vld1q_s16(quantptr + 0 * DCTSIZE);
/* Dequantize DC coefficients. */
row0 = vmulq_s16(row0, quant_row0);
/* Construct bitmap to test if all AC coefficients are 0. */
int16x8_t bitmap = vorrq_s16(row1, row2);
bitmap = vorrq_s16(bitmap, row3);
bitmap = vorrq_s16(bitmap, row4);
bitmap = vorrq_s16(bitmap, row5);
bitmap = vorrq_s16(bitmap, row6);
bitmap = vorrq_s16(bitmap, row7);
int64_t left_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 0);
int64_t right_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 1);
/* Load IDCT conversion constants. */
const int16x4_t consts = vld1_s16(jsimd_idct_ifast_neon_consts);
if (left_ac_bitmap == 0 && right_ac_bitmap == 0) {
/* All AC coefficients are zero.
* Compute DC values and duplicate into vectors.
*/
int16x8_t dcval = row0;
row1 = dcval;
row2 = dcval;
row3 = dcval;
row4 = dcval;
row5 = dcval;
row6 = dcval;
row7 = dcval;
} else if (left_ac_bitmap == 0) {
/* AC coefficients are zero for columns 0, 1, 2, and 3.
* Use DC values for these columns.
*/
int16x4_t dcval = vget_low_s16(row0);
/* Commence regular fast IDCT computation for columns 4, 5, 6, and 7. */
/* Load quantization table. */
int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE + 4);
int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE + 4);
int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE + 4);
int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE + 4);
int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE + 4);
int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE + 4);
int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE + 4);
/* Even part: dequantize DCT coefficients. */
int16x4_t tmp0 = vget_high_s16(row0);
int16x4_t tmp1 = vmul_s16(vget_high_s16(row2), quant_row2);
int16x4_t tmp2 = vmul_s16(vget_high_s16(row4), quant_row4);
int16x4_t tmp3 = vmul_s16(vget_high_s16(row6), quant_row6);
int16x4_t tmp10 = vadd_s16(tmp0, tmp2); /* phase 3 */
int16x4_t tmp11 = vsub_s16(tmp0, tmp2);
int16x4_t tmp13 = vadd_s16(tmp1, tmp3); /* phases 5-3 */
int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
tmp12 = vsub_s16(tmp12, tmp13);
tmp0 = vadd_s16(tmp10, tmp13); /* phase 2 */
tmp3 = vsub_s16(tmp10, tmp13);
tmp1 = vadd_s16(tmp11, tmp12);
tmp2 = vsub_s16(tmp11, tmp12);
/* Odd part: dequantize DCT coefficients. */
int16x4_t tmp4 = vmul_s16(vget_high_s16(row1), quant_row1);
int16x4_t tmp5 = vmul_s16(vget_high_s16(row3), quant_row3);
int16x4_t tmp6 = vmul_s16(vget_high_s16(row5), quant_row5);
int16x4_t tmp7 = vmul_s16(vget_high_s16(row7), quant_row7);
int16x4_t z13 = vadd_s16(tmp6, tmp5); /* phase 6 */
int16x4_t neg_z10 = vsub_s16(tmp5, tmp6);
int16x4_t z11 = vadd_s16(tmp4, tmp7);
int16x4_t z12 = vsub_s16(tmp4, tmp7);
tmp7 = vadd_s16(z11, z13); /* phase 5 */
int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vadd_s16(tmp11, z11_sub_z13);
int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
z5 = vadd_s16(z5, z10_add_z12);
tmp10 = vqdmulh_lane_s16(z12, consts, 0);
tmp10 = vadd_s16(tmp10, z12);
tmp10 = vsub_s16(tmp10, z5);
tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
tmp12 = vadd_s16(tmp12, z5);
tmp6 = vsub_s16(tmp12, tmp7); /* phase 2 */
tmp5 = vsub_s16(tmp11, tmp6);
tmp4 = vadd_s16(tmp10, tmp5);
row0 = vcombine_s16(dcval, vadd_s16(tmp0, tmp7));
row7 = vcombine_s16(dcval, vsub_s16(tmp0, tmp7));
row1 = vcombine_s16(dcval, vadd_s16(tmp1, tmp6));
row6 = vcombine_s16(dcval, vsub_s16(tmp1, tmp6));
row2 = vcombine_s16(dcval, vadd_s16(tmp2, tmp5));
row5 = vcombine_s16(dcval, vsub_s16(tmp2, tmp5));
row4 = vcombine_s16(dcval, vadd_s16(tmp3, tmp4));
row3 = vcombine_s16(dcval, vsub_s16(tmp3, tmp4));
} else if (right_ac_bitmap == 0) {
/* AC coefficients are zero for columns 4, 5, 6, and 7.
* Use DC values for these columns.
*/
int16x4_t dcval = vget_high_s16(row0);
/* Commence regular fast IDCT computation for columns 0, 1, 2, and 3. */
/* Load quantization table. */
int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE);
int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE);
int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE);
int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE);
int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE);
int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE);
int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE);
/* Even part: dequantize DCT coefficients. */
int16x4_t tmp0 = vget_low_s16(row0);
int16x4_t tmp1 = vmul_s16(vget_low_s16(row2), quant_row2);
int16x4_t tmp2 = vmul_s16(vget_low_s16(row4), quant_row4);
int16x4_t tmp3 = vmul_s16(vget_low_s16(row6), quant_row6);
int16x4_t tmp10 = vadd_s16(tmp0, tmp2); /* phase 3 */
int16x4_t tmp11 = vsub_s16(tmp0, tmp2);
int16x4_t tmp13 = vadd_s16(tmp1, tmp3); /* phases 5-3 */
int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
tmp12 = vsub_s16(tmp12, tmp13);
tmp0 = vadd_s16(tmp10, tmp13); /* phase 2 */
tmp3 = vsub_s16(tmp10, tmp13);
tmp1 = vadd_s16(tmp11, tmp12);
tmp2 = vsub_s16(tmp11, tmp12);
/* Odd part: dequantize DCT coefficients. */
int16x4_t tmp4 = vmul_s16(vget_low_s16(row1), quant_row1);
int16x4_t tmp5 = vmul_s16(vget_low_s16(row3), quant_row3);
int16x4_t tmp6 = vmul_s16(vget_low_s16(row5), quant_row5);
int16x4_t tmp7 = vmul_s16(vget_low_s16(row7), quant_row7);
int16x4_t z13 = vadd_s16(tmp6, tmp5); /* phase 6 */
int16x4_t neg_z10 = vsub_s16(tmp5, tmp6);
int16x4_t z11 = vadd_s16(tmp4, tmp7);
int16x4_t z12 = vsub_s16(tmp4, tmp7);
tmp7 = vadd_s16(z11, z13); /* phase 5 */
int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vadd_s16(tmp11, z11_sub_z13);
int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
z5 = vadd_s16(z5, z10_add_z12);
tmp10 = vqdmulh_lane_s16(z12, consts, 0);
tmp10 = vadd_s16(tmp10, z12);
tmp10 = vsub_s16(tmp10, z5);
tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
tmp12 = vadd_s16(tmp12, z5);
tmp6 = vsub_s16(tmp12, tmp7); /* phase 2 */
tmp5 = vsub_s16(tmp11, tmp6);
tmp4 = vadd_s16(tmp10, tmp5);
row0 = vcombine_s16(vadd_s16(tmp0, tmp7), dcval);
row7 = vcombine_s16(vsub_s16(tmp0, tmp7), dcval);
row1 = vcombine_s16(vadd_s16(tmp1, tmp6), dcval);
row6 = vcombine_s16(vsub_s16(tmp1, tmp6), dcval);
row2 = vcombine_s16(vadd_s16(tmp2, tmp5), dcval);
row5 = vcombine_s16(vsub_s16(tmp2, tmp5), dcval);
row4 = vcombine_s16(vadd_s16(tmp3, tmp4), dcval);
row3 = vcombine_s16(vsub_s16(tmp3, tmp4), dcval);
} else {
/* Some AC coefficients are non-zero; full IDCT calculation required. */
/* Load quantization table. */
int16x8_t quant_row1 = vld1q_s16(quantptr + 1 * DCTSIZE);
int16x8_t quant_row2 = vld1q_s16(quantptr + 2 * DCTSIZE);
int16x8_t quant_row3 = vld1q_s16(quantptr + 3 * DCTSIZE);
int16x8_t quant_row4 = vld1q_s16(quantptr + 4 * DCTSIZE);
int16x8_t quant_row5 = vld1q_s16(quantptr + 5 * DCTSIZE);
int16x8_t quant_row6 = vld1q_s16(quantptr + 6 * DCTSIZE);
int16x8_t quant_row7 = vld1q_s16(quantptr + 7 * DCTSIZE);
/* Even part: dequantize DCT coefficients. */
int16x8_t tmp0 = row0;
int16x8_t tmp1 = vmulq_s16(row2, quant_row2);
int16x8_t tmp2 = vmulq_s16(row4, quant_row4);
int16x8_t tmp3 = vmulq_s16(row6, quant_row6);
int16x8_t tmp10 = vaddq_s16(tmp0, tmp2); /* phase 3 */
int16x8_t tmp11 = vsubq_s16(tmp0, tmp2);
int16x8_t tmp13 = vaddq_s16(tmp1, tmp3); /* phases 5-3 */
int16x8_t tmp1_sub_tmp3 = vsubq_s16(tmp1, tmp3);
int16x8_t tmp12 = vqdmulhq_lane_s16(tmp1_sub_tmp3, consts, 1);
tmp12 = vaddq_s16(tmp12, tmp1_sub_tmp3);
tmp12 = vsubq_s16(tmp12, tmp13);
tmp0 = vaddq_s16(tmp10, tmp13); /* phase 2 */
tmp3 = vsubq_s16(tmp10, tmp13);
tmp1 = vaddq_s16(tmp11, tmp12);
tmp2 = vsubq_s16(tmp11, tmp12);
/* Odd part: dequantize DCT coefficients. */
int16x8_t tmp4 = vmulq_s16(row1, quant_row1);
int16x8_t tmp5 = vmulq_s16(row3, quant_row3);
int16x8_t tmp6 = vmulq_s16(row5, quant_row5);
int16x8_t tmp7 = vmulq_s16(row7, quant_row7);
int16x8_t z13 = vaddq_s16(tmp6, tmp5); /* phase 6 */
int16x8_t neg_z10 = vsubq_s16(tmp5, tmp6);
int16x8_t z11 = vaddq_s16(tmp4, tmp7);
int16x8_t z12 = vsubq_s16(tmp4, tmp7);
tmp7 = vaddq_s16(z11, z13); /* phase 5 */
int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vaddq_s16(tmp11, z11_sub_z13);
int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
z5 = vaddq_s16(z5, z10_add_z12);
tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
tmp10 = vaddq_s16(tmp10, z12);
tmp10 = vsubq_s16(tmp10, z5);
tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
tmp12 = vaddq_s16(tmp12, z5);
tmp6 = vsubq_s16(tmp12, tmp7); /* phase 2 */
tmp5 = vsubq_s16(tmp11, tmp6);
tmp4 = vaddq_s16(tmp10, tmp5);
row0 = vaddq_s16(tmp0, tmp7);
row7 = vsubq_s16(tmp0, tmp7);
row1 = vaddq_s16(tmp1, tmp6);
row6 = vsubq_s16(tmp1, tmp6);
row2 = vaddq_s16(tmp2, tmp5);
row5 = vsubq_s16(tmp2, tmp5);
row4 = vaddq_s16(tmp3, tmp4);
row3 = vsubq_s16(tmp3, tmp4);
}
/* Transpose rows to work on columns in pass 2. */
int16x8x2_t rows_01 = vtrnq_s16(row0, row1);
int16x8x2_t rows_23 = vtrnq_s16(row2, row3);
int16x8x2_t rows_45 = vtrnq_s16(row4, row5);
int16x8x2_t rows_67 = vtrnq_s16(row6, row7);
int32x4x2_t rows_0145_l = vtrnq_s32(vreinterpretq_s32_s16(rows_01.val[0]),
vreinterpretq_s32_s16(rows_45.val[0]));
int32x4x2_t rows_0145_h = vtrnq_s32(vreinterpretq_s32_s16(rows_01.val[1]),
vreinterpretq_s32_s16(rows_45.val[1]));
int32x4x2_t rows_2367_l = vtrnq_s32(vreinterpretq_s32_s16(rows_23.val[0]),
vreinterpretq_s32_s16(rows_67.val[0]));
int32x4x2_t rows_2367_h = vtrnq_s32(vreinterpretq_s32_s16(rows_23.val[1]),
vreinterpretq_s32_s16(rows_67.val[1]));
int32x4x2_t cols_04 = vzipq_s32(rows_0145_l.val[0], rows_2367_l.val[0]);
int32x4x2_t cols_15 = vzipq_s32(rows_0145_h.val[0], rows_2367_h.val[0]);
int32x4x2_t cols_26 = vzipq_s32(rows_0145_l.val[1], rows_2367_l.val[1]);
int32x4x2_t cols_37 = vzipq_s32(rows_0145_h.val[1], rows_2367_h.val[1]);
int16x8_t col0 = vreinterpretq_s16_s32(cols_04.val[0]);
int16x8_t col1 = vreinterpretq_s16_s32(cols_15.val[0]);
int16x8_t col2 = vreinterpretq_s16_s32(cols_26.val[0]);
int16x8_t col3 = vreinterpretq_s16_s32(cols_37.val[0]);
int16x8_t col4 = vreinterpretq_s16_s32(cols_04.val[1]);
int16x8_t col5 = vreinterpretq_s16_s32(cols_15.val[1]);
int16x8_t col6 = vreinterpretq_s16_s32(cols_26.val[1]);
int16x8_t col7 = vreinterpretq_s16_s32(cols_37.val[1]);
/* 1-D IDCT, pass 2 */
/* Even part */
int16x8_t tmp10 = vaddq_s16(col0, col4);
int16x8_t tmp11 = vsubq_s16(col0, col4);
int16x8_t tmp13 = vaddq_s16(col2, col6);
int16x8_t col2_sub_col6 = vsubq_s16(col2, col6);
int16x8_t tmp12 = vqdmulhq_lane_s16(col2_sub_col6, consts, 1);
tmp12 = vaddq_s16(tmp12, col2_sub_col6);
tmp12 = vsubq_s16(tmp12, tmp13);
int16x8_t tmp0 = vaddq_s16(tmp10, tmp13);
int16x8_t tmp3 = vsubq_s16(tmp10, tmp13);
int16x8_t tmp1 = vaddq_s16(tmp11, tmp12);
int16x8_t tmp2 = vsubq_s16(tmp11, tmp12);
/* Odd part */
int16x8_t z13 = vaddq_s16(col5, col3);
int16x8_t neg_z10 = vsubq_s16(col3, col5);
int16x8_t z11 = vaddq_s16(col1, col7);
int16x8_t z12 = vsubq_s16(col1, col7);
int16x8_t tmp7 = vaddq_s16(z11, z13); /* phase 5 */
int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vaddq_s16(tmp11, z11_sub_z13);
int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
z5 = vaddq_s16(z5, z10_add_z12);
tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
tmp10 = vaddq_s16(tmp10, z12);
tmp10 = vsubq_s16(tmp10, z5);
tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
tmp12 = vaddq_s16(tmp12, z5);
int16x8_t tmp6 = vsubq_s16(tmp12, tmp7); /* phase 2 */
int16x8_t tmp5 = vsubq_s16(tmp11, tmp6);
int16x8_t tmp4 = vaddq_s16(tmp10, tmp5);
col0 = vaddq_s16(tmp0, tmp7);
col7 = vsubq_s16(tmp0, tmp7);
col1 = vaddq_s16(tmp1, tmp6);
col6 = vsubq_s16(tmp1, tmp6);
col2 = vaddq_s16(tmp2, tmp5);
col5 = vsubq_s16(tmp2, tmp5);
col4 = vaddq_s16(tmp3, tmp4);
col3 = vsubq_s16(tmp3, tmp4);
/* Scale down by a factor of 8, narrowing to 8-bit. */
int8x16_t cols_01_s8 = vcombine_s8(vqshrn_n_s16(col0, PASS1_BITS + 3),
vqshrn_n_s16(col1, PASS1_BITS + 3));
int8x16_t cols_45_s8 = vcombine_s8(vqshrn_n_s16(col4, PASS1_BITS + 3),
vqshrn_n_s16(col5, PASS1_BITS + 3));
int8x16_t cols_23_s8 = vcombine_s8(vqshrn_n_s16(col2, PASS1_BITS + 3),
vqshrn_n_s16(col3, PASS1_BITS + 3));
int8x16_t cols_67_s8 = vcombine_s8(vqshrn_n_s16(col6, PASS1_BITS + 3),
vqshrn_n_s16(col7, PASS1_BITS + 3));
/* Clamp to range [0-255]. */
uint8x16_t cols_01 =
vreinterpretq_u8_s8
(vaddq_s8(cols_01_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
uint8x16_t cols_45 =
vreinterpretq_u8_s8
(vaddq_s8(cols_45_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
uint8x16_t cols_23 =
vreinterpretq_u8_s8
(vaddq_s8(cols_23_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
uint8x16_t cols_67 =
vreinterpretq_u8_s8
(vaddq_s8(cols_67_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
/* Transpose block to prepare for store. */
uint32x4x2_t cols_0415 = vzipq_u32(vreinterpretq_u32_u8(cols_01),
vreinterpretq_u32_u8(cols_45));
uint32x4x2_t cols_2637 = vzipq_u32(vreinterpretq_u32_u8(cols_23),
vreinterpretq_u32_u8(cols_67));
uint8x16x2_t cols_0145 = vtrnq_u8(vreinterpretq_u8_u32(cols_0415.val[0]),
vreinterpretq_u8_u32(cols_0415.val[1]));
uint8x16x2_t cols_2367 = vtrnq_u8(vreinterpretq_u8_u32(cols_2637.val[0]),
vreinterpretq_u8_u32(cols_2637.val[1]));
uint16x8x2_t rows_0426 = vtrnq_u16(vreinterpretq_u16_u8(cols_0145.val[0]),
vreinterpretq_u16_u8(cols_2367.val[0]));
uint16x8x2_t rows_1537 = vtrnq_u16(vreinterpretq_u16_u8(cols_0145.val[1]),
vreinterpretq_u16_u8(cols_2367.val[1]));
uint8x16_t rows_04 = vreinterpretq_u8_u16(rows_0426.val[0]);
uint8x16_t rows_15 = vreinterpretq_u8_u16(rows_1537.val[0]);
uint8x16_t rows_26 = vreinterpretq_u8_u16(rows_0426.val[1]);
uint8x16_t rows_37 = vreinterpretq_u8_u16(rows_1537.val[1]);
JSAMPROW outptr0 = output_buf + DCTSIZE * 0;
JSAMPROW outptr1 = output_buf + DCTSIZE * 1;
JSAMPROW outptr2 = output_buf + DCTSIZE * 2;
JSAMPROW outptr3 = output_buf + DCTSIZE * 3;
JSAMPROW outptr4 = output_buf + DCTSIZE * 4;
JSAMPROW outptr5 = output_buf + DCTSIZE * 5;
JSAMPROW outptr6 = output_buf + DCTSIZE * 6;
JSAMPROW outptr7 = output_buf + DCTSIZE * 7;
/* Store DCT block to memory. */
vst1q_lane_u64((uint64_t *)outptr0, vreinterpretq_u64_u8(rows_04), 0);
vst1q_lane_u64((uint64_t *)outptr1, vreinterpretq_u64_u8(rows_15), 0);
vst1q_lane_u64((uint64_t *)outptr2, vreinterpretq_u64_u8(rows_26), 0);
vst1q_lane_u64((uint64_t *)outptr3, vreinterpretq_u64_u8(rows_37), 0);
vst1q_lane_u64((uint64_t *)outptr4, vreinterpretq_u64_u8(rows_04), 1);
vst1q_lane_u64((uint64_t *)outptr5, vreinterpretq_u64_u8(rows_15), 1);
vst1q_lane_u64((uint64_t *)outptr6, vreinterpretq_u64_u8(rows_26), 1);
vst1q_lane_u64((uint64_t *)outptr7, vreinterpretq_u64_u8(rows_37), 1);
}
#endif

31
submodules/TelegramUI/Components/AnimationCache/ImageDCT/Sources/ImageDCT.mm Normal file
View File

@@ -0,0 +1,31 @@
#import <ImageDCT/ImageDCT.h>
#import <memory>
#include "DCT.h"
@interface ImageDCT () {
std::unique_ptr<dct::DCT> _dct;
}
@end
@implementation ImageDCT
- (instancetype _Nonnull)initWithQuality:(NSInteger)quality {
self = [super init];
if (self != nil) {
_dct = std::unique_ptr<dct::DCT>(new dct::DCT((int)quality));
}
return self;
}
- (void)forwardWithPixels:(uint8_t const * _Nonnull)pixels coefficients:(int16_t * _Nonnull)coefficients width:(NSInteger)width height:(NSInteger)height bytesPerRow:(NSInteger)bytesPerRow {
_dct->forward(pixels, coefficients, (int)width, (int)height, (int)bytesPerRow);
}
- (void)inverseWithCoefficients:(int16_t const * _Nonnull)coefficients pixels:(uint8_t * _Nonnull)pixels width:(NSInteger)width height:(NSInteger)height coefficientsPerRow:(NSInteger)coefficientsPerRow bytesPerRow:(NSInteger)bytesPerRow {
_dct->inverse(coefficients, pixels, (int)width, (int)height, (int)coefficientsPerRow, (int)bytesPerRow);
}
@end

2
submodules/TelegramUI/Components/AnimationCache/DCT/Sources/YuvConversion.m → submodules/TelegramUI/Components/AnimationCache/ImageDCT/Sources/YuvConversion.m
View File

@@ -1,4 +1,4 @@
#import <DCT/YuvConversion.h>
#import <ImageDCT/YuvConversion.h>
#import <Foundation/Foundation.h>
#import <Accelerate/Accelerate.h>

12
submodules/TelegramUI/Components/AnimationCache/Sources/ImageData.swift
View File

@@ -1,6 +1,6 @@
import Foundation
import UIKit
import DCT
import ImageDCT
final class ImagePlane {
let width: Int
@@ -134,13 +134,11 @@ extension ImageYUVA420 {
final class DctData {
let quality: Int
let dctData: Data
let idctData: Data
let dct: ImageDCT
init(quality: Int) {
self.quality = quality
self.dctData = generateForwardDctData(Int32(quality))!
self.idctData = generateInverseDctData(Int32(quality))!
self.dct = ImageDCT(quality: quality)
}
}
@@ -174,7 +172,7 @@ extension ImageYUVA420 {
targetPlane.data.withUnsafeMutableBytes { bytes in
let coefficients = bytes.baseAddress!.assumingMemoryBound(to: UInt16.self)
performForwardDct(sourcePixels, coefficients, Int32(sourcePlane.width), Int32(sourcePlane.height), Int32(sourcePlane.bytesPerRow), dctData.dctData)
dctData.dct.forward(withPixels: sourcePixels, coefficients: coefficients, width: sourcePlane.width, height: sourcePlane.height, bytesPerRow: sourcePlane.bytesPerRow)
}
}
}
@@ -217,7 +215,7 @@ extension DctCoefficientsYUVA420 {
targetPlane.data.withUnsafeMutableBytes { bytes in
let pixels = bytes.baseAddress!.assumingMemoryBound(to: UInt8.self)
performInverseDct(coefficients, pixels, Int32(sourcePlane.width), Int32(sourcePlane.height), Int32(targetPlane.bytesPerRow), Int32(sourcePlane.width), dctData.idctData)
dctData.dct.inverse(withCoefficients: coefficients, pixels: pixels, width: sourcePlane.width, height: sourcePlane.height, coefficientsPerRow: targetPlane.width, bytesPerRow: targetPlane.width)
}
}
}