mirror of
https://github.com/python-pillow/Pillow.git
synced 2025-08-20 20:24:45 +03:00
SIMD ColorLUT. First try
This commit is contained in:
Alexander
Alexander Karpinsky
parent
23451b8423
commit
16f69404f1
|
|
@ -761,7 +761,7 @@ _prepare_lut_table(PyObject *table, Py_ssize_t table_size) {
|
||||||
}
|
}
|
||||||
|
|
||||||
/* malloc check ok, max is 2 * 4 * 65**3 = 2197000 */
|
/* malloc check ok, max is 2 * 4 * 65**3 = 2197000 */
|
||||||
prepared = (INT16 *)malloc(sizeof(INT16) * table_size);
|
prepared = (INT16 *)malloc(sizeof(INT16) * table_size + 2);
|
||||||
if (!prepared) {
|
if (!prepared) {
|
||||||
if (free_table_data) {
|
if (free_table_data) {
|
||||||
free(table_data);
|
free(table_data);
|
||||||
|
|
|
||||||
|
|
@ -13,7 +13,8 @@
|
||||||
#define SCALE_BITS (32 - 8 - 6)
|
#define SCALE_BITS (32 - 8 - 6)
|
||||||
#define SCALE_MASK ((1 << SCALE_BITS) - 1)
|
#define SCALE_MASK ((1 << SCALE_BITS) - 1)
|
||||||
|
|
||||||
#define SHIFT_BITS (16 - 1)
|
#define SHIFT_BITS (16 - 2)
|
||||||
|
#define SHIFT_ROUNDING (1<<(SHIFT_BITS-1))
|
||||||
|
|
||||||
static inline UINT8
|
static inline UINT8
|
||||||
clip8(int in) {
|
clip8(int in) {
|
||||||
|
|
@ -74,10 +75,17 @@ ImagingColorLUT3D_linear(
|
||||||
+1 cells will be outside of the table.
|
+1 cells will be outside of the table.
|
||||||
With this compensation we never hit the upper cells
|
With this compensation we never hit the upper cells
|
||||||
but this also doesn't introduce any noticeable difference. */
|
but this also doesn't introduce any noticeable difference. */
|
||||||
UINT32 scale1D = (size1D - 1) / 255.0 * (1 << SCALE_BITS);
|
|
||||||
UINT32 scale2D = (size2D - 1) / 255.0 * (1 << SCALE_BITS);
|
|
||||||
UINT32 scale3D = (size3D - 1) / 255.0 * (1 << SCALE_BITS);
|
|
||||||
int size1D_2D = size1D * size2D;
|
int size1D_2D = size1D * size2D;
|
||||||
|
__m128i scale = _mm_set_epi32(0,
|
||||||
|
(size3D - 1) / 255.0 * (1<<SCALE_BITS),
|
||||||
|
(size2D - 1) / 255.0 * (1<<SCALE_BITS),
|
||||||
|
(size1D - 1) / 255.0 * (1<<SCALE_BITS));
|
||||||
|
__m128i scale_mask = _mm_set1_epi32(SCALE_MASK);
|
||||||
|
__m128i index_mul = _mm_set_epi32(0, size1D_2D, size1D, 1);
|
||||||
|
__m128i shift_rounding = _mm_set1_epi32(SHIFT_ROUNDING);
|
||||||
|
__m128i shuffle_source = _mm_set_epi8(-1,-1, -1,-1, 11,10, 5,4, 9,8, 3,2, 7,6, 1,0);
|
||||||
|
__m128i left_mask = _mm_set1_epi32(0x0000ffff);
|
||||||
|
__m128i right_mask = _mm_set1_epi32(0xffff0000);
|
||||||
int x, y;
|
int x, y;
|
||||||
ImagingSectionCookie cookie;
|
ImagingSectionCookie cookie;
|
||||||
|
|
||||||
|
|
@ -101,84 +109,85 @@ ImagingColorLUT3D_linear(
|
||||||
UINT8 *rowIn = (UINT8 *)imIn->image[y];
|
UINT8 *rowIn = (UINT8 *)imIn->image[y];
|
||||||
char *rowOut = (char *)imOut->image[y];
|
char *rowOut = (char *)imOut->image[y];
|
||||||
for (x = 0; x < imOut->xsize; x++) {
|
for (x = 0; x < imOut->xsize; x++) {
|
||||||
UINT32 index1D = rowIn[x * 4 + 0] * scale1D;
|
__m128i index = _mm_mullo_epi32(scale,
|
||||||
UINT32 index2D = rowIn[x * 4 + 1] * scale2D;
|
_mm_cvtepu8_epi32(*(__m128i *) &rowIn[x*4]));
|
||||||
UINT32 index3D = rowIn[x * 4 + 2] * scale3D;
|
__m128i shift = _mm_srli_epi32(
|
||||||
INT16 shift1D = (SCALE_MASK & index1D) >> (SCALE_BITS - SHIFT_BITS);
|
_mm_and_si128(scale_mask, index), (SCALE_BITS - SHIFT_BITS));
|
||||||
INT16 shift2D = (SCALE_MASK & index2D) >> (SCALE_BITS - SHIFT_BITS);
|
int idx = table_channels * _mm_extract_epi32(
|
||||||
INT16 shift3D = (SCALE_MASK & index3D) >> (SCALE_BITS - SHIFT_BITS);
|
_mm_hadd_epi32(_mm_hadd_epi32(
|
||||||
int idx = table_channels * table_index3D(
|
_mm_madd_epi16(index_mul, _mm_srli_epi32(index, SCALE_BITS)),
|
||||||
index1D >> SCALE_BITS,
|
_mm_setzero_si128()), _mm_setzero_si128()), 0);
|
||||||
index2D >> SCALE_BITS,
|
__m128i shift1D, shift2D, shift3D;
|
||||||
index3D >> SCALE_BITS,
|
__m128i source, left, right, result;
|
||||||
size1D,
|
__m128i leftleft, leftright, rightleft, rightright;
|
||||||
size1D_2D);
|
|
||||||
INT16 result[4], left[4], right[4];
|
shift = _mm_or_si128(
|
||||||
INT16 leftleft[4], leftright[4], rightleft[4], rightright[4];
|
_mm_sub_epi32(_mm_set1_epi32(1<<SHIFT_BITS), shift),
|
||||||
|
_mm_slli_epi32(shift, 16));
|
||||||
|
|
||||||
|
shift1D = _mm_shuffle_epi32(shift, 0x00);
|
||||||
|
shift2D = _mm_shuffle_epi32(shift, 0x55);
|
||||||
|
shift3D = _mm_shuffle_epi32(shift, 0xaa);
|
||||||
|
|
||||||
if (table_channels == 3) {
|
if (table_channels == 3) {
|
||||||
UINT32 v;
|
source = _mm_shuffle_epi8(
|
||||||
interpolate3(leftleft, &table[idx + 0], &table[idx + 3], shift1D);
|
_mm_loadu_si128((__m128i *) &table[idx + 0]), shuffle_source);
|
||||||
interpolate3(
|
leftleft = _mm_and_si128(_mm_srai_epi32(_mm_add_epi32(_mm_madd_epi16(
|
||||||
leftright,
|
source, shift1D), shift_rounding), SHIFT_BITS), left_mask);
|
||||||
&table[idx + size1D * 3],
|
|
||||||
&table[idx + size1D * 3 + 3],
|
|
||||||
shift1D);
|
|
||||||
interpolate3(left, leftleft, leftright, shift2D);
|
|
||||||
|
|
||||||
interpolate3(
|
source = _mm_shuffle_epi8(
|
||||||
rightleft,
|
_mm_loadu_si128((__m128i *) &table[idx + size1D*3]), shuffle_source);
|
||||||
&table[idx + size1D_2D * 3],
|
leftright = _mm_and_si128(_mm_slli_epi32(_mm_add_epi32(_mm_madd_epi16(
|
||||||
&table[idx + size1D_2D * 3 + 3],
|
source, shift1D), shift_rounding), 16 - SHIFT_BITS), right_mask);
|
||||||
shift1D);
|
|
||||||
interpolate3(
|
|
||||||
rightright,
|
|
||||||
&table[idx + size1D_2D * 3 + size1D * 3],
|
|
||||||
&table[idx + size1D_2D * 3 + size1D * 3 + 3],
|
|
||||||
shift1D);
|
|
||||||
interpolate3(right, rightleft, rightright, shift2D);
|
|
||||||
|
|
||||||
interpolate3(result, left, right, shift3D);
|
source = _mm_shuffle_epi8(
|
||||||
|
_mm_loadu_si128((__m128i *) &table[idx + size1D_2D*3]), shuffle_source);
|
||||||
|
rightleft = _mm_and_si128(_mm_srai_epi32(_mm_add_epi32(_mm_madd_epi16(
|
||||||
|
source, shift1D), shift_rounding), SHIFT_BITS), left_mask);
|
||||||
|
|
||||||
|
source = _mm_shuffle_epi8(
|
||||||
|
_mm_loadu_si128((__m128i *) &table[idx + size1D_2D*3 + size1D*3]), shuffle_source);
|
||||||
|
rightright = _mm_and_si128(_mm_slli_epi32(_mm_add_epi32(_mm_madd_epi16(
|
||||||
|
source, shift1D), shift_rounding), 16 - SHIFT_BITS), right_mask);
|
||||||
|
|
||||||
v = MAKE_UINT32(
|
left = _mm_and_si128(_mm_srai_epi32(_mm_add_epi32(_mm_madd_epi16(
|
||||||
clip8(result[0]),
|
_mm_or_si128(leftleft, leftright), shift2D),
|
||||||
clip8(result[1]),
|
shift_rounding), SHIFT_BITS), left_mask);
|
||||||
clip8(result[2]),
|
|
||||||
rowIn[x * 4 + 3]);
|
right = _mm_and_si128(_mm_slli_epi32(_mm_add_epi32(_mm_madd_epi16(
|
||||||
memcpy(rowOut + x * sizeof(v), &v, sizeof(v));
|
_mm_or_si128(rightleft, rightright), shift2D),
|
||||||
|
shift_rounding), 16 - SHIFT_BITS), right_mask);
|
||||||
|
|
||||||
|
result = _mm_srai_epi32(_mm_add_epi32(_mm_madd_epi16(
|
||||||
|
_mm_or_si128(left, right), shift3D),
|
||||||
|
shift_rounding), SHIFT_BITS);
|
||||||
|
|
||||||
|
result = _mm_srai_epi32(
|
||||||
|
_mm_add_epi32(_mm_set1_epi32(PRECISION_ROUNDING), result),
|
||||||
|
PRECISION_BITS);
|
||||||
|
|
||||||
|
result = _mm_packs_epi32(result, result);
|
||||||
|
rowOut[x] = _mm_cvtsi128_si32(_mm_packus_epi16(result, result));
|
||||||
}
|
}
|
||||||
|
|
||||||
if (table_channels == 4) {
|
// if (table_channels == 4) {
|
||||||
UINT32 v;
|
// interpolate4(leftleft, &table[idx + 0], &table[idx + 4], shift1D);
|
||||||
interpolate4(leftleft, &table[idx + 0], &table[idx + 4], shift1D);
|
// interpolate4(leftright, &table[idx + size1D*4],
|
||||||
interpolate4(
|
// &table[idx + size1D*4 + 4], shift1D);
|
||||||
leftright,
|
// interpolate4(left, leftleft, leftright, shift2D);
|
||||||
&table[idx + size1D * 4],
|
|
||||||
&table[idx + size1D * 4 + 4],
|
|
||||||
shift1D);
|
|
||||||
interpolate4(left, leftleft, leftright, shift2D);
|
|
||||||
|
|
||||||
interpolate4(
|
// interpolate4(rightleft, &table[idx + size1D_2D*4],
|
||||||
rightleft,
|
// &table[idx + size1D_2D*4 + 4], shift1D);
|
||||||
&table[idx + size1D_2D * 4],
|
// interpolate4(rightright, &table[idx + size1D_2D*4 + size1D*4],
|
||||||
&table[idx + size1D_2D * 4 + 4],
|
// &table[idx + size1D_2D*4 + size1D*4 + 4], shift1D);
|
||||||
shift1D);
|
// interpolate4(right, rightleft, rightright, shift2D);
|
||||||
interpolate4(
|
|
||||||
rightright,
|
|
||||||
&table[idx + size1D_2D * 4 + size1D * 4],
|
|
||||||
&table[idx + size1D_2D * 4 + size1D * 4 + 4],
|
|
||||||
shift1D);
|
|
||||||
interpolate4(right, rightleft, rightright, shift2D);
|
|
||||||
|
|
||||||
interpolate4(result, left, right, shift3D);
|
// interpolate4(result, left, right, shift3D);
|
||||||
|
|
||||||
v = MAKE_UINT32(
|
// rowOut[x] = MAKE_UINT32(
|
||||||
clip8(result[0]),
|
// clip8(result[0]), clip8(result[1]),
|
||||||
clip8(result[1]),
|
// clip8(result[2]), clip8(result[3]));
|
||||||
clip8(result[2]),
|
// }
|
||||||
clip8(result[3]));
|
|
||||||
memcpy(rowOut + x * sizeof(v), &v, sizeof(v));
|
|
||||||
}
|
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
ImagingSectionLeave(&cookie);
|
ImagingSectionLeave(&cookie);
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue
Block a user