blender
/
blender
136
407
Fork 589
Code Issues 6.3k Pull Requests 779 Projects 19 Wiki Activity

VSE: add Bicubic filtering option, and optimize bicubic performance #117100

Merged
Aras Pranckevicius merged 4 commits from aras_p/blender:vse_bicubic into main 2024-01-15 16:38:49 +01:00
Conversation 5 Commits 4 Files Changed 14 +136 -179
1 changed files with 136 additions and 179 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit bbf391aa98 - Show all commits

315
source/blender/blenlib/intern/math_interp.cc
View File

@ -12,141 +12,149 @@
#include "BLI_math_base.h"
#include "BLI_math_interp.hh"
#include "BLI_math_vector.h"
#include "BLI_math_vector_types.hh"
#include "BLI_simd.h"
#include "BLI_strict_flags.h"
#if BLI_HAVE_SSE2 && defined(__SSE4_1__)
# include <smmintrin.h> /* _mm_floor_ps */
#endif
/**************************************************************************
* INTERPOLATIONS
*
* Reference and docs:
* http://wiki.blender.org/index.php/User:Damiles#Interpolations_Algorithms
***************************************************************************/
/* BICUBIC Interpolation functions
* More info: http://wiki.blender.org/index.php/User:Damiles#Bicubic_pixel_interpolation
* function assumes out to be zero'ed, only does RGBA */
static float P(float k)
/* Cubic B-Spline coefficients. f is offset from texel center in pixel space.
* This is Mitchell-Netravali filter with B=1, C=0 parameters. */
static blender::float4 cubic_bspline_coefficients(float f)
{
float p1, p2, p3, p4;
p1 = max_ff(k + 2.0f, 0.0f);
p2 = max_ff(k + 1.0f, 0.0f);
p3 = max_ff(k, 0.0f);
p4 = max_ff(k - 1.0f, 0.0f);
return (float)(1.0f / 6.0f) *
(p1 * p1 * p1 - 4.0f * p2 * p2 * p2 + 6.0f * p3 * p3 * p3 - 4.0f * p4 * p4 * p4);
float f2 = f * f;
float f3 = f2 * f;
float w3 = f3 / 6.0f;
float w0 = -w3 + f2 * 0.5f - f * 0.5f + 1.0f / 6.0f;
float w1 = f3 * 0.5f - f2 * 1.0f + 2.0f / 3.0f;
float w2 = 1.0f - w0 - w1 - w3;
return blender::float4(w0, w1, w2, w3);
}
#if 0
/* older, slower function, works the same as above */
static float P(float k)
{
return (float)(1.0f / 6.0f) *
(pow(MAX2(k + 2.0f, 0), 3.0f) - 4.0f * pow(MAX2(k + 1.0f, 0), 3.0f) +
6.0f * pow(MAX2(k, 0), 3.0f) - 4.0f * pow(MAX2(k - 1.0f, 0), 3.0f));
}
#endif
#if BLI_HAVE_SSE2
# if defined(__SSE4_1__)
# include <smmintrin.h> /* _mm_floor_ps */
# endif
static void vector_from_float(const float *data, float vector[4], int components)
BLI_INLINE __m128 floor_simd(__m128 v)
{
if (components == 1) {
vector[0] = data[0];
}
else if (components == 3) {
copy_v3_v3(vector, data);
}
else {
copy_v4_v4(vector, data);
}
# if defined(__SSE4_1__) || defined(__ARM_NEON) && defined(WITH_SSE2NEON)
/* If we're on SSE4 or ARM NEON, just use the simple floor() way. */
__m128 v_floor = _mm_floor_ps(v);
# else
/* The hard way: truncate, for negative inputs this will round towards zero.
* Then compare with input, and subtract 1 for the inputs that were
* negative. */
__m128 v_trunc = _mm_cvtepi32_ps(_mm_cvttps_epi32(v));
__m128 v_neg = _mm_cmplt_ps(v, v_trunc);
__m128 v_floor = _mm_sub_ps(v_trunc, _mm_and_ps(v_neg, _mm_set1_ps(1.0f)));
# endif
return v_floor;
}
static void vector_from_byte(const uchar *data, float vector[4], int components)
BLI_INLINE void bicubic_interpolation_uchar_simd(
const uchar *src_buffer, uchar *output, int width, int height, float u, float v)
{
if (components == 1) {
vector[0] = data[0];
}
else if (components == 3) {
vector[0] = data[0];
vector[1] = data[1];
vector[2] = data[2];
}
else {
vector[0] = data[0];
vector[1] = data[1];
vector[2] = data[2];
vector[3] = data[3];
}
}
__m128 uv = _mm_set_ps(0, 0, v, u);
__m128 uv_floor = floor_simd(uv);
__m128i i_uv = _mm_cvttps_epi32(uv_floor);
/* BICUBIC INTERPOLATION */
BLI_INLINE void bicubic_interpolation(const uchar *byte_buffer,
const float *float_buffer,
uchar *byte_output,
float *float_output,
int width,
int height,
int components,
float u,
float v)
{
int i, j, n, m, x1, y1;
float a, b, w, wx, wy[4], out[4];
/* sample area entirely outside image? */
if (ceil(u) < 0 || floor(u) > width - 1 || ceil(v) < 0 || floor(v) > height - 1) {
if (float_output) {
copy_vn_fl(float_output, components, 0.0f);
}
if (byte_output) {
copy_vn_uchar(byte_output, components, 0);
}
/* Sample area entirely outside image?
* We check if any of (iu+1, iv+1, width, height) < (0, 0, iu+1, iv+1). */
__m128i i_uv_1 = _mm_add_epi32(i_uv, _mm_set_epi32(0, 0, 1, 1));
__m128i cmp_a = _mm_or_si128(i_uv_1, _mm_set_epi32(height, width, 0, 0));
__m128i cmp_b = _mm_shuffle_epi32(i_uv_1, _MM_SHUFFLE(1, 0, 3, 2));
__m128i invalid = _mm_cmplt_epi32(cmp_a, cmp_b);
if (_mm_movemask_ps(_mm_castsi128_ps(invalid)) != 0) {
memset(output, 0, 4);
return;
}
i = (int)floor(u);
j = (int)floor(v);
a = u - (float)i;
b = v - (float)j;
__m128 frac_uv = _mm_sub_ps(uv, uv_floor);
zero_v4(out);
/* Calculate pixel weights. */
blender::float4 wx = cubic_bspline_coefficients(_mm_cvtss_f32(frac_uv));
blender::float4 wy = cubic_bspline_coefficients(
_mm_cvtss_f32(_mm_shuffle_ps(frac_uv, frac_uv, 1)));
/* Optimized and not so easy to read */
/* Read 4x4 source pixels and blend them. */
__m128 out = _mm_setzero_ps();
int iu = _mm_cvtsi128_si32(i_uv);
int iv = _mm_cvtsi128_si32(_mm_shuffle_epi32(i_uv, 1));
/* avoid calling multiple times */
wy[0] = P(b - (-1));
wy[1] = P(b - 0);
wy[2] = P(b - 1);
wy[3] = P(b - 2);
for (int n = 0; n < 4; n++) {
int y1 = iv + n - 1;
CLAMP(y1, 0, height - 1);
for (int m = 0; m < 4; m++) {
for (n = -1; n <= 2; n++) {
x1 = i + n;
CLAMP(x1, 0, width - 1);
wx = P((float)n - a);
for (m = -1; m <= 2; m++) {
float data[4];
int x1 = iu + m - 1;
CLAMP(x1, 0, width - 1);
float w = wx[m] * wy[n];
y1 = j + m;
CLAMP(y1, 0, height - 1);
/* Normally we could do this:
* `w = P(n-a) * P(b-m);`
* except that would call `P()` 16 times per pixel therefor `pow()` 64 times,
* better pre-calculate these. */
w = wx * wy[m + 1];
const uchar *data = src_buffer + (width * y1 + x1) * 4;
/* Load 4 bytes and expand into 4-lane SIMD. */
__m128i sample_i = _mm_castps_si128(_mm_load_ss((const float *)data));
sample_i = _mm_unpacklo_epi8(sample_i, _mm_setzero_si128());
sample_i = _mm_unpacklo_epi16(sample_i, _mm_setzero_si128());
if (float_output) {
const float *float_data = float_buffer + width * y1 * components + components * x1;
/* Accumulate into out with weight. */
out = _mm_add_ps(out, _mm_mul_ps(_mm_cvtepi32_ps(sample_i), _mm_set1_ps(w)));
}
}
vector_from_float(float_data, data, components);
}
else {
const uchar *byte_data = byte_buffer + width * y1 * components + components * x1;
/* Pack and write to destination: pack to 16 bit signed, then to 8 bit
* unsigned, then write resulting 32-bit value. */
out = _mm_add_ps(out, _mm_set1_ps(0.5f));
__m128i rgba32 = _mm_cvttps_epi32(out);
__m128i rgba16 = _mm_packs_epi32(rgba32, _mm_setzero_si128());
__m128i rgba8 = _mm_packus_epi16(rgba16, _mm_setzero_si128());
_mm_store_ss((float *)output, _mm_castsi128_ps(rgba8));
}
#endif /* BLI_HAVE_SSE2 */
vector_from_byte(byte_data, data, components);
}
template<typename T>
static void bicubic_interpolation(
const T *src_buffer, T *output, int width, int height, int components, float u, float v)
{
using namespace blender;
#if BLI_HAVE_SSE2
if constexpr (std::is_same_v<T, uchar>) {
if (components == 4) {
bicubic_interpolation_uchar_simd(src_buffer, output, width, height, u, v);
return;
}
}
#endif
int iu = (int)floor(u);
int iv = (int)floor(v);
/* Sample area entirely outside image? */
if (iu + 1 < 0 || iu > width - 1 || iv + 1 < 0 || iv > height - 1) {
memset(output, 0, size_t(components) * sizeof(T));
return;
}
float frac_u = u - (float)iu;
float frac_v = v - (float)iv;
float4 out{0.0f};
/* Calculate pixel weights. */
float4 wx = cubic_bspline_coefficients(frac_u);
float4 wy = cubic_bspline_coefficients(frac_v);
/* Read 4x4 source pixels and blend them. */
for (int n = 0; n < 4; n++) {
int y1 = iv + n - 1;
CLAMP(y1, 0, height - 1);
for (int m = 0; m < 4; m++) {
int x1 = iu + m - 1;
CLAMP(x1, 0, width - 1);
float w = wx[m] * wy[n];
const T *data = src_buffer + (width * y1 + x1) * components;
if (components == 1) {
out[0] += data[0] * w;
@ -165,72 +173,32 @@ BLI_INLINE void bicubic_interpolation(const uchar *byte_buffer,
}
}
/* Done with optimized part */
#if 0
/* older, slower function, works the same as above */
for (n = -1; n <= 2; n++) {
for (m = -1; m <= 2; m++) {
x1 = i + n;
y1 = j + m;
if (x1 > 0 && x1 < width && y1 > 0 && y1 < height) {
float data[4];
if (float_output) {
const float *float_data = float_buffer + width * y1 * components + components * x1;
vector_from_float(float_data, data, components);
}
else {
const uchar *byte_data = byte_buffer + width * y1 * components + components * x1;
vector_from_byte(byte_data, data, components);
}
if (components == 1) {
out[0] += data[0] * P(n - a) * P(b - m);
}
else if (components == 3) {
out[0] += data[0] * P(n - a) * P(b - m);
out[1] += data[1] * P(n - a) * P(b - m);
out[2] += data[2] * P(n - a) * P(b - m);
}
else {
out[0] += data[0] * P(n - a) * P(b - m);
out[1] += data[1] * P(n - a) * P(b - m);
out[2] += data[2] * P(n - a) * P(b - m);
out[3] += data[3] * P(n - a) * P(b - m);
}
}
}
}
#endif
if (float_output) {
/* Write result. */
if constexpr (std::is_same_v<T, float>) {
if (components == 1) {
float_output[0] = out[0];
output[0] = out[0];
}
else if (components == 3) {
copy_v3_v3(float_output, out);
copy_v3_v3(output, out);
}
else {
copy_v4_v4(float_output, out);
copy_v4_v4(output, out);
}
}
else {
if (components == 1) {
byte_output[0] = (uchar)(out[0] + 0.5f);
output[0] = (uchar)(out[0] + 0.5f);
}
else if (components == 3) {
byte_output[0] = (uchar)(out[0] + 0.5f);
byte_output[1] = (uchar)(out[1] + 0.5f);
byte_output[2] = (uchar)(out[2] + 0.5f);
output[0] = (uchar)(out[0] + 0.5f);
output[1] = (uchar)(out[1] + 0.5f);
output[2] = (uchar)(out[2] + 0.5f);
}
else {
byte_output[0] = (uchar)(out[0] + 0.5f);
byte_output[1] = (uchar)(out[1] + 0.5f);
byte_output[2] = (uchar)(out[2] + 0.5f);
byte_output[3] = (uchar)(out[3] + 0.5f);
output[0] = (uchar)(out[0] + 0.5f);
output[1] = (uchar)(out[1] + 0.5f);
output[2] = (uchar)(out[2] + 0.5f);
output[3] = (uchar)(out[3] + 0.5f);
}
}
}
@ -238,13 +206,13 @@ BLI_INLINE void bicubic_interpolation(const uchar *byte_buffer,
void BLI_bicubic_interpolation_fl(
const float *buffer, float *output, int width, int height, int components, float u, float v)
{
bicubic_interpolation(NULL, buffer, NULL, output, width, height, components, u, v);
bicubic_interpolation<float>(buffer, output, width, height, components, u, v);
}
void BLI_bicubic_interpolation_char(
const uchar *buffer, uchar *output, int width, int height, float u, float v)
{
bicubic_interpolation(buffer, NULL, output, NULL, width, height, 4, u, v);
bicubic_interpolation<uchar>(buffer, output, width, height, 4, u, v);
}
/* BILINEAR INTERPOLATION */
@ -381,18 +349,7 @@ void BLI_bilinear_interpolation_char(
* later making sure that the result is set to zero for that sample. */
__m128 uvuv = _mm_set_ps(v, u, v, u);
# if defined(__SSE4_1__) || defined(__ARM_NEON) && defined(WITH_SSE2NEON)
/* If we're on SSE4 or ARM NEON, just use the simple floor() way. */
__m128 uvuv_floor = _mm_floor_ps(uvuv);
# else
/* The hard way: truncate, for negative inputs this will round towards zero.
* Then compare with input UV, and subtract 1 for the inputs that were
* negative. */
__m128 uv_trunc = _mm_cvtepi32_ps(_mm_cvttps_epi32(uvuv));
__m128 uv_neg = _mm_cmplt_ps(uvuv, uv_trunc);
__m128 uvuv_floor = _mm_sub_ps(uv_trunc, _mm_and_ps(uv_neg, _mm_set1_ps(1.0f)));
# endif
__m128 uvuv_floor = floor_simd(uvuv);
/* x1, y1, x2, y2 */
__m128i xy12 = _mm_add_epi32(_mm_cvttps_epi32(uvuv_floor), _mm_set_epi32(1, 1, 0, 0));