1
mirror of https://github.com/mpv-player/mpv synced 2024-11-14 22:48:35 +01:00
mpv/video/out/opengl/nnedi3.c

220 lines
7.9 KiB
C
Raw Normal View History

vo_opengl: implement NNEDI3 prescaler Implement NNEDI3, a neural network based deinterlacer. The shader is reimplemented in GLSL and supports both 8x4 and 8x6 sampling window now. This allows the shader to be licensed under LGPL2.1 so that it can be used in mpv. The current implementation supports uploading the NN weights (up to 51kb with placebo setting) in two different way, via uniform buffer object or hard coding into shader source. UBO requires OpenGL 3.1, which only guarantee 16kb per block. But I find that 64kb seems to be a default setting for recent card/driver (which nnedi3 is targeting), so I think we're fine here (with default nnedi3 setting the size of weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the "intBitsToFloat()" built-in function. This is necessary to precisely represent these weights in GLSL. I tried several human readable floating point number format (with really high precision as for single precision float), but for some reason they are not working nicely, bad pixels (with NaN value) could be produced with some weights set. We could also add support to upload these weights with texture, just for compatibility reason (etc. upscaling a still image with a low end graphics card). But as I tested, it's rather slow even with 1D texture (we probably had to use 2D texture due to dimension size limitation). Since there is always better choice to do NNEDI3 upscaling for still image (vapoursynth plugin), it's not implemented in this commit. If this turns out to be a popular demand from the user, it should be easy to add it later. For those who wants to optimize the performance a bit further, the bottleneck seems to be: 1. overhead to upload and access these weights, (in particular, the shader code will be regenerated for each frame, it's on CPU though). 2. "dot()" performance in the main loop. 3. "exp()" performance in the main loop, there are various fast implementation with some bit tricks (probably with the help of the intBitsToFloat function). The code is tested with nvidia card and driver (355.11), on Linux. Closes #2230
2015-10-28 02:37:55 +01:00
/*
* This file is part of mpv.
*
* mpv is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* mpv is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License along
* with mpv. If not, see <http://www.gnu.org/licenses/>.
*
* You can alternatively redistribute this file and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*/
#include "nnedi3.h"
#include <assert.h>
#include <stdint.h>
#include <float.h>
#include <libavutil/bswap.h>
#include "video.h"
#define GLSL(x) gl_sc_add(sc, #x "\n");
#define GLSLF(...) gl_sc_addf(sc, __VA_ARGS__)
#define GLSLH(x) gl_sc_hadd(sc, #x "\n");
#define GLSLHF(...) gl_sc_haddf(sc, __VA_ARGS__)
const struct nnedi3_opts nnedi3_opts_def = {
.neurons = 1,
.window = 0,
.upload = NNEDI3_UPLOAD_UBO,
};
#define OPT_BASE_STRUCT struct nnedi3_opts
const struct m_sub_options nnedi3_conf = {
.opts = (const m_option_t[]) {
OPT_CHOICE("neurons", neurons, 0,
({"16", 0},
{"32", 1},
{"64", 2},
{"128", 3})),
OPT_CHOICE("window", window, 0,
({"8x4", 0},
{"8x6", 1})),
OPT_CHOICE("upload", upload, 0,
({"ubo", NNEDI3_UPLOAD_UBO},
{"shader", NNEDI3_UPLOAD_SHADER})),
{0}
},
.size = sizeof(struct nnedi3_opts),
.defaults = &nnedi3_opts_def,
};
const static char nnedi3_weights[40320 * 4 + 1] =
#include "video/out/opengl/nnedi3_weights.inc"
;
const int nnedi3_weight_offsets[9] =
{0, 1088, 3264, 7616, 16320, 17920, 21120, 27520, 40320};
const int nnedi3_neurons[4] = {16, 32, 64, 128};
const int nnedi3_window_width[2] = {8, 8};
const int nnedi3_window_height[2] = {4, 6};
const float* get_nnedi3_weights(const struct nnedi3_opts *conf, int *size)
{
int idx = conf->window * 4 + conf->neurons;
const int offset = nnedi3_weight_offsets[idx];
*size = (nnedi3_weight_offsets[idx + 1] - offset) * 4;
return (const float*)(nnedi3_weights + offset * 4);
}
void pass_nnedi3(GL *gl, struct gl_shader_cache *sc, int planes, int tex_num,
int step, float tex_mul, const struct nnedi3_opts *conf,
vo_opengl: implement NNEDI3 prescaler Implement NNEDI3, a neural network based deinterlacer. The shader is reimplemented in GLSL and supports both 8x4 and 8x6 sampling window now. This allows the shader to be licensed under LGPL2.1 so that it can be used in mpv. The current implementation supports uploading the NN weights (up to 51kb with placebo setting) in two different way, via uniform buffer object or hard coding into shader source. UBO requires OpenGL 3.1, which only guarantee 16kb per block. But I find that 64kb seems to be a default setting for recent card/driver (which nnedi3 is targeting), so I think we're fine here (with default nnedi3 setting the size of weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the "intBitsToFloat()" built-in function. This is necessary to precisely represent these weights in GLSL. I tried several human readable floating point number format (with really high precision as for single precision float), but for some reason they are not working nicely, bad pixels (with NaN value) could be produced with some weights set. We could also add support to upload these weights with texture, just for compatibility reason (etc. upscaling a still image with a low end graphics card). But as I tested, it's rather slow even with 1D texture (we probably had to use 2D texture due to dimension size limitation). Since there is always better choice to do NNEDI3 upscaling for still image (vapoursynth plugin), it's not implemented in this commit. If this turns out to be a popular demand from the user, it should be easy to add it later. For those who wants to optimize the performance a bit further, the bottleneck seems to be: 1. overhead to upload and access these weights, (in particular, the shader code will be regenerated for each frame, it's on CPU though). 2. "dot()" performance in the main loop. 3. "exp()" performance in the main loop, there are various fast implementation with some bit tricks (probably with the help of the intBitsToFloat function). The code is tested with nvidia card and driver (355.11), on Linux. Closes #2230
2015-10-28 02:37:55 +01:00
struct gl_transform *transform)
{
assert(0 <= step && step < 2);
if (!conf)
conf = &nnedi3_opts_def;
const int neurons = nnedi3_neurons[conf->neurons];
const int width = nnedi3_window_width[conf->window];
const int height = nnedi3_window_height[conf->window];
const int offset = nnedi3_weight_offsets[conf->window * 4 + conf->neurons];
const uint32_t *weights = (const int*)(nnedi3_weights + offset * 4);
GLSLF("// nnedi3 (tex %d, step %d, neurons %d, window %dx%d, mode %d)\n",
tex_num, step + 1, neurons, width, height, conf->upload);
// This is required since each row will be encoded into vec4s
assert(width % 4 == 0);
const int sample_count = width * height / 4;
if (conf->upload == NNEDI3_UPLOAD_UBO) {
char buf[32];
snprintf(buf, sizeof(buf), "vec4 weights[%d];",
neurons * (sample_count * 2 + 1));
gl_sc_uniform_buffer(sc, "NNEDI3_WEIGHTS", buf, 0);
if (gl->glsl_version < 140)
gl_sc_enable_extension(sc, "GL_ARB_uniform_buffer_object");
vo_opengl: implement NNEDI3 prescaler Implement NNEDI3, a neural network based deinterlacer. The shader is reimplemented in GLSL and supports both 8x4 and 8x6 sampling window now. This allows the shader to be licensed under LGPL2.1 so that it can be used in mpv. The current implementation supports uploading the NN weights (up to 51kb with placebo setting) in two different way, via uniform buffer object or hard coding into shader source. UBO requires OpenGL 3.1, which only guarantee 16kb per block. But I find that 64kb seems to be a default setting for recent card/driver (which nnedi3 is targeting), so I think we're fine here (with default nnedi3 setting the size of weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the "intBitsToFloat()" built-in function. This is necessary to precisely represent these weights in GLSL. I tried several human readable floating point number format (with really high precision as for single precision float), but for some reason they are not working nicely, bad pixels (with NaN value) could be produced with some weights set. We could also add support to upload these weights with texture, just for compatibility reason (etc. upscaling a still image with a low end graphics card). But as I tested, it's rather slow even with 1D texture (we probably had to use 2D texture due to dimension size limitation). Since there is always better choice to do NNEDI3 upscaling for still image (vapoursynth plugin), it's not implemented in this commit. If this turns out to be a popular demand from the user, it should be easy to add it later. For those who wants to optimize the performance a bit further, the bottleneck seems to be: 1. overhead to upload and access these weights, (in particular, the shader code will be regenerated for each frame, it's on CPU though). 2. "dot()" performance in the main loop. 3. "exp()" performance in the main loop, there are various fast implementation with some bit tricks (probably with the help of the intBitsToFloat function). The code is tested with nvidia card and driver (355.11), on Linux. Closes #2230
2015-10-28 02:37:55 +01:00
} else if (conf->upload == NNEDI3_UPLOAD_SHADER) {
// Somehow necessary for hard coding approach.
GLSLH(#pragma optionNV(fastprecision on))
}
GLSLHF("float nnedi3(sampler2D tex, vec2 pos, vec2 tex_size, int plane, float tex_mul) {\n");
vo_opengl: implement NNEDI3 prescaler Implement NNEDI3, a neural network based deinterlacer. The shader is reimplemented in GLSL and supports both 8x4 and 8x6 sampling window now. This allows the shader to be licensed under LGPL2.1 so that it can be used in mpv. The current implementation supports uploading the NN weights (up to 51kb with placebo setting) in two different way, via uniform buffer object or hard coding into shader source. UBO requires OpenGL 3.1, which only guarantee 16kb per block. But I find that 64kb seems to be a default setting for recent card/driver (which nnedi3 is targeting), so I think we're fine here (with default nnedi3 setting the size of weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the "intBitsToFloat()" built-in function. This is necessary to precisely represent these weights in GLSL. I tried several human readable floating point number format (with really high precision as for single precision float), but for some reason they are not working nicely, bad pixels (with NaN value) could be produced with some weights set. We could also add support to upload these weights with texture, just for compatibility reason (etc. upscaling a still image with a low end graphics card). But as I tested, it's rather slow even with 1D texture (we probably had to use 2D texture due to dimension size limitation). Since there is always better choice to do NNEDI3 upscaling for still image (vapoursynth plugin), it's not implemented in this commit. If this turns out to be a popular demand from the user, it should be easy to add it later. For those who wants to optimize the performance a bit further, the bottleneck seems to be: 1. overhead to upload and access these weights, (in particular, the shader code will be regenerated for each frame, it's on CPU though). 2. "dot()" performance in the main loop. 3. "exp()" performance in the main loop, there are various fast implementation with some bit tricks (probably with the help of the intBitsToFloat function). The code is tested with nvidia card and driver (355.11), on Linux. Closes #2230
2015-10-28 02:37:55 +01:00
if (step == 0) {
*transform = (struct gl_transform){{{1.0,0.0}, {0.0,2.0}}, {0.0,-0.5}};
GLSLH(if (fract(pos.y * tex_size.y) < 0.5)
return texture(tex, pos + vec2(0, 0.25) / tex_size)[plane] * tex_mul;)
vo_opengl: implement NNEDI3 prescaler Implement NNEDI3, a neural network based deinterlacer. The shader is reimplemented in GLSL and supports both 8x4 and 8x6 sampling window now. This allows the shader to be licensed under LGPL2.1 so that it can be used in mpv. The current implementation supports uploading the NN weights (up to 51kb with placebo setting) in two different way, via uniform buffer object or hard coding into shader source. UBO requires OpenGL 3.1, which only guarantee 16kb per block. But I find that 64kb seems to be a default setting for recent card/driver (which nnedi3 is targeting), so I think we're fine here (with default nnedi3 setting the size of weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the "intBitsToFloat()" built-in function. This is necessary to precisely represent these weights in GLSL. I tried several human readable floating point number format (with really high precision as for single precision float), but for some reason they are not working nicely, bad pixels (with NaN value) could be produced with some weights set. We could also add support to upload these weights with texture, just for compatibility reason (etc. upscaling a still image with a low end graphics card). But as I tested, it's rather slow even with 1D texture (we probably had to use 2D texture due to dimension size limitation). Since there is always better choice to do NNEDI3 upscaling for still image (vapoursynth plugin), it's not implemented in this commit. If this turns out to be a popular demand from the user, it should be easy to add it later. For those who wants to optimize the performance a bit further, the bottleneck seems to be: 1. overhead to upload and access these weights, (in particular, the shader code will be regenerated for each frame, it's on CPU though). 2. "dot()" performance in the main loop. 3. "exp()" performance in the main loop, there are various fast implementation with some bit tricks (probably with the help of the intBitsToFloat function). The code is tested with nvidia card and driver (355.11), on Linux. Closes #2230
2015-10-28 02:37:55 +01:00
GLSLHF("#define GET(i, j) "
"(texture(tex, pos+vec2((i)-(%f),(j)-(%f)+0.25)/tex_size)[plane]*tex_mul)\n",
vo_opengl: implement NNEDI3 prescaler Implement NNEDI3, a neural network based deinterlacer. The shader is reimplemented in GLSL and supports both 8x4 and 8x6 sampling window now. This allows the shader to be licensed under LGPL2.1 so that it can be used in mpv. The current implementation supports uploading the NN weights (up to 51kb with placebo setting) in two different way, via uniform buffer object or hard coding into shader source. UBO requires OpenGL 3.1, which only guarantee 16kb per block. But I find that 64kb seems to be a default setting for recent card/driver (which nnedi3 is targeting), so I think we're fine here (with default nnedi3 setting the size of weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the "intBitsToFloat()" built-in function. This is necessary to precisely represent these weights in GLSL. I tried several human readable floating point number format (with really high precision as for single precision float), but for some reason they are not working nicely, bad pixels (with NaN value) could be produced with some weights set. We could also add support to upload these weights with texture, just for compatibility reason (etc. upscaling a still image with a low end graphics card). But as I tested, it's rather slow even with 1D texture (we probably had to use 2D texture due to dimension size limitation). Since there is always better choice to do NNEDI3 upscaling for still image (vapoursynth plugin), it's not implemented in this commit. If this turns out to be a popular demand from the user, it should be easy to add it later. For those who wants to optimize the performance a bit further, the bottleneck seems to be: 1. overhead to upload and access these weights, (in particular, the shader code will be regenerated for each frame, it's on CPU though). 2. "dot()" performance in the main loop. 3. "exp()" performance in the main loop, there are various fast implementation with some bit tricks (probably with the help of the intBitsToFloat function). The code is tested with nvidia card and driver (355.11), on Linux. Closes #2230
2015-10-28 02:37:55 +01:00
width / 2.0 - 1, (height - 1) / 2.0);
} else {
*transform = (struct gl_transform){{{2.0,0.0}, {0.0,1.0}}, {-0.5,0.0}};
GLSLH(if (fract(pos.x * tex_size.x) < 0.5)
return texture(tex, pos + vec2(0.25, 0) / tex_size)[plane] * tex_mul;)
vo_opengl: implement NNEDI3 prescaler Implement NNEDI3, a neural network based deinterlacer. The shader is reimplemented in GLSL and supports both 8x4 and 8x6 sampling window now. This allows the shader to be licensed under LGPL2.1 so that it can be used in mpv. The current implementation supports uploading the NN weights (up to 51kb with placebo setting) in two different way, via uniform buffer object or hard coding into shader source. UBO requires OpenGL 3.1, which only guarantee 16kb per block. But I find that 64kb seems to be a default setting for recent card/driver (which nnedi3 is targeting), so I think we're fine here (with default nnedi3 setting the size of weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the "intBitsToFloat()" built-in function. This is necessary to precisely represent these weights in GLSL. I tried several human readable floating point number format (with really high precision as for single precision float), but for some reason they are not working nicely, bad pixels (with NaN value) could be produced with some weights set. We could also add support to upload these weights with texture, just for compatibility reason (etc. upscaling a still image with a low end graphics card). But as I tested, it's rather slow even with 1D texture (we probably had to use 2D texture due to dimension size limitation). Since there is always better choice to do NNEDI3 upscaling for still image (vapoursynth plugin), it's not implemented in this commit. If this turns out to be a popular demand from the user, it should be easy to add it later. For those who wants to optimize the performance a bit further, the bottleneck seems to be: 1. overhead to upload and access these weights, (in particular, the shader code will be regenerated for each frame, it's on CPU though). 2. "dot()" performance in the main loop. 3. "exp()" performance in the main loop, there are various fast implementation with some bit tricks (probably with the help of the intBitsToFloat function). The code is tested with nvidia card and driver (355.11), on Linux. Closes #2230
2015-10-28 02:37:55 +01:00
GLSLHF("#define GET(i, j) "
"(texture(tex, pos+vec2((j)-(%f)+0.25,(i)-(%f))/tex_size)[plane]*tex_mul)\n",
vo_opengl: implement NNEDI3 prescaler Implement NNEDI3, a neural network based deinterlacer. The shader is reimplemented in GLSL and supports both 8x4 and 8x6 sampling window now. This allows the shader to be licensed under LGPL2.1 so that it can be used in mpv. The current implementation supports uploading the NN weights (up to 51kb with placebo setting) in two different way, via uniform buffer object or hard coding into shader source. UBO requires OpenGL 3.1, which only guarantee 16kb per block. But I find that 64kb seems to be a default setting for recent card/driver (which nnedi3 is targeting), so I think we're fine here (with default nnedi3 setting the size of weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the "intBitsToFloat()" built-in function. This is necessary to precisely represent these weights in GLSL. I tried several human readable floating point number format (with really high precision as for single precision float), but for some reason they are not working nicely, bad pixels (with NaN value) could be produced with some weights set. We could also add support to upload these weights with texture, just for compatibility reason (etc. upscaling a still image with a low end graphics card). But as I tested, it's rather slow even with 1D texture (we probably had to use 2D texture due to dimension size limitation). Since there is always better choice to do NNEDI3 upscaling for still image (vapoursynth plugin), it's not implemented in this commit. If this turns out to be a popular demand from the user, it should be easy to add it later. For those who wants to optimize the performance a bit further, the bottleneck seems to be: 1. overhead to upload and access these weights, (in particular, the shader code will be regenerated for each frame, it's on CPU though). 2. "dot()" performance in the main loop. 3. "exp()" performance in the main loop, there are various fast implementation with some bit tricks (probably with the help of the intBitsToFloat function). The code is tested with nvidia card and driver (355.11), on Linux. Closes #2230
2015-10-28 02:37:55 +01:00
(height - 1) / 2.0, width / 2.0 - 1);
}
GLSLHF("vec4 samples[%d];\n", sample_count);
for (int y = 0; y < height; y++)
for (int x = 0; x < width; x += 4) {
GLSLHF("samples[%d] = vec4(GET(%d, %d), GET(%d, %d),"
"GET(%d, %d), GET(%d, %d));\n",
(y * width + x) / 4, x, y, x+1, y, x+2, y, x+3, y);
}
GLSLHF("float sum = 0, sumsq = 0;"
"for (int i = 0; i < %d; i++) {"
"sum += dot(samples[i], vec4(1.0));"
"sumsq += dot(samples[i], samples[i]);"
"}\n", sample_count);
GLSLHF("float mstd0 = sum / %d.0;\n"
"float mstd1 = sumsq / %d.0 - mstd0 * mstd0;\n"
"float mstd2 = mix(0, inversesqrt(mstd1), mstd1 >= %.12e);\n"
"mstd1 *= mstd2;\n",
width * height, width * height, FLT_EPSILON);
GLSLHF("float vsum = 0, wsum = 0, sum1, sum2;\n");
if (conf->upload == NNEDI3_UPLOAD_SHADER) {
GLSLH(#define T(x) intBitsToFloat(x))
GLSLH(#define W(i,w0,w1,w2,w3) dot(samples[i],vec4(T(w0),T(w1),T(w2),T(w3))))
GLSLHF("#define WS(w0,w1) "
"sum1 = exp(sum1 * mstd2 + T(w0));"
"sum2 = sum2 * mstd2 + T(w1);"
"wsum += sum1;"
"vsum += sum1*(sum2/(1+abs(sum2)));\n");
for (int n = 0; n < neurons; n++) {
const uint32_t *weights_ptr = weights + (sample_count * 2 + 1) * 4 * n;
for (int s = 0; s < 2; s++) {
GLSLHF("sum%d", s + 1);
for (int i = 0; i < sample_count; i++) {
GLSLHF("%cW(%d,%d,%d,%d,%d)", i == 0 ? '=' : '+', i,
(int)av_le2ne32(weights_ptr[0]),
(int)av_le2ne32(weights_ptr[1]),
(int)av_le2ne32(weights_ptr[2]),
(int)av_le2ne32(weights_ptr[3]));
weights_ptr += 4;
}
GLSLHF(";");
}
GLSLHF("WS(%d,%d);\n", (int)av_le2ne32(weights_ptr[0]),
(int)av_le2ne32(weights_ptr[1]));
}
} else if (conf->upload == NNEDI3_UPLOAD_UBO) {
GLSLH(int idx = 0;)
GLSLHF("for (int n = 0; n < %d; n++) {\n", neurons);
for (int s = 0; s < 2; s++) {
GLSLHF("sum%d = 0;\n"
"for (int i = 0; i < %d; i++) {"
"sum%d += dot(samples[i], weights[idx++]);"
"}\n",
s + 1, sample_count, s + 1);
}
GLSLH(sum1 = exp(sum1 * mstd2 + weights[idx][0]);
sum2 = sum2 * mstd2 + weights[idx++][1];
wsum += sum1;
vsum += sum1*(sum2/(1+abs(sum2)));)
GLSLHF("}\n");
}
GLSLH(return clamp(mstd0 + 5.0 * vsum / wsum * mstd1, 0, 1);)
GLSLHF("}\n"); // nnedi3
GLSL(vec4 color = vec4(1.0);)
for (int i = 0; i < planes; i++) {
GLSLF("color[%d] = nnedi3(texture%d, texcoord%d, texture_size%d, %d, %f);\n",
i, tex_num, tex_num, tex_num, i, tex_mul);
vo_opengl: implement NNEDI3 prescaler Implement NNEDI3, a neural network based deinterlacer. The shader is reimplemented in GLSL and supports both 8x4 and 8x6 sampling window now. This allows the shader to be licensed under LGPL2.1 so that it can be used in mpv. The current implementation supports uploading the NN weights (up to 51kb with placebo setting) in two different way, via uniform buffer object or hard coding into shader source. UBO requires OpenGL 3.1, which only guarantee 16kb per block. But I find that 64kb seems to be a default setting for recent card/driver (which nnedi3 is targeting), so I think we're fine here (with default nnedi3 setting the size of weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the "intBitsToFloat()" built-in function. This is necessary to precisely represent these weights in GLSL. I tried several human readable floating point number format (with really high precision as for single precision float), but for some reason they are not working nicely, bad pixels (with NaN value) could be produced with some weights set. We could also add support to upload these weights with texture, just for compatibility reason (etc. upscaling a still image with a low end graphics card). But as I tested, it's rather slow even with 1D texture (we probably had to use 2D texture due to dimension size limitation). Since there is always better choice to do NNEDI3 upscaling for still image (vapoursynth plugin), it's not implemented in this commit. If this turns out to be a popular demand from the user, it should be easy to add it later. For those who wants to optimize the performance a bit further, the bottleneck seems to be: 1. overhead to upload and access these weights, (in particular, the shader code will be regenerated for each frame, it's on CPU though). 2. "dot()" performance in the main loop. 3. "exp()" performance in the main loop, there are various fast implementation with some bit tricks (probably with the help of the intBitsToFloat function). The code is tested with nvidia card and driver (355.11), on Linux. Closes #2230
2015-10-28 02:37:55 +01:00
}
}