]> git.sesse.net Git - movit/blobdiff - ycbcr_input.cpp
Small refactoring in YCbCrInput.
[movit] / ycbcr_input.cpp
index 65103713df923aa59d38317e183564d04e58e446..ee341f0124de5985da6c2f0d5eb7dcb37ea18a64 100644 (file)
-#include <string.h>
-#include <assert.h>
-
+#include <Eigen/Core>
 #include <Eigen/LU>
+#include <epoxy/gl.h>
+#include <assert.h>
+#include <stdio.h>
+#include <string.h>
 
-#include "ycbcr_input.h"
+#include "effect_util.h"
+#include "resource_pool.h"
 #include "util.h"
-#include "opengl.h"
+#include "ycbcr_input.h"
 
 using namespace Eigen;
-
-YCbCrInput::YCbCrInput(const ImageFormat &image_format,
-                       const YCbCrFormat &ycbcr_format,
-                       unsigned width, unsigned height)
-       : image_format(image_format),
-         ycbcr_format(ycbcr_format),
-         needs_update(false),
-         needs_pbo_recreate(false),
-         finalized(false),
-         needs_mipmaps(false),
-         width(width),
-         height(height)
+using namespace std;
+
+namespace movit {
+
+namespace {
+
+// OpenGL has texel center in (0.5, 0.5), but different formats have
+// chroma in various other places. If luma samples are X, the chroma
+// sample is *, and subsampling is 3x3, the situation with chroma
+// center in (0.5, 0.5) looks approximately like this:
+//
+//   X   X
+//     *   
+//   X   X
+//
+// If, on the other hand, chroma center is in (0.0, 0.5) (common
+// for e.g. MPEG-4), the figure changes to:
+//
+//   X   X
+//   *      
+//   X   X
+//
+// In other words, (0.0, 0.0) means that the chroma sample is exactly
+// co-sited on top of the top-left luma sample. Note, however, that
+// this is _not_ 0.5 texels to the left, since the OpenGL's texel center
+// is in (0.5, 0.5); it is in (0.25, 0.25). In a sense, the four luma samples
+// define a square where chroma position (0.0, 0.0) is in texel position
+// (0.25, 0.25) and chroma position (1.0, 1.0) is in texel position (0.75, 0.75)
+// (the outer border shows the borders of the texel itself, ie. from
+// (0, 0) to (1, 1)):
+//
+//  ---------
+// |         |
+// |  X---X  |
+// |  | * |  |
+// |  X---X  |
+// |         |
+//  ---------
+//
+// Also note that if we have no subsampling, the square will have zero
+// area and the chroma position does not matter at all.
+float compute_chroma_offset(float pos, unsigned subsampling_factor, unsigned resolution)
 {
-       pbos[0] = pbos[1] = pbos[2] = 0;
-       texture_num[0] = texture_num[1] = texture_num[2] = 0;
-
-       assert(width % ycbcr_format.chroma_subsampling_x == 0);
-       pitch[0] = widths[0] = width;
-       pitch[1] = widths[1] = width / ycbcr_format.chroma_subsampling_x;
-       pitch[2] = widths[2] = width / ycbcr_format.chroma_subsampling_x;
-
-       assert(height % ycbcr_format.chroma_subsampling_y == 0);
-       heights[0] = height;
-       heights[1] = height / ycbcr_format.chroma_subsampling_y;
-       heights[2] = height / ycbcr_format.chroma_subsampling_y;
-
-       register_int("needs_mipmaps", &needs_mipmaps);
+       float local_chroma_pos = (0.5 + pos * (subsampling_factor - 1)) / subsampling_factor;
+       return (0.5 - local_chroma_pos) / resolution;
 }
 
-YCbCrInput::~YCbCrInput()
+// Given <ycbcr_format>, compute the values needed to turn Y'CbCr into R'G'B';
+// first subtract the returned offset, then left-multiply the returned matrix
+// (the scaling is already folded into it).
+void compute_ycbcr_matrix(YCbCrFormat ycbcr_format, float* offset, Matrix3d* ycbcr_to_rgb)
 {
-       if (pbos[0] != 0) {
-               glDeleteBuffers(3, pbos);
-               check_error();
-       }
-       if (texture_num[0] != 0) {
-               glDeleteTextures(3, texture_num);
-               check_error();
-       }
-}
-
-void YCbCrInput::finalize()
-{
-       glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
-       check_error();
-
-       // Create PBOs to hold the textures holding the input image, and then the texture itself.
-       glGenBuffers(3, pbos);
-       check_error();
-       glGenTextures(3, texture_num);
-       check_error();
-
-       for (unsigned channel = 0; channel < 3; ++channel) {
-               glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbos[channel]);
-               check_error();
-               glBufferData(GL_PIXEL_UNPACK_BUFFER_ARB, pitch[channel] * heights[channel], NULL, GL_STREAM_DRAW);
-               check_error();
-               glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
-               check_error();
-               
-               glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
-               check_error();
-               glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-               check_error();
-               glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
-               check_error();
-               glTexImage2D(GL_TEXTURE_2D, 0, GL_LUMINANCE8, widths[channel], heights[channel], 0, GL_LUMINANCE, GL_UNSIGNED_BYTE, NULL);
-               check_error();
-               glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
-               check_error();
-       }
-
-       needs_update = true;
-       finalized = true;
-}
-       
-void YCbCrInput::set_gl_state(GLuint glsl_program_num, const std::string& prefix, unsigned *sampler_num)
-{
-       for (unsigned channel = 0; channel < 3; ++channel) {
-               glActiveTexture(GL_TEXTURE0 + *sampler_num + channel);
-               check_error();
-               glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
-               check_error();
-
-               if (needs_update || needs_pbo_recreate) {
-                       // Copy the pixel data into the PBO.
-                       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbos[channel]);
-                       check_error();
-
-                       if (needs_pbo_recreate) {
-                               // The pitch has changed; we need to reallocate this PBO.
-                               glBufferData(GL_PIXEL_UNPACK_BUFFER_ARB, pitch[channel] * heights[channel], NULL, GL_STREAM_DRAW);
-                               check_error();
-                       }
-
-                       void *mapped_pbo = glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_WRITE_ONLY);
-                       memcpy(mapped_pbo, pixel_data[channel], pitch[channel] * heights[channel]);
-
-                       glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB);
-                       check_error();
-
-                       // Re-upload the texture from the PBO.
-                       glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
-                       check_error();
-                       glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, widths[channel], heights[channel], GL_LUMINANCE, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
-                       check_error();
-                       glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
-                       check_error();
-                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
-                       check_error();
-                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
-                       check_error();
-                       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
-                       check_error();
-               }
-       }
-
-       // Bind samplers.
-       set_uniform_int(glsl_program_num, prefix, "tex_y", *sampler_num + 0);
-       set_uniform_int(glsl_program_num, prefix, "tex_cb", *sampler_num + 1);
-       set_uniform_int(glsl_program_num, prefix, "tex_cr", *sampler_num + 2);
-
-       *sampler_num += 3;
-       needs_update = false;
-       needs_pbo_recreate = false;
-}
-
-std::string YCbCrInput::output_fragment_shader()
-{
-       float coeff[3], offset[3], scale[3];
+       double coeff[3], scale[3];
 
        switch (ycbcr_format.luma_coefficients) {
        case YCBCR_REC_601:
@@ -153,6 +79,14 @@ std::string YCbCrInput::output_fragment_shader()
                coeff[1] = 0.7152;
                coeff[2] = 0.0722;
                break;
+
+       case YCBCR_REC_2020:
+               // Rec. 2020, page 4.
+               coeff[0] = 0.2627;
+               coeff[1] = 0.6780;
+               coeff[2] = 0.0593;
+               break;
+
        default:
                assert(false);
        }
@@ -166,7 +100,7 @@ std::string YCbCrInput::output_fragment_shader()
                scale[1] = 1.0;
                scale[2] = 1.0;
        } else {
-               // Rec. 601, page 4; Rec. 709, page 19.
+               // Rec. 601, page 4; Rec. 709, page 19; Rec. 2020, page 4.
                offset[0] = 16.0 / 255.0;
                offset[1] = 128.0 / 255.0;
                offset[2] = 128.0 / 255.0;
@@ -193,46 +127,136 @@ std::string YCbCrInput::output_fragment_shader()
        rgb_to_ycbcr(2,2) = -coeff[2] * cr_fac;
 
        // Inverting the matrix gives us what we need to go from YCbCr back to RGB.
-       Matrix3d ycbcr_to_rgb = rgb_to_ycbcr.inverse();
+       *ycbcr_to_rgb = rgb_to_ycbcr.inverse();
+
+       // Fold in the scaling.
+       *ycbcr_to_rgb *= Map<const Vector3d>(scale).asDiagonal();
+}
+
+}  // namespace
+
+YCbCrInput::YCbCrInput(const ImageFormat &image_format,
+                       const YCbCrFormat &ycbcr_format,
+                       unsigned width, unsigned height)
+       : image_format(image_format),
+         ycbcr_format(ycbcr_format),
+         width(width),
+         height(height),
+         resource_pool(NULL)
+{
+       pbos[0] = pbos[1] = pbos[2] = 0;
+       texture_num[0] = texture_num[1] = texture_num[2] = 0;
+
+       assert(width % ycbcr_format.chroma_subsampling_x == 0);
+       pitch[0] = widths[0] = width;
+       pitch[1] = widths[1] = width / ycbcr_format.chroma_subsampling_x;
+       pitch[2] = widths[2] = width / ycbcr_format.chroma_subsampling_x;
+
+       assert(height % ycbcr_format.chroma_subsampling_y == 0);
+       heights[0] = height;
+       heights[1] = height / ycbcr_format.chroma_subsampling_y;
+       heights[2] = height / ycbcr_format.chroma_subsampling_y;
+
+       pixel_data[0] = pixel_data[1] = pixel_data[2] = NULL;
+}
+
+YCbCrInput::~YCbCrInput()
+{
+       for (unsigned channel = 0; channel < 3; ++channel) {
+               if (texture_num[channel] != 0) {
+                       resource_pool->release_2d_texture(texture_num[channel]);
+               }
+       }
+}
+
+void YCbCrInput::set_gl_state(GLuint glsl_program_num, const string& prefix, unsigned *sampler_num)
+{
+       for (unsigned channel = 0; channel < 3; ++channel) {
+               glActiveTexture(GL_TEXTURE0 + *sampler_num + channel);
+               check_error();
 
-       std::string frag_shader;
+               if (texture_num[channel] == 0) {
+                       // (Re-)upload the texture.
+                       texture_num[channel] = resource_pool->create_2d_texture(GL_R8, widths[channel], heights[channel]);
+                       glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
+                       check_error();
+                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+                       check_error();
+                       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbos[channel]);
+                       check_error();
+                       glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
+                       check_error();
+                       glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
+                       check_error();
+                       glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, widths[channel], heights[channel], GL_RED, GL_UNSIGNED_BYTE, pixel_data[channel]);
+                       check_error();
+                       glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
+                       check_error();
+                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+                       check_error();
+                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+                       check_error();
+               } else {
+                       glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
+                       check_error();
+               }
+       }
+
+       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
+       check_error();
+
+       // Bind samplers.
+       set_uniform_int(glsl_program_num, prefix, "tex_y", *sampler_num + 0);
+       set_uniform_int(glsl_program_num, prefix, "tex_cb", *sampler_num + 1);
+       set_uniform_int(glsl_program_num, prefix, "tex_cr", *sampler_num + 2);
+
+       *sampler_num += 3;
+}
+
+string YCbCrInput::output_fragment_shader()
+{
+       float offset[3];
+       Matrix3d ycbcr_to_rgb;
+       compute_ycbcr_matrix(ycbcr_format, offset, &ycbcr_to_rgb);
+
+       string frag_shader;
 
        frag_shader = output_glsl_mat3("PREFIX(inv_ycbcr_matrix)", ycbcr_to_rgb);
+       frag_shader += output_glsl_vec3("PREFIX(offset)", offset[0], offset[1], offset[2]);
 
-       char buf[256];
-       sprintf(buf, "const vec3 PREFIX(offset) = vec3(%.8f, %.8f, %.8f);\n",
-               offset[0], offset[1], offset[2]);
-       frag_shader += buf;
-
-       sprintf(buf, "const vec3 PREFIX(scale) = vec3(%.8f, %.8f, %.8f);\n",
-               scale[0], scale[1], scale[2]);
-       frag_shader += buf;
-
-       // OpenGL has texel center in (0.5, 0.5), but different formats have
-       // chroma in various other places. If luma samples are X, the chroma
-       // sample is *, and subsampling is 3x3, the situation with chroma
-       // center in (0.5, 0.5) looks approximately like this:
-       //
-       //   X     X
-       //      *   
-       //   X     X
-       //
-       // If, on the other hand, chroma center is in (0.0, 0.5) (common
-       // for e.g. MPEG-4), the figure changes to:
-       //
-       //   X     X
-       //   *      
-       //   X     X
-       //
-       // Obviously, the chroma plane here needs to be moved to the left,
-       // which means _adding_ 0.5 to the texture coordinates when sampling
-       // chroma.
-       float chroma_offset_x = (0.5f - ycbcr_format.chroma_x_position) / widths[1];
-       float chroma_offset_y = (0.5f - ycbcr_format.chroma_y_position) / heights[1];
-       sprintf(buf, "const vec2 PREFIX(chroma_offset) = vec2(%.8f, %.8f);\n",
-               chroma_offset_x, chroma_offset_y);
-       frag_shader += buf;
+       float cb_offset_x = compute_chroma_offset(
+               ycbcr_format.cb_x_position, ycbcr_format.chroma_subsampling_x, widths[1]);
+       float cb_offset_y = compute_chroma_offset(
+               ycbcr_format.cb_y_position, ycbcr_format.chroma_subsampling_y, heights[1]);
+       frag_shader += output_glsl_vec2("PREFIX(cb_offset)", cb_offset_x, cb_offset_y);
+
+       float cr_offset_x = compute_chroma_offset(
+               ycbcr_format.cr_x_position, ycbcr_format.chroma_subsampling_x, widths[2]);
+       float cr_offset_y = compute_chroma_offset(
+               ycbcr_format.cr_y_position, ycbcr_format.chroma_subsampling_y, heights[2]);
+       frag_shader += output_glsl_vec2("PREFIX(cr_offset)", cr_offset_x, cr_offset_y);
 
        frag_shader += read_file("ycbcr_input.frag");
        return frag_shader;
 }
+
+void YCbCrInput::invalidate_pixel_data()
+{
+       for (unsigned channel = 0; channel < 3; ++channel) {
+               if (texture_num[channel] != 0) {
+                       resource_pool->release_2d_texture(texture_num[channel]);
+                       texture_num[channel] = 0;
+               }
+       }
+}
+
+bool YCbCrInput::set_int(const std::string& key, int value)
+{
+       if (key == "needs_mipmaps") {
+               // We currently do not support this.
+               return (value == 0);
+       }
+       return Effect::set_int(key, value);
+}
+
+}  // namespace movit