]> git.sesse.net Git - movit/commitdiff
Add input support for packed 10-bit Y'CbCr.
authorSteinar H. Gunderson <sgunderson@bigfoot.com>
Tue, 14 Feb 2017 17:50:07 +0000 (18:50 +0100)
committerSteinar H. Gunderson <sgunderson@bigfoot.com>
Tue, 14 Feb 2017 17:50:07 +0000 (18:50 +0100)
ycbcr_input.cpp
ycbcr_input.h
ycbcr_input_test.cpp

index 4a0ec0ea12b904e2a41f73b774a2c8da00fe43e6..f8df0c11bd687bcd862896dcd5b9a72e86e52a09 100644 (file)
@@ -19,10 +19,12 @@ namespace movit {
 YCbCrInput::YCbCrInput(const ImageFormat &image_format,
                        const YCbCrFormat &ycbcr_format,
                        unsigned width, unsigned height,
-                       YCbCrInputSplitting ycbcr_input_splitting)
+                       YCbCrInputSplitting ycbcr_input_splitting,
+                       GLenum type)
        : image_format(image_format),
          ycbcr_format(ycbcr_format),
          ycbcr_input_splitting(ycbcr_input_splitting),
+         type(type),
          width(width),
          height(height),
          resource_pool(NULL)
@@ -69,12 +71,20 @@ void YCbCrInput::set_gl_state(GLuint glsl_program_num, const string& prefix, uns
                if (texture_num[channel] == 0 && (pbos[channel] != 0 || pixel_data[channel] != NULL)) {
                        GLenum format, internal_format;
                        if (channel == 0 && ycbcr_input_splitting == YCBCR_INPUT_INTERLEAVED) {
-                               format = GL_RGB;
-                               internal_format = GL_RGB8;
+                               if (type == GL_UNSIGNED_INT_2_10_10_10_REV) {
+                                       format = GL_RGBA;
+                                       internal_format = GL_RGB10_A2;
+                               } else {
+                                       assert(type == GL_UNSIGNED_BYTE);
+                                       format = GL_RGB;
+                                       internal_format = GL_RGB8;
+                               }
                        } else if (channel == 1 && ycbcr_input_splitting == YCBCR_INPUT_SPLIT_Y_AND_CBCR) {
+                               assert(type == GL_UNSIGNED_BYTE);
                                format = GL_RG;
                                internal_format = GL_RG8;
                        } else {
+                               assert(type == GL_UNSIGNED_BYTE);
                                format = GL_RED;
                                internal_format = GL_R8;
                        }
@@ -91,7 +101,7 @@ void YCbCrInput::set_gl_state(GLuint glsl_program_num, const string& prefix, uns
                        check_error();
                        glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
                        check_error();
-                       glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, widths[channel], heights[channel], format, GL_UNSIGNED_BYTE, pixel_data[channel]);
+                       glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, widths[channel], heights[channel], format, type, pixel_data[channel]);
                        check_error();
                        glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
                        check_error();
index 22208e006209de2430ab5a8ae0b4874ac74a7085..31ba42bcbbacb9daa9e3fd1cadd9a80b6dee4a19 100644 (file)
@@ -1,10 +1,20 @@
 #ifndef _MOVIT_YCBCR_INPUT_H
 #define _MOVIT_YCBCR_INPUT_H 1
 
-// YCbCrInput is for handling planar or 4:4:4 interleaved 8-bit Y'CbCr
-// (also sometimes, usually rather imprecisely, called “YUV”), which is typically
-// what you get from a video decoder. It upsamples planes as needed, using the
-// default linear upsampling OpenGL gives you.
+// YCbCrInput is for handling Y'CbCr (also sometimes, usually rather
+// imprecisely, called “YUV”), which is typically what you get from a video
+// decoder. It supports these formats:
+//
+//   * 8-bit planar Y'CbCr, possibly subsampled (e.g. 4:2:0).
+//   * 8-bit semiplanar Y'CbCr (Y' in one plane, CbCr in another),
+//     possibly subsampled.
+//   * 8-bit interleaved (chunked) Y'CbCr, no subsampling (4:4:4 only).
+//   * 10-bit interleaved (chunked) Y'CbCr packed into 32-bit words
+//     (10:10:10:2), no subsampling (4:4:4 only).
+//
+// For the former case, it upsamples planes as needed, using the default linear
+// upsampling OpenGL gives you. Note that YCbCr422InterleavedInput supports the
+// important special case of 8-bit 4:2:2 interleaved.
 
 #include <epoxy/gl.h>
 #include <assert.h>
@@ -20,9 +30,7 @@ namespace movit {
 
 class ResourcePool;
 
-// Whether the data is planar (Y', Cb and Cr in one texture each)
-// or not. Note that YCbCr422InterleavedInput supports the important special
-// case of 4:2:2 interleaved.
+// Whether the data is planar (Y', Cb and Cr in one texture each) or not.
 enum YCbCrInputSplitting {
        // The standard, default case; Y', Cb and Cr in one texture each.
        YCBCR_INPUT_PLANAR,
@@ -41,10 +49,13 @@ enum YCbCrInputSplitting {
 
 class YCbCrInput : public Input {
 public:
+       // Type can be GL_UNSIGNED_BYTE for 8-bit, or GL_UNSIGNED_INT_2_10_10_10_REV
+       // for 10-bit (YCBCR_INPUT_INTERLEAVED only).
        YCbCrInput(const ImageFormat &image_format,
                   const YCbCrFormat &ycbcr_format,
                   unsigned width, unsigned height,
-                  YCbCrInputSplitting ycbcr_input_splitting = YCBCR_INPUT_PLANAR);
+                  YCbCrInputSplitting ycbcr_input_splitting = YCBCR_INPUT_PLANAR,
+                  GLenum type = GL_UNSIGNED_BYTE);
        ~YCbCrInput();
 
        virtual std::string effect_type_id() const { return "YCbCrInput"; }
@@ -75,12 +86,22 @@ public:
        // the pointer (and PBO, if set) has to be valid at the time of the render call.
        void set_pixel_data(unsigned channel, const unsigned char *pixel_data, GLuint pbo = 0)
        {
+               assert(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_2_10_10_10_REV);
                assert(channel >= 0 && channel < num_channels);
                this->pixel_data[channel] = pixel_data;
                this->pbos[channel] = pbo;
                invalidate_pixel_data();
        }
 
+       void set_pixel_data(unsigned channel, const uint32_t *pixel_data, GLuint pbo = 0)
+       {
+               assert(type == GL_UNSIGNED_INT_2_10_10_10_REV);
+               assert(channel == 0);
+               this->pixel_data[channel] = reinterpret_cast<const unsigned char *>(pixel_data);
+               this->pbos[channel] = pbo;
+               invalidate_pixel_data();
+       }
+
        void invalidate_pixel_data();
 
        // Note: Sets pitch to width, so even if your pitch is unchanged,
@@ -142,6 +163,7 @@ private:
        YCbCrFormat ycbcr_format;
        GLuint num_channels;
        YCbCrInputSplitting ycbcr_input_splitting;
+       GLenum type;
        GLuint pbos[3], texture_num[3];
        GLint uniform_tex_y, uniform_tex_cb, uniform_tex_cr;
 
index 019cc07010aad40256382e16299310ac5b406c1f..7792b289ef6c1e43e8fd06cdc12dab700e412a46 100644 (file)
@@ -798,4 +798,64 @@ TEST(YCbCrInputTest, NoData) {
        // Don't care what the output was, just that it does not crash.
 }
 
+TEST(YCbCrInputTest, TenBitInterleaved) {
+       const int width = 1;
+       const int height = 5;
+
+       // Pure-color inputs, calculated using formulas 3.2, 3.3 and 3.4 from
+       // Rec. 709. (Except the first two, which are obvious given the 64–940
+       // range of luminance.)
+       unsigned expanded_data[width * height * 3] = {
+                 64, 512, 512,
+                940, 512, 512,
+                250, 409, 960,
+                691, 167, 105,
+                127, 960, 471,
+       };
+       float expected_data[4 * width * height] = {
+               0.0, 0.0, 0.0, 1.0,
+               1.0, 1.0, 1.0, 1.0,
+               1.0, 0.0, 0.0, 1.0,
+               0.0, 1.0, 0.0, 1.0,
+               0.0, 0.0, 1.0, 1.0,
+       };
+       float out_data[4 * width * height];
+
+       // Pack 32:32:32 to 10:10:10:2.
+       uint32_t data[width * height];
+       for (unsigned i = 0; i < width * height; ++i) {
+               data[i] =
+                        expanded_data[i * 3 + 0]        |
+                       (expanded_data[i * 3 + 1] << 10) |
+                       (expanded_data[i * 3 + 2] << 20);
+       }
+
+       EffectChainTester tester(NULL, width, height);
+
+       ImageFormat format;
+       format.color_space = COLORSPACE_sRGB;
+       format.gamma_curve = GAMMA_sRGB;
+
+       YCbCrFormat ycbcr_format;
+       ycbcr_format.luma_coefficients = YCBCR_REC_709;
+       ycbcr_format.full_range = false;
+       ycbcr_format.num_levels = 1024;  // 10-bit.
+       ycbcr_format.chroma_subsampling_x = 1;
+       ycbcr_format.chroma_subsampling_y = 1;
+       ycbcr_format.cb_x_position = 0.5f;
+       ycbcr_format.cb_y_position = 0.5f;
+       ycbcr_format.cr_x_position = 0.5f;
+       ycbcr_format.cr_y_position = 0.5f;
+
+       YCbCrInput *input = new YCbCrInput(format, ycbcr_format, width, height, YCBCR_INPUT_INTERLEAVED, GL_UNSIGNED_INT_2_10_10_10_REV);
+       input->set_pixel_data(0, data);
+       tester.get_chain()->add_input(input);
+
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_sRGB);
+
+       // We can set much tighter limits on this than 8-bit Y'CbCr;
+       // even tighter than the default limits.
+       expect_equal(expected_data, out_data, 4 * width, height, 0.002, 0.0003);
+}
+
 }  // namespace movit