If not using VA-API zerocopy, don't write extra copy textures.

[nageru] / theme.cpp

diff --git a/theme.cpp b/theme.cpp
index bcdd6f37bad5d69cd6462f6d0732064c21896a9d..0cc498e8ba0a4297be74bb320440f6c75cba2f7f 100644 (file)
--- a/theme.cpp
+++ b/theme.cpp
@@ -265,19 +265,32 @@ int EffectChain_finalize(lua_State* L)
                // happens in a pass not run by Movit (see ChromaSubsampler::subsample_chroma()).
                output_ycbcr_format.chroma_subsampling_x = 1;
                output_ycbcr_format.chroma_subsampling_y = 1;
+
+               // This will be overridden if HDMI/SDI output is in force.
                if (global_flags.ycbcr_rec709_coefficients) {
                        output_ycbcr_format.luma_coefficients = YCBCR_REC_709;
                } else {
                        output_ycbcr_format.luma_coefficients = YCBCR_REC_601;
                }
+
                output_ycbcr_format.full_range = false;
-               output_ycbcr_format.num_levels = 256;
+               output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
+
+               GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
+
+               chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, output_ycbcr_format, YCBCR_OUTPUT_SPLIT_Y_AND_CBCR, type);
 
-               chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, output_ycbcr_format, YCBCR_OUTPUT_SPLIT_Y_AND_CBCR);
-               chain->set_dither_bits(8);
+               // If we're using zerocopy video encoding (so the destination
+               // Y texture is owned by VA-API and will be unavailable for
+               // display), add a copy, where we'll only be using the Y component.
+               if (global_flags.use_zerocopy) {
+                       chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, output_ycbcr_format, YCBCR_OUTPUT_INTERLEAVED, type);  // Add a copy where we'll only be using the Y component.
+               }
+               chain->set_dither_bits(global_flags.x264_bit_depth > 8 ? 16 : 8);
                chain->set_output_origin(OUTPUT_ORIGIN_TOP_LEFT);
+       } else {
+               chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
        }
-       chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 
        chain->finalize();
        return 0;
@@ -603,8 +616,9 @@ LiveInputWrapper::LiveInputWrapper(Theme *theme, EffectChain *chain, bool overri
        // Perhaps 601 was only to indicate the subsampling positions, not the
        // colorspace itself? Tested with a Lenovo X1 gen 3 as input.
        YCbCrFormat input_ycbcr_format;
-       input_ycbcr_format.chroma_subsampling_x = 2;
+       input_ycbcr_format.chroma_subsampling_x = global_flags.ten_bit_input ? 1 : 2;
        input_ycbcr_format.chroma_subsampling_y = 1;
+       input_ycbcr_format.num_levels = global_flags.ten_bit_input ? 1024 : 256;
        input_ycbcr_format.cb_x_position = 0.0;
        input_ycbcr_format.cr_x_position = 0.0;
        input_ycbcr_format.cb_y_position = 0.5;
@@ -628,10 +642,12 @@ LiveInputWrapper::LiveInputWrapper(Theme *theme, EffectChain *chain, bool overri
                num_inputs = 1;
        }
        for (unsigned i = 0; i < num_inputs; ++i) {
+               // When using 10-bit input, we're converting to interleaved through v210Converter.
+               YCbCrInputSplitting splitting = global_flags.ten_bit_input ? YCBCR_INPUT_INTERLEAVED : YCBCR_INPUT_SPLIT_Y_AND_CBCR;
                if (override_bounce) {
-                       inputs.push_back(new NonBouncingYCbCrInput(inout_format, input_ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR));
+                       inputs.push_back(new NonBouncingYCbCrInput(inout_format, input_ycbcr_format, global_flags.width, global_flags.height, splitting));
                } else {
-                       inputs.push_back(new YCbCrInput(inout_format, input_ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR));
+                       inputs.push_back(new YCbCrInput(inout_format, input_ycbcr_format, global_flags.width, global_flags.height, splitting));
                }
                chain->add_input(inputs.back());
        }
@@ -680,8 +696,12 @@ void LiveInputWrapper::connect_signal(int signal_num)
                        userdata = (const PBOFrameAllocator::Userdata *)frame.frame->userdata;
                }
 
-               inputs[i]->set_texture_num(0, userdata->tex_y[frame.field_number]);
-               inputs[i]->set_texture_num(1, userdata->tex_cbcr[frame.field_number]);
+               if (global_flags.ten_bit_input) {
+                       inputs[i]->set_texture_num(0, userdata->tex_444[frame.field_number]);
+               } else {
+                       inputs[i]->set_texture_num(0, userdata->tex_y[frame.field_number]);
+                       inputs[i]->set_texture_num(1, userdata->tex_cbcr[frame.field_number]);
+               }
                inputs[i]->set_width(userdata->last_width[frame.field_number]);
                inputs[i]->set_height(userdata->last_height[frame.field_number]);
 
@@ -963,12 +983,28 @@ int Theme::map_signal(int signal_num)
        if (signal_to_card_mapping.count(signal_num)) {
                return signal_to_card_mapping[signal_num];
        }
-       if (signal_num >= int(num_cards)) {
-               fprintf(stderr, "WARNING: Theme asked for input %d, but we only have %u card(s).\n", signal_num, num_cards);
-               fprintf(stderr, "Mapping to card %d instead.\n", signal_num % num_cards);
+
+       int card_index;
+       if (global_flags.output_card != -1 && num_cards > 1) {
+               // Try to exclude the output card from the default card_index.
+               card_index = signal_num % (num_cards - 1);
+               if (card_index >= global_flags.output_card) {
+                        ++card_index;
+               }
+               if (signal_num >= int(num_cards - 1)) {
+                       fprintf(stderr, "WARNING: Theme asked for input %d, but we only have %u input card(s) (card %d is busy with output).\n",
+                               signal_num, num_cards - 1, global_flags.output_card);
+                       fprintf(stderr, "Mapping to card %d instead.\n", card_index);
+               }
+       } else {
+               card_index = signal_num % num_cards;
+               if (signal_num >= int(num_cards)) {
+                       fprintf(stderr, "WARNING: Theme asked for input %d, but we only have %u card(s).\n", signal_num, num_cards);
+                       fprintf(stderr, "Mapping to card %d instead.\n", card_index);
+               }
        }
-       signal_to_card_mapping[signal_num] = signal_num % num_cards;
-       return signal_num % num_cards;
+       signal_to_card_mapping[signal_num] = card_index;
+       return card_index;
 }
 
 void Theme::set_signal_mapping(int signal_num, int card_num)