From 4c2edecd7dc83caaaa37c797d66d9077e105eaee Mon Sep 17 00:00:00 2001 From: Niklas Haas Date: Sat, 10 Feb 2018 22:42:11 +0100 Subject: vo_gpu: refactor HDR peak detection algorithm The major changes are as follows: 1. Use `uint32_t` instead of `unsigned int` for the SSBO size calculation. This doesn't really matter, since a too-big buffer will still work just fine, but since `uint` is a 32-bit integer by definition this is the correct way to do it. 2. Pre-divide the frame_sum by the num_wg immediately at the end of a frame. This change was made to prevent overflow. At 4K screen size, this code is currently already very at risk of overflow, especially once I started playing with longer averaging sizes. Pre-dividing this out makes it just about fit into 32-bit even for worst-case PQ content. (It's technically also faster and easier this way, so I should have done it to begin with). Rename `frame_sum` to `frame_avg` to clearly signal the change in semantics. 3. Implement a scene transition detection algorithm. This basically compares the current frame's average brightness against the (averaged) value of the past frames. If it exceeds a threshold, which I experimentally configured, we reset the peak detection SSBO's state immediately - so that it just contains the current frame. This prevents annoying "eye adaptation"-like effects on scene transitions. 4. As a result of the previous change, we can now use a much larger buffer size by default, which results in a more stable and less flickery result. I experimented with values between 20 and 256 and settled on the new value of 64. (I also switched to a power-of-2 array size, because I like powers of two) --- video/out/gpu/video.c | 18 +++++++++--------- video/out/gpu/video.h | 2 +- video/out/gpu/video_shaders.c | 37 +++++++++++++++++++++++++++++++------ 3 files changed, 41 insertions(+), 16 deletions(-) diff --git a/video/out/gpu/video.c b/video/out/gpu/video.c index 9bf7baeb77..c27004e63b 100644 --- a/video/out/gpu/video.c +++ b/video/out/gpu/video.c @@ -2448,13 +2448,13 @@ static void pass_colormanage(struct gl_video *p, struct mp_colorspace src, bool bool detect_peak = p->opts.compute_hdr_peak >= 0 && mp_trc_is_hdr(src.gamma); if (detect_peak && !p->hdr_peak_ssbo) { struct { - unsigned int counter; - unsigned int frame_idx; - unsigned int frame_num; - unsigned int frame_max[PEAK_DETECT_FRAMES+1]; - unsigned int frame_sum[PEAK_DETECT_FRAMES+1]; - unsigned int total_max; - unsigned int total_sum; + uint32_t counter; + uint32_t frame_idx; + uint32_t frame_num; + uint32_t frame_max[PEAK_DETECT_FRAMES+1]; + uint32_t frame_sum[PEAK_DETECT_FRAMES+1]; + uint32_t total_max; + uint32_t total_sum; } peak_ssbo = {0}; struct ra_buf_params params = { @@ -2479,9 +2479,9 @@ static void pass_colormanage(struct gl_video *p, struct mp_colorspace src, bool "uint frame_idx;" "uint frame_num;" "uint frame_max[%d];" - "uint frame_sum[%d];" + "uint frame_avg[%d];" "uint total_max;" - "uint total_sum;", + "uint total_avg;", PEAK_DETECT_FRAMES + 1, PEAK_DETECT_FRAMES + 1 ); diff --git a/video/out/gpu/video.h b/video/out/gpu/video.h index 71666059f9..dad6d447f6 100644 --- a/video/out/gpu/video.h +++ b/video/out/gpu/video.h @@ -96,7 +96,7 @@ enum tone_mapping { }; // How many frames to average over for HDR peak detection -#define PEAK_DETECT_FRAMES 20 +#define PEAK_DETECT_FRAMES 63 struct gl_video_opts { int dumb_mode; diff --git a/video/out/gpu/video_shaders.c b/video/out/gpu/video_shaders.c index 8e33255390..eb3f287236 100644 --- a/video/out/gpu/video_shaders.c +++ b/video/out/gpu/video_shaders.c @@ -559,6 +559,10 @@ void pass_inverse_ootf(struct gl_shader_cache *sc, enum mp_csp_light light, floa // under a typical presentation gamma of about 2.0. static const float sdr_avg = 0.25; +// The threshold for which to consider an average luminance difference to be +// a sign of a scene change. +static const int scene_threshold = 0.2 * MP_REF_WHITE; + static void hdr_update_peak(struct gl_shader_cache *sc) { // For performance, we want to do as few atomic operations on global @@ -578,14 +582,16 @@ static void hdr_update_peak(struct gl_shader_cache *sc) GLSL(if (gl_LocalInvocationIndex == 0) {) GLSL( uint wg_avg = wg_sum / (gl_WorkGroupSize.x * gl_WorkGroupSize.y);) GLSL( atomicMax(frame_max[frame_idx], wg_avg);) - GLSL( atomicAdd(frame_sum[frame_idx], wg_avg);) + GLSL( atomicAdd(frame_avg[frame_idx], wg_avg);) GLSL(}) + const float refi = 1.0 / MP_REF_WHITE; + // Update the sig_peak/sig_avg from the old SSBO state GLSL(uint num_wg = gl_NumWorkGroups.x * gl_NumWorkGroups.y;) GLSL(if (frame_num > 0) {) - GLSLF(" float peak = float(total_max) / (%f * float(frame_num));\n", MP_REF_WHITE); - GLSLF(" float avg = float(total_sum) / (%f * float(frame_num * num_wg));\n", MP_REF_WHITE); + GLSLF(" float peak = %f * float(total_max) / float(frame_num);\n", refi); + GLSLF(" float avg = %f * float(total_avg) / float(frame_num);\n", refi); GLSLF(" sig_peak = max(1.0, peak);\n"); GLSLF(" sig_avg = max(%f, avg);\n", sdr_avg); GLSL(}); @@ -594,12 +600,31 @@ static void hdr_update_peak(struct gl_shader_cache *sc) GLSL(memoryBarrierBuffer();) GLSL(barrier();) GLSL(if (gl_LocalInvocationIndex == 0 && atomicAdd(counter, 1) == num_wg - 1) {) + + // Since we sum up all the workgroups, we also still need to divide the + // average by the number of work groups GLSL( counter = 0;) + GLSL( frame_avg[frame_idx] /= num_wg;) + GLSL( uint cur_max = frame_max[frame_idx];) + GLSL( uint cur_avg = frame_avg[frame_idx];) + + // Scene change detection + GLSL( int diff = int(frame_num * cur_avg) - int(total_avg);) + GLSLF(" if (abs(diff) > frame_num * %d) {\n", scene_threshold); + GLSL( frame_num = 0;) + GLSL( total_max = total_avg = 0;) + GLSLF(" for (uint i = 0; i < %d; i++)\n", PEAK_DETECT_FRAMES+1); + GLSL( frame_max[i] = frame_avg[i] = 0;) + GLSL( frame_max[frame_idx] = cur_max;) + GLSL( frame_avg[frame_idx] = cur_avg;) + GLSL( }) + // Add the current frame, then subtract and reset the next frame GLSLF(" uint next = (frame_idx + 1) %% %d;\n", PEAK_DETECT_FRAMES+1); - GLSL( total_max += frame_max[frame_idx] - frame_max[next];) - GLSL( total_sum += frame_sum[frame_idx] - frame_sum[next];) - GLSL( frame_max[next] = frame_sum[next] = 0;) + GLSL( total_max += cur_max - frame_max[next];) + GLSL( total_avg += cur_avg - frame_avg[next];) + GLSL( frame_max[next] = frame_avg[next] = 0;) + // Update the index and count GLSL( frame_idx = next;) GLSLF(" frame_num = min(frame_num + 1, %d);\n", PEAK_DETECT_FRAMES); -- cgit v1.2.3