vo_opengl: add option for caching shaders on disk

Mostly because of ANGLE (sadly).

The implementation became unpleasantly big, but at least it's relatively
self-contained.

I'm not sure to what degree shaders from different drivers are
compatible as in whether a driver would randomly misbehave if it's fed
a binary created by another driver. The useless binayFormat parameter
won't help it, as they can probably easily clash. As usual, OpenGL is
pretty shit here.

6 files changed, 142 insertions, 12 deletions

diff --git a/video/out/opengl/common.c b/video/out/opengl/common.c
index 5762f44440..271214954a 100644
--- a/video/out/opengl/common.c
+++ b/video/out/opengl/common.c
@@ -316,6 +316,16 @@ static const struct gl_functions gl_functions[] = {
             {0}
         },
     },
+    {
+        .ver_core = 410,
+        .ver_es_core = 300,
+        .extension = "GL_ARB_get_program_binary",
+        .functions = (const struct gl_function[]) {
+            DEF_FN(GetProgramBinary),
+            DEF_FN(ProgramBinary),
+            {0}
+        },
+    },
     // Swap control, always an OS specific extension
     // The OSX code loads this manually.
     {
diff --git a/video/out/opengl/common.h b/video/out/opengl/common.h
index 11cc37ba8a..3eb2a8ecf8 100644
--- a/video/out/opengl/common.h
+++ b/video/out/opengl/common.h
@@ -150,6 +150,10 @@ struct GL {
     void (GLAPIENTRY *GetShaderiv)(GLuint, GLenum, GLint *);
     void (GLAPIENTRY *GetProgramInfoLog)(GLuint, GLsizei, GLsizei *, GLchar *);
     void (GLAPIENTRY *GetProgramiv)(GLenum, GLenum, GLint *);
+    void (GLAPIENTRY *GetProgramBinary)(GLuint, GLsizei, GLsizei *, GLenum *,
+                                        void *);
+    void (GLAPIENTRY *ProgramBinary)(GLuint, GLenum, const void *, GLsizei);
+
     const GLubyte* (GLAPIENTRY *GetStringi)(GLenum, GLuint);
     void (GLAPIENTRY *BindAttribLocation)(GLuint, GLuint, const GLchar *);
     void (GLAPIENTRY *BindFramebuffer)(GLenum, GLuint);
diff --git a/video/out/opengl/utils.c b/video/out/opengl/utils.c
index c080f71299..7e8680fff2 100644
--- a/video/out/opengl/utils.c
+++ b/video/out/opengl/utils.c
@@ -23,7 +23,15 @@
 #include <stdarg.h>
 #include <assert.h>
 
+#include <libavutil/sha.h>
+#include <libavutil/intreadwrite.h>
+#include <libavutil/mem.h>
+
+#include "osdep/io.h"
+
 #include "common/common.h"
+#include "options/path.h"
+#include "stream/stream.h"
 #include "formats.h"
 #include "utils.h"
 
@@ -488,6 +496,10 @@ struct gl_shader_cache {
 
     // temporary buffers (avoids frequent reallocations)
     bstr tmp[5];
+
+    // For the disk-cache.
+    char *cache_dir;
+    struct mpv_global *global; // can be NULL
 };
 
 struct gl_shader_cache *gl_sc_create(GL *gl, struct mp_log *log)
@@ -817,6 +829,14 @@ static void update_uniform(GL *gl, struct sc_entry *e, struct sc_uniform *u, int
     }
 }
 
+void gl_sc_set_cache_dir(struct gl_shader_cache *sc, struct mpv_global *global,
+                         const char *dir)
+{
+    talloc_free(sc->cache_dir);
+    sc->cache_dir = talloc_strdup(sc, dir);
+    sc->global = global;
+}
+
 static void compile_attach_shader(struct gl_shader_cache *sc, GLuint program,
                                   GLenum type, const char *source)
 {
@@ -882,18 +902,10 @@ static void link_shader(struct gl_shader_cache *sc, GLuint program)
         sc->error_state = true;
 }
 
-static GLuint create_program(struct gl_shader_cache *sc, const char *vertex,
-                             const char *frag)
+static GLuint compile_program(struct gl_shader_cache *sc, const char *vertex,
+                              const char *frag)
 {
     GL *gl = sc->gl;
-    MP_VERBOSE(sc, "recompiling a shader program:\n");
-    if (sc->header_text.len) {
-        MP_VERBOSE(sc, "header:\n");
-        mp_log_source(sc->log, MSGL_V, sc->header_text.start);
-        MP_VERBOSE(sc, "body:\n");
-    }
-    if (sc->text.len)
-        mp_log_source(sc->log, MSGL_V, sc->text.start);
     GLuint prog = gl->CreateProgram();
     compile_attach_shader(sc, prog, GL_VERTEX_SHADER, vertex);
     compile_attach_shader(sc, prog, GL_FRAGMENT_SHADER, frag);
@@ -906,6 +918,105 @@ static GLuint create_program(struct gl_shader_cache *sc, const char *vertex,
     return prog;
 }
 
+static GLuint load_program(struct gl_shader_cache *sc, const char *vertex,
+                           const char *frag)
+{
+    GL *gl = sc->gl;
+
+    MP_VERBOSE(sc, "new shader program:\n");
+    if (sc->header_text.len) {
+        MP_VERBOSE(sc, "header:\n");
+        mp_log_source(sc->log, MSGL_V, sc->header_text.start);
+        MP_VERBOSE(sc, "body:\n");
+    }
+    if (sc->text.len)
+        mp_log_source(sc->log, MSGL_V, sc->text.start);
+
+    if (!sc->cache_dir || !sc->cache_dir[0] || !gl->ProgramBinary)
+        return compile_program(sc, vertex, frag);
+
+    // Try to load it from a disk cache, or compiling + saving it.
+
+    GLuint prog = 0;
+    void *tmp = talloc_new(NULL);
+    char *dir = mp_get_user_path(tmp, sc->global, sc->cache_dir);
+
+    struct AVSHA *sha = av_sha_alloc();
+    if (!sha)
+        abort();
+    av_sha_init(sha, 256);
+
+    av_sha_update(sha, vertex, strlen(vertex) + 1);
+    av_sha_update(sha, frag, strlen(frag) + 1);
+
+    // In theory, the array could change order, breaking old binaries.
+    for (int n = 0; sc->vao->entries[n].name; n++) {
+        av_sha_update(sha, sc->vao->entries[n].name,
+                      strlen(sc->vao->entries[n].name) + 1);
+    }
+
+    uint8_t hash[256 / 8];
+    av_sha_final(sha, hash);
+    av_free(sha);
+
+    char hashstr[256 / 8 * 2 + 1];
+    for (int n = 0; n < 256 / 8; n++)
+        snprintf(hashstr + n * 2, sizeof(hashstr) - n * 2, "%02X", hash[n]);
+
+    const char *header = "mpv shader cache v1\n";
+    size_t header_size = strlen(header) + 4;
+
+    char *filename = mp_path_join(tmp, dir, hashstr);
+    if (stat(filename, &(struct stat){0}) == 0) {
+        MP_VERBOSE(sc, "Trying to load shader from disk...\n");
+        struct bstr cachedata = stream_read_file(filename, tmp, sc->global,
+                                                 1000000000); // 1 GB
+        if (cachedata.len > header_size) {
+            GLenum format = AV_RL32(cachedata.start + header_size - 4);
+            prog = gl->CreateProgram();
+            gl_check_error(gl, sc->log, "before loading program");
+            gl->ProgramBinary(prog, format, cachedata.start + header_size,
+                                            cachedata.len - header_size);
+            gl->GetError(); // discard potential useless error
+            GLint status = 0;
+            gl->GetProgramiv(prog, GL_LINK_STATUS, &status);
+            if (!status) {
+                gl->DeleteProgram(prog);
+                prog = 0;
+            }
+        }
+        MP_VERBOSE(sc, "Loading cached shader %s.\n", prog ? "ok" : "failed");
+    }
+
+    if (!prog) {
+        prog = compile_program(sc, vertex, frag);
+
+        GLint size = 0;
+        gl->GetProgramiv(prog, GL_PROGRAM_BINARY_LENGTH, &size);
+        uint8_t *buffer = talloc_size(tmp, size + header_size);
+        GLsizei actual_size = 0;
+        GLenum binary_format = 0;
+        gl->GetProgramBinary(prog, size, &actual_size, &binary_format,
+                             buffer + header_size);
+        memcpy(buffer, header, header_size - 4);
+        AV_WL32(buffer + header_size - 4, binary_format);
+
+        if (actual_size) {
+            mp_mkdirp(dir);
+
+            MP_VERBOSE(sc, "Writing shader cache file: %s\n", filename);
+            FILE *out = fopen(filename, "wb");
+            if (out) {
+                fwrite(buffer, header_size + actual_size, 1, out);
+                fclose(out);
+            }
+        }
+    }
+
+    talloc_free(tmp);
+    return prog;
+}
+
 #define ADD(x, ...) bstr_xappend_asprintf(sc, (x), __VA_ARGS__)
 #define ADD_BSTR(x, s) bstr_xappend(sc, (x), (s))
 
@@ -1030,7 +1141,7 @@ void gl_sc_generate(struct gl_shader_cache *sc)
     }
     // build vertex shader from vao and cache the locations of the uniform variables
     if (!entry->gl_shader) {
-        entry->gl_shader = create_program(sc, vert->start, frag->start);
+        entry->gl_shader = load_program(sc, vert->start, frag->start);
         entry->num_uniforms = 0;
         for (int n = 0; n < sc->num_uniforms; n++) {
             struct sc_cached_uniform un = {
diff --git a/video/out/opengl/utils.h b/video/out/opengl/utils.h
index aa936f6e35..95eb1c4fea 100644
--- a/video/out/opengl/utils.h
+++ b/video/out/opengl/utils.h
@@ -171,6 +171,9 @@ void gl_sc_set_vao(struct gl_shader_cache *sc, struct gl_vao *vao);
 void gl_sc_enable_extension(struct gl_shader_cache *sc, char *name);
 void gl_sc_generate(struct gl_shader_cache *sc);
 void gl_sc_reset(struct gl_shader_cache *sc);
+struct mpv_global;
+void gl_sc_set_cache_dir(struct gl_shader_cache *sc, struct mpv_global *global,
+                         const char *dir);
 
 struct gl_timer;
 
diff --git a/video/out/opengl/video.c b/video/out/opengl/video.c
index 38d3d413e2..251a3d4f8b 100644
--- a/video/out/opengl/video.c
+++ b/video/out/opengl/video.c
@@ -389,7 +389,7 @@ const struct m_sub_options gl_video_conf = {
         OPT_SUBSTRUCT("", icc_opts, mp_icc_conf, 0),
         OPT_CHOICE("opengl-early-flush", early_flush, 0,
                    ({"no", 0}, {"yes", 1}, {"auto", -1})),
-
+        OPT_STRING("opengl-shader-cache-dir", shader_cache_dir, 0),
         {0}
     },
     .size = sizeof(struct gl_video_opts),
@@ -3362,6 +3362,7 @@ static void reinit_from_options(struct gl_video *p)
 
     check_gl_features(p);
     uninit_rendering(p);
+    gl_sc_set_cache_dir(p->sc, p->global, p->opts.shader_cache_dir);
     gl_video_setup_hooks(p);
     reinit_osd(p);
 
diff --git a/video/out/opengl/video.h b/video/out/opengl/video.h
index 3b5f452bf4..236ab067c4 100644
--- a/video/out/opengl/video.h
+++ b/video/out/opengl/video.h
@@ -134,6 +134,7 @@ struct gl_video_opts {
     int tex_pad_x, tex_pad_y;
     struct mp_icc_opts *icc_opts;
     int early_flush;
+    char *shader_cache_dir;
 };
 
 extern const struct m_sub_options gl_video_conf;