vo_opengl: refactor tex_upload to ra_buf_pool

Also refactors the usage of tex_upload to make ra_tex_upload_pbo a
RA-internal thing again.

ra_buf_pool has the main advantage of being dynamically sized depending
on buf_poll, so for OpenGL we'll end up only using one buffer (when not
persistently mapping) - while for vulkan we'll use as many as necessary,
which depends on the swapchain depth anyway.

video/out/opengl/osd.c

@@ -54,7 +54,6 @@ struct mpgl_osd_part {
     enum sub_bitmap_format format;
     int change_id;
     struct ra_tex *texture;
-    struct tex_upload pbo;
     int w, h;
     int num_subparts;
     int prev_num_subparts;
@@ -71,7 +70,6 @@ struct mpgl_osd {
     const struct ra_format *fmt_table[SUBBITMAP_COUNT];
     bool formats[SUBBITMAP_COUNT];
     bool change_flag; // for reporting to API user only
-    bool want_pbo;
     // temporary
     int stereo_mode;
     struct mp_osd_res osd_res;
@@ -79,7 +77,7 @@ struct mpgl_osd {
 };
 
 struct mpgl_osd *mpgl_osd_init(struct ra *ra, struct mp_log *log,
-                               struct osd_state *osd, bool want_pbo)
+                               struct osd_state *osd)
 {
     struct mpgl_osd *ctx = talloc_ptrtype(NULL, ctx);
     *ctx = (struct mpgl_osd) {
@@ -88,7 +86,6 @@ struct mpgl_osd *mpgl_osd_init(struct ra *ra, struct mp_log *log,
         .ra = ra,
         .change_flag = true,
         .scratch = talloc_zero_size(ctx, 1),
-        .want_pbo = want_pbo,
     };
 
     ctx->fmt_table[SUBBITMAP_LIBASS] = ra_find_unorm_format(ra, 1, 1);
@@ -111,7 +108,6 @@ void mpgl_osd_destroy(struct mpgl_osd *ctx)
     for (int n = 0; n < MAX_OSD_PARTS; n++) {
         struct mpgl_osd_part *p = ctx->parts[n];
         ra_tex_free(ctx->ra, &p->texture);
-        tex_upload_uninit(ctx->ra, &p->pbo);
     }
     talloc_free(ctx);
 }
@@ -180,7 +176,7 @@ static bool upload_osd(struct mpgl_osd *ctx, struct mpgl_osd_part *osd,
         .stride = imgs->packed->stride[0],
     };
 
-    ok = tex_upload(ra, &osd->pbo, ctx->want_pbo, &params);
+    ok = ra->fns->tex_upload(ra, &params);
 
 done:
     return ok;

video/out/opengl/osd.h

@@ -9,7 +9,7 @@
 #include "sub/osd.h"
 
 struct mpgl_osd *mpgl_osd_init(struct ra *ra, struct mp_log *log,
-                               struct osd_state *osd, bool want_pbo);
+                               struct osd_state *osd);
 void mpgl_osd_destroy(struct mpgl_osd *ctx);
 
 void mpgl_osd_generate(struct mpgl_osd *ctx, struct mp_osd_res res, double pts,

video/out/opengl/ra.h

@@ -30,6 +30,11 @@ struct ra {
     // formats should have a lower index. (E.g. GLES3 should put rg8 before la.)
     struct ra_format **formats;
     int num_formats;
+
+    // Accelerate texture uploads via an extra PBO even when
+    // RA_CAP_DIRECT_UPLOAD is supported. This is basically only relevant for
+    // OpenGL. Set by the RA user.
+    bool use_pbo;
 };
 
 enum {

video/out/opengl/ra_gl.c

@@ -15,6 +15,7 @@ struct ra_gl {
 
 // For ra_tex.priv
 struct ra_tex_gl {
+    struct ra_buf_pool pbo; // for ra.use_pbo
     bool own_objects;
     GLenum target;
     GLuint texture; // 0 if no texture data associated
@@ -218,6 +219,8 @@ static void gl_tex_destroy(struct ra *ra, struct ra_tex *tex)
     GL *gl = ra_gl_get(ra);
     struct ra_tex_gl *tex_gl = tex->priv;
 
+    ra_buf_pool_uninit(ra, &tex_gl->pbo);
+
     if (tex_gl->own_objects) {
         if (tex_gl->fbo)
             gl->DeleteFramebuffers(1, &tex_gl->fbo);
@@ -435,6 +438,9 @@ static bool gl_tex_upload(struct ra *ra,
     assert(tex->params.host_mutable);
     assert(!params->buf || !params->src);
 
+    if (ra->use_pbo && !params->buf)
+        return ra_tex_upload_pbo(ra, &tex_gl->pbo, params);
+
     const void *src = params->src;
     if (buf) {
         gl->BindBuffer(GL_PIXEL_UNPACK_BUFFER, buf_gl->buffer);

video/out/opengl/utils.c

@@ -39,6 +39,91 @@ void gl_transform_ortho_fbodst(struct gl_transform *t, struct fbodst fbo)
     gl_transform_ortho(t, 0, fbo.tex->params.w, 0, fbo.tex->params.h * y_dir);
 }
 
+void ra_buf_pool_uninit(struct ra *ra, struct ra_buf_pool *pool)
+{
+    for (int i = 0; i < pool->num_buffers; i++)
+        ra_buf_free(ra, &pool->buffers[i]);
+
+    talloc_free(pool->buffers);
+    *pool = (struct ra_buf_pool){0};
+}
+
+static bool ra_buf_params_compatible(const struct ra_buf_params *new,
+                                     const struct ra_buf_params *old)
+{
+    return new->type == old->type &&
+           new->size <= old->size &&
+           new->host_mapped  == old->host_mapped &&
+           new->host_mutable == old->host_mutable;
+}
+
+static bool ra_buf_pool_grow(struct ra *ra, struct ra_buf_pool *pool)
+{
+    struct ra_buf *buf = ra_buf_create(ra, &pool->current_params);
+    if (!buf)
+        return false;
+
+    MP_TARRAY_INSERT_AT(NULL, pool->buffers, pool->num_buffers, pool->index, buf);
+    MP_VERBOSE(ra, "Resized buffer pool to size %d\n", pool->num_buffers);
+    return true;
+}
+
+struct ra_buf *ra_buf_pool_get(struct ra *ra, struct ra_buf_pool *pool,
+                               const struct ra_buf_params *params)
+{
+    assert(!params->initial_data);
+
+    if (!ra_buf_params_compatible(params, &pool->current_params)) {
+        ra_buf_pool_uninit(ra, pool);
+        pool->current_params = *params;
+    }
+
+    // Make sure we have at least one buffer available
+    if (!pool->buffers && !ra_buf_pool_grow(ra, pool))
+        return NULL;
+
+    // Make sure the next buffer is available for use
+    if (!ra->fns->buf_poll(ra, pool->buffers[pool->index]) &&
+        !ra_buf_pool_grow(ra, pool))
+    {
+        return NULL;
+    }
+
+    struct ra_buf *buf = pool->buffers[pool->index++];
+    pool->index %= pool->num_buffers;
+
+    return buf;
+}
+
+bool ra_tex_upload_pbo(struct ra *ra, struct ra_buf_pool *pbo,
+                       const struct ra_tex_upload_params *params)
+{
+    if (params->buf)
+        return ra->fns->tex_upload(ra, params);
+
+    struct ra_tex *tex = params->tex;
+    size_t row_size = tex->params.dimensions == 2 ? params->stride :
+                      tex->params.w * tex->params.format->pixel_size;
+
+    struct ra_buf_params bufparams = {
+        .type = RA_BUF_TYPE_TEX_UPLOAD,
+        .size = row_size * tex->params.h * tex->params.d,
+        .host_mutable = true,
+    };
+
+    struct ra_buf *buf = ra_buf_pool_get(ra, pbo, &bufparams);
+    if (!buf)
+        return false;
+
+    ra->fns->buf_update(ra, buf, 0, params->src, bufparams.size);
+
+    struct ra_tex_upload_params newparams = *params;
+    newparams.buf = buf;
+    newparams.src = NULL;
+
+    return ra->fns->tex_upload(ra, &newparams);
+}
+
 // Create a texture and a FBO using the texture as color attachments.
 //  fmt: texture internal format
 // If the parameters are the same as the previous call, do not touch it.
@@ -120,63 +205,6 @@ void fbotex_uninit(struct fbotex *fbo)
     }
 }
 
-bool tex_upload(struct ra *ra, struct tex_upload *pbo, bool want_pbo,
-                const struct ra_tex_upload_params *params)
-{
-    if (!(ra->caps & RA_CAP_DIRECT_UPLOAD))
-        want_pbo = true;
-
-    if (!want_pbo || params->buf)
-        return ra->fns->tex_upload(ra, params);
-
-    struct ra_tex *tex = params->tex;
-    size_t row_size = tex->params.dimensions == 2 ? params->stride :
-                      tex->params.w * tex->params.format->pixel_size;
-    size_t needed_size = row_size * tex->params.h * tex->params.d;
-
-    if (needed_size > pbo->buffer_size)
-        tex_upload_uninit(ra, pbo);
-
-    if (!pbo->buffers[0]) {
-        struct ra_buf_params bufparams = {
-            .type = RA_BUF_TYPE_TEX_UPLOAD,
-            .size = needed_size,
-            .host_mutable = true,
-        };
-
-        pbo->buffer_size = bufparams.size;
-        for (int i = 0; i < NUM_PBO_BUFFERS; i++) {
-            pbo->buffers[i] = ra_buf_create(ra, &bufparams);
-            if (!pbo->buffers[i])
-                return false;
-        }
-    }
-
-    struct ra_buf *buf = pbo->buffers[pbo->index++];
-    pbo->index %= NUM_PBO_BUFFERS;
-
-    if (!ra->fns->buf_poll(ra, buf)) {
-        MP_WARN(ra, "Texture upload buffer was not free to use! Try "
-                    "increasing NUM_PBO_BUFFERS.\n");
-        return false;
-    }
-
-    ra->fns->buf_update(ra, buf, 0, params->src, needed_size);
-
-    struct ra_tex_upload_params newparams = *params;
-    newparams.buf = buf;
-    newparams.src = NULL;
-
-    return ra->fns->tex_upload(ra, &newparams);
-}
-
-void tex_upload_uninit(struct ra *ra, struct tex_upload *pbo)
-{
-    for (int i = 0; i < NUM_PBO_BUFFERS; i++)
-        ra_buf_free(ra, &pbo->buffers[i]);
-    *pbo = (struct tex_upload){0};
-}
-
 struct timer_pool {
     struct ra *ra;
     ra_timer *timer;

video/out/opengl/utils.h

@@ -69,6 +69,25 @@ struct fbodst {
 
 void gl_transform_ortho_fbodst(struct gl_transform *t, struct fbodst fbo);
 
+// A pool of buffers, which can grow as needed
+struct ra_buf_pool {
+    struct ra_buf_params current_params;
+    struct ra_buf **buffers;
+    int num_buffers;
+    int index;
+};
+
+void ra_buf_pool_uninit(struct ra *ra, struct ra_buf_pool *pool);
+
+// Note: params->initial_data is *not* supported
+struct ra_buf *ra_buf_pool_get(struct ra *ra, struct ra_buf_pool *pool,
+                               const struct ra_buf_params *params);
+
+// Helper that wraps ra_tex_upload using texture upload buffers to ensure that
+// params->buf is always set. This is intended for RA-internal usage.
+bool ra_tex_upload_pbo(struct ra *ra, struct ra_buf_pool *pbo,
+                       const struct ra_tex_upload_params *params);
+
 struct fbotex {
     struct ra *ra;
     struct ra_tex *tex;
@@ -83,21 +102,6 @@ bool fbotex_change(struct fbotex *fbo, struct ra *ra, struct mp_log *log,
 #define FBOTEX_FUZZY_H 2
 #define FBOTEX_FUZZY (FBOTEX_FUZZY_W | FBOTEX_FUZZY_H)
 
-#define NUM_PBO_BUFFERS 3
-
-// A wrapper around tex_upload that uses PBOs internally if requested or
-// required
-struct tex_upload {
-    size_t buffer_size;
-    struct ra_buf *buffers[NUM_PBO_BUFFERS];
-    int index;
-};
-
-bool tex_upload(struct ra *ra, struct tex_upload *pbo, bool want_pbo,
-                const struct ra_tex_upload_params *params);
-
-void tex_upload_uninit(struct ra *ra, struct tex_upload *pbo);
-
 // A wrapper around ra_timer that does result pooling, averaging etc.
 struct timer_pool;
 

video/out/opengl/video.c

@@ -84,7 +84,6 @@ static const struct ra_renderpass_input vertex_vao[] = {
 
 struct texplane {
     struct ra_tex *tex;
-    struct tex_upload pbo;
     int w, h;
     bool flipped;
 };
@@ -494,7 +493,7 @@ static void reinit_osd(struct gl_video *p)
     mpgl_osd_destroy(p->osd);
     p->osd = NULL;
     if (p->osd_state)
-        p->osd = mpgl_osd_init(p->ra, p->log, p->osd_state, p->opts.pbo);
+        p->osd = mpgl_osd_init(p->ra, p->log, p->osd_state);
 }
 
 static void uninit_rendering(struct gl_video *p)
@@ -987,7 +986,6 @@ static void uninit_video(struct gl_video *p)
     for (int n = 0; n < p->plane_count; n++) {
         struct texplane *plane = &vimg->planes[n];
         ra_tex_free(p->ra, &plane->tex);
-        tex_upload_uninit(p->ra, &plane->pbo);
     }
     *vimg = (struct video_image){0};
 
@@ -3291,7 +3289,7 @@ static bool pass_upload_image(struct gl_video *p, struct mp_image *mpi, uint64_t
             MP_VERBOSE(p, "DR enabled: %s\n", p->using_dr_path ? "yes" : "no");
         }
 
-        if (!tex_upload(p->ra, &plane->pbo, p->opts.pbo, &params)) {
+        if (!p->ra->fns->tex_upload(p->ra, &params)) {
             timer_pool_stop(p->upload_timer);
             goto error;
         }
@@ -3300,7 +3298,9 @@ static bool pass_upload_image(struct gl_video *p, struct mp_image *mpi, uint64_t
             mapped->mpi = mp_image_new_ref(mpi);
     }
     timer_pool_stop(p->upload_timer);
-    const char *mode = p->using_dr_path ? "DR" : p->opts.pbo ? "PBO" : "naive";
+
+    bool using_pbo = p->ra->use_pbo || !(p->ra->caps & RA_CAP_DIRECT_UPLOAD);
+    const char *mode = p->using_dr_path ? "DR" : using_pbo ? "PBO" : "naive";
     pass_describe(p, "upload frame (%s)", mode);
     pass_record(p, timer_pool_measure(p->upload_timer));
 
@@ -3639,6 +3639,7 @@ static void reinit_from_options(struct gl_video *p)
     check_gl_features(p);
     uninit_rendering(p);
     gl_sc_set_cache_dir(p->sc, p->opts.shader_cache_dir);
+    p->ra->use_pbo = p->opts.pbo;
     gl_video_setup_hooks(p);
     reinit_osd(p);