vo_gpu: implement VO_DR_FLAG_HOST_CACHED

For OpenGL, this is based on simply comparing GL_VENDOR strings against
a list of allowed vendors.

Co-authored-by: Nicolas F. <ovdev@fratti.ch>
Co-authored-by: Niklas Haas <git@haasn.dev>

video/out/gpu/ra.h

@@ -80,6 +80,7 @@ enum {
     RA_CAP_FRAGCOORD      = 1 << 10, // supports reading from gl_FragCoord
     RA_CAP_PARALLEL_COMPUTE  = 1 << 11, // supports parallel compute shaders
     RA_CAP_NUM_GROUPS     = 1 << 12, // supports gl_NumWorkGroups
+    RA_CAP_SLOW_DR        = 1 << 13, // direct rendering is assumed to be slow
 };
 
 enum ra_ctype {

video/out/gpu/video.c

@@ -4290,6 +4290,13 @@ static void gl_video_dr_free_buffer(void *opaque, uint8_t *data)
 struct mp_image *gl_video_get_image(struct gl_video *p, int imgfmt, int w, int h,
                                     int stride_align, int flags)
 {
+    if (flags & VO_DR_FLAG_HOST_CACHED) {
+        if (p->ra->caps & RA_CAP_SLOW_DR) {
+            MP_VERBOSE(p, "DR path suspected slow/uncached, disabling..");
+            return NULL;
+        }
+    }
+
     if (!gl_video_check_format(p, imgfmt))
         return NULL;
 

video/out/opengl/common.c

@@ -25,6 +25,7 @@
 #include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
+#include <strings.h>
 #include <stdbool.h>
 #include <math.h>
 #include <assert.h>
@@ -47,6 +48,18 @@ static bool is_software_gl(GL *gl)
            strcmp(renderer, "Apple Software Renderer") == 0;
 }
 
+// This guesses whether our DR path is fast or slow
+static bool is_fast_dr(GL *gl)
+{
+    const char *vendor = gl->GetString(GL_VENDOR);
+    if (!vendor)
+        return false;
+
+    return strcasecmp(vendor, "AMD") == 0 ||
+           strcasecmp(vendor, "NVIDIA Corporation") == 0 ||
+           strcasecmp(vendor, "ATI Technologies Inc.") == 0;    // AMD on Windows
+}
+
 static void GLAPIENTRY dummy_glBindFramebuffer(GLenum target, GLuint framebuffer)
 {
     assert(framebuffer == 0);
@@ -650,6 +663,9 @@ void mpgl_load_functions2(GL *gl, void *(*get_fn)(void *ctx, const char *n),
         mp_verbose(log, "Detected suspected software renderer.\n");
     }
 
+    if (!is_fast_dr(gl))
+        gl->mpgl_caps |= MPGL_CAP_SLOW_DR;
+
     // GL_ARB_compute_shader & GL_ARB_shader_image_load_store
     if (gl->DispatchCompute && gl->BindImageTexture)
         gl->mpgl_caps |= MPGL_CAP_COMPUTE_SHADER;

video/out/opengl/common.h

@@ -59,6 +59,7 @@ enum {
     MPGL_CAP_COMPUTE_SHADER     = (1 << 23),    // GL_ARB_compute_shader & GL_ARB_shader_image_load_store
     MPGL_CAP_NESTED_ARRAY       = (1 << 24),    // GL_ARB_arrays_of_arrays
 
+    MPGL_CAP_SLOW_DR            = (1 << 29),    // direct rendering is assumed to be slow
     MPGL_CAP_SW                 = (1 << 30),    // indirect or sw renderer
 };
 

video/out/opengl/ra_gl.c

@@ -103,6 +103,7 @@ static int ra_init_gl(struct ra *ra, GL *gl)
         {RA_CAP_COMPUTE,            MPGL_CAP_COMPUTE_SHADER},
         {RA_CAP_NUM_GROUPS,         MPGL_CAP_COMPUTE_SHADER},
         {RA_CAP_NESTED_ARRAY,       MPGL_CAP_NESTED_ARRAY},
+        {RA_CAP_SLOW_DR,            MPGL_CAP_SLOW_DR},
     };
 
     for (int i = 0; i < MP_ARRAY_SIZE(caps_map); i++) {

video/out/placebo/ra_pl.c

@@ -52,6 +52,10 @@ struct ra *ra_create_pl(pl_gpu gpu, struct mp_log *log)
     if (gpu->limits.max_variables)
         ra->caps |= RA_CAP_GLOBAL_UNIFORM;
 #endif
+#if PL_API_VER >= 234
+    if (!gpu->limits.host_cached)
+        ra->caps |= RA_CAP_SLOW_DR;
+#endif
 
     if (gpu->limits.max_tex_1d_dim)
         ra->caps |= RA_CAP_TEX_1D;