tools/guest-diag: tag BGRA capture records as GZF2 fmt 33

sdl_cap.c wrote the same 32 in the GZF2 record format field for every
capture path, so gzf2_video.py's auto mode decoded ARGB8888/X-surface
frames (BGRA byte order on little-endian) as RGBA and swapped red and
blue. Make the field the documented pixel-order tag: 32 = RGBA8 from
the GL swap path (glReadPixels GL_RGBA), 33 = BGRA8 from
SDL_RenderReadPixels(ARGB8888) and X window surfaces; xgrab.c tags its
ZPixmap BGRA records 33 as well.

gzf2_video.py continues to accept both tags (and old all-32 captures
via --pix-fmt bgra), checks ffmpeg/ffprobe before indexing, and keeps
the final resampled endpoint so a 90-frame 60 fps capture encodes as
90 frames instead of 88.

3 files changed, 36 insertions, 15 deletions

diff --git a/tools/guest-diag/gzf2_video.py b/tools/guest-diag/gzf2_video.py
index aa88aa0..8635ca0 100644
--- a/tools/guest-diag/gzf2_video.py
+++ b/tools/guest-diag/gzf2_video.py
@@ -9,9 +9,11 @@
 #
 # Format: [0..8) = "GZCAPV01", zero pad to 4096; then records:
 #   4B "GZF2" | u32 w | u32 h | u32 stride | u32 fmt | u64 ts_ns | u64 seq |
-#   stride*h bytes, top-down.  fmt 32 = RGBA8 (GL swap path), 33 = BGRA8
-#   (SDL_RenderReadPixels ARGB8888 / X surface).  A trailing
-#   "GZF2" + u64 0xffffffffffffffff marks a clean stop.
+#   stride*h bytes, top-down.  fmt is a pixel-order tag: 32 = RGBA8 (GL swap
+#   path, glReadPixels GL_RGBA), 33 = BGRA8 (SDL_RenderReadPixels ARGB8888 /
+#   X surface).  A trailing "GZF2" + u64 0xffffffffffffffff marks a clean stop.
+#   Captures written before the 33 tag existed recorded BGRA paths as 32;
+#   decode those with --pix-fmt bgra.
 #
 # Usage:
 #   gzf2_video.py CAPTURE OUT.mp4 [--fps 60] [--skip 5] [--duration 20]
@@ -22,6 +24,7 @@ import bisect
 import hashlib
 import json
 import os
+import shutil
 import struct
 import subprocess
 import sys
@@ -89,7 +92,13 @@ def pick_frames(recs, fps, skip_s, duration_s):
     cand = [r for r in recs if r["ts"] >= t0] or recs
     ts = [r["ts"] for r in cand]
     span = (ts[-1] - ts[0]) / 1e9
-    n = int(duration_s * fps) if duration_s else max(1, int(span * fps))
+    if duration_s:
+        n = max(1, int(round(duration_s * fps)))
+    else:
+        # Nearest-frame sampling must include both endpoints: a capture that
+        # spans k frame periods yields k+1 output frames (int() alone drops
+        # the last frame and can lose one more to float formatting).
+        n = max(1, int(round(span * fps)) + 1)
     picks, held = [], 0
     for i in range(n):
         want = ts[0] + int(i * 1e9 / fps)
@@ -179,10 +188,16 @@ def main():
                     default="libx264")
     ap.add_argument("--crf", type=int, default=18)
     ap.add_argument("--pix-fmt", choices=["auto", "rgba", "bgra"],
-                    default="auto", help="auto uses record fmt (33=BGRA)")
+                    default="auto",
+                    help="auto uses the record tag (32=RGBA, 33=BGRA); "
+                         "old BGRA captures tagged 32 need bgra")
     ap.add_argument("--info-only", action="store_true")
     args = ap.parse_args()
 
+    for tool in ("ffmpeg", "ffprobe"):
+        if not shutil.which(tool):
+            die(f"{tool} not found in PATH")
+
     recs, clean, corrupt = index_capture(args.capture)
     uniq = len({r["md5"] for r in recs})
     span = (recs[-1]["ts"] - recs[0]["ts"]) / 1e9
diff --git a/tools/guest-diag/sdl_cap.c b/tools/guest-diag/sdl_cap.c
index 5b9c096..f875ed0 100644
--- a/tools/guest-diag/sdl_cap.c
+++ b/tools/guest-diag/sdl_cap.c
@@ -11,9 +11,12 @@
 // File layout (little endian):
 //   [0]   8B  magic "GZCAPV01"  (+ 4088 B reserved zero)
 //   [4096] frame records:
-//         4B "GZF2" | u32 w | u32 h | u32 stride | u32 bpp
+//         4B "GZF2" | u32 w | u32 h | u32 stride | u32 fmt
 //         u64 ts_ns (CLOCK_MONOTONIC) | u64 seq | stride*h bytes (top-down)
-//   A trailing record with seq = 0xFFFFFFFFFFFFFFFF marks clean stop.
+//   fmt is a pixel-order tag (not bits per pixel): 32 = RGBA8 from the GL
+//   swap path (glReadPixels GL_RGBA), 33 = BGRA8 from SDL_RenderReadPixels
+//   ARGB8888 or an X window surface.  A trailing record with
+//   seq = 0xFFFFFFFFFFFFFFFF marks clean stop.
 //
 // Env: GZCAP (path, default /dev/shm/gzframes.bin), GZCAP_MAXF (0=unlimited)
 #define _GNU_SOURCE
@@ -26,6 +29,8 @@
 #include <unistd.h>
 
 #define ARGB8888 0x16362004u
+#define GZF2_FMT_RGBA 32
+#define GZF2_FMT_BGRA 33
 
 typedef struct SDL_Renderer SDL_Renderer;
 typedef struct SDL_Window SDL_Window;
@@ -87,7 +92,7 @@ __attribute__((constructor)) static void cap_init(void)
         if (h)
             real_glReadPixels = dlsym(h, "glReadPixels");
     }
-    fprintf(stderr, "[gzcap] init pid=%d path=%s present=%p updwin=%p swap=%p read=%p gros=%p\n",
+    fprintf(stderr, "[gzcap] init pid=%d path=%s present=%p updwin=%p swap=%p read=%p gros=%p readpix=%p\n",
             (int)getpid(), p, (void *)real_RenderPresent, (void *)real_UpdateWindowSurface,
             (void *)real_GL_SwapWindow, (void *)real_RenderReadPixels,
             (void *)real_GetRendererOutputSize, (void *)real_glReadPixels);
@@ -102,7 +107,7 @@ __attribute__((constructor)) static void cap_init(void)
     fflush(cap_fp);
 }
 
-static void cap_record(const void *px, int w, int h, int stride, int bpp)
+static void cap_record(const void *px, int w, int h, int stride, int fmt)
 {
     struct timespec ts;
     uint64_t tns, seq;
@@ -126,7 +131,7 @@ static void cap_record(const void *px, int w, int h, int stride, int bpp)
     rh[0] = (uint32_t)w;
     rh[1] = (uint32_t)h;
     rh[2] = (uint32_t)stride;
-    rh[3] = (uint32_t)bpp;
+    rh[3] = (uint32_t)fmt;
     fwrite(m, 1, 4, cap_fp);
     fwrite(rh, 4, 4, cap_fp);
     fwrite(&tns, 8, 1, cap_fp);
@@ -161,7 +166,7 @@ int SDL_RenderPresent(SDL_Renderer *r)
             }
             if (buf && real_RenderReadPixels(r, NULL, ARGB8888, buf,
                                              w * 4) == 0)
-                cap_record(buf, w, h, w * 4, 32);
+                cap_record(buf, w, h, w * 4, GZF2_FMT_BGRA);
         }
     }
     if (real_RenderPresent)
@@ -180,7 +185,7 @@ int SDL_UpdateWindowSurface(SDL_Window *win)
 
         if (s && s->pixels && s->w > 0 && s->h > 0 && s->pitch > 0 &&
             s->w <= 4096 && s->h <= 4096)
-            cap_record(s->pixels, s->w, s->h, s->pitch, 32);
+            cap_record(s->pixels, s->w, s->h, s->pitch, GZF2_FMT_BGRA);
     }
     if (real_UpdateWindowSurface)
         return real_UpdateWindowSurface(win);
@@ -221,7 +226,7 @@ void SDL_GL_SwapWindow(SDL_Window *win)
                             memcpy(tmp + (size_t)y * w * 4,
                                    gbuf + (size_t)(h - 1 - y) * w * 4,
                                    (size_t)w * 4);
-                        cap_record(tmp, w, h, w * 4, 32);
+                        cap_record(tmp, w, h, w * 4, GZF2_FMT_RGBA);
                         free(tmp);
                     }
                 }
diff --git a/tools/guest-diag/xgrab.c b/tools/guest-diag/xgrab.c
index eb605d6..bca23c0 100644
--- a/tools/guest-diag/xgrab.c
+++ b/tools/guest-diag/xgrab.c
@@ -3,8 +3,9 @@
 // Pure client-side XGetImage on a software X server (Xvfb): no GL, no driver,
 // no present-path hooks. Same record format as sdl_cap.c:
 //   [0]   8B "GZCAPV01" + zero pad to 4096
-//   [4096..] records: 4B "GZF2" | u32 w | u32 h | u32 stride | u32 bpp |
+//   [4096..] records: 4B "GZF2" | u32 w | u32 h | u32 stride | u32 fmt |
 //                     u64 ts_ns | u64 seq | stride*h bytes (top-down BGRA)
+//   fmt is a pixel-order tag: 33 = BGRA8 (ZPixmap on little-endian X).
 // Trailing record seq=0xFFFFFFFFFFFFFFFF on clean stop.
 // Usage: xgrab <display> <out> <maxframes|0> <pace_us|0>
 #include <X11/Xlib.h>
@@ -83,7 +84,7 @@ int main(int argc, char **argv)
         rh[0] = (uint32_t)r_w;
         rh[1] = (uint32_t)r_h;
         rh[2] = (uint32_t)r_w * 4;
-        rh[3] = 32;
+        rh[3] = 33; /* GZF2 pixel-order tag: BGRA8 */
         fwrite(m, 1, 4, f);
         fwrite(rh, 4, 4, f);
         fwrite(&tns, 8, 1, f);