From fac42e34629c5b13f48941ef6b4a5c95c92ab216 Mon Sep 17 00:00:00 2001
From: Musa Mahmood <Musasmahmood@gmail.com>
Date: Thu, 8 Oct 2026 00:22:12 +0100
Subject: [PATCH] [0.0.2] OCR: run each CUDA engine once when it is made, as
 cuBLASLt 13.8 crashed after an engine destroyed unused

---
 check_targets.jai      |  2 +-
 first.jai              |  2 +-
 modules/OCR/models.jai | 24 ++++++++++++++++++++++++
 modules/OCR/module.jai |  1 +
 4 files changed, 27 insertions(+), 2 deletions(-)

diff --git a/check_targets.jai b/check_targets.jai
index 5f1229e..8e0f7b3 100644
--- a/check_targets.jai
+++ b/check_targets.jai
@@ -29,7 +29,7 @@ APP_NAME        :: "mcap";
 APP_FOLDER_NAME :: "mcap";
 GIT_BRANCH    :: "check";
 GIT_HASH      :: "-";
-VERSION       :: "0.0.1";
+VERSION       :: "0.0.2";
 BUILD_DATE    :: "-";
 DEBUG         :: true;
 COMPILE_TESTS :: true;
diff --git a/first.jai b/first.jai
index c864eac..c50da55 100644
--- a/first.jai
+++ b/first.jai
@@ -8,7 +8,7 @@ APP_NAME              :: "mcap";
 APP_FOLDER_NAME       :: "mcap";
 APP_BUNDLE_IDENTIFIER :: "com.example.mcap";
 
-VERSION :: "0.0.1";
+VERSION :: "0.0.2";
 // Where everything goes: bin/, or bin-headless/ for builds made with no_crash_dialog, so automated
 // builds never write over the ones you run by hand.
 build_directory := "bin";
diff --git a/modules/OCR/models.jai b/modules/OCR/models.jai
index feb2d52..9ab6db7 100644
--- a/modules/OCR/models.jai
+++ b/modules/OCR/models.jai
@@ -117,6 +117,30 @@ ocr_release_sessions :: (engine: *Ocr_Engine) {
   ocr_release_session(*engine.orientation);
 }
 
+// Runs each of a CUDA engine's sessions once, on a small input, before it is used or destroyed.
+// With cuBLASLt 13.8 (CUDA 13, onnxruntime 1.30.0, RTX 5080 laptop), the process's first CUDA
+// engine destroyed without ever having run left the next engine to crash in cublasLtCtxInit, with a
+// DEP access violation, on its first run; an engine that has run once leaves nothing broken. It also
+// moves the first run's setup into creation. On failure, the reason.
+ocr_warm_up :: (engine: *Ocr_Engine) -> ok: bool, error: string {
+  Warm_Up :: struct { session: *Ocr_Session; height, width: s64; }
+  runs := Warm_Up.[
+    .{*engine.det, 64, 64},
+    .{*engine.rec, engine.model.rec_height, engine.model.rec_height * 2},
+    .{*engine.orientation, engine.orientation_height, engine.orientation_width},
+  ];
+  for runs {
+    if !it.session.session  continue;   // no orientation classifier
+    shape := s64.[1, 3, it.height, it.width];
+    input := NewArray(3 * it.height * it.width, float, initialized = true,, context.default_allocator);
+    defer array_free(input,, context.default_allocator);
+    out, ok, error := ocr_run(it.session, engine.memory_info, input, shape);
+    if !ok  return false, tprint("The first run on the GPU failed: %", error);
+    ocr_release_output(*out);
+  }
+  return true, "";
+}
+
 // A session for the model file; with CUDA when gpu is set. On failure, the reason.
 ocr_create_session :: (env: *OrtEnv, path: string, gpu: bool, intra_op_threads: s32, ocr_options: Ocr_Options) -> Ocr_Session, ok: bool, error: string {
   result: Ocr_Session;
diff --git a/modules/OCR/module.jai b/modules/OCR/module.jai
index 94d0b9b..549c0cd 100644
--- a/modules/OCR/module.jai
+++ b/modules/OCR/module.jai
@@ -118,6 +118,7 @@ ocr_create :: (models_folder: string, model_name: string, prefer_gpu := true, op
     // A few CPU threads prepare images and decode results while the GPU runs the networks.
     engine.workers = max(options.gpu_workers, 1);
     sessions_ok, sessions_error := ocr_create_sessions(engine, models_folder, true, 2);
+    if sessions_ok  sessions_ok, sessions_error = ocr_warm_up(engine);
     if sessions_ok {
       engine.provider = .CUDA;
       engine.rec_batch = max(engine.model.rec_batch_gpu, 1);