fix(amf): disable LTR by default to match FFmpeg amfenc behavior (#631)

* fix(amf): disable LTR by default to match FFmpeg amfenc behavior

PR #630 enabled native AMF LTR with 4 slots by default to improve weak-network RFI recovery, but this introduced a regression: static screen regions retain encoder artifacts (color blocks) until motion forces a refresh.

Root cause: with LTR_MODE_RESET_UNUSED + ULTRA_LOW_LATENCY usage, the IDR-time LTR baseline (slot 0) is permanently referenced by P-frames for static macroblocks. The slot-0 quantization noise from IDR is never overwritten, so static regions inherit it forever; only motion forces a fresh intra block.

FFmpeg's libavcodec/amfenc.c does NOT use the native AMF LTR API at all (it only optionally uses Pre-Analysis LTR which is a different mechanism). That is why the legacy FFmpeg-AMF path never exhibited this artifact.

Fix: revert default amd_ltr_frames to 0 so out-of-the-box behavior matches FFmpeg. Users who actually need RFI on lossy networks can opt in via amd_ltr_frames>=1; the slot rotation logic from #630 still applies in that case.

Refs #630

* docs(amf): explain LTR opt-in trade-off and dead-code rationale

Add inline comments at the three LTR enable sites and the effective_ltr_slots calc explaining: (1) why we keep the implementation despite default-disabled, (2) the IDR storm risk if RFI is removed entirely, (3) the static-region color-block trade-off when users opt in.

Author: qiin2333

src/amf/amf_config.h

@@ -55,8 +55,11 @@ namespace amf {
     // Enforce HRD
     std::optional<int> enforce_hrd;
 
-    // Number of LTR frames for RFI
-    int max_ltr_frames = 4;
+    // Number of LTR frames for RFI (0 = disabled, matches FFmpeg amfenc behavior).
+    // When enabled, static screen regions may retain encoder artifacts from the
+    // baseline LTR frame until motion forces a refresh; only opt in when the
+    // network actually needs reference-frame invalidation recovery.
+    int max_ltr_frames = 0;
 
     // --- Pre-Analysis sub-system ---
     // PAQ mode (AMF_PA_PAQ_MODE_ENUM): 0=none, 1=CAQ

src/amf/amf_d3d11.cpp

@@ -152,7 +152,23 @@ namespace amf {
       encoder->SetProperty(AMF_VIDEO_ENCODER_INPUT_QUEUE_SIZE, (amf_int64) 1);
       encoder->SetProperty(AMF_VIDEO_ENCODER_QUERY_TIMEOUT, (amf_int64) 1);
 
-      // LTR for RFI
+      // LTR for RFI (Reference Frame Invalidation, weak-network recovery).
+      //
+      // Disabled by default (max_ltr_frames == 0) to match FFmpeg amfenc behavior:
+      // FFmpeg's libavcodec/amfenc.c never sets MAX_LTR_FRAMES / LTR_MODE, so static
+      // screen regions are not pinned to a baseline LTR frame and never accumulate
+      // color-block artifacts.
+      //
+      // Trade-off when the user opts in (amd_ltr_frames >= 1):
+      //   + On lossy links, client-side reference invalidation can recover by
+      //     sending a P-frame referencing a known-good LTR slot instead of a full
+      //     IDR. IDRs are 10-20x larger than P-frames and themselves more likely
+      //     to be lost on weak networks, which can cascade into an "IDR storm".
+      //   - Static regions may inherit the IDR-time quantization noise of the
+      //     baseline LTR slot until motion forces a fresh intra block.
+      //
+      // The slot rotation / IDR-baseline preservation logic below (PR #630) only
+      // takes effect when LTR is opted in.
       max_ltr_frames = config.max_ltr_frames;
       if (max_ltr_frames > 0) {
         encoder->SetProperty(AMF_VIDEO_ENCODER_MAX_LTR_FRAMES, (amf_int64) max_ltr_frames);
@@ -217,6 +233,8 @@ namespace amf {
         encoder->SetProperty(AMF_VIDEO_ENCODER_HEVC_PROFILE, (amf_int64) AMF_VIDEO_ENCODER_HEVC_PROFILE_MAIN);
       }
 
+      // LTR for RFI - see H.264 block above for detailed trade-off rationale.
+      // Disabled by default; opt-in via amd_ltr_frames config.
       max_ltr_frames = config.max_ltr_frames;
       if (max_ltr_frames > 0) {
         encoder->SetProperty(AMF_VIDEO_ENCODER_HEVC_MAX_LTR_FRAMES, (amf_int64) max_ltr_frames);
@@ -298,6 +316,8 @@ namespace amf {
         encoder->SetProperty(AMF_VIDEO_ENCODER_AV1_AQ_MODE, (amf_int64) *config.pa_paq_mode);
       }
 
+      // LTR for RFI - see H.264 block above for detailed trade-off rationale.
+      // Disabled by default; opt-in via amd_ltr_frames config.
       max_ltr_frames = config.max_ltr_frames;
       if (max_ltr_frames > 0) {
         encoder->SetProperty(AMF_VIDEO_ENCODER_AV1_MAX_LTR_FRAMES, (amf_int64) max_ltr_frames);
@@ -595,7 +615,11 @@ namespace amf {
 
 
 
-    // Clamp effective LTR slots to what the encoder actually reserves
+    // Clamp effective LTR slots to what the encoder actually reserves.
+    // When max_ltr_frames == 0 (default), the entire LTR/RFI subsystem becomes
+    // a no-op: the IDR baseline marking, slot rotation, and invalidate handling
+    // below all gate on `effective_ltr_slots > 0`. The fallback for client-side
+    // invalidate_ref_frames in this case is force_idr=true (see video.cpp).
     effective_ltr_slots = (max_ltr_frames > 0) ? std::min(max_ltr_frames, MAX_LTR_SLOTS) : 0;
 
     // Reset LTR state

src/config.h

@@ -73,7 +73,7 @@ namespace config {
       std::optional<int> amd_vbaq;
       int amd_coder;
       int amd_qvbr_quality = 23;  // QVBR quality level 1-51 (lower=better, default=23)
-      int amd_ltr_frames = 4;  // LTR frames for RFI (0=disabled, default=4)
+      int amd_ltr_frames = 0;  // LTR frames for RFI (0=disabled by default; matches FFmpeg amfenc behavior to avoid static-region color blocks)
       int amd_slices_per_frame = 0;  // Slices/tiles per frame (0=client decides, 1-4=minimum)
     } amd;