[Perf] (MG_Pipe): a vertex array keeps its buffer and index bindings on the server, a small uniform change crosses as a span, and unchanged context values skip their record (rd12 split arms +5-10%).

MobileGL/MG_Impl/Pipe/OwnedDrawInputs.h

@@ -59,6 +59,7 @@ namespace MobileGL::MG_Pipe {
                 binding.Res = info.IndexResource;
                 binding.IndexSize = info.IndexSize;
                 MGPipeRouteSetIndexBuffer(binding);
+                MGPipeVertexInputEmitterInstance().NoteIndexBufferRoutedElsewhere(m_context);
                 m_indexBindingChanged = true;
             }
 

MobileGL/MG_Impl/Pipe/PipeFill.cpp

@@ -1422,6 +1422,7 @@ namespace MobileGL::MG_Pipe {
         // refusal, and the refusal asserts (PipeApply.cpp's ResolveVertexElements), i.e. it
         // stops a verify build.
         MGPipeVertexInputEmitter& emitter = MGPipeVertexInputEmitterInstance();
+        emitter.ForgetVao(lifetimeId);
         const Bool published = emitter.RecordIsPublished(handle);
         if (published) {
             MGPHandleOnly only{};
@@ -3152,6 +3153,18 @@ namespace MobileGL::MG_Pipe {
         // server-role-only fixture) - nothing is emitted and the fields keep being pulled, byte
         // for byte as before (G1).
         Uint64 EmitContextValues(GLContext& ctx) {
+            // P14: nothing the record reads has moved since the record whose hash the suppressor
+            // still holds was built (same context, same input generation): it would be dropped.
+            static MGPContextValues s_lastValues{};
+            static Uint64 s_lastHash = 0;
+            static Uint64 s_lastContextId = 0;
+            static Uint64 s_lastGeneration = 0;
+            auto& suppressor = MGPipeSetHashSuppressorInstance();
+            const Uint64 contextId = ctx.GetTextureContextId();
+            const Uint64 generation = ctx.GetContextValuesGeneration();
+            if (s_lastHash != 0 && contextId == s_lastContextId && generation == s_lastGeneration &&
+                suppressor.LastEmitted(MGPipeSuppressorSlot::SetContextValues) == s_lastHash)
+                return 0;
             MGPContextValues values{};
             values.ActiveTextureUnit = static_cast<Uint32>(ctx.GetActiveTextureUnit());
             values.MaxTouchedTextureUnit = static_cast<Uint32>(ctx.GetMaxTouchedTextureUnit());
@@ -3174,9 +3187,8 @@ namespace MobileGL::MG_Pipe {
             // is byte-equal to the record whose hash the suppressor still holds, the suppressor
             // would drop it: skip the hash. (An InvalidateAll changes what the suppressor holds,
             // so a reset still re-emits.)
-            static MGPContextValues s_lastValues{};
-            static Uint64 s_lastHash = 0;
-            auto& suppressor = MGPipeSetHashSuppressorInstance();
+            s_lastContextId = contextId;
+            s_lastGeneration = generation;
             if (s_lastHash != 0 && std::memcmp(&values, &s_lastValues, sizeof(values)) == 0 &&
                 suppressor.LastEmitted(MGPipeSuppressorSlot::SetContextValues) == s_lastHash)
                 return 0;

MobileGL/MG_Impl/Pipe/ProgramEmit.h

@@ -218,6 +218,37 @@ namespace MobileGL::MG_Pipe {
             m_lastConstants = MGPGlobalConstants{};
             m_lastConstants.ShaderCso = cso;
             m_lastConstants.Version = version;
+            // P14: the server holds this program's previous image while the latch names it (every
+            // path that could drop or replace the applier's block clears the latch), so only the
+            // span that moved goes out - when it is small enough to be worth it.
+            const Uint8* image = static_cast<const Uint8*>(program->GetUBOData());
+            if (cso == m_constantsCso && m_constantsShadow.size() == size && size <= 0xffffu) {
+                Uint32 first = 0;
+                while (first < size && image[first] == m_constantsShadow[first]) ++first;
+                Uint32 last = size;
+                while (last > first && image[last - 1] == m_constantsShadow[last - 1]) --last;
+                // Nothing moved but the version: one byte carries the new version.
+                if (first == size) {
+                    first = 0;
+                    last = 1;
+                }
+                const Uint32 length = last - first;
+                if (length * 2 <= size) {
+                    m_lastConstants.Range = (first << 16) | length;
+                    m_lastConstants.Blob.Seg = kMGHostSpanSegNone;
+                    m_lastConstants.Blob.Offset = reinterpret_cast<Uint64>(image + first);
+                    m_lastConstants.Blob.Size = 0;
+                    MGPipeRouteSetGlobalConstants(m_lastConstants, image + first, static_cast<Uint64>(length));
+                    std::memcpy(m_constantsShadow.data() + first, image + first, length);
+                    m_constantsVersion = version;
+                    ++m_constantSets;
+                    ++m_constantSpans;
+                    if (MG_Util::PipeStats::Enabled()) {
+                        MG_Util::PipeStats::AddBytes(MG_Util::PipeStats::ByteClass::CsoBlobBytes, length);
+                    }
+                    return bytes + sizeof(MGPGlobalConstants) + length;
+                }
+            }
             // THE ONE BLOB RULE: Size 0 means "this record does not declare its blob" - which
             // is what a monolith emission is - and the bytes ride beside it as a companion
             // pointer. Offset carries the staging address for diagnostics only; nothing reads
@@ -233,6 +264,7 @@ namespace MobileGL::MG_Pipe {
                                           static_cast<Uint64>(size));
             m_constantsCso = cso;
             m_constantsVersion = version;
+            m_constantsShadow.assign(image, image + size);
             ++m_constantSets;
             if (MG_Util::PipeStats::Enabled()) {
                 MG_Util::PipeStats::AddBytes(MG_Util::PipeStats::ByteClass::CsoBlobBytes, size);
@@ -705,6 +737,10 @@ namespace MobileGL::MG_Pipe {
         MGPipeHandle m_dispatchCso = kMGPipeNullHandle;
         MGPipeHandle m_constantsCso = kMGPipeNullHandle;
         Uint32 m_constantsVersion = kMGPipeGlobalConstantsNeverUploaded;
+        // The image last sent for m_constantsCso (whole or patched by spans): what the server
+        // holds while the latch names it.
+        Vector<Uint8> m_constantsShadow;
+        Uint64 m_constantSpans = 0;
 
         Uint64 m_creates = 0;
         Uint64 m_binds = 0;

MobileGL/MG_Impl/Pipe/SetHashSuppressor.h

@@ -103,6 +103,13 @@ namespace MobileGL::MG_Pipe {
             return true;
         }
 
+        // The server now holds the set `contentHash` names without this slot having emitted it
+        // (P14: bind_vertex_elements re-applies the array's kept set). Same 0 -> 1 remap as
+        // ShouldEmit, so the next ShouldEmit of the same hash answers false.
+        void Latch(MGPipeSuppressorSlot slot, Uint64 contentHash) {
+            m_lastEmitted[static_cast<SizeT>(slot)] = contentHash == 0 ? 1 : contentHash;
+        }
+
         // A context change or a server reset: what the server has is no longer what this
         // slot last emitted, so the next resolved set must go out whatever it hashes to.
         void Invalidate(MGPipeSuppressorSlot slot) { m_lastEmitted[static_cast<SizeT>(slot)] = 0; }

MobileGL/MG_Impl/Pipe/VertexInputEmit.h

@@ -181,13 +181,26 @@ namespace MobileGL::MG_Pipe {
             Uint64 bytes = 0;
             const Bool configMoved = !latch.Published || latch.ConfigVersion != configVersion ||
                                      latch.Gen != handle.Gen;
-            if (configMoved) bytes += EmitCreate(*vao, handle, latch, configVersion);
+            if (configMoved) {
+                bytes += EmitCreate(*vao, handle, latch, configVersion);
+                // The applier drops the record's kept bindings at the same create.
+                memo.StoredBuffersValid = false;
+                memo.StoredIndexValid = false;
+            }
             if (lifetimeId != m_boundLifetimeId || m_boundHandle != handle) {
                 MGPipeRouteBindVertexElements(HandleOnly(handle));
                 ++m_binds;
                 bytes += sizeof(MGPHandleOnly);
                 m_boundHandle = handle;
                 m_boundLifetimeId = lifetimeId;
+                // P14: the bind re-applies whatever the record keeps (PipeApply.h,
+                // MGPipeVertexElementsRecord). The mirror below says what that is; where it does
+                // not know, the next set goes out whatever it hashes to.
+                auto& suppressor = MGPipeSetHashSuppressorInstance();
+                if (memo.StoredBuffersValid) suppressor.Latch(MGPipeSuppressorSlot::SetVertexBuffers, memo.StoredBuffersHash);
+                else suppressor.Invalidate(MGPipeSuppressorSlot::SetVertexBuffers);
+                m_serverIndexKnown = memo.StoredIndexValid;
+                m_serverIndexRes = memo.StoredIndexRes;
             }
             return bytes;
         }
@@ -287,6 +300,14 @@ namespace MobileGL::MG_Pipe {
             m_lastBuffers.ContentHash = hash;
             MGPipeRouteSetVertexBuffers(m_lastBuffers, m_entries.data());
             ++m_bufferSets;
+            // The bound array now keeps this set (the applier's rule, mirrored): a draw-time
+            // client-array set is the array's too, but it names private buffers, so the mirror
+            // forgets rather than remembers it.
+            if (vao && vao->GetLifetimeId() == m_boundLifetimeId) {
+                VaoMemo& bound = MemoFor(*vao);
+                bound.StoredBuffersValid = ownedClientBuffers == nullptr && count <= kMGPipeVaoStoredVertexBuffers;
+                bound.StoredBuffersHash = hash;
+            }
             return sizeof(MGPVertexBuffers) + static_cast<Uint64>(count) * sizeof(MGPVertexBuffer);
         }
 
@@ -316,11 +337,29 @@ namespace MobileGL::MG_Pipe {
                     MGPipeResourceTrackerInstance().NoteBoundAs(m_lastIndex.Res, BufferTarget::Index);
                 }
             }
+            // P14: the server already holds this binding - re-applied by the bind from the
+            // array's kept one, or sent last time - so there is nothing to send.
+            if (m_serverIndexKnown && m_serverIndexRes == m_lastIndex.Res) return 0;
             MGPipeRouteSetIndexBuffer(m_lastIndex);
             ++m_indexSets;
+            m_serverIndexKnown = true;
+            m_serverIndexRes = m_lastIndex.Res;
+            if (vao && vao->GetLifetimeId() == m_boundLifetimeId) {
+                VaoMemo& bound = MemoFor(*vao);
+                bound.StoredIndexValid = true;
+                bound.StoredIndexRes = m_lastIndex.Res;
+            }
             return sizeof(MGPIndexBuffer);
         }
 
+        // A set_index_buffer this emitter did not send (OwnedDrawInputs' private element
+        // buffer): the server's binding and the bound array's kept one are no longer known.
+        void NoteIndexBufferRoutedElsewhere(GLContext& ctx) {
+            m_serverIndexKnown = false;
+            const auto& vao = ctx.GetBoundVertexArray();
+            if (vao && vao->GetLifetimeId() == m_boundLifetimeId) MemoFor(*vao).StoredIndexValid = false;
+        }
+
         // ---- what a unit case reads. None of it costs a copy: the emitter builds INTO
         // these and hands the applier the same pointers. ----
         const Array<MGPVertexAttribWire, kMGPipeMaxVertexAttribs>& LastAttributes() const { return m_attributes; }
@@ -394,16 +433,19 @@ namespace MobileGL::MG_Pipe {
             }
             m_boundHandle = kMGPipeNullHandle;
             m_boundLifetimeId = 0;
+            m_serverIndexKnown = false;
         }
 
         void ResetCounters() { m_creates = m_binds = m_bufferSets = m_indexSets = 0; }
 
     private:
         // P14: what EmitVertexElements / EmitVertexBuffers / EmitIndexBuffer resolve from a VAO,
-        // direct-mapped on the VAO's lifetime id (monotonic, never reused) and dropped whole at an
-        // allocator reset. A colliding VAO simply takes the entry over. It changes no record: the
-        // same handles and entries are emitted, only found without the per-draw walk and lookups.
-        static constexpr SizeT kVaoMemoEntries = 1024;
+        // one entry per live VAO keyed on its lifetime id (monotonic, never reused), erased at the
+        // VAO's death and re-initialised after an allocator reset. A direct-mapped table was tried
+        // first and lost most entries to collisions: a Minecraft frame draws ~1,500 chunk arrays
+        // whose ids spread well past any fixed table. The handles and entries it keeps change no
+        // record; the kept-binding mirror below does decide what is sent, and is only ever
+        // believed when the server provably holds the same.
         static constexpr SizeT kVaoMemoBuffers = 8;
         struct VaoMemo {
             Uint64 LifetimeId = 0;
@@ -418,11 +460,29 @@ namespace MobileGL::MG_Pipe {
             Uint32 HashBaseInstance = 0;
             Uint64 Hash = 0;
             Array<MGPVertexBuffer, kVaoMemoBuffers> Buffers{};
+            // The mirror of what the applier's record for this array keeps (PipeApply.h's
+            // MGPipeVertexElementsRecord): valid only while the server provably keeps exactly
+            // this. A memo entry taken over by another array starts invalid, which only costs a
+            // re-send.
+            Bool StoredBuffersValid = false;
+            Uint64 StoredBuffersHash = 0;
+            Bool StoredIndexValid = false;
+            MGPipeHandle StoredIndexRes = kMGPipeNullHandle;
         };
+        // What the server's index binding is, when known (P14).
+        Bool m_serverIndexKnown = false;
+        MGPipeHandle m_serverIndexRes = kMGPipeNullHandle;
+        // The returned reference stays valid until the next MemoFor of a DIFFERENT array (an
+        // insert may grow the table) or ForgetVao.
         VaoMemo& MemoFor(const VertexArrayObject& vao) {
             const Uint64 lifetimeId = vao.GetLifetimeId();
             const Uint64 epoch = MGPipeSlots().ResetEpoch();
-            VaoMemo& memo = m_vaoMemo[lifetimeId % kVaoMemoEntries];
+            if (m_vaoMemo.empty() || (m_vaoMemoUsed + 1) * 2 > m_vaoMemo.size()) GrowVaoMemo();
+            const SizeT mask = m_vaoMemo.size() - 1;
+            SizeT i = VaoMemoHome(lifetimeId, mask);
+            while (m_vaoMemo[i].LifetimeId != 0 && m_vaoMemo[i].LifetimeId != lifetimeId) i = (i + 1) & mask;
+            VaoMemo& memo = m_vaoMemo[i];
+            if (memo.LifetimeId == 0) ++m_vaoMemoUsed;
             if (memo.LifetimeId != lifetimeId || memo.SlotEpoch != epoch) {
                 memo = VaoMemo{};
                 memo.LifetimeId = lifetimeId;
@@ -430,7 +490,52 @@ namespace MobileGL::MG_Pipe {
             }
             return memo;
         }
-        Array<VaoMemo, kVaoMemoEntries> m_vaoMemo{};
+        static SizeT VaoMemoHome(Uint64 lifetimeId, SizeT mask) {
+            return static_cast<SizeT>((lifetimeId * 0x9E3779B97F4A7C15ull) >> 32) & mask;
+        }
+        void GrowVaoMemo() {
+            Vector<VaoMemo> old;
+            old.swap(m_vaoMemo);
+            m_vaoMemo.resize(old.empty() ? 1024 : old.size() * 2);
+            m_vaoMemoUsed = 0;
+            const SizeT mask = m_vaoMemo.size() - 1;
+            for (const VaoMemo& entry : old) {
+                if (entry.LifetimeId == 0) continue;
+                SizeT i = VaoMemoHome(entry.LifetimeId, mask);
+                while (m_vaoMemo[i].LifetimeId != 0) i = (i + 1) & mask;
+                m_vaoMemo[i] = entry;
+                ++m_vaoMemoUsed;
+            }
+        }
+        Vector<VaoMemo> m_vaoMemo;
+        SizeT m_vaoMemoUsed = 0;
+
+    public:
+        // The VAO died (MGPipeEmitVertexElementsDestroyAndFree): its entry goes, by backward-shift
+        // deletion so every other entry stays reachable from its home.
+        void ForgetVao(Uint64 lifetimeId) {
+            if (m_vaoMemo.empty() || lifetimeId == 0) return;
+            const SizeT mask = m_vaoMemo.size() - 1;
+            SizeT i = VaoMemoHome(lifetimeId, mask);
+            while (m_vaoMemo[i].LifetimeId != lifetimeId) {
+                if (m_vaoMemo[i].LifetimeId == 0) return;
+                i = (i + 1) & mask;
+            }
+            SizeT hole = i;
+            for (SizeT j = (hole + 1) & mask; m_vaoMemo[j].LifetimeId != 0; j = (j + 1) & mask) {
+                const SizeT home = VaoMemoHome(m_vaoMemo[j].LifetimeId, mask);
+                // Move j into the hole when its home is not strictly inside (hole, j].
+                const Bool between = hole <= j ? (home > hole && home <= j) : (home > hole || home <= j);
+                if (!between) {
+                    m_vaoMemo[hole] = m_vaoMemo[j];
+                    hole = j;
+                }
+            }
+            m_vaoMemo[hole] = VaoMemo{};
+            --m_vaoMemoUsed;
+        }
+
+    private:
 
         struct Latch {
             // The PER-CONTEXT half: "has this emitter told THIS server about this handle's

MobileGL/MG_Pipe/MGPipeTypes.h

@@ -1140,10 +1140,16 @@ namespace MobileGL::MG_Pipe {
     MGP_ASSERT_POD(MGPProgramBindings, 32);
 
     // Covers the DEFAULT UNIFORM BLOCK only (D6).
+    //
+    // Range (P14, was Pad0 and the size did not move): 0 means the bytes are the whole block.
+    // Otherwise it is (offset << 16) | length, and the bytes are only that span of the block,
+    // written over the image the program's record already holds - which the client sends only
+    // while it knows the server holds its previous image for the same ShaderCso (the latch that
+    // suppresses an unchanged block). A Minecraft chunk draw moves one vec3 of a ~230-byte block.
     struct MGPGlobalConstants {
         MGPipeHandle ShaderCso;
         Uint32 Version;
-        Uint32 Pad0;
+        Uint32 Range;
         MGPBlobRef Blob;
     };
     MGP_ASSERT_POD(MGPGlobalConstants, 40);

MobileGL/MG_Pipe/PipeApply.cpp

@@ -2951,6 +2951,10 @@ namespace MobileGL::MG_Pipe {
         record.AttributeCount = desc.AttributeCount;
         record.BindingPointCount = desc.BindingPointCount;
         record.Live = true;
+        // A new configuration starts without kept bindings (see the record): the client drops
+        // its mirror of them at the same create.
+        record.StoredBuffersValid = false;
+        record.StoredIndexValid = false;
         // Serial 0 means "never created", so the first create of an identity lands on 1 and a
         // backend twin that has synced nothing can never accidentally match a live record.
         ++record.ContentSerial;
@@ -2973,6 +2977,14 @@ namespace MobileGL::MG_Pipe {
         const MGPipeVertexElementsRecord* record = ResolveVertexElements("bind_vertex_elements", handle.Handle);
         if (record == nullptr) return;
         MGPipeApplier().BoundVertexElements = handle.Handle;
+        // The array's kept bindings come back with it, through the same two applies their
+        // records take, so every serial and field moves exactly as it would for the records.
+        // (The apply writes the same entries back over themselves: harmless, and no copy.)
+        if (record->StoredBuffersValid) {
+            const MGPVertexBuffers header = record->StoredBuffersHeader;
+            MGPipeApplySetVertexBuffers(header, record->StoredBuffers.data());
+        }
+        if (record->StoredIndexValid) MGPipeApplySetIndexBuffer(record->StoredIndex);
     }
 
     void MGPipeApplyDeleteVertexElements(const MGPHandleOnly& handle) {
@@ -3043,6 +3055,15 @@ namespace MobileGL::MG_Pipe {
         // would keep the previous shift.
         MGPipeApplier().VertexFetchBaseInstance = hdr.BaseInstance;
         ++MGPipeApplier().VertexBuffersSerial;
+        // Kept by the bound array (MGPipeVertexElementsRecord's P14 note). A re-apply from the
+        // bind writes back what it read, which is harmless.
+        if (MGPipeVertexElementsRecord* bound = FindVertexElements(MGPipeApplier().BoundVertexElements)) {
+            bound->StoredBuffersValid = hdr.Start == 0 && hdr.Count <= kMGPipeVaoStoredVertexBuffers;
+            if (bound->StoredBuffersValid) {
+                bound->StoredBuffersHeader = hdr;
+                for (Uint32 i = 0; i < hdr.Count; ++i) bound->StoredBuffers[i] = tail[i];
+            }
+        }
     }
 
     void MGPipeApplySetIndexBuffer(const MGPIndexBuffer& record) {
@@ -3058,6 +3079,11 @@ namespace MobileGL::MG_Pipe {
         //     configuration's, which is exactly what the backend's two separate compares need.
         MGPipeApplier().IndexBuffer = record;
         ++MGPipeApplier().IndexBufferSerial;
+        // Kept by the bound array (MGPipeVertexElementsRecord's P14 note).
+        if (MGPipeVertexElementsRecord* bound = FindVertexElements(MGPipeApplier().BoundVertexElements)) {
+            bound->StoredIndex = record;
+            bound->StoredIndexValid = true;
+        }
     }
 
     void MGPipeDeriveRenderStateFields(PipeInputs& inputs) {
@@ -3949,10 +3975,21 @@ namespace MobileGL::MG_Pipe {
         // bounded there - so this call can only ever allocate what the create already declared,
         // and the blob rule has a real field to cross-check against for once.
         const Uint32 size = stored->Desc.GlobalUboSize;
+        // A span (MGPGlobalConstants::Range) lands on the block the record already holds.
+        const Uint32 spanOffset = record.Range >> 16;
+        const Uint32 spanLength = record.Range & 0xffffu;
         const char* fault = nullptr;
         if (record.Version == ~Uint32{0}) {
             fault = "the version is the backends' never-uploaded sentinel, which no record may carry";
-        } else if (record.Blob.Size != 0 && record.Blob.Size != size) {
+        } else if (record.Range != 0 && (spanLength == 0 || spanOffset + spanLength > size)) {
+            fault = "the span runs outside the program's default uniform block";
+        } else if (record.Range != 0 && stored->GlobalConstants.size() != size) {
+            fault = "a span arrived for a block the record does not hold";
+        } else if (record.Range != 0 && record.Blob.Size != 0 && record.Blob.Size != spanLength) {
+            fault = "the declared blob length is not the span's length";
+        } else if (record.Range != 0 && bytes == nullptr) {
+            fault = "a span carries no bytes";
+        } else if (record.Range == 0 && record.Blob.Size != 0 && record.Blob.Size != size) {
             fault = "the declared blob length is not the program's own default uniform block size";
         } else if (size > kMGPipeMaxGlobalConstantsBytes) {
             fault = "the program's default uniform block is larger than any program may declare";
@@ -3968,7 +4005,9 @@ namespace MobileGL::MG_Pipe {
             return;
         }
 
-        if (size == 0) {
+        if (record.Range != 0) {
+            std::memcpy(stored->GlobalConstants.data() + spanOffset, bytes, spanLength);
+        } else if (size == 0) {
             stored->GlobalConstants.clear();
         } else {
             const Uint8* image = static_cast<const Uint8*>(bytes);

MobileGL/MG_Pipe/PipeApply.h

@@ -500,6 +500,11 @@ namespace MobileGL::MG_Pipe {
     // The 32 is GL's MAX_VERTEX_ATTRIBS as MobileGL advertises it (kMGPipeMaxVertexAttribs,
     // MGPipeTypes.h), which is also the bound the record's two declared counts are checked
     // against before the blob is unpacked.
+    // The widest vertex-buffer set a vertex-elements record keeps as its own (see below). A
+    // wider set is not kept, and the client, which mirrors this rule, then re-sends it after
+    // every bind of that array.
+    inline constexpr Uint32 kMGPipeVaoStoredVertexBuffers = 16;
+
     struct MGPipeVertexElementsRecord {
         Uint32 Gen = 0;
         Bool Live = false;
@@ -511,6 +516,25 @@ namespace MobileGL::MG_Pipe {
         // including a RE-create on the same handle, which is how a configuration change
         // travels (the handle is minted per frontend VAO and Gen moves only on slot reuse).
         Uint64 ContentSerial = 0;
+        // P14: THE ARRAY'S OWN BUFFER AND ELEMENT BINDINGS, as GL keeps them on the vertex array
+        // object. The last set_vertex_buffers (window starting at 0, at most
+        // kMGPipeVaoStoredVertexBuffers entries) and the last set_index_buffer applied while this
+        // record was bound are kept here, and bind_vertex_elements re-applies them, exactly as if
+        // the two records had arrived again - so a client switching between arrays whose
+        // bindings did not move sends the bind alone. create_vertex_elements on the handle drops
+        // both (the client drops its mirror at the same create), so a kept set never outlives the
+        // configuration it was sent for. The client's mirror (VertexInputEmit.h) only ever
+        // believes a kept set the server also keeps; where it is unsure it re-sends.
+        Bool StoredBuffersValid = false;
+        Bool StoredIndexValid = false;
+        MGPVertexBuffers StoredBuffersHeader{};
+        Array<MGPVertexBuffer, kMGPipeVaoStoredVertexBuffers> StoredBuffers{};
+        MGPIndexBuffer StoredIndex{};
+        // P14: a backend's own derivation from this record (Magma's resolved vertex input
+        // layout), opaque here. It lives and dies with the record - a re-created or recycled
+        // record starts without it - and the backend keys it on ContentSerial plus the inputs it
+        // read, so it is never believed for a configuration it was not built from.
+        mutable SharedPtr<void> BackendMemo;
     };
 
     // set_framebuffer_state's record, HELD PER FRAMEBUFFER OBJECT and indexed by the handle's

MobileGL/MG_Pipe/PipeFields.def

@@ -159,7 +159,7 @@
     F(Name) F(Binding)
 
 #define MGP_FIELDS_MGPGlobalConstants(F) \
-    F(ShaderCso) F(Version) F(Blob)
+    F(ShaderCso) F(Version) F(Range) F(Blob)
 
 #define MGP_FIELDS_MGPAttribValue(F) \
     F(Location) F(ValueClass) F(FloatView) F(IntView) F(UintView)

MobileGL/MG_Pipe/PipeWireLayout.h

@@ -35,7 +35,9 @@ namespace MobileGL::MG_Pipe {
         kMGPipeWireLayoutMembers, sizeof(kMGPipeWireLayoutMembers) / sizeof(WireLayoutMember));
     // Bump when a fixed-layout pipe carrier keeps its bytes but changes their meaning. The
     // mutable-level extent carrier reuses BufOffset/BufSize only on named image respecifies.
-    inline constexpr Uint32 kMGPipeResourceRespecifyExtentCarrierRevision = 1;
+    // 2 (P14): bind_vertex_elements re-applies the array's kept vertex-buffer and index
+    // bindings (PipeApply.h, MGPipeVertexElementsRecord), which a revision-1 peer would not.
+    inline constexpr Uint32 kMGPipeResourceRespecifyExtentCarrierRevision = 2;
 
     constexpr Uint64 WireCatalogueDigest() {
         Uint64 hash = 1469598103934665603ull;

MobileGL/MG_Pipe/generated/PipeWireLayout.inc

@@ -221,6 +221,7 @@ inline constexpr WireLayoutMember kMGPipeWireLayoutMembers[] = {
     {"MGPGlobalConstants", 0, sizeof(MGPGlobalConstants)},
     {"MGPGlobalConstants.ShaderCso", offsetof(MGPGlobalConstants, ShaderCso), sizeof(((MGPGlobalConstants*)nullptr)->ShaderCso)},
     {"MGPGlobalConstants.Version", offsetof(MGPGlobalConstants, Version), sizeof(((MGPGlobalConstants*)nullptr)->Version)},
+    {"MGPGlobalConstants.Range", offsetof(MGPGlobalConstants, Range), sizeof(((MGPGlobalConstants*)nullptr)->Range)},
     {"MGPGlobalConstants.Blob", offsetof(MGPGlobalConstants, Blob), sizeof(((MGPGlobalConstants*)nullptr)->Blob)},
     {"MGPAttribValue", 0, sizeof(MGPAttribValue)},
     {"MGPAttribValue.Location", offsetof(MGPAttribValue, Location), sizeof(((MGPAttribValue*)nullptr)->Location)},

MobileGL/MG_Remote/Transport/StreamLink.cpp

@@ -673,10 +673,18 @@ namespace MobileGL::MG_Remote::Transport {
         auto& x = *m_impl;
         if (!x.Client() || span.Offset > x.memory->StageBytes() || span.Size > x.memory->StageBytes() - span.Offset)
             Corrupt("staged span outside window");
-        if (!x.stage.empty() && x.stage.back().Offset + x.stage.back().Size == span.Offset)
-            x.stage.back().Size += span.Size;
-        else
-            x.stage.push_back(span);
+        // Adjacent once the producer's 8-byte allocation granule is counted (StageAllocate rounds
+        // every blob up): one span then carries both, padding included, instead of one Stage
+        // message per small blob - a Minecraft chunk draw stages a 12-byte uniform span.
+        if (!x.stage.empty()) {
+            auto& back = x.stage.back();
+            const std::uint64_t end = back.Offset + back.Size;
+            if (span.Offset >= end && span.Offset - end < 8) {
+                back.Size = span.Offset + span.Size - back.Offset;
+                return;
+            }
+        }
+        x.stage.push_back(span);
     }
     MobileGLResult StreamLink::Flush() {
         auto& x = *m_impl;

MobileGL/MG_State/GLState/BufferState/BufferState.cpp

@@ -16,6 +16,7 @@ namespace MobileGL::MG_State::GLState {
             const auto target = static_cast<SizeT>(GlobalBufferTargets[i]);
             if (target < m_bindingSlotIndex.size()) m_bindingSlotIndex[target] = static_cast<Int16>(i);
         }
+        ++m_contextValuesGeneration;
         for (SizeT i = 0; i < m_touchedBindPointCount.size(); ++i) {
             m_touchedBindPointCount[i] = 0;
         }

MobileGL/MG_State/GLState/BufferState/BufferState.h

@@ -51,9 +51,14 @@ namespace MobileGL::MG_State::GLState {
             auto it = std::find(BufferBindPointTargets.begin(), BufferBindPointTargets.end(), target);
             if (it == BufferBindPointTargets.end()) return;
             auto slot = std::distance(BufferBindPointTargets.begin(), it);
-            if (static_cast<SizeT>(index) + 1 > m_touchedBindPointCount[slot])
+            if (static_cast<SizeT>(index) + 1 > m_touchedBindPointCount[slot]) {
                 m_touchedBindPointCount[slot] = static_cast<SizeT>(index) + 1;
+                ++m_contextValuesGeneration;
+            }
         }
+        // Moves whenever a touched binding-point count does: the buffer half of
+        // GLContext::GetContextValuesGeneration.
+        Uint64 GetContextValuesGeneration() const { return m_contextValuesGeneration; }
         SizeT GetTouchedBindPointCount(const BufferTarget target) const {
             auto it = std::find(BufferBindPointTargets.begin(), BufferBindPointTargets.end(), target);
             if (it == BufferBindPointTargets.end()) return 0;
@@ -113,5 +118,6 @@ namespace MobileGL::MG_State::GLState {
         Array<Array<BindingSlotRange1D<BufferObject>, BufferBindingPointCount>, BufferBindPointTargets.size()>
             m_bufferBindPointTargets;
         Array<SizeT, BufferBindPointTargets.size()> m_touchedBindPointCount;
+        Uint64 m_contextValuesGeneration = 0;
     };
 } // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/Core.cpp

@@ -1436,6 +1436,7 @@ namespace MobileGL::MG_State {
                     point.ClearRange();
                 }
             }
+            ++m_transformFeedbackValuesGeneration;
             m_transformFeedbackActive = object.active;
             m_transformFeedbackPaused = object.paused;
             m_transformFeedbackPrimitiveMode = object.primitiveMode;

MobileGL/MG_State/GLState/Core.h

@@ -437,6 +437,7 @@ namespace MobileGL {
                 // feedback object currently bound to GL_TRANSFORM_FEEDBACK; see the object
                 // block further down for how a bind swaps them.
                 void BeginTransformFeedback(GLenum primitiveMode, const SharedPtr<ProgramObject>& program) {
+                    ++m_transformFeedbackValuesGeneration;
                     m_transformFeedbackActive = true;
                     m_transformFeedbackPaused = false;
                     m_transformFeedbackPrimitiveMode = primitiveMode;
@@ -446,6 +447,7 @@ namespace MobileGL {
                     m_transformFeedbackInputPrimitives = 0;
                 }
                 void EndTransformFeedback() {
+                    ++m_transformFeedbackValuesGeneration;
                     m_transformFeedbackActive = false;
                     m_transformFeedbackPaused = false;
                     m_transformFeedbackProgram.reset();
@@ -456,7 +458,10 @@ namespace MobileGL {
                 }
                 Bool IsTransformFeedbackActive() const { return m_transformFeedbackActive; }
                 Bool IsTransformFeedbackPaused() const { return m_transformFeedbackPaused; }
-                void SetTransformFeedbackPaused(Bool paused) { m_transformFeedbackPaused = paused; }
+                void SetTransformFeedbackPaused(Bool paused) {
+                    ++m_transformFeedbackValuesGeneration;
+                    m_transformFeedbackPaused = paused;
+                }
                 GLenum GetTransformFeedbackPrimitiveMode() const { return m_transformFeedbackPrimitiveMode; }
                 const SharedPtr<ProgramObject>& GetTransformFeedbackProgram() const {
                     return m_transformFeedbackProgram;
@@ -484,8 +489,17 @@ namespace MobileGL {
                 // Vertices already captured since BeginTransformFeedback (drives the
                 // buffer-capacity clamp on the primitives-written accounting).
                 void AddTransformFeedbackCapturedVertices(Uint64 vertices) {
+                    ++m_transformFeedbackValuesGeneration;
                     m_transformFeedbackCapturedVertices += vertices;
                 }
+                // P14: moves whenever any input of set_context_values does (the active texture
+                // unit, the two touched high-water marks, the transform feedback state of the bound
+                // object). Each half only ever counts up, so the sum moves exactly when one does;
+                // the emitter skips rebuilding the record while it holds.
+                Uint64 GetContextValuesGeneration() const {
+                    return m_textureState.GetContextValuesGeneration() + m_bufferState.GetContextValuesGeneration() +
+                           m_transformFeedbackValuesGeneration;
+                }
                 Uint64 GetTransformFeedbackCapturedVertices() const { return m_transformFeedbackCapturedVertices; }
                 // Raw assembled input primitives fed to the capture stage since Begin
                 // (pre-clamp; drives the GS strip capture-order fixup at EndTF).
@@ -655,6 +669,7 @@ namespace MobileGL {
                 Array<CurrentVertexAttributeValue, VertexArrayObject::MAX_VERTEX_ATTRIBS> m_currentVertexAttributes{};
                 Bool m_transformFeedbackActive = false;
                 Bool m_transformFeedbackPaused = false;
+                Uint64 m_transformFeedbackValuesGeneration = 0;
                 GLenum m_transformFeedbackPrimitiveMode = GL_POINTS;
                 SharedPtr<ProgramObject> m_transformFeedbackProgram;
                 Uint64 m_anyVertexAttribDefaultGeneration = 0;

MobileGL/MG_State/GLState/TextureState/TextureState.cpp

@@ -98,6 +98,7 @@ namespace MobileGL::MG_State::GLState {
     }
 
     void TextureState::SetActiveTextureUnit(Int unit) {
+        if (unit != m_activeTextureUnit) ++m_contextValuesGeneration;
         m_activeTextureUnit = unit;
     }
 } // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/TextureState/TextureState.h

@@ -84,6 +84,7 @@ namespace MobileGL::MG_State::GLState {
         void NoteUnitTouched(Int unit, Bool bindingChanged = true) {
             if (unit > m_maxTouchedUnit && unit < MAX_TEXTURE_IMAGE_UNITS) {
                 m_maxTouchedUnit = unit;
+                ++m_contextValuesGeneration;
                 // Push-on-mutation (MG_Pipe/PipeMutation.h): the high-water mark is a pushed
                 // PipeInputs field, and a bind reached from inside a verb - a backend binding
                 // its own synthesised fallback texture - would otherwise leave the block
@@ -105,6 +106,9 @@ namespace MobileGL::MG_State::GLState {
             }
         }
         Int GetMaxTouchedUnit() const { return m_maxTouchedUnit; }
+        // Moves whenever the active unit or the touched high-water mark does: the texture half of
+        // GLContext::GetContextValuesGeneration.
+        Uint64 GetContextValuesGeneration() const { return m_contextValuesGeneration; }
         // High-water mark of IMAGE units ever bound by glBindImageTexture, and it is a SECOND
         // mark rather than a widening of the one above because the two index different
         // namespaces: GL 4.6 core 8.22's image units are their own array, and folding them
@@ -198,6 +202,7 @@ namespace MobileGL::MG_State::GLState {
         Uint64 m_samplingResolutionGeneration = 0;
         Int m_maxTouchedUnit = -1;
         Int m_activeTextureUnit = 0;
+        Uint64 m_contextValuesGeneration = 0;
         Array<TextureUnit, MAX_TEXTURE_IMAGE_UNITS> m_textureUnits;
         Array<ImageTextureBinding, MAX_TEXTURE_IMAGE_UNITS> m_imageTextureBindings;
         // One default texture object (external name 0) per target, created with the context and

MobileGL/MG_Test/Pipe/ResourceEmitTest.cpp

@@ -852,6 +852,59 @@ namespace {
         EXPECT_TRUE(MGPipeHandleIsNull(MGPipeApplier().BoundVertexElements));
     }
 
+    // P14: a vertex-elements record keeps the last vertex-buffer set and index binding applied
+    // while it was bound, and bind_vertex_elements re-applies them through the same two applies
+    // - every field and both serials move as if the records had arrived again. A re-create of
+    // the handle drops what it kept (the client drops its mirror at the same create).
+    TEST(ResourceEmit, AVertexElementsRecordKeepsItsBindingsAndABindReappliesThem) {
+        ApplierGuard guard;
+        const MGPipeHandle resA{7, 3};
+        const MGPipeHandle resB{8, 1};
+        const MGPipeHandle csoA{2, 1};
+        const MGPipeHandle csoB{3, 1};
+        const ElementsBlob elementsA = MakeElements(csoA, 2, 1, true);
+        const ElementsBlob elementsB = MakeElements(csoB, 2, 1, true);
+        MGPipeApplyCreateVertexElements(elementsA.Desc, elementsA.Data());
+        MGPipeApplyCreateVertexElements(elementsB.Desc, elementsB.Data());
+
+        const auto setBuffers = [](MGPipeHandle res, Uint32 stride, Uint32 baseInstance) {
+            MGPVertexBuffers hdr{};
+            hdr.Count = 1;
+            hdr.BaseInstance = baseInstance;
+            MGPVertexBuffer entry{};
+            entry.Res = res;
+            entry.Stride = stride;
+            MGPipeApplySetVertexBuffers(hdr, &entry);
+        };
+        MGPipeApplyBindVertexElements(ElementsHandle(csoA));
+        setBuffers(resA, 12, 0);
+        MGPipeApplySetIndexBuffer(MGPIndexBuffer{resA, 0, 0, 0});
+        MGPipeApplyBindVertexElements(ElementsHandle(csoB));
+        setBuffers(resB, 20, 0);
+        MGPipeApplySetIndexBuffer(MGPIndexBuffer{resB, 0, 0, 0});
+
+        // Back to A with nothing sent but the bind: A's own bindings come back.
+        const Uint64 vertexSerial = MGPipeApplier().VertexBuffersSerial;
+        const Uint64 indexSerial = MGPipeApplier().IndexBufferSerial;
+        MGPipeApplyBindVertexElements(ElementsHandle(csoA));
+        EXPECT_EQ(MGPipeApplier().VertexBufferCount, 1u);
+        EXPECT_EQ(MGPipeApplier().VertexBuffers[0].Res, resA);
+        EXPECT_EQ(MGPipeApplier().VertexBuffers[0].Stride, 12u);
+        EXPECT_EQ(MGPipeApplier().IndexBuffer.Res, resA);
+        EXPECT_GT(MGPipeApplier().VertexBuffersSerial, vertexSerial) << "a re-applied set must move its serial";
+        EXPECT_GT(MGPipeApplier().IndexBufferSerial, indexSerial);
+        MGPipeApplyBindVertexElements(ElementsHandle(csoB));
+        EXPECT_EQ(MGPipeApplier().VertexBuffers[0].Res, resB);
+        EXPECT_EQ(MGPipeApplier().IndexBuffer.Res, resB);
+
+        // A re-create drops what A kept: binding it again leaves B's bindings in place.
+        MGPipeApplyCreateVertexElements(elementsA.Desc, elementsA.Data());
+        MGPipeApplyBindVertexElements(ElementsHandle(csoA));
+        EXPECT_EQ(MGPipeApplier().VertexBuffers[0].Res, resB);
+        EXPECT_EQ(MGPipeApplier().IndexBuffer.Res, resB);
+        EXPECT_EQ(MGPipeApplier().RefusedVertexInputCalls, 0u);
+    }
+
     // C1's observable. A call that names a record this applier does not have is a DEFINED
     // no-op - nothing stored, nothing dispatched, no serial moved - because the teardown order
     // makes exactly one such sequence legal (release the records, then every ~BufferObject

MobileGL/MG_Test/Pipe/VertexInputEmitTest.cpp

@@ -557,6 +557,60 @@ namespace {
         EXPECT_EQ(Emitter().CreateCount(), 3u);
     }
 
+    // P14: GL keeps a vertex array's buffer and element bindings ON THE ARRAY, and so does the
+    // applier's record (MGPipeVertexElementsRecord): a switch back to an array whose bindings did
+    // not move sends the bind alone - no set_vertex_buffers, no set_index_buffer. A binding that
+    // did move still goes out after the bind.
+    TEST(VertexInputEmit, ASwitchBetweenArraysWithUnchangedBindingsSendsOnlyTheBind) {
+        EmitterScope scope;
+        const SharedPtr<BufferObject> vertices = Ctx().CreateBufferObject(1);
+        vertices->Respecify(256, nullptr);
+        const SharedPtr<BufferObject> elements = Ctx().CreateBufferObject(2);
+        elements->Respecify(64, nullptr);
+        const SharedPtr<BufferObject> other = Ctx().CreateBufferObject(3);
+        other->Respecify(256, nullptr);
+        const auto configure = [&](Uint name, const SharedPtr<BufferObject>& buffer) {
+            const SharedPtr<VertexArrayObject> vao = MakeVao(name);
+            vao->SetAttributeFormat(0, 4, DataType::Float32, false, 16, 0, false);
+            vao->BindAttributeBuffer(0, buffer);
+            vao->EnableAttribute(0);
+            vao->GetIndexBufferBindingSlot().Bind(elements);
+            return vao;
+        };
+        const auto validate = [&] {
+            Uint64 bytes = Emitter().EmitVertexElements(Ctx());
+            bytes += Emitter().EmitVertexBuffers(Ctx(), 0);
+            bytes += Emitter().EmitIndexBuffer(Ctx());
+            return bytes;
+        };
+        const SharedPtr<VertexArrayObject> a = configure(1, vertices);
+        ASSERT_GT(validate(), 0u);
+        const SharedPtr<VertexArrayObject> b = configure(2, other);
+        ASSERT_GT(validate(), 0u);
+        const Uint64 sets = Emitter().VertexBufferSetCount();
+        const Uint64 indexSets = Emitter().IndexBufferSetCount();
+
+        for (int i = 0; i < 4; ++i) {
+            Ctx().BindVertexArray(1);
+            EXPECT_EQ(validate(), sizeof(MGPHandleOnly)) << "a switch to an unchanged array sent more than its bind";
+            Ctx().BindVertexArray(2);
+            EXPECT_EQ(validate(), sizeof(MGPHandleOnly));
+        }
+        EXPECT_EQ(Emitter().VertexBufferSetCount(), sets);
+        EXPECT_EQ(Emitter().IndexBufferSetCount(), indexSets);
+
+        // A moved binding is sent after the bind, and kept from then on.
+        Ctx().BindVertexArray(1);
+        a->BindAttributeBuffer(0, other);
+        EXPECT_GT(validate(), sizeof(MGPHandleOnly));
+        EXPECT_EQ(Emitter().VertexBufferSetCount(), sets + 1);
+        Ctx().BindVertexArray(2);
+        EXPECT_EQ(validate(), sizeof(MGPHandleOnly));
+        Ctx().BindVertexArray(1);
+        EXPECT_EQ(validate(), sizeof(MGPHandleOnly));
+        EXPECT_EQ(Emitter().VertexBufferSetCount(), sets + 1);
+    }
+
     // ============================ C-1: the death path ============================
     //
     // THE LEAK THIS RULES OUT, and why it is a client case rather than a backend one. Every