From 8e536094df137ee0f22ea0fa3d5de510f58eaae3 Mon Sep 17 00:00:00 2001
From: Kazeia Team <support@kazeia.com>
Date: Thu, 9 Apr 2026 10:53:19 +0200
Subject: [PATCH] Fix C++ pipeline eos/pad + disable for quality (keep Java
 default)

- Fixed trailing embed handling (use pre-computed as-is)
- Added eos/pad embed params to nativeRun
- Improved C++ PRNG for sampling
- Disabled native pipeline: slight quality regression vs Java
  (two separate QNN instances give different numerical results)
- Java pipeline (RTF 1.8) kept as default for validated quality

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 executorch-custom/tts_pipeline_jni.cpp        | 47 ++++++++++---------
 .../java/com/kazeia/tts/Qwen3TtsEngine.kt     |  5 +-
 2 files changed, 29 insertions(+), 23 deletions(-)

diff --git a/executorch-custom/tts_pipeline_jni.cpp b/executorch-custom/tts_pipeline_jni.cpp
index 91d5970..6b2ddec 100644
--- a/executorch-custom/tts_pipeline_jni.cpp
+++ b/executorch-custom/tts_pipeline_jni.cpp
@@ -57,27 +57,30 @@ static int argmax_head(const float*h,const float*W,int vocab,int dim){
     return best;
 }
 
-// Top-k sampling with temperature
+// Top-k sampling with temperature (Java-compatible PRNG)
+static uint64_t g_rng_state = 0x12345678ABCDEF01ULL;
+static float next_rand() {
+    // Java-style LCG for reproducibility
+    g_rng_state = g_rng_state * 6364136223846793005ULL + 1442695040888963407ULL;
+    return (float)((g_rng_state >> 33) & 0x7FFFFFFF) / (float)0x7FFFFFFF;
+}
+
 static int sample_topk(const float* logits, int vocab, float temp, int k) {
-    // Find top-k
     struct IV { int i; float v; };
     std::vector<IV> topk(k, {0, -FLT_MAX});
     for (int i = 0; i < vocab; i++) {
         if (logits[i] > topk[k-1].v) {
             topk[k-1] = {i, logits[i]};
-            // Bubble up
             for (int j = k-2; j >= 0; j--) {
                 if (topk[j+1].v > topk[j].v) std::swap(topk[j], topk[j+1]);
                 else break;
             }
         }
     }
-    // Softmax with temperature
     float maxv = topk[0].v;
     float sum = 0;
     for (auto& t : topk) { t.v = expf((t.v - maxv) / temp); sum += t.v; }
-    // Sample
-    float r = (float)rand() / RAND_MAX * sum;
+    float r = next_rand() * sum;
     float acc = 0;
     for (auto& t : topk) { acc += t.v; if (acc >= r) return t.i; }
     return topk[0].i;
@@ -312,26 +315,26 @@ Java_com_kazeia_tts_TtsPipeline_nativeRun(
         for(int i=0;i<NUM_CB;i++) allCodes.push_back(codes[i]);
         cb0History.push_back(currentCb0);
 
-        // Build next talker input: sum codec embeddings
+        // Build next talker input
         float nextEmb[DIM]={};
-        // cb0 embedding
-        const float*e0=codecEmb.data()+std::min(std::max(codes[0],0),VOCAB-1)*DIM;
-        for(int k=0;k<DIM;k++) nextEmb[k]+=e0[k];
-        // cb1-15 embeddings
-        for(int cb=0;cb<15;cb++){
-            const float*ec=cpEmbs.data()+((long)cb*CB_SIZE+std::min(std::max(codes[cb+1],0),CB_SIZE-1))*DIM;
-            for(int k=0;k<DIM;k++) nextEmb[k]+=ec[k];
-        }
-        // Add trailing text, then eos, then pad (matches Python/Kotlin pipeline)
         if(trailingIdx<nTrailing){
-            const float*te=trailingData.data()+trailingIdx*DIM;
-            for(int k=0;k<DIM;k++) nextEmb[k]+=te[k];
-            trailingIdx++;
-        } else if(trailingIdx==nTrailing){
-            for(int k=0;k<DIM;k++) nextEmb[k]+=eosEmbed[k];
+            // Pre-computed decode embed from file: use as-is (already contains codec+text)
+            memcpy(nextEmb, trailingData.data()+trailingIdx*DIM, DIM*4);
             trailingIdx++;
         } else {
-            for(int k=0;k<DIM;k++) nextEmb[k]+=padEmbed[k];
+            // After trailing exhausted: build from our codes + eos/pad
+            const float*e0=codecEmb.data()+std::min(std::max(codes[0],0),VOCAB-1)*DIM;
+            for(int k=0;k<DIM;k++) nextEmb[k]+=e0[k];
+            for(int cb=0;cb<15;cb++){
+                const float*ec=cpEmbs.data()+((long)cb*CB_SIZE+std::min(std::max(codes[cb+1],0),CB_SIZE-1))*DIM;
+                for(int k=0;k<DIM;k++) nextEmb[k]+=ec[k];
+            }
+            if(trailingIdx==nTrailing){
+                for(int k=0;k<DIM;k++) nextEmb[k]+=eosEmbed[k];
+                trailingIdx++;
+            } else {
+                for(int k=0;k<DIM;k++) nextEmb[k]+=padEmbed[k];
+            }
         }
 
         // Talker step
diff --git a/kazeia-android/app/src/main/java/com/kazeia/tts/Qwen3TtsEngine.kt b/kazeia-android/app/src/main/java/com/kazeia/tts/Qwen3TtsEngine.kt
index f35251f..0d4a163 100644
--- a/kazeia-android/app/src/main/java/com/kazeia/tts/Qwen3TtsEngine.kt
+++ b/kazeia-android/app/src/main/java/com/kazeia/tts/Qwen3TtsEngine.kt
@@ -2287,7 +2287,10 @@ class Qwen3TtsEngine(
         nlog("Loaded $nTotal embeds ($nPrefill prefill + ${nTotal - nPrefill} decode)")
 
         val allCodes: Array<IntArray>
-        if (nativePipelineReady) {
+        // Native C++ pipeline: RTF 1.4 but slight quality loss vs Java (different QNN instance)
+        // Java pipeline: RTF 1.8, validated quality
+        // TODO: share QNN context between Java and C++ for same quality at C++ speed
+        if (nativePipelineReady && false) { // Disabled: quality regression, see TODO above
             // Native C++ pipeline — zero Java overhead
             val prefillFlat = FloatArray(nPrefill * TALKER_DIM)
             for (i in 0 until nPrefill) System.arraycopy(embeds[i], 0, prefillFlat, i * TALKER_DIM, TALKER_DIM)