Add optimized FP16 performance numbers (xOS-v4.34)

README.md

@@ -39,6 +39,15 @@ SigLIP is a dual-encoder model consisting of a vision encoder and a text encoder
 
 ## Performance
 The following performance metrics were measured on the Renesas X5H board.
+
+### Optimized (xOS-v4.34)
+TAPI kernel optimizations: L2/L1 tiling re-tuning, Q/K/V copy segment elimination, and pooling tiling. These optimizations reduce L3 DDR bandwidth without changing model accuracy.
+
+| Precision | Device | VE Latency (ms) | TE Latency (ms) | NPU DDR (MB) |
+|:---:|:---:|:---:|:---:|:---:|
+| FP16 | NPX6 | 396.8 | 43.1 | 1708.04 |
+
+### Unoptimized baseline (xOS-v4.32)
 | Precision | Device | VE Latency (ms) | TE Latency (ms) | NPU DDR (MB) |
 |:---:|:---:|:---:|:---:|:---:|
 | FP16 | NPX6 | 1229.9 | 46.1 | 1714.88 |
@@ -68,7 +77,7 @@ Both variants evaluated on 9 images (5 vandalism, 4 benign) x 10 text prompts:
 ## Prerequisites
 To run the model, you need:
 
-1.  **Renesas X5H Board with SDK v4.32.0**
+1.  **Renesas X5H Board with SDK v4.32.0 or v4.34.0**
 2.  **Hugging Face CLI**: For downloading the model and installer.
 
 ## Board Setup
@@ -100,7 +109,12 @@ This loads kernel modules and starts all 14 NPU firmware cores.
 ## Deployment
 
 ### SigLIP-SO400M-patch14-384 (FP16)
-1. Download the installer **siglip-runner-3.0.0-Linux.sh** from **Files and versions** tab under **fp16/binaries/rcar-x5hv1/xOS-v4.32/** folder.
+
+Two versions are available:
+- **xOS-v4.34 (optimized)**: 3.1x faster VE encoding via TAPI kernel tiling optimizations
+- **xOS-v4.32 (baseline)**: Original unoptimized build
+
+1. Download the installer **siglip-runner-3.0.0-Linux.sh** from **Files and versions** tab under **fp16/binaries/rcar-x5hv1/xOS-v4.34/** (optimized) or **fp16/binaries/rcar-x5hv1/xOS-v4.32/** (baseline).
 2. Copy the installer to the X5H board and run the installer.
    ```bash
    bash ./siglip-runner-3.0.0-Linux.sh --prefix=./ --exclude-subdir --skip-license

fp16/benchmarks/x5h_npu.yaml

@@ -24,7 +24,14 @@ benchmark:
   parameters:
     batch_size: 1
 
+# Optimized (xOS-v4.34) — TAPI kernel tiling + segment elimination
 performance:
+  ve_encode_ms: 396.8
+  te_encode_ms: 43.1
+  total_ms: 3311.7
+
+# Unoptimized baseline (xOS-v4.32)
+performance_baseline:
   ve_encode_ms: 1229.9
   te_encode_ms: 46.1
   total_ms: 3298.6
@@ -32,7 +39,14 @@ performance:
 metrics:
   accuracy: null
 
+# Optimized (xOS-v4.34)
 memory:
+  ve_npu_ddr_mb: 915.30
+  te_npu_ddr_mb: 792.74
+  total_npu_ddr_mb: 1708.04
+
+# Unoptimized baseline (xOS-v4.32)
+memory_baseline:
   ve_npu_ddr_mb: 920.75
   te_npu_ddr_mb: 794.12
   total_npu_ddr_mb: 1714.88