QNN: LR-ASPP MobileNetV3 a16w8, one context binary per Hexagon version

README.md

@@ -19,6 +19,26 @@ Upstream model: [LRASPP](https://docs.pytorch.org/vision/main/models/lraspp.html
 | `coreml/lraspp_mobilenet_v3_large_coreml_fp16.pte` | coreml | fp16 |
 | `xnnpack/lraspp_mobilenet_v3_large_xnnpack_fp32.pte` | xnnpack | fp32 |
 | `xnnpack/lraspp_mobilenet_v3_large_xnnpack_int8.pte` | xnnpack | int8 |
+| `qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v69.pte` | qnn | a16w8 |
+| `qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v73.pte` | qnn | a16w8 |
+| `qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v75.pte` | qnn | a16w8 |
+| `qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v79.pte` | qnn | a16w8 |
+| `qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v81.pte` | qnn | a16w8 |
+
+## QNN variants
+
+A QNN `.pte` embeds an HTP context binary compiled for one Hexagon version, so
+there is one file per version and the runtime picks the one matching the device.
+These return a class index per pixel, `(1, 520, 520)` int32, rather than
+`(1, 21, 520, 520)` logits.
+
+| Hexagon | Example SoC |
+| --- | --- |
+| v69 | Snapdragon 8 Gen 1 |
+| v73 | Snapdragon 8 Gen 2 |
+| v75 | Snapdragon 8 Gen 3 |
+| v79 | Snapdragon 8 Elite |
+| v81 | Snapdragon 8 Elite Gen 5 |
 
 ## Unshipped backends
 

qnn/config.json

@@ -0,0 +1,167 @@
+{
+  "$schema": "https://huggingface.co/software-mansion/react-native-executorch-spec/resolve/main/config.schema.json",
+  "model": "lraspp_mobilenet_v3_large",
+  "family": "lraspp",
+  "capabilities": [
+    "semantic-segmentation"
+  ],
+  "backend": "qnn",
+  "license": "bsd-3-clause",
+  "variants": [
+    {
+      "file": "lraspp_mobilenet_v3_large_qnn_a16w8_v69.pte",
+      "precision": "a16w8",
+      "quantized": true,
+      "default": false,
+      "methods": {
+        "forward": {
+          "inputs": [
+            {
+              "shape": [
+                1,
+                3,
+                520,
+                520
+              ],
+              "dtype": "float32"
+            }
+          ],
+          "outputs": [
+            {
+              "shape": [
+                1,
+                520,
+                520
+              ],
+              "dtype": "int32"
+            }
+          ]
+        }
+      }
+    },
+    {
+      "file": "lraspp_mobilenet_v3_large_qnn_a16w8_v73.pte",
+      "precision": "a16w8",
+      "quantized": true,
+      "default": false,
+      "methods": {
+        "forward": {
+          "inputs": [
+            {
+              "shape": [
+                1,
+                3,
+                520,
+                520
+              ],
+              "dtype": "float32"
+            }
+          ],
+          "outputs": [
+            {
+              "shape": [
+                1,
+                520,
+                520
+              ],
+              "dtype": "int32"
+            }
+          ]
+        }
+      }
+    },
+    {
+      "file": "lraspp_mobilenet_v3_large_qnn_a16w8_v75.pte",
+      "precision": "a16w8",
+      "quantized": true,
+      "default": false,
+      "methods": {
+        "forward": {
+          "inputs": [
+            {
+              "shape": [
+                1,
+                3,
+                520,
+                520
+              ],
+              "dtype": "float32"
+            }
+          ],
+          "outputs": [
+            {
+              "shape": [
+                1,
+                520,
+                520
+              ],
+              "dtype": "int32"
+            }
+          ]
+        }
+      }
+    },
+    {
+      "file": "lraspp_mobilenet_v3_large_qnn_a16w8_v79.pte",
+      "precision": "a16w8",
+      "quantized": true,
+      "default": false,
+      "methods": {
+        "forward": {
+          "inputs": [
+            {
+              "shape": [
+                1,
+                3,
+                520,
+                520
+              ],
+              "dtype": "float32"
+            }
+          ],
+          "outputs": [
+            {
+              "shape": [
+                1,
+                520,
+                520
+              ],
+              "dtype": "int32"
+            }
+          ]
+        }
+      }
+    },
+    {
+      "file": "lraspp_mobilenet_v3_large_qnn_a16w8_v81.pte",
+      "precision": "a16w8",
+      "quantized": true,
+      "default": true,
+      "methods": {
+        "forward": {
+          "inputs": [
+            {
+              "shape": [
+                1,
+                3,
+                520,
+                520
+              ],
+              "dtype": "float32"
+            }
+          ],
+          "outputs": [
+            {
+              "shape": [
+                1,
+                520,
+                520
+              ],
+              "dtype": "int32"
+            }
+          ]
+        }
+      }
+    }
+  ]
+}

qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v69.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cafb0690fc815c9ddc18db2b21fdd263ce1b56757d2f66995b534b2919ebb502
+size 7169536

qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v73.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5c2bed25b9dac0844cc68b655bc1333e61e62547f3041fc427e66b5c447008d
+size 4920832

qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v75.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:413185caca4f5013355b5ef384cd463474dcfa57a4ef9d014c12f7ab0b1af7c9
+size 4920832

qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v79.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fadada5a79b0d079aaf4c647a4ef57ce5ea8b3f3672a2cfb6404f2ca2118532c
+size 4945408

qnn/lraspp_mobilenet_v3_large_qnn_a16w8_v81.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aec83444a2b492313449456d5f1d89c088fb93566de23290bba222c7411eac48
+size 4953600