Merge branch 'main' into Add-support-for-elu

Author: agrima1304

.ci/scripts/setup-vulkan-linux-deps.sh

@@ -43,7 +43,7 @@ install_vulkan_sdk() {
   export PATH="${PATH}:${_vulkan_sdk_dir}/${VULKAN_SDK_VERSION}/x86_64/bin/"
 }
 
-VULKAN_SDK_VERSION="1.3.296.0"
+VULKAN_SDK_VERSION="1.4.321.1"
 
 install_swiftshader
 install_vulkan_sdk "${VULKAN_SDK_VERSION}"

.github/workflows/apple-perf.yml

@@ -416,7 +416,7 @@ jobs:
       - set-parameters
     secrets: inherit
     with:
-      runner: macos-latest-xlarge
+      runner: macos-14-xlarge
       python-version: '3.11'
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}

.github/workflows/apple.yml

@@ -49,7 +49,7 @@ jobs:
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     secrets: inherit
     with:
-      runner: macos-latest-xlarge
+      runner: macos-14-xlarge
       python-version: '3.11'
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
@@ -136,7 +136,7 @@ jobs:
     needs: set-version
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     with:
-      runner: macos-latest-xlarge
+      runner: macos-14-xlarge
       python-version: '3.11'
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
@@ -276,7 +276,7 @@ jobs:
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     secrets: inherit
     with:
-      runner: macos-latest-xlarge
+      runner: macos-14-xlarge
       python-version: '3.11'
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}

.github/workflows/build-presets.yml

@@ -22,7 +22,7 @@ jobs:
     with:
       job-name: build
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      runner: macos-latest-xlarge
+      runner: macos-14-xlarge
       python-version: 3.12
       submodules: recursive
       timeout: 90

.github/workflows/build-wheels-macos.yml

@@ -63,7 +63,7 @@ jobs:
       post-script: ${{ matrix.post-script }}
       package-name: ${{ matrix.package-name }}
       # Meta's macOS runners do not have Xcode, so use GitHub's runners.
-      runner-type: macos-latest-xlarge
+      runner-type: macos-14-xlarge
       setup-miniconda: true
       smoke-test-script: ${{ matrix.smoke-test-script }}
       trigger-event: ${{ github.event_name }}

.github/workflows/pull.yml

@@ -855,13 +855,14 @@ jobs:
         .ci/scripts/setup-linux.sh --build-tool "cmake"
 
         # Install test requirements
-        pip install -r backends/nxp/requirements-tests.txt
+        pip install -r backends/nxp/requirements-tests-pypi.txt
+        pip install -r backends/nxp/requirements-tests-eiq.txt
 
         # Run pytest
         PYTHON_EXECUTABLE=python bash backends/nxp/run_unittests.sh
 
         # Run aot examples:
-        PYTHON_EXECUTABLE=python bash examples/nxp/run_aot_example.sh cifar10        
+        PYTHON_EXECUTABLE=python bash examples/nxp/run_aot_example.sh cifar10
         PYTHON_EXECUTABLE=python bash examples/nxp/run_aot_example.sh mobilenetv2
 
 
@@ -901,6 +902,34 @@ jobs:
         done
 
 
+  test-vulkan-operators-linux:
+    name: test-vulkan-operators-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    with:
+      runner: linux.2xlarge
+      docker-image: ci-image:executorch-ubuntu-22.04-clang12
+      submodules: 'recursive'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        set -eux
+
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        # Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
+        source .ci/scripts/setup-vulkan-linux-deps.sh
+
+        # Setup python
+        PYTHON_EXECUTABLE=python \
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_VULKAN=ON" \
+        .ci/scripts/setup-linux.sh --build-tool "cmake"
+
+        PYTHON_EXECUTABLE=python bash backends/vulkan/test/custom_ops/build_and_run.sh add
 
   nxp-build-test:
     name: nxp-build-test

.github/workflows/trunk.yml

@@ -435,7 +435,7 @@ jobs:
     name: test-coreml-delegate
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     with:
-      runner: macos-latest-xlarge
+      runner: macos-14-xlarge
       python-version: '3.11'
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}

backends/cadence/aot/TARGETS

@@ -604,3 +604,17 @@ python_unittest(
         "//later:lib",
     ],
 )
+
+python_unittest(
+    name = "test_ref_implementations",
+    srcs = [
+        "tests/test_ref_implementations.py",
+    ],
+    supports_static_listing = False,
+    typing = True,
+    deps = [
+        ":typing_stubs",
+        "//executorch/backends/cadence/aot:ref_implementations",
+        "//caffe2:torch",
+    ]
+)

backends/cadence/aot/ops_registrations.py

@@ -16,6 +16,7 @@
     get_im2row_output_size,
 )
 from executorch.exir.scalar_type import ScalarType
+from torch._meta_registrations import _linalg_svd_meta
 from torch.library import Library, register_fake
 
 lib = Library("cadence", "DEF")
@@ -250,6 +251,12 @@
     "int in_zero_point, bool channel_last=False) -> (Tensor out)"
 )
 lib.define("linalg_vector_norm(Tensor X) -> (Tensor Y)")
+lib.define(
+    "linalg_svd(Tensor A, bool full_matrices=False, bool compute_uv=True, str? driver=None) -> (Tensor U, Tensor S, Tensor Vh)"
+)
+lib.define(
+    "linalg_svd.out(Tensor A, bool full_matrices=False, bool compute_uv=True, str? driver=None, *, Tensor(a!) U, Tensor(b!) S, Tensor(c!) Vh) -> (Tensor(a!) U, Tensor(b!) S, Tensor(c!) Vh)"
+)
 lib.define(
     "transposed_im2row(Tensor input, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride, "
     "int[2] output_padding, Tensor in_zero_point, bool channel_last=False) -> (Tensor out)"
@@ -1576,6 +1583,26 @@ def linalg_vector_norm_meta(
     return X.new_empty([], dtype=X.dtype)
 
 
+@register_fake("cadence::linalg_svd")
+def linalg_svd_meta(
+    A: torch.Tensor,
+    full_matrices: bool = False,
+    compute_uv: bool = True,
+    driver: Optional[str] = None,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    # Based on the _linalg_svd meta implementation, but ensuring contiguous strides
+
+    # Get the shapes from the original meta function
+    U, S, Vh = _linalg_svd_meta(A, full_matrices, compute_uv, driver)
+
+    # Create new tensors with contiguous strides to fix the non-contiguous issue
+    U_contiguous = A.new_empty(U.shape, dtype=A.dtype).contiguous()
+    S_contiguous = A.new_empty(S.shape, dtype=A.dtype).contiguous()
+    Vh_contiguous = A.new_empty(Vh.shape, dtype=A.dtype).contiguous()
+
+    return U_contiguous, S_contiguous, Vh_contiguous
+
+
 @register_fake("cadence::requantize")
 def requantize_meta(
     input: torch.Tensor,

backends/cadence/aot/ref_implementations.py

@@ -20,6 +20,42 @@
 }
 
 
+@impl(m, "quantize_per_tensor")
+def quantize_per_tensor(
+    input: torch.Tensor,
+    scale: float,
+    zero_point: int,
+    quant_min: int,
+    quant_max: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    """
+    Quantizes a floating-point tensor to an integral tensor.
+
+    Args:
+        - input (Tensor): input tensor
+        - scale (float): Quantization scale. Derived from the ratio
+            between the min/max of the floating-point tensor and the
+            min/max of the quantized range.
+        - zero_point (int): The point which represents 0 in the quantized
+            range. For example, consider the floating point range [-1., 2.] and
+            quantized integer range [-7, 7]. In this case, 0 is 1/3 of way from
+            -1. to 2. So, the point that represents 0 in the quantized range should
+            be 1/3 of the way from [-7, 7]. This ends up being -2 in the integer space.
+        - quant_min (int): The smallest value in the quantized domain. Unused since scale
+            is already provided.
+        - quant_max (int): The largest value in the quantized domain. Unused since scale
+            is already provided.
+        - dtype (torch.dtype): The type of the output tensor
+    """
+    supported_quant_types = [torch.int8, torch.int16, torch.int32]
+    if dtype not in supported_quant_types:
+        raise ValueError(
+            f"Unsupported dtype to quantize to. Supported dtypes must be one of {supported_quant_types}"
+        )
+    return torch.round(input / scale + zero_point).to(dtype)
+
+
 @impl(m, "requantize")
 def requantize(
     input: torch.Tensor,