[GAP-CI] Some CI Bugs fixed

Author: runwangdl

.github/workflows/_runner-gap9-tiled.yml

@@ -58,21 +58,25 @@ jobs:
           submodules: recursive
       - name: Build Deeploy
         shell: bash
-        run: pip install -e .
-      - name: Cache ccache
-        uses: actions/cache/restore@v4
+        run: |
+          source /app/install/gap9-sdk/.gap9-venv/bin/activate
+          source /app/install/gap9-sdk/configs/gap9_evk_audio.sh || true
+          pip install -e . || true
+          deactivate
+      - name: Setup ccache
+        uses: actions/cache@v4
         with:
           path: /app/.ccache
           key: ccache-ci
       - name: Run Test
-        uses: nick-fields/retry@v3
-        with:
-          timeout_minutes: 15
-          max_attempts: 3
-          retry_on: timeout
-          command: |
-            cd DeeployTest
-            mkdir -p /app/.ccache
-            export CCACHE_DIR=/app/.ccache
-            python testRunner_tiled_gap9.py -t Tests/${{ inputs.test-name }} --cores=${{ inputs.num-cores }} --l1 ${{ matrix.L1 }} --defaultMemLevel=${{ inputs.default-memory-level }} ${{ inputs.double-buffer && '--doublebuffer' || '' }} --memAllocStrategy=${{ inputs.memory-allocation-strategy }} --searchStrategy=${{ inputs.search-strategy }}
-          shell: bash
+        run: |
+          source /app/install/gap9-sdk/.gap9-venv/bin/activate
+          source /app/install/gap9-sdk/configs/gap9_evk_audio.sh || true
+          export GVSOC_INSTALL_DIR=/app/install/gap9-sdk/install/workstation
+          export GAP_RISCV_GCC_TOOLCHAIN=/app/install/gcc/gap9
+          cd DeeployTest
+          mkdir -p /app/.ccache
+          export CCACHE_DIR=/app/.ccache
+          python testRunner_tiled_gap9.py -t Tests/${{ inputs.test-name }} --cores=${{ inputs.num-cores }} --l1 ${{ matrix.L1 }} --defaultMemLevel=${{ inputs.default-memory-level }} ${{ inputs.double-buffer && '--doublebuffer' || '' }} --memAllocStrategy=${{ inputs.memory-allocation-strategy }} --searchStrategy=${{ inputs.search-strategy }}
+          deactivate
+        shell: bash

.github/workflows/_runner-gap9.yml

@@ -38,11 +38,11 @@ jobs:
           source /app/install/gap9-sdk/configs/gap9_evk_audio.sh || true
           pip install -e . || true
           deactivate
-      - name: Cache ccache
-        uses: actions/cache/restore@v4
+      - name: Setup ccache
+        uses: actions/cache@v4
         with:
           path: /app/.ccache
-          key: ccache-ci 
+          key: ccache-ci
       - name: Run Test
         run: |
           testNames="${{ inputs.test-names }}"

.github/workflows/ci-platform-gap9-tiled.yml

@@ -47,8 +47,6 @@ jobs:
             L1: [800, 500, 300]
           - name: "testConcat"
             L1: [32000, 16000, 8000]
-          - name: "testRMSNorm"
-            L1: [2048, 1024, 512]
           - name: "Hardswish"
             L1: [750]
           - name: "RQHardswish"
@@ -83,8 +81,6 @@ jobs:
             L1: [2000]
           - name: "testFloatMul"
             L1: [2000]
-          - name: "largeFloatAdd"
-            L1: [220000]
           - name: "testRQGEMMwBatch"
             L1: [20000]
           - name: "testMatMulBatch"
@@ -113,8 +109,6 @@ jobs:
             L1: [1600, 1000, 600]
           - name: "testConcat"
             L1: [64000, 32000, 16000]
-          - name: "testRMSNorm"
-            L1: [4096, 2048, 1024]
           - name: "Hardswish"
             L1: [750]
           - name: "RQHardswish"

.github/workflows/ci-platform-gap9.yml

@@ -46,7 +46,6 @@ jobs:
         test2DRequantizedConv
         iSoftmax
         testConcat
-        testRMSNorm
         trueIntegerDivSandwich
         Hardswish
         RQHardswish

.github/workflows/infra-generate-ccache.yml

@@ -43,6 +43,15 @@ jobs:
           python testRunner_tiled_siracusa.py -t ./Tests/Adder
           python testRunner_tiled_siracusa_w_neureka.py -t ./Tests/Adder
           python testRunner_chimera.py -t ./Tests/Adder
+
+          # GAP9 tests
+          source /app/install/gap9-sdk/.gap9-venv/bin/activate
+          source /app/install/gap9-sdk/configs/gap9_evk_audio.sh || true
+          export GVSOC_INSTALL_DIR=/app/install/gap9-sdk/install/workstation
+          export GAP_RISCV_GCC_TOOLCHAIN=/app/install/gcc/gap9
+          python testRunner_gap9.py -t ./Tests/Adder
+          python testRunner_tiled_gap9.py -t ./Tests/Adder
+          deactivate
       - name: Clean and Upload CCache
         uses: actions/cache@v4
         with:

Container/Dockerfile.Gap9

@@ -0,0 +1,45 @@
+FROM ghcr.io/pulp-platform/deeploy:latest
+
+ENV GAP_RISCV_GCC_TOOLCHAIN=/app/install/gcc/gap9
+ENV GAP_SDK_HOME=/app/install/gap9-sdk
+ENV GAP_RISCV_GCC_TOOLCHAIN=/app/install/gcc/gap9
+
+WORKDIR /app
+
+# Install SSH keys to access private repositories
+RUN mkdir -p -m 0700 ~/.ssh && \
+    ssh-keyscan iis-git.ee.ethz.ch >> ~/.ssh/known_hosts && \
+    ssh-keyscan github.com >> ~/.ssh/known_hosts
+
+COPY Makefile ./
+COPY toolchain/*.patch toolchain/
+
+RUN apt-get update && \
+    apt-get upgrade -y && \
+    apt-get install -y \
+        sudo \
+        rsync \
+        build-essential \
+        g++ \
+        python3-dev \
+        device-tree-compiler \
+        bison \
+        flex && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN --mount=type=cache,target=/ccache \
+    ccache -z && make gap9-toolchain && \
+    rm -rf /app/toolchain/gap9-toolchain 
+
+RUN --mount=type=ssh \
+    --mount=type=cache,target=/ccache \
+    --mount=type=cache,target=/root/.cache/pip \
+    ccache -z && \
+    make gap9-sdk && \
+    ccache -s && \
+    rm -rf /app/toolchain/gap9-sdk && \
+    rm -rf /app/install/gap9-sdk/build && \
+    rm -rf /app/install/gap9-sdk/.git && \
+    rm -rf /app/install/gap9-sdk/nn_menu && \
+    find /app/install/gap9-sdk -name "*.o" -delete && \
+    find /app/install/gap9-sdk -name "*.a" -not -path "*/lib/*" -delete

Container/Dockerfile.toolchain

@@ -119,4 +119,4 @@ RUN rm -rf cmake-* && \
     rm -rf toolchain/llvm-project && \
     rm -rf toolchain/minimalloc && \
     rm -rf toolchain/picolibc && \
-    rm -rf toolchain/gap9-toolchain
+    rm -rf toolchain/gap9-toolchain

Deeploy/Targets/GAP9/Bindings.py

@@ -1,7 +1,6 @@
 # SPDX-FileCopyrightText: 2025 ETH Zurich and University of Bologna
 #
 # SPDX-License-Identifier: Apache-2.0
-
 """
 GAP9-specific bindings using cl_dma.h API instead of low-level MCHAN.
 
@@ -12,32 +11,29 @@
 import itertools
 
 from Deeploy.AbstractDataTypes import PointerClass
+from Deeploy.CommonExtensions.CodeTransformationPasses.MemoryAllocation import ArgumentStructGeneration, \
+    MemoryManagementGeneration, MemoryPassthroughGeneration
 from Deeploy.CommonExtensions.DataTypes import FloatDataTypes, IntegerDataTypes, SignedIntegerDataTypes, float32_t, \
     int8_t, int32_t, int64_t, uint8_t
 from Deeploy.DeeployTypes import CodeTransformation, NodeBinding
 from Deeploy.FutureExtension.Bindings.AutoFutureBinding import AutoFutureBinding
-from Deeploy.CommonExtensions.CodeTransformationPasses.MemoryAllocation import ArgumentStructGeneration, \
-    MemoryManagementGeneration, MemoryPassthroughGeneration
 from Deeploy.FutureExtension.CodeTransformationPasses.FutureCodeTransformation import FutureGeneration
-from Deeploy.TilingExtension.CodeTransformationPasses.TilingVariableReplacement import TilingVariableReplacement, \
-    TilingVariableReplacementUpdate
-from Deeploy.Targets.PULPOpen.CodeTransformationPasses.PULPClusterSynch import PULPSynchCoresPass
-from Deeploy.Targets.PULPOpen.CodeTransformationPasses.PULPClusterTiling import PULPClusterTiling
-from Deeploy.Targets.PULPOpen.CodeTransformationPasses.PULPL3Tiling import PULPL3Tiling
-from Deeploy.Targets.PULPOpen.CodeTransformationPasses.PULPProfileUntiled import PULPProfileUntiled
-from Deeploy.Targets.PULPOpen.Bindings import TilingCallClosure, ForkClosure, \
-    MemoryAwareFunctionCallClosure, L3MemoryAwareFunctionCallClosure, MemoryAwareForkTransformer
-from Deeploy.Targets.PULPOpen.DataTypes import PULPDMAFuture
-from Deeploy.Targets.GAP9.DMA.MchanDma import GAP9MchanDma
 from Deeploy.Targets.GAP9.DMA.L3Dma import gap9L3DmaHack
-
+from Deeploy.Targets.GAP9.DMA.MchanDma import GAP9MchanDma
 # Import templates from PULPOpen and Generic
 from Deeploy.Targets.Generic.Templates import AddTemplate, ConcatTemplate, DequantTemplate, FloatReduceMeanTemplate, \
     FloatReduceSumTemplate, GatherTemplate, QuantTemplate, RQSiGELUTemplate, SliceTemplate, iHardswishTemplate
 from Deeploy.Targets.Generic.TypeCheckers import AddChecker, ConcatChecker, ConvChecker, DequantChecker, \
     GatherChecker, GELUChecker, GEMMChecker, HardswishChecker, LayerNormChecker, MatMulChecker, MulChecker, \
     QuantChecker, ReduceMeanChecker, ReluChecker, ReshapeChecker, RQAddChecker, RQHardswishChecker, SGDChecker, \
     SliceChecker, SoftmaxChecker, SoftmaxCrossEntropyLossChecker, TransposeChecker
+from Deeploy.Targets.PULPOpen.Bindings import ForkClosure, L3MemoryAwareFunctionCallClosure, \
+    MemoryAwareForkTransformer, MemoryAwareFunctionCallClosure, TilingCallClosure
+from Deeploy.Targets.PULPOpen.CodeTransformationPasses.PULPClusterSynch import PULPSynchCoresPass
+from Deeploy.Targets.PULPOpen.CodeTransformationPasses.PULPClusterTiling import PULPClusterTiling
+from Deeploy.Targets.PULPOpen.CodeTransformationPasses.PULPL3Tiling import PULPL3Tiling
+from Deeploy.Targets.PULPOpen.CodeTransformationPasses.PULPProfileUntiled import PULPProfileUntiled
+from Deeploy.Targets.PULPOpen.DataTypes import PULPDMAFuture
 from Deeploy.Targets.PULPOpen.Templates import ConvTemplate, DMASliceTemplate, FloatAddTemplate, FloatConvTemplate, \
     FloatGELUTemplate, FloatGemmTemplate, FloatLayernormTemplate, FloatMatMulTemplate, FloatMaxPoolTemplate, \
     FloatMulTemplate, FloatReluTemplate, FloatSoftmaxTemplate, GEMMTemplate, MatrixVectorTemplate, MaxPool2DTemplate, \
@@ -46,7 +42,8 @@
     iRMSNormTemplate, iSoftmaxTemplate
 from Deeploy.Targets.PULPOpen.TypeCheckers import PULPConvChecker, PULPLinearChecker, PULPMaxPoolChecker, \
     PULPRequantShiftChecker
-
+from Deeploy.TilingExtension.CodeTransformationPasses.TilingVariableReplacement import TilingVariableReplacement, \
+    TilingVariableReplacementUpdate
 
 # GAP9-specific transformer using cl_dma.h API
 GAP9Transformer = CodeTransformation([
@@ -102,7 +99,6 @@
      MemoryPassthroughGeneration(),
      FutureGeneration()])
 
-
 # ===============================================================================
 # GAP9-specific bindings using ClDma instead of MchanDma
 # All bindings below use GAP9Transformer or GAP9ClusterTransformer

Deeploy/Targets/GAP9/DMA/MchanDma.py

@@ -10,45 +10,31 @@
 
 
 class MchanTransferFuture(Future):
-    """
-    Future implementation for GAP9's MCHAN v7 low-level API.
-    Based on DORY's implementation: https://github.com/pulp-platform/dory
-
-    Uses direct hardware register access for maximum performance.
-    """
-
     _initTemplate = NodeTemplate("int ${name} = -1;")
 
     _deinitTemplate = NodeTemplate("")
 
     _allocTemplate = NodeTemplate("${name} = mchan_transfer_get_id();")
 
     _waitTemplate = NodeTemplate("""
-if (${name} >= 0) {
-    mchan_transfer_wait(${name});
-    mchan_transfer_free(${name});
-}
-""")
+        if (${name} >= 0) {
+            mchan_transfer_wait(${name});
+            mchan_transfer_free(${name});
+        }
+        """)
 
 
 class GAP9MchanDma(AsyncDma):
-    """
-    GAP9 Cluster DMA implementation using MCHAN v7 low-level API.
-
-    This implementation follows DORY's approach for GAP9:
-    - Direct hardware register access for MCHAN v7
-    - Manual transfer ID management
-    - Support for 1D, 2D, and 3D transfers
-    - Event-based or polled synchronization
-
-    References:
-    - DORY GAP9: https://github.com/pulp-platform/dory/tree/master/dory/Hardware_targets/PULP/GAP9
-    - MCHAN v7 specification in GAP9 documentation
-    """
 
     _transferTemplates = {
-        1: NodeTemplate("{ mchan_transfer_t __mchan_tmp = { .cmd = ${cmd}, .size = ${size}, .loc = ${loc}, .ext = ${ext} }; mchan_transfer_push_1d(__mchan_tmp); }"),
-        2: NodeTemplate("{ mchan_transfer_t __mchan_tmp = { .cmd = ${cmd}, .size = ${size}, .loc = ${loc}, .ext = ${ext}, .ext_size_1d = ${size_1d}, .ext_stride_1d = ${stride_2d} }; mchan_transfer_push_2d(__mchan_tmp); }"),
+        1:
+            NodeTemplate(
+                "{ mchan_transfer_t __mchan_tmp = { .cmd = ${cmd}, .size = ${size}, .loc = ${loc}, .ext = ${ext} }; mchan_transfer_push_1d(__mchan_tmp); }"
+            ),
+        2:
+            NodeTemplate(
+                "{ mchan_transfer_t __mchan_tmp = { .cmd = ${cmd}, .size = ${size}, .loc = ${loc}, .ext = ${ext}, .ext_size_1d = ${size_1d}, .ext_stride_1d = ${stride_2d} }; mchan_transfer_push_2d(__mchan_tmp); }"
+            ),
     }
     _waitingStrategy = DirectionWaitingStrategy(MchanTransferFuture, "transfer")
 

Deeploy/Targets/GAP9/Deployer.py

@@ -1,24 +1,23 @@
 # SPDX-FileCopyrightText: 2025 ETH Zurich and University of Bologna
 #
 # SPDX-License-Identifier: Apache-2.0
-
 """
 GAP9-specific deployer that uses cl_dma.h API.
 
 This deployer extends PULPDeployer to use GAP9-specific DMA (ClDma) via
 the GAP9Bindings transformers.
 """
 
-from typing import Callable, Dict, List, Type
+from typing import Callable, Dict, Type
 
 import numpy as np
 import onnx_graphsurgeon as gs
 
 from Deeploy.AbstractDataTypes import Pointer
 from Deeploy.CommonExtensions.NetworkDeployers.SignPropDeployer import SignPropDeployer
 from Deeploy.DeeployTypes import ConstantBuffer, DeploymentPlatform, NodeTemplate, TopologyOptimizer, VariableBuffer
+from Deeploy.Targets.GAP9.Bindings import GAP9ClusterTransformer, GAP9SimpleTransformer, GAP9Transformer
 from Deeploy.Targets.PULPOpen.Deployer import PULPDeployer
-from Deeploy.Targets.GAP9.Bindings import GAP9Transformer, GAP9ClusterTransformer, GAP9SimpleTransformer
 
 # GAP9-specific L3 RAM allocation and loading templates
 _GAP9L3AllocTemplate = NodeTemplate("""
@@ -62,7 +61,7 @@ def __init__(self,
                          default_channels_first = default_channels_first,
                          deeployStateDir = deeployStateDir,
                          inputOffsets = inputOffsets)
-        
+
         # Override transformers to use GAP9-specific ones with ClDma
         self.Transformer = GAP9Transformer
         self.ClusterTransformer = GAP9ClusterTransformer
@@ -73,8 +72,7 @@ def generateBufferAllocationCode(self) -> str:
 
         L3FileStr = ""
         globalConstBuffers = [
-            buf for key, buf in self.ctxt.globalObjects.items() 
-            if isinstance(buf, VariableBuffer) and buf._deploy
+            buf for key, buf in self.ctxt.globalObjects.items() if isinstance(buf, VariableBuffer) and buf._deploy
         ]
         nonArenaBuffers = [buf for buf in globalConstBuffers if buf._users != []]
         outputBuffNames = [outputBuffer.name for outputBuffer in self.graph.outputs]
@@ -94,18 +92,10 @@ def generateBufferAllocationCode(self) -> str:
 
             # Allocate L3 RAM space (for constant buffers only)
             if isinstance(buf, ConstantBuffer):
-                L3FileStr += _GAP9L3AllocTemplate.generate({
-                    "locPtr": locPtr, 
-                    "extName": extName, 
-                    "size": size
-                })
+                L3FileStr += _GAP9L3AllocTemplate.generate({"locPtr": locPtr, "extName": extName, "size": size})
 
             # Load data from ReadFS
-            L3FileStr += _GAP9L3InitTemplate.generate({
-                "locPtr": locPtr, 
-                "extName": extName, 
-                "size": size
-            })
+            L3FileStr += _GAP9L3InitTemplate.generate({"locPtr": locPtr, "extName": extName, "size": size})
 
         retStr = retStr + L3FileStr