IRON host runtime abstraction

.github/workflows/buildAndTestRyzenAI.yml

@@ -61,7 +61,7 @@ jobs:
       matrix:
         runner_type: [ amd7940hs, amdhx370 ]
     env:
-      IRON_CACHE_HOME: ${{ github.workspace }}/iron-cache-${{ matrix.runner_type }}-${{ github.run_id }}
+      NPU_CACHE_HOME: ${{ github.workspace }}/iron-cache-${{ matrix.runner_type }}-${{ github.run_id }}
       PIP_CACHE_DIR: ${{ github.workspace }}/.pip-cache-${{ matrix.runner_type }}-${{ github.run_id }}
     steps:
       - uses: actions/checkout@v4
@@ -130,18 +130,21 @@ jobs:
             $CMAKE_ARGS
 
           # Create runner-specific cache directory
-          rm -rf $IRON_CACHE_HOME
-          mkdir $IRON_CACHE_HOME
+          rm -rf $NPU_CACHE_HOME
+          mkdir $NPU_CACHE_HOME
+
+          # Set number of contexts to maintain in cache per process
+          export XRT_CONTEXT_CACHE_SIZE=3
 
           ninja install
           ninja check-aie
           ninja check-aie-concurrency
           popd
 
-      - name: Cleanup IRON_CACHE_HOME
+      - name: Cleanup NPU_CACHE_HOME
         if: always()
         run: |
-          rm -rf $IRON_CACHE_HOME
+          rm -rf $NPU_CACHE_HOME
           rm -rf $PIP_CACHE_DIR
 
   build-quick-setup:
@@ -152,7 +155,7 @@ jobs:
       matrix:
         runner_type: [ amd7940hs, amdhx370 ]
     env:
-      IRON_CACHE_HOME: ${{ github.workspace }}/iron-cache-${{ matrix.runner_type }}-${{ github.run_id }}
+      NPU_CACHE_HOME: ${{ github.workspace }}/iron-cache-${{ matrix.runner_type }}-${{ github.run_id }}
       PIP_CACHE_DIR: ${{ github.workspace }}/.pip-cache-${{ matrix.runner_type }}-${{ github.run_id }}
     steps:
       - uses: actions/checkout@v4
@@ -200,17 +203,20 @@ jobs:
             LIT_OPTS="-j12 $LIT_OPTS"
           fi
 
+          # Set number of contexts to maintain in cache per process
+          export XRT_CONTEXT_CACHE_SIZE=3
+
           # Create runner-specific cache directory
-          rm -rf $IRON_CACHE_HOME
-          mkdir $IRON_CACHE_HOME
+          rm -rf $NPU_CACHE_HOME
+          mkdir $NPU_CACHE_HOME
 
           ninja install
           ninja check-reference-designs
           ninja check-programming-guide
           popd
 
-      - name: Cleanup IRON_CACHE_HOME
+      - name: Cleanup NPU_CACHE_HOME
         if: always()
         run: |
-          rm -rf $IRON_CACHE_HOME
+          rm -rf $NPU_CACHE_HOME
           rm -rf $PIP_CACHE_DIR

.github/workflows/buildAndTestVitis.yml

@@ -61,7 +61,7 @@ jobs:
       matrix:
         runner_type: [ ubuntu-vitis ]
     env:
-      IRON_CACHE_HOME: ${{ github.workspace }}/iron-cache-${{ matrix.runner_type }}-${{ github.run_id }}
+      NPU_CACHE_HOME: ${{ github.workspace }}/iron-cache-${{ matrix.runner_type }}-${{ github.run_id }}
     steps:
       - uses: actions/checkout@v4
         with:
@@ -125,13 +125,13 @@ jobs:
             $CMAKE_ARGS
 
           # Create runner-specific cache directory
-          rm -rf $IRON_CACHE_HOME
-          mkdir $IRON_CACHE_HOME
+          rm -rf $NPU_CACHE_HOME
+          mkdir $NPU_CACHE_HOME
 
           ninja install
           ninja check-aie
           popd
 
-      - name: Cleanup IRON_CACHE_HOME
+      - name: Cleanup NPU_CACHE_HOME
         if: always()
-        run: rm -rf $IRON_CACHE_HOME
+        run: rm -rf $NPU_CACHE_HOME

programming_examples/basic/matrix_multiplication/makefile-common

@@ -196,13 +196,13 @@ run: ${targetname}.exe ${xclbin_target}
 trace: ${targetname}.exe ${trace_xclbin_target} ${insts_target}
 	export XRT_HACK_UNSECURE_LOADING_XCLBIN=1 && \
 	${powershell} ./$< -x ${trace_xclbin_target} -i ${insts_target} -k MLIR_AIE -M $M -K $K -N $N ${runargs} -t ${trace_size}
-	${srcdir}/../../../../python/utils/parse_trace.py --input trace.txt --mlir ${trace_mlir_target} --output trace_mm.json
+	${srcdir}/../../../../python/utils/trace/parse.py --input trace.txt --mlir ${trace_mlir_target} --output trace_mm.json
 
 #	${powershell} ./$< -x ${trace_xclbin_target} -i ${insts_target} -k MLIR_AIE -M $M -K $K -N $N -v 1 --warmup 0 --iters 1 -t ${trace_size}
 
 .PHONY: parse_trace
 parse_trace:
-	${srcdir}/../../../../python/utils/parse_trace.py --input trace.txt --mlir ${trace_mlir_target} --output trace_mm.json
+	${srcdir}/../../../../python/utils/trace/parse.py --input trace.txt --mlir ${trace_mlir_target} --output trace_mm.json
 
 .PHONY: clean
 clean: clean_trace

programming_examples/basic/matrix_multiplication/single_core/single_core.py

@@ -12,7 +12,6 @@
 from aie.dialects.aie import *
 from aie.dialects.aiex import *
 import aie.utils.trace as trace_utils
-from aie.utils.trace import PortEvent
 from aie.iron.controlflow import range_
 from aie.iron.dtype import str_to_dtype
 
@@ -280,28 +279,28 @@ def sequence(A, B, C):
                         trace_size=trace_size,
                         coretile_events=[
                             # captures input A (PORT_RUNNING_0, at port number 1, master for inputs)
-                            trace_utils.PortEvent(
-                                trace_utils.CoreEvent.PORT_RUNNING_0,
+                            trace_utils.events.PortEvent(
+                                trace_utils.events.CoreEvent.PORT_RUNNING_0,
                                 port_number=1,
                                 master=True,
                             ),
                             # captures input B (PORT_RUNNING_1, at port number 2, master for inputs)
-                            trace_utils.PortEvent(
-                                trace_utils.CoreEvent.PORT_RUNNING_1,
+                            trace_utils.events.PortEvent(
+                                trace_utils.events.CoreEvent.PORT_RUNNING_1,
                                 port_number=2,
                                 master=True,
                             ),
                             # captures output C (PORT_RUNNING_2, at port number 1, slave for outputs)
-                            trace_utils.PortEvent(
-                                trace_utils.CoreEvent.PORT_RUNNING_2,
+                            trace_utils.events.PortEvent(
+                                trace_utils.events.CoreEvent.PORT_RUNNING_2,
                                 port_number=1,
                                 master=False,
                             ),
-                            trace_utils.CoreEvent.INSTR_EVENT_0,
-                            trace_utils.CoreEvent.INSTR_EVENT_1,
-                            trace_utils.CoreEvent.MEMORY_STALL,
-                            trace_utils.CoreEvent.LOCK_STALL,
-                            trace_utils.CoreEvent.INSTR_VECTOR,
+                            trace_utils.events.CoreEvent.INSTR_EVENT_0,
+                            trace_utils.events.CoreEvent.INSTR_EVENT_1,
+                            trace_utils.events.CoreEvent.MEMORY_STALL,
+                            trace_utils.events.CoreEvent.LOCK_STALL,
+                            trace_utils.events.CoreEvent.INSTR_VECTOR,
                         ],
                     )
 

programming_examples/basic/packet_switch/test.cpp

@@ -113,10 +113,13 @@ int main(int argc, const char *argv[]) {
       ref = srcVecA[i] + 2; // ref for the second input packet
     }
     if (*(bufOut + i) != ref) {
-      if (errors < 10) {
+      if (errors < 100) {
         std::cout << "Error in output " << i << "; Input: " << srcVecA[i]
                   << "; Output: " << *(bufOut + i) << " != reference:" << ref
                   << std::endl;
+      } else if (errors == 100) {
+        std::cout << "..." << std::endl;
+        std::cout << "[Errors truncated]" << std::endl;
       }
       errors++;
     }

programming_examples/basic/passthrough_kernel/Makefile

@@ -82,13 +82,13 @@ run_py: build/final_${data_size}.xclbin build/insts_${data_size}.bin
 
 trace: ${targetname}_${data_size}.exe build/final_trace_${data_size}.xclbin build/insts_${data_size}.bin
 	${powershell} ./$< -x build/final_trace_${data_size}.xclbin -i build/insts_${data_size}.bin -k MLIR_AIE -t ${trace_size}
-	${srcdir}/../../../python/utils/parse_trace.py --input trace.txt --mlir build/aie2_trace_lineBased_8b_${data_size}.mlir --output trace_passthrough_kernel.json
-	${srcdir}/../../../python/utils/get_trace_summary.py --input trace_passthrough_kernel.json
+	${srcdir}/../../../python/utils/trace/parse.py --input trace.txt --mlir build/aie2_trace_lineBased_8b_${data_size}.mlir --output trace_passthrough_kernel.json
+	${srcdir}/../../../python/utils/trace/get_trace_summary.py --input trace_passthrough_kernel.json
 
 trace_py: build/final_trace_${data_size}.xclbin build/insts_${data_size}.bin
 	${powershell} python3 ${srcdir}/test.py -x build/final_trace_${data_size}.xclbin -i build/insts_${data_size}.bin -k MLIR_AIE -t ${trace_size} -i1s ${in1_size} -os ${out_size}
-	${srcdir}/../../../python/utils/parse_trace.py --input trace.txt --mlir build/aie2_trace_lineBased_8b_${data_size}.mlir --output trace_${targetname}.json
-	${srcdir}/../../../python/utils/get_trace_summary.py --input trace_${targetname}.json
+	${srcdir}/../../../python/utils/trace/parse.py --input trace.txt --mlir build/aie2_trace_lineBased_8b_${data_size}.mlir --output trace_${targetname}.json
+	${srcdir}/../../../python/utils/trace/get_trace_summary.py --input trace_${targetname}.json
 
 clean_trace:
 	rm -rf tmpTrace trace.txt parse*json trace*json

programming_examples/basic/passthrough_kernel/run_strix_makefile.lit

@@ -9,4 +9,4 @@
 // RUN: make -f %S/Makefile devicename=npu2 
 // RUN: %run_on_npu2% make -f %S/Makefile run devicename=npu2
 // RUN: make -f %S/Makefile clean
-// RUN: env CHESS=false %run_on_npu2% make -f %S/Makefile trace devicename=npu2
+// RUN: env CHESS=false %run_on_npu2% make -f %S/Makefile trace devicename=npu2

programming_examples/basic/passthrough_kernel/test.py

@@ -7,13 +7,13 @@
 # (c) Copyright 2024 Advanced Micro Devices, Inc. or its affiliates
 import numpy as np
 import sys
-import aie.utils.xrt as xrt_utils
 import aie.utils.test as test_utils
+import aie.iron as iron
+from aie.utils import DefaultNPURuntime
 
 
 def main(opts):
     in1_size = int(opts.in1_size)  # in bytes
-    in2_size = int(opts.in2_size)  # in bytes
     out_size = int(opts.out_size)  # in bytes
 
     # --------------------------------------------------------------------------
@@ -36,28 +36,23 @@ def main(opts):
     assert out_size == in1_size
 
     # Initialize data
-    in1_data = np.arange(0, in1_volume, dtype=in1_dtype)
-    out_data = np.zeros([out_volume], dtype=out_dtype)
-
-    # Define reference data
-    ref = in1_data
+    ref = np.arange(0, in1_volume, dtype=in1_dtype)
+    in1 = iron.tensor(ref, dtype=in1_dtype)
+    out = iron.zeros([out_volume], dtype=out_dtype)
 
     # --------------------------------------------------------------------------
 
     print("Running...\n")
-    res = xrt_utils.setup_and_run_aie(
-        in1_dtype,
-        None,
-        out_dtype,
-        in1_data,
-        None,
-        out_data,
-        in1_volume,
-        None,
-        out_volume,
-        ref,
-        opts,
+    npu_opts = test_utils.create_npu_kernel(opts)
+    res = DefaultNPURuntime.run_test(
+        npu_opts.npu_kernel,
+        [in1, out],
+        {1: ref},
+        verify=npu_opts.verify,
+        verbosity=npu_opts.verbosity,
     )
+    if res == 0:
+        print("\nPASS!\n")
     sys.exit(res)
 
 

programming_examples/basic/passthrough_pykernel/passthrough_pykernel.ipynb

@@ -24,7 +24,9 @@
     "from aie.helpers.dialects.func import func\n",
     "from aie.iron.controlflow import range_\n",
     "\n",
-    "from aie.utils.xrt import setup_aie, execute as execute_on_aie\n",
+    "from aie.utils import NPUKernel, DefaultNPURuntime, get_current_device\n",
+    "import aie.iron as iron\n",
+    "from aie.iron.device import NPU2\n",
     "import aie.utils.test as test_utils"
    ]
   },
@@ -69,7 +71,11 @@
     "    N = vector_size\n",
     "    lineWidthInBytes = N // 4  # chop input in 4 sub-tensors\n",
     "\n",
-    "    @device(AIEDevice.npu1_1col)\n",
+    "    d = get_current_device()\n",
+    "    if d is None:\n",
+    "        raise RuntimeError(\"Could not determine current device\")\n",
+    "\n",
+    "    @device(d.resolve())\n",
     "    def device_body():\n",
     "        # define types\n",
     "        line_ty = np.ndarray[(lineWidthInBytes,), np.dtype[np.uint8]]\n",
@@ -106,20 +112,15 @@
     "\n",
     "        @runtime_sequence(vector_ty, vector_ty, vector_ty)\n",
     "        def sequence(inTensor, outTensor, notUsed):\n",
-    "            npu_dma_memcpy_nd(\n",
-    "                metadata=of_in,\n",
-    "                bd_id=0,\n",
-    "                mem=inTensor,\n",
-    "                sizes=[1, 1, 1, N],\n",
-    "                issue_token=True,\n",
+    "            in_task = shim_dma_single_bd_task(\n",
+    "                of_in, inTensor, sizes=[1, 1, 1, N], issue_token=True\n",
     "            )\n",
-    "            npu_dma_memcpy_nd(\n",
-    "                metadata=of_out,\n",
-    "                bd_id=1,\n",
-    "                mem=outTensor,\n",
-    "                sizes=[1, 1, 1, N],\n",
+    "            out_task = shim_dma_single_bd_task(\n",
+    "                of_out, outTensor, sizes=[1, 1, 1, N], issue_token=True\n",
     "            )\n",
-    "            dma_wait(of_in, of_out)"
+    "\n",
+    "            dma_start_task(in_task, out_task)\n",
+    "            dma_await_task(in_task, out_task)"
    ]
   },
   {
@@ -181,19 +182,22 @@
    "source": [
     "dtype = np.uint8\n",
     "\n",
-    "app = setup_aie(\n",
+    "npu_kernel = NPUKernel(\n",
     "    \"notebook_build/notebook.xclbin\",\n",
     "    \"notebook_build/notebook_insts.bin\",\n",
-    "    VECTOR_SIZE,\n",
-    "    dtype,\n",
-    "    None,\n",
-    "    None,\n",
-    "    VECTOR_SIZE,\n",
-    "    dtype,\n",
+    "    kernel_name=\"MLIR_AIE\",\n",
     ")\n",
+    "kernel_handle = DefaultNPURuntime.load(npu_kernel)\n",
+    "\n",
     "input = np.arange(1, VECTOR_SIZE + 1, dtype=dtype)\n",
+    "in_tensor = iron.tensor(input, dtype=dtype)\n",
+    "out_tensor = iron.zeros(VECTOR_SIZE, dtype=dtype)\n",
+    "\n",
+    "buffers = [in_tensor, out_tensor]\n",
+    "\n",
     "print(\"Running...\")\n",
-    "aie_output = execute_on_aie(app, input)\n",
+    "DefaultNPURuntime.run(kernel_handle, buffers)\n",
+    "aie_output = out_tensor.numpy()\n",
     "\n",
     "# Copy output results and verify they are correct\n",
     "errors = 0\n",

programming_examples/basic/passthrough_pykernel/test.py

@@ -7,8 +7,9 @@
 # (c) Copyright 2024 Advanced Micro Devices, Inc. or its affiliates
 import numpy as np
 import sys
-from aie.utils.xrt import setup_aie, execute
 import aie.utils.test as test_utils
+import aie.iron as iron
+from aie.utils import DefaultNPURuntime
 
 
 def main(opts):
@@ -17,34 +18,21 @@ def main(opts):
     data_size = int(opts.size)
     dtype = np.uint8
 
-    app = setup_aie(
-        opts.xclbin,
-        opts.instr,
-        data_size,
-        dtype,
-        None,
-        None,
-        data_size,
-        dtype,
+    input_data = np.arange(1, data_size + 1, dtype=dtype)
+    in1 = iron.tensor(input_data, dtype=dtype)
+    out = iron.zeros(data_size, dtype=dtype)
+
+    npu_opts = test_utils.create_npu_kernel(opts)
+    res = DefaultNPURuntime.run_test(
+        npu_opts.npu_kernel,
+        [in1, out],
+        {1: input_data},
+        verify=npu_opts.verify,
+        verbosity=npu_opts.verbosity,
     )
-    input = np.arange(1, data_size + 1, dtype=dtype)
-    aie_output = execute(app, input)
-
-    # Copy output results and verify they are correct
-    errors = 0
-    if opts.verify:
-        if opts.verbosity >= 1:
-            print("Verifying results ...")
-        e = np.equal(input, aie_output)
-        errors = np.size(e) - np.count_nonzero(e)
-
-    if not errors:
+    if res == 0:
         print("\nPASS!\n")
-        exit(0)
-    else:
-        print("\nError count: ", errors)
-        print("\nFailed.\n")
-        exit(-1)
+    sys.exit(res)
 
 
 if __name__ == "__main__":