Update trace scripts and vector_scalar_mul to use new scripts (#1853)

programming_examples/basic/vector_scalar_mul/aie2.py

@@ -51,9 +51,10 @@ def device_body():
         )
         of_out = object_fifo("out", ComputeTile2, ShimTile, buffer_depth, tile_ty)
 
-        # Set up a circuit-switched flow from core to shim for tracing information
+        # Set up a packet-switched flow from core to shim for tracing information
+        tiles_to_trace = [ComputeTile2]
         if trace_size > 0:
-            flow(ComputeTile2, WireBundle.Trace, 0, ShimTile, WireBundle.DMA, 1)
+            trace_utils.configure_packet_tracing_flow(tiles_to_trace, ShimTile)
 
         # Set up compute tiles
 
@@ -77,13 +78,10 @@ def core_body():
         def sequence(A, F, C):
 
             if trace_size > 0:
-                trace_utils.configure_simple_tracing_aie2(
-                    ComputeTile2,
-                    ShimTile,
-                    ddr_id=2,
-                    size=trace_size,
-                    offset=N_in_bytes,
+                trace_utils.configure_packet_tracing_aie2(
+                    tiles_to_trace, ShimTile, trace_size, N_in_bytes
                 )
+
             npu_dma_memcpy_nd(
                 metadata=of_in, bd_id=1, mem=A, sizes=[1, 1, 1, N], issue_token=True
             )

programming_examples/basic/vector_scalar_mul/test.py

@@ -7,7 +7,8 @@
 # (c) Copyright 2024 Advanced Micro Devices, Inc. or its affiliates
 import numpy as np
 import sys
-from aie.utils.xrt import setup_aie, execute as execute_on_aie
+import time
+from aie.utils.xrt import setup_aie, write_out_trace, execute
 import aie.utils.test as test_utils
 
 
@@ -18,6 +19,10 @@ def main(opts):
     vector_dtype = np.int16
     scalar_dtype = np.int32
     scale_factor = 3
+    size_out = data_size * 2
+    print("output buffer size: " + str(size_out))
+
+    enable_trace = opts.trace_size > 0
 
     app = setup_aie(
         opts.xclbin,
@@ -28,12 +33,34 @@ def main(opts):
         scalar_dtype,
         data_size,
         vector_dtype,
+        enable_trace=enable_trace,
+        trace_size=opts.trace_size,
     )
     input_vector = np.arange(1, data_size + 1, dtype=vector_dtype)
     input_factor = np.array([3], dtype=scalar_dtype)
-    aie_output = execute_on_aie(app, input_vector, input_factor)
+    # aie_output = execute_on_aie(app, input_vector, input_factor)
+
+    start = time.time_ns()
+    full_output = execute(app, input_vector, input_factor)
+    stop = time.time_ns()
+    npu_time = stop - start
+    print("npu_time: ", npu_time)
+
+    # aie_output = full_output[:size_out].view(np.int8)
+    # aie_output = full_output[:size_out].view(np.uint8)
+    aie_output = full_output[:size_out].view(np.int16)
+    if enable_trace:
+        trace_buffer = full_output[size_out:].view(np.uint32)
+
     ref = np.arange(1, data_size + 1, dtype=vector_dtype) * scale_factor
 
+    if enable_trace:
+        # trace_buffer = full_output[3920:]
+        print("trace_buffer shape: ", trace_buffer.shape)
+        print("trace_buffer dtype: ", trace_buffer.dtype)
+        # write_out_trace(trace_buffer, str(opts.trace_file))
+        write_out_trace(trace_buffer, "trace.txt")
+
     # Copy output results and verify they are correct
     errors = 0
     if opts.verify:

programming_examples/utils/parse_trace.py

@@ -3,6 +3,7 @@
 import argparse
 import sys
 import re
+
 from aie.utils.trace_events_enum import CoreEvent, MemEvent, PLEvent, MemTileEvent
 
 # Number of different trace types, currently 4
@@ -13,7 +14,8 @@
 NumTraceTypes = 4
 NUM_EVENTS = 8  # number of events we can view per trace
 
-DEBUG = False
+# DEBUG = False
+# DEBUG = True
 
 
 def parse_args():
@@ -23,6 +25,7 @@ def parse_args():
     parser.add_argument(
         "--colshift", help="column shift adjustment to source mlir", required=False
     )
+    parser.add_argument("--debug", help="debug mode", required=False)
     # TODO tracelabels removed since we can have multiple sets of labels for each pkt_type & loc combination
     # parser.add_argument('--tracelabels',
     #         nargs='+',
@@ -610,7 +613,7 @@ def parse_mlir_trace_events(lines):
                     row = int(result.group(3 * i2 + 3))
                 elif var == "column":
                     col = int(result.group(3 * i2 + 3)) + colshift
-                    col = 1 if col == 0 else col
+                    # col = 1 if col == 0 else col
                 elif var == "value":
                     if result.group(3 * i2 + 2) == "0x":
                         value = int(result.group(3 * i2 + 3), 16)
@@ -757,6 +760,10 @@ def lookup_event_name_by_type(trace_type, code):
         events_enum = CoreEvent
     elif trace_type == 1:  # Mem traces
         events_enum = MemEvent
+    elif trace_type == 2:  # Shim traces
+        events_enum = PLEvent
+    elif trace_type == 3:  # MemTile traces
+        events_enum = MemTileEvent
     if events_enum is not None and code in set(x.value for x in events_enum):
         event = events_enum(code).name
     else:
@@ -871,6 +878,10 @@ def setup_trace_metadata(trace_events, pid_events):
 
 opts = parse_args()
 
+DEBUG = opts.debug
+if DEBUG:
+    print("Debug mode enable\n")
+
 # set colshift based on optional argument
 colshift = int(opts.colshift) if opts.colshift else 0