zigzag-dse 3.7.2__tar.gz → 3.7.4__tar.gz

{zigzag_dse-3.7.2/zigzag_dse.egg-info → zigzag_dse-3.7.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: zigzag-dse
-Version: 3.7.2
+Version: 3.7.4
 Summary: ZigZag - Deep Learning Hardware Design Space Exploration
 Author-email: Arne Symons <arne.symons@kuleuven.be>, Linyan Mei <linyan.mei@kuleuven.be>
 License: MIT License

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/pyproject.toml

@@ -14,7 +14,7 @@ namespaces = true  # true by default
 
 [project]
 name = "zigzag-dse"
-version = "3.7.2"
+version = "3.7.4"
 description = "ZigZag - Deep Learning Hardware Design Space Exploration"
 readme = "README.md"
 authors = [{ name = "Arne Symons", email = "arne.symons@kuleuven.be" }, { name = "Linyan Mei", email = "linyan.mei@kuleuven.be" }]
@@ -50,7 +50,7 @@ Homepage = "https://github.com/KULeuven-MICAS/zigzag"
 realpython = "zigzag.__main__:main"
 
 [tool.bumpver]
-current_version = "3.7.2"
+current_version = "3.7.4"
 version_pattern = "MAJOR.MINOR.PATCH"
 commit_message = "bump version {old_version} -> {new_version}"
 commit = true

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_origin/test_ascend_like.py

@@ -9,8 +9,8 @@ workloads = (
 
 # Expected energy and latency for each workload defined above
 ens_lats = {
-    "zigzag/inputs/workload/resnet18.onnx": (1940385435.76, 3028219.0),
-    "zigzag/inputs/workload/resnet18.yaml": (2449578483.76, 3974885.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1940385435.76, 3032311.0),
+    "zigzag/inputs/workload/resnet18.yaml": (2449578483.76, 3980000.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_origin/test_edge_tpu_like.py

@@ -9,8 +9,8 @@ workloads = (
 
 # Expected energy and latency for each workload defined above
 ens_lats = {
-    "zigzag/inputs/workload/resnet18.onnx": (1906107856.91, 2868613.5),
-    "zigzag/inputs/workload/resnet18.yaml": (2413348670.67, 3641431.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1903229377.55, 2887804.5),
+    "zigzag/inputs/workload/resnet18.yaml": (2413348670.67, 3741293.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_origin/test_meta_prototype_like.py

@@ -9,8 +9,8 @@ workloads = (
 
 # Expected energy and latency for each workload defined above
 ens_lats = {
-    "zigzag/inputs/workload/resnet18.onnx": (1873660858.24, 2679987.16),
-    "zigzag/inputs/workload/resnet18.yaml": (2395695243.0, 3275453.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1863284164.16, 2700449.66),
+    "zigzag/inputs/workload/resnet18.yaml": (2373454257.72, 3316037.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_origin/test_tesla_npu_like.py

@@ -9,8 +9,8 @@ workloads = (
 
 # Expected energy and latency for each workload defined above
 ens_lats = {
-    "zigzag/inputs/workload/resnet18.onnx": (1844670026.55, 2706356.5),
-    "zigzag/inputs/workload/resnet18.yaml": (2352110509.45, 3274219.66),
+    "zigzag/inputs/workload/resnet18.onnx": (1859923747.33, 2719709.75),
+    "zigzag/inputs/workload/resnet18.yaml": (2363901208.14, 3313105.91),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_origin/test_tpu_like.py

@@ -9,8 +9,8 @@ workloads = (
 
 # Expected energy and latency for each workload defined above
 ens_lats = {
-    "zigzag/inputs/workload/resnet18.onnx": (1871436380.66, 3685888.5),
-    "zigzag/inputs/workload/resnet18.yaml": (2610625982.61, 5516489.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1871436380.66, 3703960.5),
+    "zigzag/inputs/workload/resnet18.yaml": (2615172831.38, 5584991.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_exploit_data_locality/test_ascend_like.py

@@ -11,8 +11,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (5649553476, 8605524),
     "zigzag/inputs/workload/mobilenetv2.onnx": (1880536921, 6472147),
-    "zigzag/inputs/workload/resnet18.onnx": (1813915548.48, 3015266.0),
-    "zigzag/inputs/workload/resnet18.yaml": (2338798423.04, 3958170.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1813915548.47, 3017312.0),
+    "zigzag/inputs/workload/resnet18.yaml": (2338798423.04, 3961239.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_exploit_data_locality/test_edge_tpu_like.py

@@ -11,8 +11,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (5568602396.684999, 8134431),
     "zigzag/inputs/workload/mobilenetv2.onnx": (735250234.1699998, 2417321.0),
-    "zigzag/inputs/workload/resnet18.onnx": (1783892066.11, 2951044.5),
-    "zigzag/inputs/workload/resnet18.yaml": (2115121959.8699996, 3620941.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1783892066.11, 2957953.5),
+    "zigzag/inputs/workload/resnet18.yaml": (2115121959.87, 3706985.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_exploit_data_locality/test_meta_prototype_like.py

@@ -11,8 +11,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (5679636177, 8279493),
     "zigzag/inputs/workload/mobilenetv2.onnx": (901087682, 2596085),
-    "zigzag/inputs/workload/resnet18.onnx": (1721342830.56, 2680789.5),
-    "zigzag/inputs/workload/resnet18.yaml": (2226399589.40, 3265450.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1710966136.48, 2689737.0),
+    "zigzag/inputs/workload/resnet18.yaml": (2204158604.12, 3295822.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_exploit_data_locality/test_tesla_npu_like.py

@@ -11,8 +11,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (6039993369, 8373281),
     "zigzag/inputs/workload/mobilenetv2.onnx": (930699349, 1955657),
-    "zigzag/inputs/workload/resnet18.onnx": (1709857690.01, 2701244.5),
-    "zigzag/inputs/workload/resnet18.yaml": (2201631649.06, 3262988.33),
+    "zigzag/inputs/workload/resnet18.onnx": (1725111410.78, 2709223.75),
+    "zigzag/inputs/workload/resnet18.yaml": (2213422347.76, 3296500.58),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_exploit_data_locality/test_tpu_like.py

@@ -11,8 +11,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (5475639039, 8963572),
     "zigzag/inputs/workload/mobilenetv2.onnx": (952681462, 21868059),
-    "zigzag/inputs/workload/resnet18.onnx": (1735241704.2240005, 3756282.50),
-    "zigzag/inputs/workload/resnet18.yaml": (2298383687.30, 5493246.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1735241704.22, 3762552.5),
+    "zigzag/inputs/workload/resnet18.yaml": (2302362179.97, 5548138.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_mix_spatial_mapping/test_ascend_like.py

@@ -13,8 +13,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (5664863934, 8479644),
     "zigzag/inputs/workload/mobilenetv2.onnx": (920740293, 3814440),
-    "zigzag/inputs/workload/resnet18.onnx": (1811468387.52, 3117931.33),
-    "zigzag/inputs/workload/resnet18.yaml": (2349013378.88, 3714899.66),
+    "zigzag/inputs/workload/resnet18.onnx": (1811468387.52, 3119977.33),
+    "zigzag/inputs/workload/resnet18.yaml": (2349013378.88, 3718548.66),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_mix_spatial_mapping/test_edge_tpu_like.py

@@ -13,8 +13,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (6159980160, 8337351),
     "zigzag/inputs/workload/mobilenetv2.onnx": (742114179, 2421959),
-    "zigzag/inputs/workload/resnet18.onnx": (1735517944, 3834226.0),
-    "zigzag/inputs/workload/resnet18.yaml": (2029477205, 4509461.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1735517943.87, 3834226.0),
+    "zigzag/inputs/workload/resnet18.yaml": (2029477205.31, 4586293.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_mix_spatial_mapping/test_meta_prototype_like.py

@@ -13,8 +13,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (5681909132.480001, 8279495.0),
     "zigzag/inputs/workload/mobilenetv2.onnx": (909647916.68, 2602479.0),
-    "zigzag/inputs/workload/resnet18.onnx": (1746339277.28, 2685566.66),
-    "zigzag/inputs/workload/resnet18.yaml": (2250370155.48, 3247268.33),
+    "zigzag/inputs/workload/resnet18.onnx": (1746642408.16, 2688913.66),
+    "zigzag/inputs/workload/resnet18.yaml": (2249214190.04, 3278925.33),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_mix_spatial_mapping/test_tesla_npu_like.py

@@ -13,8 +13,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (6006136982.778, 8290892.0),
     "zigzag/inputs/workload/mobilenetv2.onnx": (947736166.5380002, 1857838.0),
-    "zigzag/inputs/workload/resnet18.onnx": (1611048997.07, 2295226.5),
-    "zigzag/inputs/workload/resnet18.yaml": (2101519570.14, 2841397.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1599110638.5, 2318166.5),
+    "zigzag/inputs/workload/resnet18.yaml": (2084306723.24, 2889873.0),
 }
 
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/tests/main/test_with_mix_spatial_mapping/test_tpu_like.py

@@ -13,8 +13,8 @@ workloads = (
 ens_lats = {
     "zigzag/inputs/workload/alexnet.onnx": (5589583403, 8671208),
     "zigzag/inputs/workload/mobilenetv2.onnx": (932571650, 7304307),
-    "zigzag/inputs/workload/resnet18.onnx": (1759424218.3160002, 4250030.0),
-    "zigzag/inputs/workload/resnet18.yaml": (2191903732.78, 4601363.0),
+    "zigzag/inputs/workload/resnet18.onnx": (1759424218.32, 4256300.0),
+    "zigzag/inputs/workload/resnet18.yaml": (2191903732.78, 4657608.0),
 }
 
 

zigzag_dse-3.7.4/zigzag/__init__.py

@@ -0,0 +1 @@
+__version__ = "3.7.4"

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/zigzag/cost_model/cost_model.py

@@ -10,10 +10,10 @@ import numpy as np
 from zigzag.cost_model.port_activity import PortActivity, PortBeginOrEndActivity
 from zigzag.datatypes import ArrayType, Constants, LayerOperand, MemoryOperand
 from zigzag.hardware.architecture.accelerator import Accelerator
-from zigzag.hardware.architecture.memory_instance import MemoryInstance
+from zigzag.hardware.architecture.memory_level import MemoryLevel
 from zigzag.hardware.architecture.memory_port import MemoryPort
 from zigzag.hardware.architecture.operational_array import OperationalArray
-from zigzag.mapping.data_movement import AccessEnergy, DataDirection, MemoryAccesses
+from zigzag.mapping.data_movement import AccessEnergy, DataDirection, FourWayDataMoving, MemoryAccesses
 from zigzag.mapping.mapping import Mapping
 from zigzag.mapping.spatial_mapping_internal import SpatialMappingInternal
 from zigzag.mapping.temporal_mapping import TemporalMapping
@@ -1141,24 +1141,37 @@ class CostModelEvaluation(CostModelEvaluationABC):
         self.mac_utilization1 = mac_utilization1
         self.mac_utilization2 = mac_utilization2
 
-    def get_total_inst_bandwidth(self, memory_instance: MemoryInstance) -> MemoryAccesses:
-        """Given a cost model evaluation and a memory instance, compute the memory's total instantaneous bandwidth
-        required throughout the execution of the layer that corresponds to this CME. Returns empty bandwidth
-        requirements if the given memory instance is not included in this CME's memory hierarchy.
+    def get_total_inst_bandwidth(self, memory_level: MemoryLevel) -> FourWayDataMoving[int]:
+        """Given a cost model evaluation and a memory level, compute the memory's total instantaneous bandwidth
+        required throughout the execution of the layer that corresponds to this CME. Raises an assertion error
+        if the given memory level is not included in this CME's memory hierarchy.
         NOTE: this function is used in Stream
         """
-        # Check which operands require offchip memory throughout the computation
-        offchip_mem_operands: list[MemoryOperand] = []
-        for op, memory_levels in self.mem_hierarchy_dict.items():
-            last_mem_level = memory_levels[-1]
-            if last_mem_level.memory_instance == memory_instance:
-                offchip_mem_operands.append(op)
-        # Obtain the required instantaneous bandwidth to/from offchip for these operands
-        total_inst_bw = MemoryAccesses(0, 0, 0, 0)
-        for mem_op in offchip_mem_operands:
+        # Check that the given MemoryInstance is part of the memory hierarchy of this CME
+        assert (
+            memory_level in self.mem_level_list
+        ), f"Memory level {memory_level} is not part of the memory hierarchy of this cost model evaluation."
+        # Obtain the required instantaneous bandwidth to/from the memory for its operands during computation
+        total_inst_bw = FourWayDataMoving(0, 0, 0, 0)
+        for mem_op in memory_level.operands:
             layer_op = self.layer.memory_operand_links.mem_to_layer_op(mem_op)
-            inst_bw_4way = self.mapping.unit_mem_data_movement[layer_op][-1].req_mem_bw_inst
+            req_bw_4way = self.mapping.unit_mem_data_movement[layer_op][-1].req_mem_bw_inst
+            inst_bw_4way = FourWayDataMoving(
+                req_bw_4way.rd_out_to_low,
+                req_bw_4way.wr_in_by_low,
+                req_bw_4way.rd_out_to_high,
+                req_bw_4way.wr_in_by_high,
+            )
             total_inst_bw += inst_bw_4way
+        # Add the potential borrowed bandwidth by the loading phases (onloading and offloading)
+        # As this bandwidth is already aggregated across all directions for a single port,
+        # we put it in the WR_IN_BY_HIGH direction if the output memory operand is present in the memory level.
+        # Otherwise, we put it in the RD_OUT_TO_LOW direction.
+        for port in memory_level.port_list:
+            if Constants.OUTPUT_MEM_OP in memory_level.operands:
+                total_inst_bw.wr_in_by_high += self.total_loading_bw_during_computation[port]
+            else:
+                total_inst_bw.rd_out_to_low += self.total_loading_bw_during_computation[port]
         return total_inst_bw
 
     def __str__(self):

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/zigzag/mapping/data_movement.py

@@ -84,6 +84,9 @@ class FourWayDataMoving(Generic[T], metaclass=ABCMeta):
             self.wr_in_by_high * other,
         )
 
+    def __add__(self, other: "FourWayDataMoving[T]"):
+        return self._add_with_type(other, type(self))
+
     def __repr__(self):
         return (
             f"4waydatamoving (rd ^: {self.rd_out_to_high}, wr v: {self.wr_in_by_high}, "
@@ -124,14 +127,14 @@ class DataMovePattern:
 
     def __init__(self, operand: LayerOperand, mem_level: int):
         self.name = operand.name + str(mem_level)
-        self.data_elem_move_count = MemoryAccesses(0, 0, 0, 0)
-        self.data_precision = MemoryAccesses(0, 0, 0, 0)
-        self.req_mem_bw_aver = MemoryAccesses(0, 0, 0, 0)
-        self.req_mem_bw_inst = MemoryAccesses(0, 0, 0, 0)
-        self.data_trans_period = MemoryAccesses(0, 0, 0, 0)
-        self.data_trans_period_count = MemoryAccesses(0, 0, 0, 0)
-        self.data_trans_amount_per_period = MemoryAccesses(0, 0, 0, 0)
-        self.inst_data_trans_window = MemoryAccesses(0, 0, 0, 0)
+        self.data_elem_move_count = MemoryAccesses(0, 0, 0, 0)  # elements
+        self.data_precision = MemoryAccesses(0, 0, 0, 0)  # bits/element
+        self.req_mem_bw_aver = MemoryAccesses(0, 0, 0, 0)  # bits/cycle
+        self.req_mem_bw_inst = MemoryAccesses(0, 0, 0, 0)  # bits/cycle
+        self.data_trans_period = MemoryAccesses(0, 0, 0, 0)  # cycles
+        self.data_trans_period_count = MemoryAccesses(0, 0, 0, 0)  # periods
+        self.data_trans_amount_per_period = MemoryAccesses(0, 0, 0, 0)  # elements
+        self.inst_data_trans_window = MemoryAccesses(0, 0, 0, 0)  # cycles
 
     def set_data_elem_move_count(
         self,

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/zigzag/mapping/mapping.py

@@ -481,9 +481,10 @@ class Mapping:
         # e.g. for the DTL of Global Buffer (Weight) talking to spatially unrolled Weight Reg File,
         # each Weight Reg File's required write BW is indicated by "_L",
         # while Global Buffer (Weight)'s required read BW is indicate by "_H"
+        # The unit of these variables is bits per cycle
         req_mem_bw_low_raw = {
             op: [
-                data_each_level_unrolled[op][lv] / cycle_each_level[op][lv]
+                data_each_level_unrolled[op][lv] * self.data_precision_dict[op][lv] / cycle_each_level[op][lv]
                 for lv in range(self.spatial_mapping.arch_level[op])
             ]
             for op in self.operand_list

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/zigzag/visualization/results/plot_cme.py

@@ -31,10 +31,11 @@ GROUP_SPACING = 1
 
 
 def shorten_onnx_layer_name(name: str):
-    """Names generated in the ONNX format are quite long (e.g. `layer1/layer1.0/conv2/Conv). This function extracts
-    the most informative part"""
+    """Names generated in the ONNX format are quite long (e.g. `layer1/layer1.0/conv2/Conv`). This function extracts
+    the most informative part (`conv2/Conv` in this case)"""
     try:
-        return name.split("/")[-2]
+        last_two = name.split("/")[-2:]
+        return "/".join(last_two)
     except IndexError:
         return name
 

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4/zigzag_dse.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: zigzag-dse
-Version: 3.7.2
+Version: 3.7.4
 Summary: ZigZag - Deep Learning Hardware Design Space Exploration
 Author-email: Arne Symons <arne.symons@kuleuven.be>, Linyan Mei <linyan.mei@kuleuven.be>
 License: MIT License

{zigzag_dse-3.7.2 → zigzag_dse-3.7.4}/zigzag_dse.egg-info/top_level.txt

@@ -1,5 +1,6 @@
 build
 dist
 lab1
+rooflines
 tests
 zigzag

zigzag_dse-3.7.2/zigzag/__init__.py

@@ -1 +0,0 @@
-__version__ = "3.7.2"