MLBEDSW-6263: Use separate tensors for double buffering

Uses separate tensors for the individual weight buffers
in case of weight double buffering.

Each weight buffer tensor gets its own individual live range.

This patch is a clone of a previously reverted patch, but with some
additional bug fixes applied.

Signed-off-by: Rickard Bolin <rickard.bolin@arm.com>
Change-Id: I868c70d15821eb9f1399186f2da6e7345f6ee343

1 files changed, 45 insertions, 25 deletions

diff --git a/ethosu/vela/scheduler.py b/ethosu/vela/scheduler.py
index 3cfde28a..67b890e1 100644
--- a/ethosu/vela/scheduler.py
+++ b/ethosu/vela/scheduler.py
@@ -107,7 +107,7 @@ class SchedulerOpInfo:
         self.ofm_depth_slices: List[int] = [0, stripe.depth]
         self.npu_weights_tensor: Optional[NpuWeightTensor] = None
         self.npu_scales_tensor: Optional[NpuWeightTensor] = None
-        self.buffered_weight_tensor: Optional[Tensor] = None
+        self.buffered_weight_tensors: List[Tensor] = []
         self.cycles: Optional[CycleCost] = None
         self.slack_buffering_cycles = 0
         self.slack_buffering_memory = 0
@@ -131,9 +131,8 @@ class SchedulerOpInfo:
         res += f"\t\tIFM2 Stripe  = {self.stripe_input2}\n"
         res += f"\t\tOFM Stripe   = {self.stripe}\n"
         res += f"\t\tEncoded Weights = {self.npu_weights_tensor and len(self.npu_weights_tensor.buffer)} bytes\n"
-        res += (
-            f"\t\tWeight buffer = {self.buffered_weight_tensor and self.buffered_weight_tensor.storage_size()} bytes\n"
-        )
+        for idx, tens in enumerate(self.buffered_weight_tensors):
+            res += f"\t\tWeight buffer{idx + 1} = {tens.storage_size()} bytes\n"
         res += f"\t\tDepth slices = {self.ofm_depth_slices}\n"
         res += f"\t\tAssigned Cascade = {self.cascade}"
         return res
@@ -734,7 +733,7 @@ class Scheduler:
                     # Chosen buffering might not fit at all, iterate until it does
                     # or until the minimum usable slice size is reached
                     if (
-                        encoded_weights.max_range_bytes <= half_buffer_limit
+                        encoded_weights.double_buffer_size() <= buffer_limit_bytes
                         or prebuffer_depth == ArchitectureFeatures.OFMSplitDepth
                     ):
                         break
@@ -751,24 +750,42 @@ class Scheduler:
                 cost.slack_buffering_cycles = tail_cycles.op_cycles
 
         # Determine whether the weights need to be double buffered
-        weight_buffer_size = min(len(encoded_weights.buffer), encoded_weights.max_range_bytes)
+        weight_buffer_size = min(len(encoded_weights.buffer), encoded_weights.max_range_bytes())
 
         # Only buffer weights if there's still space left for the buffer
         if weight_buffer_size <= buffer_limit_bytes:
             assert weight_buffer_size % 16 == 0
             # Determine whether to double buffer or single buffer
-            if (weight_buffer_size * 2 <= buffer_limit_bytes) and (weight_buffer_size < len(encoded_weights.buffer)):
-                weight_buffer_size = weight_buffer_size * 2
+            double_buffer_size = encoded_weights.double_buffer_size()
+            if (double_buffer_size <= buffer_limit_bytes) and (weight_buffer_size < len(encoded_weights.buffer)):
                 weight_tensor_purpose = TensorSubPurpose.DoubleBuffer
             else:
                 weight_tensor_purpose = TensorSubPurpose.Standard
 
-            cost.buffered_weight_tensor = self.buffer_tensor(
-                encoded_weights, weight_tensor_purpose, weight_buffer_size, weight_tensor.name
-            )
+            cost.buffered_weight_tensors = [
+                self.buffer_tensor(
+                    encoded_weights,
+                    weight_tensor_purpose,
+                    encoded_weights.double_buffer_sizes[0],
+                    weight_tensor.name + "_buffer",
+                )
+            ]
+            if weight_tensor_purpose == TensorSubPurpose.DoubleBuffer:
+                buf2 = self.buffer_tensor(
+                    encoded_weights,
+                    weight_tensor_purpose,
+                    encoded_weights.double_buffer_sizes[1],
+                    weight_tensor.name + "_buffer2",
+                )
+                cost.buffered_weight_tensors.append(buf2)
+
+            last_used_buffer_idx = len(cost.ofm_depth_slices) % len(cost.buffered_weight_tensors)
+            weight_buffer_size = encoded_weights.double_buffer_sizes[last_used_buffer_idx]
+
             if ref_cost.cascade == 0:
-                # Determine if the lifetime can be extended and pre-buffer weights under the previous operation
-                cost.buffered_weight_tensor.pre_buffer = weight_buffer_size < slack_memory
+                # Determine if the lifetime can be extended and pre-buffer the first weight buffer
+                # under the previous operation
+                cost.buffered_weight_tensors[0].pre_buffer = encoded_weights.double_buffer_size() < slack_memory
 
             cost.slack_buffering_memory -= weight_buffer_size
         else:
@@ -781,7 +798,7 @@ class Scheduler:
         cost.npu_scales_tensor = encoded_scales
 
     def buffer_tensor(self, src_tensor: Tensor, sub_purpose: TensorSubPurpose, buffer_size: int, name: str) -> Tensor:
-        buffered_weight_tensor = Tensor([1, 1, 1, buffer_size], DataType.uint8, name + "_buffer")
+        buffered_weight_tensor = Tensor([1, 1, 1, buffer_size], DataType.uint8, name)
         buffered_weight_tensor.src_tensor = src_tensor
         buffered_weight_tensor.mem_area = self.arch.fast_storage_mem_area
         buffered_weight_tensor.mem_type = MemType.Scratch_fast
@@ -823,11 +840,16 @@ class Scheduler:
             # Create a cost entry with the new stripe
             cost = sched_op.create_scheduler_info(self.nng, stripe)
 
-            if ref_cost[sched_op].buffered_weight_tensor:
+            weight_tensor = cost.npu_weights_tensor
+            for idx, buffered_tens in enumerate(ref_cost[sched_op].buffered_weight_tensors):
                 # If the weights are buffered in the reference schedule they should be in the new proposal
-                weight_tensor = cost.npu_weights_tensor
-                cost.buffered_weight_tensor = self.buffer_tensor(
-                    weight_tensor, TensorSubPurpose.Standard, len(weight_tensor.buffer), weight_tensor.name
+                cost.buffered_weight_tensors.append(
+                    self.buffer_tensor(
+                        weight_tensor,
+                        buffered_tens.sub_purpose,
+                        weight_tensor.double_buffer_sizes[idx],
+                        buffered_tens.name,
+                    )
                 )
 
             # Estimate performance
@@ -856,9 +878,7 @@ class Scheduler:
                 peak_mem_usage = max(cascade_info.mem_usage, peak_mem_usage)
             else:
                 # This Op is not part of a cascade - calculate the memory usage
-                op_weight_buffer = 0
-                if cost[sched_op].buffered_weight_tensor:
-                    op_weight_buffer = cost[sched_op].buffered_weight_tensor.storage_size()
+                op_weight_buffer = sum(tens.storage_size() for tens in cost[sched_op].buffered_weight_tensors)
 
                 op_mem_usage = (
                     sched_op.ifm_size_in_bytes()
@@ -1013,8 +1033,8 @@ class Scheduler:
         weight_ops = {}
         for sched_op in self.sched_ops:
             cost = self.sg.schedule.cost_map[sched_op]
-            if cost.buffered_weight_tensor:
-                weight_ops[cost.buffered_weight_tensor] = sched_op
+            for tens in cost.buffered_weight_tensors:
+                weight_ops[tens] = sched_op
 
         # Filter out weight buffer live ranges
         weight_lrs = []
@@ -1088,8 +1108,8 @@ class Scheduler:
             sched_op.parent_ps.block_config = op_info.block_config.old_style_representation()
 
             # Ensure that the src_tensor reference is set correctly
-            if op_info.buffered_weight_tensor:
-                op_info.buffered_weight_tensor.src_tensor = op_info.npu_weights_tensor
+            for tens in op_info.buffered_weight_tensors:
+                tens.src_tensor = op_info.npu_weights_tensor
 
     def use_fast_storage_for_feature_maps(self, schedule, staging_limit):
         max_mem_usage = []