commit b5df773e92051004158046b0ed2c7b802198de6e
author Tim Hall <tim.hall@arm.com> Wed May 04 16:20:43 2022 +0100
committer Tim Hall <tim.hall@arm.com> Wed May 04 16:26:09 2022 +0100
tree 7d738a28a63b66a20f379acbdfc6c3e7c4a98a61
parent 95b07c1c0fed6a985607131e59a593786d40b389

Revert "MLBEDSW-6263: Use separate tensors for double buffering"

This reverts commit cc5f4de1c35ba44fca7ff6295c6ae846f8242344.

Signed-off-by: Tim Hall <tim.hall@arm.com>
Change-Id: I0fa5babfe9ad9ec668720d04fe1c16d9a9092131

ethosu/vela/high_level_command_to_npu_op.py

diff --git a/ethosu/vela/high_level_command_to_npu_op.py b/ethosu/vela/high_level_command_to_npu_op.py
index e6bfc1c..3a78d6f 100644
--- a/ethosu/vela/high_level_command_to_npu_op.py
+++ b/ethosu/vela/high_level_command_to_npu_op.py
@@ -68,6 +68,7 @@
 from .tensor import Tensor
 from .tensor import TensorFormat
 from .tensor import TensorPurpose
+from .tensor import TensorSubPurpose
 from .weight_compressor import NpuWeightTensor
 from .weight_compressor import WeightKey
 
@@ -201,15 +202,9 @@
     return mem_limits
 
 
-def get_upscale(op: Operation) -> NpuResamplingMode:
-    upscale = NpuResamplingMode.NONE
-    if op.type == Op.ResizeBilinear:
-        # perform nearest neighbor upscale
-        upscale = NpuResamplingMode.NEAREST
-    elif op.type == Op.Conv2DBackpropInputSwitchedBias:
-        # perform insert zero upscale
-        upscale = NpuResamplingMode.TRANSPOSE
-    return upscale
+def get_double_buffer_offset(arch: ArchitectureFeatures, range_index: int, core: int) -> int:
+    """Returns 0 if the first half of a double buffer should be used, 1 if the second half should be used"""
+    return ((range_index - core) // arch.ncores) % 2
 
 
 def get_ifm_depth(npu_block_type: NpuBlockType, ifm_box: Box, ofm_box: Box) -> int:
@@ -319,13 +314,20 @@
         key = WeightKey(core, weight_box.start_coord[-1])
         if key in w_tensor_src.encoded_ranges:
             weight_range = w_tensor_src.encoded_ranges[key]
-            if weight_tensor == w_tensor_src:
-                # Straight from source tensor
-                address = weight_tensor.address + weight_range.offset
-            else:
-                # Weight buffered tensor
+            if weight_tensor.sub_purpose == TensorSubPurpose.DoubleBuffer:
+                assert weight_tensor != w_tensor_src
+                # Double buffered inside weight_tensor
                 address = weight_tensor.address + core_offset
+                address += get_double_buffer_offset(arch, weight_range.index, core) * w_tensor_src.max_range_bytes
                 core_offset += round_up(weight_range.total_bytes, 16)
+            else:
+                if weight_tensor == w_tensor_src:
+                    # Straight from source tensor
+                    address = weight_tensor.address + weight_range.offset
+                else:
+                    # Single buffered inside weight tensor
+                    address = weight_tensor.address + core_offset
+                    core_offset += round_up(weight_range.total_bytes, 16)
 
             # Location of weights in tensor
             addr_range = NpuAddressRange(
@@ -524,7 +526,13 @@
                 if core == 0:
                     weight_range = cmd.in_tensor.encoded_ranges[key]
                     src_addr = cmd.in_tensor.address + weight_range.offset
-                    dest_addr = cmd.out_tensor.address
+
+                    if cmd.out_tensor.sub_purpose == TensorSubPurpose.DoubleBuffer:
+                        dest_addr = cmd.out_tensor.address + cmd.in_tensor.max_range_bytes * (
+                            get_double_buffer_offset(arch, weight_range.index, core)
+                        )
+                    else:
+                        dest_addr = cmd.out_tensor.address
     else:
         start_coord = cmd.box.start_coord
         src_addr = cmd.in_tensor.address_for_coordinate(start_coord)