commit eeb85154b00a9864d0d63e382e9c80ca8e294d5d
author patrik.gustavsson <patrik.gustavsson@arm.com> Mon Dec 21 17:10:40 2020 +0000
committer patrik.gustavsson <patrik.gustavsson@arm.com> Mon Dec 21 17:10:40 2020 +0000
tree 85ddd620916565aa8565d072b764ca4918b405a1
parent df0a5905177f3a1b836076bc3f9f39b2e86f1794

Revert "Revert "MLBEDSW-3645 4D class for op ifm/ofm shapes""

This reverts commit df0a5905177f3a1b836076bc3f9f39b2e86f1794.

Reason for revert: <INSERT REASONING HERE>

Change-Id: I891c66fb29db9d25e942947e8d1c29a10610de51

ethosu/vela/shared_buffer_allocation.py

diff --git a/ethosu/vela/shared_buffer_allocation.py b/ethosu/vela/shared_buffer_allocation.py
index 1f027d6..d8faf36 100644
--- a/ethosu/vela/shared_buffer_allocation.py
+++ b/ethosu/vela/shared_buffer_allocation.py
@@ -32,6 +32,7 @@
 from .operation import NpuBlockType
 from .range_set import MemoryRangeSet
 from .register_command_stream_util import to_kernel
+from .shape4d import Shape4D
 from .tensor import MemArea
 
 
@@ -195,14 +196,14 @@
         ifm_bits = ifm_tensor.dtype.size_in_bits()
         ifm_shape = ps.primary_op.ifm_shapes[0]
 
-        if ifm_shape != []:
-            ifm_depth = ifm_shape[-1]
+        if ifm_tensor.shape != []:
+            ifm_depth = ifm_shape.depth
 
         if is_elementwise:
             ifm_count = 2
             if ifm_tensor.shape == []:  # Scalar in ifm1
                 assert ifm2_tensor
-                ifm_depth = ps.primary_op.ifm_shapes[1][-1]
+                ifm_depth = ps.primary_op.ifm_shapes[1].depth
                 ifm_count = 1
             elif not ifm2_tensor or ifm2_tensor.shape == []:  # Scalar in ifm2
                 ifm_count = 1
@@ -251,7 +252,7 @@
         ifm_bits=ifm_bits,
         ifm_depth=ifm_depth,
         ifm_count=ifm_count,
-        ofm_shape=ofm_shape,
+        ofm_shape=Shape4D(ofm_shape),
     )
 
 
@@ -265,14 +266,9 @@
 
     # Constrain the search space if the OFM is smaller than the max block size
     # - Add other block search constraints here if required
-    if len(alloc.ofm_shape) <= 2:
-        max_block_height = max_block_width = alloc.ofm_shape[0]
-    else:
-        max_block_width = alloc.ofm_shape[-2]
-        max_block_height = alloc.ofm_shape[-3]
-
-    # Common block depth
-    max_block_depth = alloc.ofm_shape[-1]
+    max_block_width = alloc.ofm_shape.width
+    max_block_height = alloc.ofm_shape.height
+    max_block_depth = alloc.ofm_shape.depth
 
     # Constrain to valid ranges before search
     max_block_width = min(arch.ofm_block_max.width, max_block_width)