Fix MV2 runtime pool sizes and skip Corstone-300 in CI

Corstone-320: redirect zephyr,sram to the 4 MiB ISRAM so the ~3 MiB of
allocator pools (method + Ethos-U scratch) fit alongside code.  Bump pool
sizes to the 1.5 MiB defaults — verified locally that the build links at
FLASH 11% / RAM 78% / MODEL_DDR 23%.

Corstone-300: 2 MiB ISRAM cannot hold the ~2.9 MiB of runtime pools MV2
needs regardless of pool tuning.  Document this in the board conf and skip
the MV2 ethos-u55 matrix entry in trunk CI.  The hello-executorch sample
already validates the Corstone-300 + Ethos-U55 pipeline end-to-end.

Co-authored-by: Claude <noreply@anthropic.com>

Author: psiddh

.github/workflows/trunk.yml

@@ -191,6 +191,11 @@ jobs:
             continue
           fi
 
+          if [[ ${TARGET} == "ethos-u55" ]]; then
+            echo "---- Skipping MV2 for ${TARGET} (Corstone-300 ISRAM too small for MV2 runtime pools) ----"
+            continue
+          fi
+
           echo "---- MV2 ${TARGET} ----"
           rm -Rf build
 

zephyr/samples/mv2-ethosu/boards/mps3_corstone300_fvp.conf

@@ -12,7 +12,10 @@ CONFIG_ETHOS_U=y
 # DDR on the FVP so no SRAM copy is needed.
 CONFIG_ET_ARM_MODEL_PTE_DMA_ACCESSIBLE=y
 
-# Corstone-300 has 2 MiB ISRAM. Reduce pool sizes to fit within budget
-# alongside stack, heap, and runtime buffers.
+# Corstone-300 has 2 MiB ISRAM.  With the model in DDR the method pool
+# needs ~1.4 MiB (752 KiB planned buffer + 602 KiB input tensor).
+# The remaining ISRAM is not enough for the Ethos-U scratch (~1.5 MiB),
+# so MV2 inference will fail at runtime on this board; Corstone-320 is
+# the supported target for MV2.  These sizes keep the link step green.
 CONFIG_EXECUTORCH_METHOD_ALLOCATOR_POOL_SIZE=786432
 CONFIG_EXECUTORCH_TEMP_ALLOCATOR_POOL_SIZE=786432

zephyr/samples/mv2-ethosu/boards/mps4_corstone320_fvp.conf

@@ -12,7 +12,7 @@ CONFIG_ETHOS_U=y
 # DDR on the FVP so no SRAM copy is needed.
 CONFIG_ET_ARM_MODEL_PTE_DMA_ACCESSIBLE=y
 
-# Corstone-320 has 4 MiB SRAM. Reduce pool sizes from the 1.5 MiB defaults
-# to leave headroom for stack, heap, and runtime buffers.
-CONFIG_EXECUTORCH_METHOD_ALLOCATOR_POOL_SIZE=786432
-CONFIG_EXECUTORCH_TEMP_ALLOCATOR_POOL_SIZE=786432
+# With zephyr,sram redirected to the 4 MiB ISRAM (shared with code),
+# the default 1.5 MiB pools fit alongside ~460 KiB of .text and overhead.
+CONFIG_EXECUTORCH_METHOD_ALLOCATOR_POOL_SIZE=1572864
+CONFIG_EXECUTORCH_TEMP_ALLOCATOR_POOL_SIZE=1572864

zephyr/samples/mv2-ethosu/boards/mps4_corstone320_fvp.overlay

@@ -5,11 +5,18 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-/* The model PTE blob is placed in DDR (0x7000_0000) via a linker snippet
- * to avoid overflowing SRAM with the full MobileNetV2 model.
+/* Override zephyr,sram to the 4 MiB ISRAM (0x3100_0000) so the allocator
+ * pools (~3 MiB total) fit alongside code in the same region.  The default
+ * sram@12000000 is only 2 MiB which is too small for MV2.
+ *
+ * The model PTE blob is placed in DDR (0x7000_0000) via a linker snippet.
  * The Ethos-U can DMA from DDR on the Corstone-320 FVP.
  */
 / {
+	chosen {
+		zephyr,sram = &isram;
+	};
+
 	model_ddr: memory@70000000 {
 		compatible = "zephyr,memory-region", "mmio-sram";
 		reg = <0x70000000 DT_SIZE_M(16)>;