MLECO-3847: Optimisation flag for TensorFlow Lite Micro
Overriding the default THIRD_PARTY_KERNEL_OPTIMIZATION_LEVEL flag
for comilation of Arm CMSIS-NN library. GNU linker script also
needed to be modified with this change because of increase in
the code footprint.
Change-Id: I65c76fcaf4b6421533086a0905b2e6f0048b46fd
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>
diff --git a/scripts/cmake/platforms/mps3/sse-300/mps3-sse-300.ld b/scripts/cmake/platforms/mps3/sse-300/mps3-sse-300.ld
index 880a233..2d72ed9 100644
--- a/scripts/cmake/platforms/mps3/sse-300/mps3-sse-300.ld
+++ b/scripts/cmake/platforms/mps3/sse-300/mps3-sse-300.ld
@@ -1,5 +1,5 @@
/*
- * SPDX-FileCopyrightText: Copyright 2021 Arm Limited and/or its affiliates <open-source-office@arm.com>
+ * SPDX-FileCopyrightText: Copyright 2021, 2023 Arm Limited and/or its affiliates <open-source-office@arm.com>
* SPDX-License-Identifier: Apache-2.0
*
* Licensed under the Apache License, Version 2.0 (the "License");
@@ -74,12 +74,17 @@
KEEP(*(.vectors))
/**
- * All code goes here, with one exception of
- * all_ops_resolver object file. This code
- * instead placed on BRAM. See comment in the
- * BRAM section for details.
- **/
- *(EXCLUDE_FILE(*all_ops_resolver.o *hal.c.obj) .text*)
+ * Any code that is not time sensitive can be excluded from here.
+ * This code is instead placed on BRAM. See comment in the BRAM
+ * section for details.
+ */
+ *(EXCLUDE_FILE(*all_ops_resolver.o
+ *hal.c.obj
+ *_allocator.o
+ *flatbuffer*.o
+ *lcd*.obj
+ *timing_adapter.c.obj)
+ .text*)
KEEP(*(.init))
KEEP(*(.fini))
@@ -238,6 +243,14 @@
. = ALIGN(4);
*hal.c.obj (*.text*)
. = ALIGN(4);
+ *_allocator.o (*.text*)
+ . = ALIGN(4);
+ *flatbuffer*.o (*.text*)
+ . = ALIGN(4);
+ *lcd*.obj (*.text*)
+ . = ALIGN(4);
+ *timing_adapter.* (*.text*)
+ . = ALIGN(4);
__data_end__ = .;
} > BRAM
diff --git a/scripts/cmake/platforms/mps3/sse-310/mps3-sse-310.sct b/scripts/cmake/platforms/mps3/sse-310/mps3-sse-310.sct
index c46fae4..61f2aba 100644
--- a/scripts/cmake/platforms/mps3/sse-310/mps3-sse-310.sct
+++ b/scripts/cmake/platforms/mps3/sse-310/mps3-sse-310.sct
@@ -1,4 +1,4 @@
-; SPDX-FileCopyrightText: Copyright 2021 Arm Limited and/or its affiliates <open-source-office@arm.com>
+; SPDX-FileCopyrightText: Copyright 2021, 2023 Arm Limited and/or its affiliates <open-source-office@arm.com>
; SPDX-License-Identifier: Apache-2.0
;
; Licensed under the Apache License, Version 2.0 (the "License");
@@ -30,13 +30,13 @@
LOAD_REGION_0 0x11000000 0x00200000
{
;-----------------------------------------------------
- ; 0.5 MiB of SRAM/BRAM region. Our vector table also
+ ; First 640K of SRAM/BRAM. Our vector table also
; resides here as the default INITSVTOR is 0x11000000.
; We currently do not use the ITCM for any code, but
; could potentially put some critical code in there
; if we need to.
;-----------------------------------------------------
- bram.bin 0x11000000 0x00080000
+ bram.bin 0x11000000 0x000A0000
{
*.o (RESET, +First)
* (InRoot$$Sections)
@@ -46,10 +46,10 @@
}
;-----------------------------------------------------
- ; Next 0.5 MiB of SRAM/BRAM region for RO, RW and ZI
+ ; Next 384K of SRAM/BRAM region for RO, RW and ZI
; data, 8 byte aligned.
;-----------------------------------------------------
- data.bin 0x11080000 ALIGN 8 0x00080000
+ data.bin 0x110A0000 ALIGN 8 0x00060000
{
; Any RO-DATA
.ANY (+RO-DATA)
diff --git a/scripts/cmake/platforms/simple_platform/simple_platform.ld b/scripts/cmake/platforms/simple_platform/simple_platform.ld
index 40b8ab6..492f6da 100644
--- a/scripts/cmake/platforms/simple_platform/simple_platform.ld
+++ b/scripts/cmake/platforms/simple_platform/simple_platform.ld
@@ -1,5 +1,5 @@
/*
- * SPDX-FileCopyrightText: Copyright 2021 Arm Limited and/or its affiliates <open-source-office@arm.com>
+ * SPDX-FileCopyrightText: Copyright 2021,2023 Arm Limited and/or its affiliates <open-source-office@arm.com>
* SPDX-License-Identifier: Apache-2.0
*
* Licensed under the Apache License, Version 2.0 (the "License");
@@ -67,12 +67,17 @@
KEEP(*(.vectors))
/**
- * All code goes here, with one exception of
- * all_ops_resolver object file. This code
- * instead placed on BRAM. See comment in the
- * BRAM section for details.
- **/
- *(EXCLUDE_FILE(*all_ops_resolver.o *hal.c.obj) .text*)
+ * Any code that is not time sensitive can be excluded from here.
+ * This code is instead placed on BRAM. See comment in the BRAM
+ * section for details.
+ */
+ *(EXCLUDE_FILE(*all_ops_resolver.o
+ *hal.c.obj
+ *_allocator.o
+ *flatbuffer*.o
+ *lcd*.obj
+ *timing_adapter.c.obj)
+ .text*)
KEEP(*(.init))
KEEP(*(.fini))
@@ -231,6 +236,14 @@
. = ALIGN(4);
*hal.c.obj (*.text*)
. = ALIGN(4);
+ *_allocator.o (*.text*)
+ . = ALIGN(4);
+ *flatbuffer*.o (*.text*)
+ . = ALIGN(4);
+ *lcd*.obj (*.text*)
+ . = ALIGN(4);
+ *timing_adapter.* (*.text*)
+ . = ALIGN(4);
__data_end__ = .;
} > BRAM
diff --git a/scripts/cmake/tensorflow.cmake b/scripts/cmake/tensorflow.cmake
index 29d5bce..ea3b320 100644
--- a/scripts/cmake/tensorflow.cmake
+++ b/scripts/cmake/tensorflow.cmake
@@ -1,5 +1,5 @@
#----------------------------------------------------------------------------
-# SPDX-FileCopyrightText: Copyright 2021-2022 Arm Limited and/or its affiliates <open-source-office@arm.com>
+# SPDX-FileCopyrightText: Copyright 2021-2023 Arm Limited and/or its affiliates <open-source-office@arm.com>
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
@@ -26,8 +26,8 @@
set(TENSORFLOW_LITE_MICRO_CORE_OPTIMIZATION_LEVEL "-O0")
set(TENSORFLOW_LITE_MICRO_KERNEL_OPTIMIZATION_LEVEL "-O0")
elseif (CMAKE_BUILD_TYPE STREQUAL Release)
- set(TENSORFLOW_LITE_MICRO_CORE_OPTIMIZATION_LEVEL "-O3")
- set(TENSORFLOW_LITE_MICRO_KERNEL_OPTIMIZATION_LEVEL "-O3")
+ set(TENSORFLOW_LITE_MICRO_CORE_OPTIMIZATION_LEVEL "-Ofast")
+ set(TENSORFLOW_LITE_MICRO_KERNEL_OPTIMIZATION_LEVEL "-Ofast")
endif()
assert_defined(TENSORFLOW_LITE_MICRO_BUILD_TYPE)
@@ -118,6 +118,7 @@
$<$<BOOL:${TENSORFLOW_LITE_MICRO_CORE_OPTIMIZATION_LEVEL}>:CORE_OPTIMIZATION_LEVEL=${TENSORFLOW_LITE_MICRO_CORE_OPTIMIZATION_LEVEL}>
$<$<BOOL:${TENSORFLOW_LITE_MICRO_KERNEL_OPTIMIZATION_LEVEL}>:KERNEL_OPTIMIZATION_LEVEL=${TENSORFLOW_LITE_MICRO_KERNEL_OPTIMIZATION_LEVEL}>
+ $<$<BOOL:${TENSORFLOW_LITE_MICRO_KERNEL_OPTIMIZATION_LEVEL}>:THIRD_PARTY_KERNEL_OPTIMIZATION_LEVEL=${TENSORFLOW_LITE_MICRO_KERNEL_OPTIMIZATION_LEVEL}>
$<$<BOOL:${TENSORFLOW_LITE_MICRO_OPTIMIZED_KERNEL}>:OPTIMIZED_KERNEL_DIR=${TENSORFLOW_LITE_MICRO_OPTIMIZED_KERNEL}>
$<$<BOOL:${TENSORFLOW_LITE_MICRO_CO_PROCESSOR}>:CO_PROCESSOR=${TENSORFLOW_LITE_MICRO_CO_PROCESSOR}>