From 48ef088b1f2d594b4d756c19574ead05492d3f84 Mon Sep 17 00:00:00 2001 From: zhangyanhui Date: Mon, 8 Jul 2024 18:27:42 +0800 Subject: [PATCH] bugfix for depthwise when kernel size is longer than input_w --- .../device/cpu/kernel/nnacl/assembly/arm32/ConvDwFp32Row.S | 2 +- .../device/cpu/kernel/nnacl/assembly/arm64/ConvDwFp32Row.S | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/assembly/arm32/ConvDwFp32Row.S b/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/assembly/arm32/ConvDwFp32Row.S index 8d1d819d..19750920 100644 --- a/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/assembly/arm32/ConvDwFp32Row.S +++ b/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/assembly/arm32/ConvDwFp32Row.S @@ -35,7 +35,7 @@ asm_function ConvDwFp32Row mov r6, #4 mul r5, r5, r6 cmp r3, #0 - beq End + ble End LoopNumPixel: mov r6, r1 // input_ptr diff --git a/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/assembly/arm64/ConvDwFp32Row.S b/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/assembly/arm64/ConvDwFp32Row.S index 417aa4ce..e50e2ee5 100644 --- a/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/assembly/arm64/ConvDwFp32Row.S +++ b/mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/assembly/arm64/ConvDwFp32Row.S @@ -30,7 +30,7 @@ asm_function ConvDwFp32Row // x19 ~ x29 should be also preserved // whereas our coding style do not permit such amount of parameters cmp x3, #0 -beq End +ble End mov x9, x0 mov x12, #4 -- 2.25.1