From d2191150736dde66d79eb97e0c8ee506eef3c8fc Mon Sep 17 00:00:00 2001 From: Gunes Bayir Date: Tue, 19 Mar 2024 22:10:04 +0000 Subject: Make Cpu/Gpu/Ref scalar/vectoral S32 division consistent - Neon(TM) implementation converts integers to float and performs the division because there is no vector integer division instructions. However, leftover loop still uses integer division, which makes results inconsistent depending on where we are in the tensor. - SVE path does it in integer domain. - OpenCL(TM) does it similar to Neon(TM) vector path. - Reference implementation does it in integer domain. These differences cause intermittent mismatches. This patch ensures all follow the same logic. On the other hand, the provided Neon(TM) implementation is faster than the Fp32 converted version. Resolves: COMPMID-6925 Change-Id: Ia12606d57f40a7d331b9b698f87fd4321496b275 Signed-off-by: Gunes Bayir Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11316 Tested-by: Arm Jenkins Reviewed-by: Pablo Marquez Tello Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../kernels/elementwise_binary/generic/neon/impl.h | 26 +++++++++++----------- 1 file changed, 13 insertions(+), 13 deletions(-) (limited to 'src/cpu') diff --git a/src/cpu/kernels/elementwise_binary/generic/neon/impl.h b/src/cpu/kernels/elementwise_binary/generic/neon/impl.h index 98f7e8b949..78e3baf74b 100644 --- a/src/cpu/kernels/elementwise_binary/generic/neon/impl.h +++ b/src/cpu/kernels/elementwise_binary/generic/neon/impl.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2022 Arm Limited. + * Copyright (c) 2021-2022, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef SRC_CORE_NEON_KERNELS_ELEMENTWISE_IMPL_H -#define SRC_CORE_NEON_KERNELS_ELEMENTWISE_IMPL_H +#ifndef ACL_SRC_CPU_KERNELS_ELEMENTWISE_BINARY_GENERIC_NEON_IMPL_H +#define ACL_SRC_CPU_KERNELS_ELEMENTWISE_BINARY_GENERIC_NEON_IMPL_H #include "src/core/NEON/NEAsymm.h" @@ -198,14 +198,6 @@ inline ScalarType elementwise_arithm_op_scalar(const ScalarType &a, const Scalar case ArithmeticOperation::DIV: { res = a / b; - if (std::is_integral::value) - { - res = (b == 0) ? 0 : res; - if (static_cast(a) % static_cast(b) != 0 && ((a < 0) != (b < 0))) - { - --res; - } - } break; } case ArithmeticOperation::POWER: @@ -224,7 +216,15 @@ inline int32x4_t elementwise_arithm_op>(const int32x4_t &a, const int32x4_t &b) { - return vcvtq_s32_f32(vfloorq_f32(wrapper::vdiv(vcvtq_f32_s32(a), vcvtq_f32_s32(b)))); + int32x4_t result; + + // Neon(TM) does not have vector integer division + result[0] = a[0] / b[0]; + result[1] = a[1] / b[1]; + result[2] = a[2] / b[2]; + result[3] = a[3] / b[3]; + + return result; } template <> @@ -1313,4 +1313,4 @@ void elementwise_comp_op_quantized_signed(const ITensor *in1, const ITensor *in2 } // namespace cpu } // namespace arm_compute -#endif /* SRC_CORE_NEON_KERNELS_ELEMENTWISE_IMPL_H */ +#endif // ACL_SRC_CPU_KERNELS_ELEMENTWISE_BINARY_GENERIC_NEON_IMPL_H -- cgit v1.2.1