From 4c30de056afe8680b42723b26a2241811715b989 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Fri, 7 Jul 2023 12:01:32 +0100 Subject: Enable premultiplication for depthwise convolution with fp16 and quantized types Resolves: COMPMID-6337 Change-Id: I81542e51c9c0329f202ac8452f173b138e51a0f6 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9883 Reviewed-by: Gunes Bayir Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- .../NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp index 3a4999296a..5d7b54f235 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp @@ -42,7 +42,7 @@ size_t interleave_sve_s8q_3x3_dot::get_packed_size(const DepthwiseArgs &args) { // We store 7 vectors for every of channels. const unsigned int n = arm_gemm::roundup( - arm_gemm::iceildiv((long unsigned int) args.input_channels, + arm_gemm::iceildiv((long unsigned int) args.input_channels * args.channel_multiplier, get_vector_length(arm_gemm::VLType::SVE)), 4lu ); return n * 7 * get_vector_length(arm_gemm::VLType::SVE); -- cgit v1.2.1