From 0d0cd3e8e261cb61caa6f361014cdcc0e830a523 Mon Sep 17 00:00:00 2001 From: Borys Bradel Date: Tue, 3 Dec 2024 20:00:19 +0000 Subject: [PATCH] #13621: enable default fp32 acc for reduce --- .../operations/reduction/generic/device/reduce_op.cpp | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/ttnn/cpp/ttnn/operations/reduction/generic/device/reduce_op.cpp b/ttnn/cpp/ttnn/operations/reduction/generic/device/reduce_op.cpp index 2237d13c86a..0e474752333 100644 --- a/ttnn/cpp/ttnn/operations/reduction/generic/device/reduce_op.cpp +++ b/ttnn/cpp/ttnn/operations/reduction/generic/device/reduce_op.cpp @@ -191,8 +191,12 @@ Tensor reduce( auto is_multicore_hw = parallelization_strategy == ReduceOpParallelizationStrategy::MULTI_CORE_HW; float pad_value = reduce_math == ReduceOpMath::MAX ? -std::numeric_limits::infinity() : 0; - ttnn::DeviceComputeKernelConfig config = compute_kernel_config.value_or( - ttnn::init_device_compute_kernel_config(input_tensor.device()->arch(), std::nullopt, MathFidelity::HiFi4)); + ttnn::DeviceComputeKernelConfig config = compute_kernel_config.value_or(ttnn::init_device_compute_kernel_config( + input_tensor.device()->arch(), + std::nullopt, + MathFidelity::HiFi4, + /*default_approx_mode=*/false, + /*default_fp32_acc=*/true)); std::vector output_tensors = {Tensor(operation::get_workers_for_op_output({input_tensor}))}; if (is_multicore_hw) {