diff options
author | Michael Tyler <michael.tyler@arm.com> | 2024-06-04 15:47:37 +0100 |
---|---|---|
committer | Michael Tyler <michael.tyler@arm.com> | 2024-06-25 09:10:13 +0000 |
commit | fc94f4d23abd4bc427b701f54ad85282e9ec7872 (patch) | |
tree | 5e2980599256e2b2f4374e5beb61596fc95c9d5a /src/core/NEON/kernels/arm_gemm/barrier.hpp | |
parent | c2237ec4094c7824f8f7e61bc89504d01c5b59ff (diff) | |
download | ComputeLibrary-fc94f4d23abd4bc427b701f54ad85282e9ec7872.tar.gz |
Update CPU kernels and add mixed sign GEMM support
- Add support for mixed sign quantized convolution.
- Add support for mixed sign dequantized GEMM.
- Add SME FP16 GEMV kernel.
- Change SME vector length function to use RDSVL instead of static variable.
- Add GEMM dilation support internally (not exposed yet).
- Remove unused "get_default_activation_values" functions.
- Add SVE fixed format interleaved BF16 DOT kernel.
- Updates and optimizations to assembly kernels.
Resolves COMPMID-6926
Change-Id: I227f502502611d4cc4111c89e30c53ce94079544
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11570
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/barrier.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/barrier.hpp | 10 |
1 files changed, 9 insertions, 1 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/barrier.hpp b/src/core/NEON/kernels/arm_gemm/barrier.hpp index 8fbcddfef8..3b34e8089f 100644 --- a/src/core/NEON/kernels/arm_gemm/barrier.hpp +++ b/src/core/NEON/kernels/arm_gemm/barrier.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2020, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -39,6 +39,14 @@ private: public: barrier(unsigned int threads) : m_threads(threads), m_waiters(0), m_leavers(0) { } + // Add a move constructor because these objects might be moved around at setup time. + // Moving while the barrier is active won't work. + barrier(barrier &&other) : m_threads(other.m_threads), m_waiters(0), m_leavers(0) { + // This doesn't make it safe, but will have a chance of firing if something odd is occurring. + assert(other.m_waiters==0); + assert(other.m_leavers==0); + } + /* This isn't safe if any thread is waiting... */ void set_nthreads(unsigned int nthreads) { m_threads = nthreads; |