diff options
| author | Pablo Tello <pablo.tello@arm.com> | 2017-06-28 10:05:29 +0100 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-09-17 14:15:39 +0100 |
| commit | dcdc85ef876e854749db58ecd60c37f64a627536 (patch) | |
| tree | 8af2006f14a8b32eb1aced14def96afab89c5981 /src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp | |
| parent | fc95ed2b9900471922d93c963b263f1f506da167 (diff) | |
| download | ComputeLibrary-dcdc85ef876e854749db58ecd60c37f64a627536.tar.gz | |
COMPMID-421: Added F16 support in FC Layer.
Change-Id: I9c3ab51ae024be69c7b1d83803b1a8f60a0cdbfd
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79326
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp')
| -rw-r--r-- | src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp | 23 |
1 files changed, 22 insertions, 1 deletions
diff --git a/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp index 826a386557..f3d06ed481 100644 --- a/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.cpp @@ -45,7 +45,7 @@ NEGEMMMatrixAccumulateBiasesKernel::NEGEMMMatrixAccumulateBiasesKernel() void NEGEMMMatrixAccumulateBiasesKernel::configure(ITensor *accum, const ITensor *biases) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::QS16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(biases, accum); ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(biases, accum); ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() != 1); @@ -109,6 +109,27 @@ void NEGEMMMatrixAccumulateBiasesKernel::run(const Window &window) in0_out, in1); break; } +#ifdef ARM_COMPUTE_ENABLE_FP16 + case DataType::F16: + { + execute_window_loop(window, [&](const Coordinates & id) + { + const float16x8x2_t accum = vld2q_f16(reinterpret_cast<const float16_t *>(in0_out.ptr())); + const float16x8x2_t biases = vld2q_f16(reinterpret_cast<const float16_t *>(in1.ptr())); + const float16x8x2_t res = + { + { + vaddq_f16(accum.val[0], biases.val[0]), + vaddq_f16(accum.val[1], biases.val[1]) + } + }; + + vst2q_f16(reinterpret_cast<float16_t *>(in0_out.ptr()), res); + }, + in0_out, in1); + break; + } +#endif /* ARM_COMPUTE_ENABLE_FP16 */ case DataType::QS8: { execute_window_loop(window, [&](const Coordinates & id) |