aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
diff options
context:
space:
mode:
authorMichele Di Giorgio <michele.digiorgio@arm.com>2021-07-26 13:18:50 +0100
committerGeorgios Pinitas <georgios.pinitas@arm.com>2021-09-07 13:44:08 +0000
commitaed63ee175e0d64c934389e9d1b2edd0cb1a5cdd (patch)
tree8f025f849e863b9cdec1d6b889bc463e6c4f78d1 /src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
parent58d3c5a7df769def499806e4d26cea518add161a (diff)
downloadComputeLibrary-aed63ee175e0d64c934389e9d1b2edd0cb1a5cdd.tar.gz
Add support for non-constant weights and biases in CpuFullyConnected
Changing the approach for specifying that weights and biases tensors are non-constant by making it a member of TensorInfo rather than an option of the functions. Resolves: COMPMID-4222 Change-Id: I96e6f3868f51785c9700a3ef6a1fe7b05747862c Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com> Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6162 Tested-by: Arm Jenkins <bsgcomp@arm.com> Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp8
1 files changed, 6 insertions, 2 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp b/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
index d4348beabf..f0b4e5db9e 100644
--- a/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
@@ -201,11 +201,11 @@ public:
return _buffer_per_multi * _args._nmulti * sizeof(To) + get_col_sum_size();
}
- void pretranspose_B_array(void *buffer, const To *B, const int ldb, const int B_multi_stride) override {
+ void requantize_bias(void *in_buffer, const To *B, const int ldb, const int B_multi_stride) override {
// Column sums go on the front of the pretransposed buffer in requantized cases.
// We could optimize here in case we don't actually need to sum the columns, but this code is only run on setup.
if (std::is_same<OutputStage, Requantize32>::value) {
- col_bias = reinterpret_cast<int32_t *>(buffer);
+ col_bias = reinterpret_cast<int32_t *>(in_buffer);
Requantize32 *qp_ptr = reinterpret_cast<Requantize32 *>(&_os);
@@ -213,6 +213,10 @@ public:
compute_col_sums(*qp_ptr, _args._Nsize, _args._Ksize, B + (i * B_multi_stride), ldb, col_bias + (i * _args._Nsize), _args._Ksize, i, 0);
}
}
+ }
+
+ void pretranspose_B_array(void *buffer, const To *B, const int ldb, const int B_multi_stride) override {
+ requantize_bias(buffer, B, ldb, B_multi_stride);
// The actual transposed buffer goes after the column sums (if any)
uintptr_t buffer_int = reinterpret_cast<uintptr_t>(buffer);