aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
diff options
context:
space:
mode:
authorramelg01 <ramy.elgammal@arm.com>2022-04-07 02:42:52 +0100
committerRamy Elgammal <ramy.elgammal@arm.com>2022-04-26 15:51:22 +0000
commit8a164884dddf769643cf3b9f7f94e43cb4f3c20b (patch)
tree35958dd48b6df1a851c880dad2b2ce285671b611 /src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
parentc827e99fc46521f43719b0c2d1b6f05d66abf68c (diff)
downloadComputeLibrary-8a164884dddf769643cf3b9f7f94e43cb4f3c20b.tar.gz
Update Neon™ depthwise kernel
- Reduce duplication and simplify overall structure. - Improve multi-threaded performance by sharing more data in lower-level caches. Partially Resolves: COMPMID-5054 Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com> Change-Id: Iac747f39b21c540122fa75218762631c4d787911 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7449 Tested-by: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Andrew Mundy Reviewed-by: Sheri Zhang <sheri.zhang@arm.com> Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp51
1 files changed, 26 insertions, 25 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
index 0fde00ba37..22b6b657a5 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -34,39 +34,40 @@
namespace arm_conv {
namespace depthwise {
-void a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst_impl(const int8_t *const *, int8_t *const *, const void *, uint64_t, const arm_gemm::Requantize32&);
+void a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst_impl(unsigned int, const int8_t *const *, const int8_t *, const int32_t *, const arm_gemm::Requantize32&, const int32_t *, const int32_t *, int8_t *const *);
-struct a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst
+class a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst : public DepthwiseDepthfirstStrategy<int8_t, int8_t, int8_t, int32_t>
{
- typedef int32_t bias_type;
- typedef int8_t input_type;
- typedef int8_t weight_type;
- typedef int8_t return_type;
-
- constexpr static arm_gemm::VLType vl_type = arm_gemm::VLType::None;
-
- typedef void (*kern_type)(const int8_t *const *, int8_t *const *, const void *, uint64_t, const arm_gemm::Requantize32&);
- typedef void (*parameter_packing_fn)(unsigned int, void *, const int32_t *, const int8_t *, const arm_gemm::Requantize32 &, size_t, size_t);
- typedef size_t (*parameter_sizing_fn)(const DepthwiseArgs &);
+ using Parent = DepthwiseDepthfirstStrategy<int8_t, int8_t, int8_t, int32_t>;
+ public:
constexpr static unsigned int kernel_rows = 3;
constexpr static unsigned int kernel_cols = 3;
constexpr static unsigned int stride_rows = 1;
constexpr static unsigned int stride_cols = 1;
- constexpr static unsigned int output_rows = 2;
- constexpr static unsigned int output_cols = 2;
-
- constexpr static unsigned int input_rows = 4;
- constexpr static unsigned int input_cols = 4;
-
- constexpr static parameter_packing_fn pack_parameters = interleave_a64_s8q_3x3_dot::pack_parameters;
- constexpr static parameter_sizing_fn get_packed_size = interleave_a64_s8q_3x3_dot::get_packed_size;
-
- kern_type kernel = a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst_impl;
-
- a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst(const CPUInfo *) {}
+ a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst(const CPUInfo *) : Parent(2, 2, 3, 3, 1, 1) {}
+
+ arm_gemm::VLType get_vl_type(void) const override { return arm_gemm::VLType::None; }
+
+ Parent::KernelType kernel = a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst_impl;
+ Parent::KernelType get_kernel(void) const override { return kernel; }
+ size_t get_storage_size(const DepthwiseArgs &args) const override
+ {
+ return interleave_a64_s8q_3x3_dot::get_packed_size(args);
+ }
+
+ void pack_parameters(
+ const DepthwiseArgs &args, void *buffer, const void *biases, const arm_gemm::Requantize32 &qp,
+ const void *weights, size_t ld_weight_col, size_t ld_weight_row
+ ) const override
+ {
+ interleave_a64_s8q_3x3_dot::pack_parameters(
+ args.input_channels, buffer, reinterpret_cast<const int32_t *>(biases),
+ reinterpret_cast<const int8_t *>(weights), qp, ld_weight_col, ld_weight_row
+ );
+ }
};
} // namespace depthwise