diff options
author | David Mansell <David.Mansell@arm.com> | 2023-03-30 19:10:52 +0100 |
---|---|---|
committer | Viet-Hoa Do <viet-hoa.do@arm.com> | 2023-04-26 11:05:03 +0000 |
commit | b2758f35da97319fd15722485e9b4ba7b35c8cfa (patch) | |
tree | edec5ead897c52b027f95e470fe23e09c6b1291e /src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp | |
parent | af15076bf8148870aa41d414b73e81f3b7e86436 (diff) | |
download | ComputeLibrary-b2758f35da97319fd15722485e9b4ba7b35c8cfa.tar.gz |
Add FP16 depthwise kernels for SME2
Resolves: COMPMID-5988
Change-Id: I93e78edf31c9eec8242ccbb8c3c768f46a7c7c38
Signed-off-by: David Mansell <David.Mansell@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9456
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp | 69 |
1 files changed, 69 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp index 6b100d9d61..350e93b874 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp @@ -36,6 +36,13 @@ #if defined(__aarch64__) #if defined(ARM_COMPUTE_ENABLE_SVE) +#if defined(ARM_COMPUTE_ENABLE_SME2) +#include "kernels/sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst.hpp" +#include "kernels/sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst.hpp" +#include "kernels/sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst.hpp" +#include "kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst.hpp" +#include "kernels/sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst.hpp" +#endif // defined(ARM_COMPUTE_ENABLE_SME2) #include "kernels/sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst.hpp" #include "kernels/sve_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst.hpp" #include "kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst.hpp" @@ -82,6 +89,68 @@ namespace static const DepthwiseImplementation<__fp16, __fp16> depthwise_fp16_methods[] = { #if defined(__aarch64__) #if defined(ARM_COMPUTE_ENABLE_SVE) +#if defined(ARM_COMPUTE_ENABLE_SME2) + { + DepthwiseMethod::DEPTHFIRST, + "sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst", + constraint(is_supported<sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst>, + has_no_channel_multiplier, + cpu_has_sme2), + cycle_estimate<sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst>, + [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * { + auto strat = new sme2_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst(args.cpu_info); + return new DepthwiseDepthfirst<__fp16>(strat, args); + }, + }, + { + DepthwiseMethod::DEPTHFIRST, + "sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst", + constraint(is_supported<sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst>, + has_no_channel_multiplier, + cpu_has_sme2), + cycle_estimate<sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst>, + [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * { + auto strat = new sme2_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst(args.cpu_info); + return new DepthwiseDepthfirst<__fp16>(strat, args); + }, + }, + { + DepthwiseMethod::DEPTHFIRST, + "sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst", + constraint(is_supported<sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst>, + has_no_channel_multiplier, + cpu_has_sme2), + cycle_estimate<sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst>, + [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * { + auto strat = new sme2_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst(args.cpu_info); + return new DepthwiseDepthfirst<__fp16>(strat, args); + }, + }, + { + DepthwiseMethod::DEPTHFIRST, + "sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst", + constraint(is_supported<sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst>, + has_no_channel_multiplier, + cpu_has_sme2), + cycle_estimate<sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst>, + [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * { + auto strat = new sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst(args.cpu_info); + return new DepthwiseDepthfirst<__fp16>(strat, args); + }, + }, + { + DepthwiseMethod::DEPTHFIRST, + "sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst", + constraint(is_supported<sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst>, + has_no_channel_multiplier, + cpu_has_sme2), + cycle_estimate<sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst>, + [] (const DepthwiseArgs &args, const Nothing &) -> DepthwiseCommon<__fp16, __fp16, __fp16> * { + auto strat = new sme2_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst(args.cpu_info); + return new DepthwiseDepthfirst<__fp16>(strat, args); + }, + }, +#endif // defined(ARM_COMPUTE_ENABLE_SME2) { DepthwiseMethod::DEPTHFIRST, "sve_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst", |