Cycle estimate-based kernel selection for dot product quantized s8/u8 kernels

Resolves: COMPMID-3990

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: If840c79209940535450f4ea1cbf6b0ec646a168e
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4866
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 10 insertions, 6 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp b/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp
index 2b3e170a3b..02d9486cc6 100644
--- a/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp
+++ b/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp
@@ -314,27 +314,25 @@ template void Interleave<6, 1, VLType::None>(float *, const bfloat16 *, size_t,
 
 /* AArch64 */
 #ifdef __aarch64__
-/* FP64 */
-/* NEON/SVE implementation (height 8) */
-template void IndirectInterleave<8, 1, VLType::None>(double *, const double * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
-template void ConvolutionInterleave<8, 1, VLType::None>(double *, const double *, size_t, const convolver<double> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
-template void Interleave<8, 1, VLType::None>(double *, const double *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
-
 /* FP32 */
 /* NEON/SVE implementation (height 8) */
 template void IndirectInterleave<8, 1, VLType::None>(float *, const float * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void ConvolutionInterleave<8, 1, VLType::None>(float *, const float *, size_t, const convolver<float> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void Interleave<8, 1, VLType::None>(float *, const float *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 
+#if defined(__ARM_FEATURE_SVE) && defined(MMLA_FP32)
 /* FMMLA */
 template void IndirectInterleave<8, 2, VLType::None>(float *, const float * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void ConvolutionInterleave<8, 2, VLType::None>(float *, const float *, size_t, const convolver<float> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void Interleave<8, 2, VLType::None>(float *, const float *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
+#endif // SVE && MMLA_FP32
 
 /* FP16 */
+#if defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
 template void IndirectInterleave<8, 1, VLType::None>(__fp16 *, const __fp16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void ConvolutionInterleave<8, 1, VLType::None>(__fp16 *, const __fp16 *, size_t, const convolver<__fp16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void Interleave<8, 1, VLType::None>(__fp16 *, const __fp16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
+#endif // FP16_KERNELS ar __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
 template void IndirectInterleave<8, 1, VLType::None>(float *, const __fp16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void ConvolutionInterleave<8, 1, VLType::None>(float *, const __fp16 *, size_t, const convolver<__fp16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
@@ -342,6 +340,7 @@ template void Interleave<8, 1, VLType::None>(float *, const __fp16 *, size_t, un
 
 /* BF16 */
 /* NEON/SVE BFDOT */
+#ifdef V8P6_BF
 template void IndirectInterleave<8, 2, VLType::None>(bfloat16 *, const bfloat16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void ConvolutionInterleave<8, 2, VLType::None>(bfloat16 *, const bfloat16 *, size_t, const convolver<bfloat16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void Interleave<8, 2, VLType::None>(bfloat16 *, const bfloat16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
@@ -349,6 +348,7 @@ template void Interleave<8, 2, VLType::None>(bfloat16 *, const bfloat16 *, size_
 template void IndirectInterleave<8, 4, VLType::None>(bfloat16 *, const bfloat16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void ConvolutionInterleave<8, 4, VLType::None>(bfloat16 *, const bfloat16 *, size_t, const convolver<bfloat16> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void Interleave<8, 4, VLType::None>(bfloat16 *, const bfloat16 *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
+#endif // V8P6_BF
 
 /* NEON/SVE using FP32 kernel */
 template void IndirectInterleave<8, 1, VLType::None>(float *, const bfloat16 * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
@@ -375,10 +375,12 @@ template void IndirectInterleave<8, 4, VLType::None>(int8_t *, const int8_t * co
 template void ConvolutionInterleave<8, 4, VLType::None>(int8_t *, const int8_t *, size_t, const convolver<int8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void Interleave<8, 4, VLType::None>(int8_t *, const int8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 
+#ifdef MMLA_INT8
 /* MMLA SMMLA (height 8, block 8) */
 template void IndirectInterleave<8, 8, VLType::None>(int8_t *, const int8_t * const * const *, unsigned int, unsigned int, unsigned int y0, unsigned int ymax, unsigned int k0, unsigned int kmax, bool, int32_t);
 template void ConvolutionInterleave<8, 8, VLType::None>(int8_t *, const int8_t *, size_t, const convolver<int8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void Interleave<8, 8, VLType::None>(int8_t *, const int8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
+#endif // MMLA_INT8
 
 /* NEON SDOT (height 8, block 1) */
 template void IndirectInterleave<8, 1, VLType::None>(int16_t *, const int8_t * const * const *, unsigned int, unsigned int, unsigned int y0, unsigned int ymax, unsigned int k0, unsigned int kmax, bool, int32_t);
@@ -395,10 +397,12 @@ template void IndirectInterleave<8, 4, VLType::None>(uint8_t *, const uint8_t *
 template void ConvolutionInterleave<8, 4, VLType::None>(uint8_t *, const uint8_t *, size_t, const convolver<uint8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void Interleave<8, 4, VLType::None>(uint8_t *, const uint8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 
+#ifdef MMLA_INT8
 /* MMLA SMMLA (height 8, block 8) */
 template void IndirectInterleave<8, 8, VLType::None>(uint8_t *, const uint8_t * const * const *, unsigned int, unsigned int, unsigned int y0, unsigned int ymax, unsigned int k0, unsigned int kmax, bool, int32_t);
 template void ConvolutionInterleave<8, 8, VLType::None>(uint8_t *, const uint8_t *, size_t, const convolver<uint8_t> &, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
 template void Interleave<8, 8, VLType::None>(uint8_t *, const uint8_t *, size_t, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
+#endif // MMLA_INT8
 
 /* NEON 16-bit (height 8, block 1) */
 template void IndirectInterleave<8, 1, VLType::None>(uint16_t *, const uint8_t * const * const *, unsigned int, unsigned int, unsigned int y0, unsigned int ymax, unsigned int k0, unsigned int kmax, bool, int32_t);