Revert "Update CPU kernels and add mixed sign GEMM support"

This reverts commit fc94f4d23abd4bc427b701f54ad85282e9ec7872 and 5d6fff041ade7eb44af0945867212f3979be3d3e (because the latter fixes a build failure caused by the former)

Change-Id: I7d07fea8307e9a7033b30874bbb14ba9202b23d8
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11815
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Adnan AlSinan <adnan.alsinan@arm.com>

1 files changed, 47 insertions, 46 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
index 7c182b6777..93eecf991e 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
@@ -54,13 +54,14 @@
 
 #include "gemm_hybrid_indirect.hpp"
 #include "gemm_hybrid_quantized.hpp"
+#include "gemm_hybrid_quantized_inline.hpp"
 #include "gemm_interleaved.hpp"
 #include "gemv_pretransposed.hpp"
 #include "quantize_wrapper.hpp"
 
 namespace arm_gemm {
 
-static const GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32> gemm_quint8_methods[] =
+static const GemmImplementation<uint8_t, uint8_t, Requantize32> gemm_quint8_methods[] =
 {
 #ifdef ARM_COMPUTE_ENABLE_SVE
 #ifdef ARM_COMPUTE_ENABLE_SME2
@@ -96,69 +97,69 @@ static const GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32> gemm_qu
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_u8q_mopa_2VLx2VL, uint8_t, uint8_t>(args, qp); }
 },
 #endif // ARM_COMPUTE_ENABLE_SME2
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_u8qa_mmla_4x4VL",
     [](const GemmArgs &args, const Requantize32 &qp) { return quant_hybrid_asymmetric(qp) && args._ci->has_sve2() && args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8qa_mmla_4x4VL, uint8_t, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8qa_mmla_4x4VL, uint8_t, uint8_t, uint8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8qa_mmla_4x4VL, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8qa_mmla_4x4VL, uint8_t, uint8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_u8u32_mmla_8x3VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_hybrid_u8u32_mmla_6x4VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, uint8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, Requantize32, true>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_u8qa_dot_4x4VL",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sve2() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8qa_dot_4x4VL, uint8_t, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8qa_dot_4x4VL, uint8_t, uint8_t, uint8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8qa_dot_4x4VL, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8qa_dot_4x4VL, uint8_t, uint8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_u8u32_dot_6x4VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, uint8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, Requantize32, true>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_u8u32_dot_8x3VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve() && (args._Ksize>4); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t>(args, qp); }
 ),
 #endif // ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_u8qa_mmla_4x16",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_i8mm() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8qa_mmla_4x16, uint8_t, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8qa_mmla_4x16, uint8_t, uint8_t, uint8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8qa_mmla_4x16, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8qa_mmla_4x16, uint8_t, uint8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_interleaved_u8u32_mmla_8x12",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_hybrid_u8u32_mmla_6x16",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, uint8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, Requantize32, true>(args, qp); }
 ),
 {
     GemmMethod::GEMM_HYBRID_QUANTIZED,
@@ -179,35 +180,35 @@ GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
     "a64_gemm_u16_8x12",
     nullptr,
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Msize > 4; },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u16_8x12, uint8_t, uint8_t, uint8_t>(args, qp); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u16_8x12, uint8_t, uint8_t>(args, qp); },
 },
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_u8qa_dot_4x16",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_dotprod() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8qa_dot_4x16, uint8_t, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8qa_dot_4x16, uint8_t, uint8_t, uint8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8qa_dot_4x16, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8qa_dot_4x16, uint8_t, uint8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_u8u32_dot_6x16",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, uint8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, Requantize32, true>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_u8_8x12",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, uint8_t, uint8_t>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_u8_4x4",
     nullptr,
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_u8_4x4, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_4x4, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_u8_4x4, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_4x4, uint8_t, uint8_t>(args, qp); }
 ),
 {
     GemmMethod::QUANTIZE_WRAPPER,
@@ -226,14 +227,14 @@ GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
 };
 
 template<>
-const GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32> *gemm_implementation_list<uint8_t, uint8_t, uint8_t, Requantize32>() {
+const GemmImplementation<uint8_t, uint8_t, Requantize32> *gemm_implementation_list<uint8_t, uint8_t, Requantize32>() {
     return gemm_quint8_methods;
 }
 
-template UniqueGemmCommon<uint8_t, uint8_t, uint8_t> gemm<uint8_t, uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
-template bool has_opt_gemm<uint8_t, uint8_t, uint8_t, Requantize32>(WeightFormat &weight_format, const GemmArgs &args, const Requantize32 &os);
-template KernelDescription get_gemm_method<uint8_t, uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
-template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template UniqueGemmCommon<uint8_t, uint8_t> gemm<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template bool has_opt_gemm<uint8_t, uint8_t, Requantize32>(WeightFormat &weight_format, const GemmArgs &args, const Requantize32 &os);
+template KernelDescription get_gemm_method<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
 
 } // namespace arm_gemm