commit 33e03074c36d85de87e9032a2583b04ce8ddcd6b
author Georgios Pinitas <georgios.pinitas@arm.com> Thu Jan 14 13:43:40 2021 +0000
committer Georgios Pinitas <georgios.pinitas@arm.com> Mon Jan 18 12:39:30 2021 +0000
tree 4442ec5b5022fa5681d689e6ccf3a6423efa8a93
parent 8d5337ef18901f1b54d0c062ae7486bc5a4c6610

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>

src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
index eead592..9720c7d 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
@@ -121,39 +121,39 @@
     [](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>(args, qp); },
 },
-{
+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 args._Nsize<=256 && args._Ksize>128; },
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8qa_dot_4x16, int8_t, int8_t, Requantize32>::estimate_cycles(args, cls_a64_hybrid_u8qa_dot_4x16::get_performance_parameters(args._ci)); },
     [](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, 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 args._Nsize<=256 && args._Ksize>128; },
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, int8_t, int8_t, Requantize32, true>::estimate_cycles(args, cls_a64_hybrid_u8u32_dot_6x16::get_performance_parameters(args._ci)); },
     [](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, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_u8_8x12",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); },
-    nullptr,
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, int8_t, int8_t>::estimate_cycles(args, cls_a64_gemm_u8_8x12::get_performance_parameters(args._ci)); },
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, uint8_t, uint8_t>(args, qp); }
-},
+),
 {
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_u8_4x4",
     nullptr,
-    nullptr,
+    [](const GemmArgs &args, const Requantize32 &) { return !args._ci->has_dotprod(); },
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_4x4, uint8_t, uint8_t>(args, qp); }
 },
 {
     GemmMethod::QUANTIZE_WRAPPER,
     "quantized_wrapper",
     [](const GemmArgs &args, const Requantize32 &) { return !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args, const Requantize32 &) { return !args._ci->has_dotprod(); },
     [](const GemmArgs &args, const Requantize32 &qp) { return new QuantizeWrapper<uint8_t, uint8_t, uint32_t>(args, qp); }
 },
 {