commit 9e698d55d861b3a2a2deb9486a75dd541e928705
author David Mansell <David.Mansell@arm.com> Tue Aug 25 15:02:02 2020 +0100
committer Georgios Pinitas <georgios.pinitas@arm.com> Tue Sep 22 10:13:53 2020 +0000
tree 4f7a25d8af58a5d714971635ad4ba36850cd6d6b
parent 047c6fcd2ead657ea251a251893767aa90d6bde3

COMPMID-3757: (u)int8: Don't select the 16-bit route on A53 for cases with very few rows

Also added 2D version of the 16-bit route, and altered the selection
heuristic so that 2D mode will be used in cases where 1D mode won't
thread well.

Change-Id: I0057fde08456771dc0090ac51f50d82f8bb86044
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3903
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
index 88726b1..06e68cb 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
@@ -101,13 +101,6 @@
     [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_4x6, uint8_t, uint32_t>(args); }
 },
 {
-    GemmMethod::GEMM_INTERLEAVED,
-    "gemm_u16_12x8",
-    nullptr,
-    [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53; },
-    [](const GemmArgs &args) { return new GemmInterleaved<gemm_u16_12x8, uint8_t, uint32_t>(args); },
-},
-{
     GemmMethod::GEMM_HYBRID,
     "hybrid_u8u32_dot_16x4",
     [](const GemmArgs &args) { return args._ci->has_dotprod() && args._Ksize>=16; },
@@ -130,9 +123,24 @@
 },
 {
     GemmMethod::GEMM_INTERLEAVED_2D,
+    "gemm_u16_12x8_2d",
+    nullptr,
+    [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Msize > 4 && (args._Msize / args._maxthreads) < 8; },
+    [](const GemmArgs &args) { return new GemmInterleavedPretransposed2d<gemm_u16_12x8, uint8_t, uint32_t>(args); },
+},
+{
+    GemmMethod::GEMM_INTERLEAVED,
+    "gemm_u16_12x8_1d",
+    nullptr,
+    [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Msize > 4; },
+    [](const GemmArgs &args) { return new GemmInterleaved<gemm_u16_12x8, uint8_t, uint32_t>(args); },
+},
+{
+    GemmMethod::GEMM_INTERLEAVED_2D,
     "gemm_u8_4x4_2d",
     nullptr,
-    [](const GemmArgs &args) { return (args._maxthreads >= 8) && (args._Msize >= 8) && (args._Nsize >= 8); },
+    [](const GemmArgs &args) { return ((args._maxthreads >= 8) && (args._Msize >= 8) && (args._Nsize >= 8)) ||
+                                       ((args._Msize / args._maxthreads) < 4); },
     [](const GemmArgs &args) { return new GemmInterleavedPretransposed2d<gemm_u8_4x4, uint8_t, uint32_t>(args); }
 },
 {