commit 146138378c1587b7297d245b7177641315f6180b
author Georgios Pinitas <georgios.pinitas@arm.com> Fri Mar 01 19:07:11 2019 +0000
committer Georgios Pinitas <georgios.pinitas@arm.com> Tue Mar 19 11:53:45 2019 +0000
tree 4689218b48a0884418099f373015429f2845ceaf
parent 3dd5b6884a65c06bcb9d15589ee2dc2978e3b336

COMPMID-1995: Update RSH GEMM assembly kernels.

-Updates u8/s8 hybrid dot product kernels to work for any N and any K >=16.
-Adds hybrid FP32 kernels with generic and A55 variants.
-Adds SVE native kernels for fp16/u8/s8.

Change-Id: Ifc0eaba9e3c8ea5bb19d334e870e1b39e4e7e728
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/863
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@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 b95ca80..9321bfc 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 Arm Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -34,6 +34,7 @@
 #include "kernels/a64_gemm_u8_12x8.hpp"
 #include "kernels/a64_gemm_u8_4x4.hpp"
 #include "kernels/a64_hybrid_u8u32_dot_16x4.hpp"
+#include "kernels/sve_hybrid_u8u32_dot_4VLx4.hpp"
 #include "kernels/sve_interleaved_u8u32_dot_3VLx8.hpp"
 #include "kernels/sve_native_u8u32_dot_4VLx4.hpp"
 
@@ -42,6 +43,13 @@
 static const GemmImplementation<uint8_t, uint32_t> gemm_u8_methods[] = {
 #ifdef __ARM_FEATURE_SVE
 {
+    GemmMethod::GEMM_HYBRID,
+    "hybrid_u8u32_dot_4VLx4",
+    [](const GemmArgs<uint32_t> &args) { return args._Ksize>=16 && args._alpha==1 && !args._trA && !args._trB && args._pretransposed_hint; },
+    [](const GemmArgs<uint32_t> &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); },
+    [](const GemmArgs<uint32_t> &args) { return new GemmHybrid<hybrid_u8u32_dot_4VLx4, uint8_t, uint32_t>(args); }
+},
+{
     GemmMethod::GEMM_NATIVE,
     "native_u8u32_dot_4VLx4",
     [](const GemmArgs<uint32_t> &args) { return (args._Ksize>=16 && args._alpha==1 && !args._trA && !args._trB); },
@@ -59,7 +67,7 @@
 {
     GemmMethod::GEMM_HYBRID,
     "hybrid_u8u32_dot_16x4",
-    [](const GemmArgs<uint32_t> &args) { return args._ci->has_dotprod() && args._Ksize>=16 && (args._Ksize % 16 == 0) && (args._Nsize % 16 == 0) && !args._trA && !args._trB && args._pretransposed_hint; },
+    [](const GemmArgs<uint32_t> &args) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; },
     [](const GemmArgs<uint32_t> &args) { return args._Nsize<=256 && args._Ksize>128; },
     [](const GemmArgs<uint32_t> &args) { return new GemmHybrid<hybrid_u8u32_dot_16x4, uint8_t, uint32_t>(args); }
 },
@@ -95,7 +103,7 @@
 template UniqueGemmCommon<uint8_t, uint32_t> gemm<uint8_t, uint32_t>(const GemmArgs<uint32_t> &args);
 template KernelDescription get_gemm_method<uint8_t, uint32_t>(const GemmArgs<uint32_t> &args);
 template bool method_is_compatible<uint8_t, uint32_t>(GemmMethod method, const GemmArgs<uint32_t> &args);
-template std::vector<std::string> get_compatible_kernels<uint8_t, uint32_t> (const GemmArgs<uint32_t> &args);
+template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint32_t> (const GemmArgs<uint32_t> &args);
 
 } // namespace arm_gemm