From 146138378c1587b7297d245b7177641315f6180b Mon Sep 17 00:00:00 2001
From: Georgios Pinitas <georgios.pinitas@arm.com>
Date: Fri, 1 Mar 2019 19:07:11 +0000
Subject: 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_int8.cpp | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

(limited to 'src/core/NEON/kernels/arm_gemm/gemm_int8.cpp')

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
index 5811c2a1ce..9e49df1c28 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.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_s8_12x8.hpp"
 #include "kernels/a64_gemm_s8_4x4.hpp"
 #include "kernels/a64_hybrid_s8s32_dot_16x4.hpp"
+#include "kernels/sve_hybrid_s8s32_dot_4VLx4.hpp"
 #include "kernels/sve_interleaved_s8s32_dot_3VLx8.hpp"
 #include "kernels/sve_native_s8s32_dot_4VLx4.hpp"
 
@@ -41,6 +42,13 @@ namespace arm_gemm {
 
 static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = {
 #ifdef __ARM_FEATURE_SVE
+{
+    GemmMethod::GEMM_HYBRID,
+    "hybrid_s8s32_dot_4VLx4",
+    [](const GemmArgs<int32_t> &args) { return args._Ksize>=16 && args._alpha==1 && !args._trA && !args._trB && args._pretransposed_hint; },
+    [](const GemmArgs<int32_t> &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); },
+    [](const GemmArgs<int32_t> &args) { return new GemmHybrid<hybrid_s8s32_dot_4VLx4, int8_t, int32_t>(args); }
+},
 {
     GemmMethod::GEMM_NATIVE,
     "native_s8s32_dot_4VLx4",
@@ -59,7 +67,7 @@ static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = {
 {
     GemmMethod::GEMM_HYBRID,
     "hybrid_s8s32_dot_16x4",
-    [](const GemmArgs<int32_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<int32_t> &args) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; },
     [](const GemmArgs<int32_t> &args) { return args._Nsize<=256 && args._Ksize>128; },
     [](const GemmArgs<int32_t> &args) { return new GemmHybrid<hybrid_s8s32_dot_16x4, int8_t, int32_t>(args); }
 },
@@ -95,7 +103,7 @@ const GemmImplementation<int8_t, int32_t> *gemm_implementation_list<int8_t, int3
 template UniqueGemmCommon<int8_t, int32_t> gemm<int8_t, int32_t>(const GemmArgs<int32_t> &args);
 template KernelDescription get_gemm_method<int8_t, int32_t>(const GemmArgs<int32_t> &args);
 template bool method_is_compatible<int8_t, int32_t>(GemmMethod method, const GemmArgs<int32_t> &args);
-template std::vector<std::string> get_compatible_kernels<int8_t, int32_t> (const GemmArgs<int32_t> &args);
+template std::vector<KernelDescription> get_compatible_kernels<int8_t, int32_t> (const GemmArgs<int32_t> &args);
 
 } // namespace arm_gemm
 
-- 
cgit v1.2.1