From 1d480652b820317fc97ccbc3cb517e3b9e8be197 Mon Sep 17 00:00:00 2001
From: Georgios Pinitas <georgios.pinitas@arm.com>
Date: Wed, 23 Jan 2019 11:24:50 +0000
Subject: COMPMID-1867: Add u8 and s8 hybrid assembly kernels.

Change-Id: Ifeb005f9d18d19feff11949474cce84d9e03749c
Reviewed-on: https://review.mlplatform.org/565
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
---
 src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

(limited to '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 3c8df3f044..b95ca8016b 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
  *
@@ -27,11 +27,13 @@
 #include "gemm_common.hpp"
 #include "gemm_implementation.hpp"
 #include "gemm_interleaved.hpp"
+#include "gemm_hybrid.hpp"
 #include "gemm_native.hpp"
 
 #include "kernels/a64_gemm_u16_12x8.hpp"
 #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_interleaved_u8u32_dot_3VLx8.hpp"
 #include "kernels/sve_native_u8u32_dot_4VLx4.hpp"
 
@@ -54,6 +56,13 @@ static const GemmImplementation<uint8_t, uint32_t> gemm_u8_methods[] = {
     [](const GemmArgs<uint32_t> &args) { return new GemmInterleaved<interleaved_u8u32_dot_3VLx8, uint8_t, uint32_t>(args); }
 },
 #endif
+{
+    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._Nsize<=256 && args._Ksize>128; },
+    [](const GemmArgs<uint32_t> &args) { return new GemmHybrid<hybrid_u8u32_dot_16x4, uint8_t, uint32_t>(args); }
+},
 {
     GemmMethod::GEMM_INTERLEAVED,
     "gemm_u8_12x8",
-- 
cgit v1.2.1