From 778b95cb755880ab4adc972fbf3b7022a99b63f9 Mon Sep 17 00:00:00 2001
From: Michalis Spyrou <michalis.spyrou@arm.com>
Date: Tue, 20 Apr 2021 12:15:52 +0100
Subject: Update assembly code

This patch brings performance uplift on Cortex-A35.

Resolves: COMPMID-4316

Change-Id: I2b9c02a599373f780dd1b981b821e33bd59a3422
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5461
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
---
 src/core/NEON/kernels/arm_gemm/asmlib.hpp          |   1 +
 src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp       |  12 +-
 .../kernels/arm_gemm/kernels/a64_sgemm_8x6.hpp     |  75 ++++
 .../arm_gemm/kernels/a64_sgemm_8x6/generic.cpp     | 381 +++++++++++++++++++++
 src/core/NEON/kernels/arm_gemm/mergeresults.cpp    |   9 +-
 5 files changed, 474 insertions(+), 4 deletions(-)
 create mode 100644 src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6.hpp
 create mode 100644 src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp

(limited to 'src/core/NEON/kernels')

diff --git a/src/core/NEON/kernels/arm_gemm/asmlib.hpp b/src/core/NEON/kernels/arm_gemm/asmlib.hpp
index 00674334a9..7766656adb 100644
--- a/src/core/NEON/kernels/arm_gemm/asmlib.hpp
+++ b/src/core/NEON/kernels/arm_gemm/asmlib.hpp
@@ -32,6 +32,7 @@
 
 // "Correct" version
 #define ASM_PREFETCH(address)    "PRFM PLDL1KEEP, " address "\n"
+#define ASM_PREFETCHU(address)   "PRFUM PLDL1KEEP, " address "\n"
 #define ASM_PREFETCHL2(address)  "PRFM PLDL2KEEP, " address "\n"
 #define ASM_PREFETCHW(address)   "PRFM PSTL1KEEP, " address "\n"
 #define ASM_PREFETCHWL2(address) "PRFM PSTL2KEEP, " address "\n"
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
index 6c3743dce7..d94814fb4c 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,6 +35,7 @@
 #include "kernels/a64_hybrid_fp32_mla_6x16.hpp"
 #include "kernels/a64_hybrid_fp32_mla_8x4.hpp"
 #include "kernels/a64_sgemm_8x12.hpp"
+#include "kernels/a64_sgemm_8x6.hpp"
 #include "kernels/a64_smallK_hybrid_fp32_mla_6x4.hpp"
 #include "kernels/a64_smallK_hybrid_fp32_mla_8x4.hpp"
 
@@ -110,7 +111,14 @@ static const GemmImplementation<float, float> gemm_fp32_methods[] =
     [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float>(args); }
 },
 #endif // __ARM_FEATURE_SVE
-
+// Cortex-A35 specific kernel - use for any problem on A35, and never in any other cases.
+{
+    GemmMethod::GEMM_INTERLEAVED,
+    "a64_sgemm_8x6",
+    nullptr,
+    [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A35; },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x6, float, float>(args); }
+},
 // Arm® Neon™ hybrid methods
 {
     GemmMethod::GEMM_HYBRID,
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6.hpp
new file mode 100644
index 0000000000..c1318a2a06
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6.hpp
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 2017-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#pragma once
+
+#ifdef __aarch64__
+
+#include "../std_transforms_fixed.hpp"
+
+namespace arm_gemm {
+
+// Actual kernel implementations
+void a64_sgemm_asimd_8x6(const float *, const float *, float *, int, int, int);
+
+// 8x6 SGEMM "strategy" class.
+//
+// This describes the characteristics of a family of kernels, in terms of
+// the required interleave properties and the output block size.
+//
+// All kernels in the family must share these characteristics.  The actual
+// kernel to be used can be chosen at runtime, based on the CPU_type
+// structure.
+class cls_a64_sgemm_8x6 {
+public:
+    typedef float operand_type;
+    typedef float result_type;
+
+    typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
+
+    /* Kernel blocking parameters */
+    static unsigned int out_width() {
+        return 6;
+    }
+
+    static unsigned int out_height() {
+        return 8;
+    }
+
+    static unsigned int k_unroll() {
+        return 1;
+    }
+
+    // Use the standard fixed size transforms.
+    StdTransformsFixed<operand_type, result_type, 8, 6, 1> transforms = {};
+
+    kern_type kernel=a64_sgemm_asimd_8x6;
+
+    cls_a64_sgemm_8x6(const CPUInfo *) {
+
+    }
+};
+
+} // namespace arm_gemm
+
+#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp
new file mode 100644
index 0000000000..9b81374d2d
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp
@@ -0,0 +1,381 @@
+/*
+ * Copyright (c) 2017-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifdef __aarch64__
+
+#include <arm_neon.h>
+
+#include "../../asmlib.hpp"
+
+// Kernel implementation.
+//
+// Assume that "Apanel" points to a chunk of A blocks (each size 8xK) in read-order.
+// Assume that "Bpanel" points to a chunk of B blocks (each size 12xK) in read-order.
+// Assume that "Cpanel" points to a chunk of C output blocks (each size
+// 12x8), the chunks being arranged in a row major fashion.
+//
+// Note that the intent of this is that either ablocks or bblocks will be 1
+// - this construction allows the output loop to proceed in either order.
+
+namespace arm_gemm {
+
+void a64_sgemm_asimd_8x6(const float *Apanel, const float *Bpanel, float *Cpanel, int ablocks, int bblocks, int K) {
+    const float *a_ptr = Apanel;
+    float *c_ptr = Cpanel;
+
+    for (int yb=0; yb<ablocks; yb++) {
+        const float *a_ptr0 = a_ptr;
+        const float *b_ptr = Bpanel;
+
+        for (int xb=0; xb<bblocks; xb++) {
+            a_ptr = a_ptr0;
+            // Fix up for odd lengths - set a flag if K is odd, but make
+            // sure we round up the iteration count.
+            int oddk = (K & 1);
+            int k = ((K+1)/2) - 1;
+
+            register float32x4_t a0  asm("v0");
+            register float32x4_t a1  asm("v1");
+            register float32x4_t a2  asm("v2");
+            register float32x4_t a3  asm("v3");
+            register float32x4_t b0  asm("v4");
+            register float32x4_t b1  asm("v5");
+            register float32x4_t b2  asm("v6");
+
+            __asm __volatile (
+                // Initialize result registers, load initial operands, prime prefetches.
+                "movi	v8.2s, #0x0\n"
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "movi	v9.2s, #0x0\n"
+                "movi	v10.2s, #0x0\n"
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "movi	v11.2s, #0x0\n"
+                "movi	v12.2s, #0x0\n"
+                "movi	v13.2s, #0x0\n"
+                "movi	v14.2s, #0x0\n"
+                ASM_PREFETCH("[%[b_ptr], #64]")
+                ASM_PREFETCHU("[%[a_ptr], #52]")
+                ASM_PREFETCHU("[%[a_ptr], #116]")
+                ASM_PREFETCH("[%[b_ptr], #128]")
+                "movi	v15.2s, #0x0\n"
+                "movi	v16.2s, #0x0\n"
+                "movi	v17.2s, #0x0\n"
+                "movi	v18.2s, #0x0\n"
+                "movi	v19.2s, #0x0\n"
+                "movi	v20.2s, #0x0\n"
+                "movi	v21.2s, #0x0\n"
+                "movi	v22.2s, #0x0\n"
+                "movi	v23.2s, #0x0\n"
+                "movi	v24.2s, #0x0\n"
+                "movi	v25.2s, #0x0\n"
+                "movi	v26.2s, #0x0\n"
+                "movi	v27.2s, #0x0\n"
+                "movi	v28.2s, #0x0\n"
+                "movi	v29.2s, #0x0\n"
+                "movi	v30.2s, #0x0\n"
+                "movi	v31.2s, #0x0\n"
+
+                // Skip loop if we are doing zero iterations of it.
+                "cbz	%w[k], 4f\n"
+
+                // Loop proper
+                "1:\n"
+                "ldr	%d[b0], [%[b_ptr], #0]\n"
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "ldr	%d[b1], [%[b_ptr], #8]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #16]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                 ASM_PREFETCH("[%[b_ptr], #128]")
+                "subs	%w[k], %w[k], #1\n"
+                 ASM_PREFETCHU("[%[a_ptr], #156]")
+                "ldr	%d[b0], [%[b_ptr], #24]\n"
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+
+                "ldr	%d[b1], [%[b_ptr], #32]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #40]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                 ASM_PREFETCHU("[%[a_ptr], #188]")
+                "bne	1b\n"
+
+                // Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
+                "4:\n"
+                ASM_PREFETCH("[%[c_ptr]]")
+                ASM_PREFETCH("[%[c_ptr], #64]")
+
+                "ldr	%d[b0], [%[b_ptr]]\n"
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+
+                // Branch to alternative tail for odd K
+                "cbnz	%w[oddk], 2f\n"
+
+                // Detached final iteration (even K)
+                "ldr	%d[b1], [%[b_ptr], #8]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #16]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "ldr	%d[b0], [%[b_ptr], #24]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                 ASM_PREFETCH("[%[b_ptr], #128]")
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+
+                "ldr	%d[b1], [%[b_ptr], #-16]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #-8]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "b	3f\n"
+
+                // Detached final iteration (odd K)
+                "2:\n"
+                "ldr	%d[b1], [%[b_ptr], #8]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #16]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "add	%[b_ptr], %[b_ptr], #24\n"
+
+                // Common tail
+                "3:\n"
+                "str	d8, [%[c_ptr], #0]\n"
+                "str	d16, [%[c_ptr], #8]\n"
+                "str	d24, [%[c_ptr], #16]\n"
+                "str	d9, [%[c_ptr], #24]\n"
+                "str	d17, [%[c_ptr], #32]\n"
+                "str	d25, [%[c_ptr], #40]\n"
+                "str	d10, [%[c_ptr], #48]\n"
+                "str	d18, [%[c_ptr], #56]\n"
+                "str	d26, [%[c_ptr], #64]\n"
+                "str	d11, [%[c_ptr], #72]\n"
+                "str	d19, [%[c_ptr], #80]\n"
+                "str	d27, [%[c_ptr], #88]\n"
+                "str	d12, [%[c_ptr], #96]\n"
+                "str	d20, [%[c_ptr], #104]\n"
+                "str	d28, [%[c_ptr], #112]\n"
+                "str	d13, [%[c_ptr], #120]\n"
+                "str	d21, [%[c_ptr], #128]\n"
+                "str	d29, [%[c_ptr], #136]\n"
+                "str	d14, [%[c_ptr], #144]\n"
+                "str	d22, [%[c_ptr], #152]\n"
+                "str	d30, [%[c_ptr], #160]\n"
+                "str	d15, [%[c_ptr], #168]\n"
+                "str	d23, [%[c_ptr], #176]\n"
+                "str	d31, [%[c_ptr], #184]\n"
+                "add	%[c_ptr], %[c_ptr], #192\n"
+
+            :
+              [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
+              [a0] "+w" (a0), [a1] "+w" (a1), [a2] "+w" (a2), [a3] "+w" (a3),
+              [b0] "+w" (b0), [b1] "+w" (b1), [b2] "+w" (b2), [k] "+r" (k)
+            : [oddk] "r" (oddk)
+            : "x20", "x21", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18",
+              "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "cc", "memory"
+            );
+        }
+    }
+}
+
+} // namespace arm_gemm
+
+#endif
diff --git a/src/core/NEON/kernels/arm_gemm/mergeresults.cpp b/src/core/NEON/kernels/arm_gemm/mergeresults.cpp
index 8ca947a6dc..adcdc88fd9 100644
--- a/src/core/NEON/kernels/arm_gemm/mergeresults.cpp
+++ b/src/core/NEON/kernels/arm_gemm/mergeresults.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 Arm Limited.
+ * Copyright (c) 2017-2018, 2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,7 +25,6 @@
 /* As some of the merges need these headers, but are all included in the
  * arm_gemm namespace, put these headers here.  */
 #include <algorithm>
-#include <limits>
 
 #include <arm_neon.h>
 
@@ -97,6 +96,12 @@ void MergeResults(Tout * out, const Tin * in, int ldc, int y0, int ymax, int x0,
 
 #include "merges/list.hpp"
 
+/* Cortex-A53 8x6 SGEMM kernel uses a templated merge as the optimized merge
+ * generator cannot cope with the width (6) not being a multiple of VL (4). */
+#ifdef __aarch64__
+template void MergeResults<6u, 8u, false, float, float>(float *, float const*, int, int, int, int, int, float const *, Activation, bool);
+#endif
+
 #if defined(__aarch64__) && defined(__ARM_FP16_ARGS)
 template void MergeResults<12u, 8u, false, float, __fp16>(__fp16*, float const*, int, int, int, int, int, __fp16 const*, Activation, bool);
 #endif
-- 
cgit v1.2.1