Add amx option in CMakeLists; add amx detection and register header file

Signed-off-by: Molly Sophia <[email protected]>

Author: MollySophia

CMakeLists.txt

@@ -150,6 +150,10 @@ if(NOT (CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "
     endif()
 endif()
 
+if(MACOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm")
+    option(NCNN_APPLE_AMX "optimize apple silicon platforms with apple amx" ON)
+endif()
+
 if((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm")
     OR (APPLE AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
     OR (CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm|aarch64)")

cmake/ncnn_add_layer.cmake

@@ -341,6 +341,9 @@ macro(ncnn_add_layer class)
             if(NCNN_ARM86SVEF32MM)
                 ncnn_add_arch_opt_source(${class} svef32mm "-march=armv8.6-a+fp16+dotprod+sve+f32mm")
             endif()
+            if(NCNN_APPLE_AMX)
+                ncnn_add_arch_opt_source(${class} amx " ")
+            endif()
         endif()
     endif()
 

src/cpu.cpp

@@ -2378,6 +2378,21 @@ int cpu_support_arm_svef32mm()
 #endif
 }
 
+int cpu_support_arm_amx()
+{
+    try_initialize_global_cpu_info();
+#if __aarch64__ && __APPLE__
+    return g_hw_cpufamily == CPUFAMILY_ARM_FIRESTORM_ICESTORM
+           || g_hw_cpufamily == CPUFAMILY_ARM_AVALANCHE_BLIZZARD
+           || g_hw_cpufamily == CPUFAMILY_ARM_IBIZA
+           || g_hw_cpufamily == CPUFAMILY_ARM_LOBOS
+           || g_hw_cpufamily == CPUFAMILY_ARM_PALMA;
+
+#else
+    return 0;
+#endif
+}
+
 int cpu_support_x86_avx()
 {
     try_initialize_global_cpu_info();

src/cpu.h

@@ -80,6 +80,8 @@ NCNN_EXPORT int cpu_support_arm_svebf16();
 NCNN_EXPORT int cpu_support_arm_svei8mm();
 // svef32mm = aarch64 svef32mm
 NCNN_EXPORT int cpu_support_arm_svef32mm();
+// amx = aarch64 apple amx
+NCNN_EXPORT int cpu_support_arm_amx();
 
 // avx = x86 avx
 NCNN_EXPORT int cpu_support_x86_avx();

src/layer/arm/amx_usability.h

@@ -0,0 +1,50 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#ifndef AMX_USABILITY_H
+#define AMX_USABILITY_H
+
+// From https://github.com/corsix/amx/blob/main/aarch64.h
+#define AMX_NOP_OP_IMM5(op, imm5) \
+    __asm("nop\nnop\nnop\n.word (0x201000 + (%0 << 5) + %1)" : : "i"(op), "i"(imm5) : "memory")
+
+#define AMX_OP_GPR(op, gpr) \
+    __asm(".word (0x201000 + (%0 << 5) + 0%1 - ((0%1 >> 4) * 6))" : : "i"(op), "r"((uint64_t)(gpr)) : "memory")
+
+#define AMX_LDX(gpr)    AMX_OP_GPR( 0, gpr)
+#define AMX_LDY(gpr)    AMX_OP_GPR( 1, gpr)
+#define AMX_STX(gpr)    AMX_OP_GPR( 2, gpr)
+#define AMX_STY(gpr)    AMX_OP_GPR( 3, gpr)
+#define AMX_LDZ(gpr)    AMX_OP_GPR( 4, gpr)
+#define AMX_STZ(gpr)    AMX_OP_GPR( 5, gpr)
+#define AMX_LDZI(gpr)   AMX_OP_GPR( 6, gpr)
+#define AMX_STZI(gpr)   AMX_OP_GPR( 7, gpr)
+#define AMX_EXTRX(gpr)  AMX_OP_GPR( 8, gpr)
+#define AMX_EXTRY(gpr)  AMX_OP_GPR( 9, gpr)
+#define AMX_FMA64(gpr)  AMX_OP_GPR(10, gpr)
+#define AMX_FMS64(gpr)  AMX_OP_GPR(11, gpr)
+#define AMX_FMA32(gpr)  AMX_OP_GPR(12, gpr)
+#define AMX_FMS32(gpr)  AMX_OP_GPR(13, gpr)
+#define AMX_MAC16(gpr)  AMX_OP_GPR(14, gpr)
+#define AMX_FMA16(gpr)  AMX_OP_GPR(15, gpr)
+#define AMX_FMS16(gpr)  AMX_OP_GPR(16, gpr)
+#define AMX_SET()       AMX_NOP_OP_IMM5(17, 0)
+#define AMX_CLR()       AMX_NOP_OP_IMM5(17, 1)
+#define AMX_VECINT(gpr) AMX_OP_GPR(18, gpr)
+#define AMX_VECFP(gpr)  AMX_OP_GPR(19, gpr)
+#define AMX_MATINT(gpr) AMX_OP_GPR(20, gpr)
+#define AMX_MATFP(gpr)  AMX_OP_GPR(21, gpr)
+#define AMX_GENLUT(gpr) AMX_OP_GPR(22, gpr)
+
+#endif // AMX_USABILITY_H