neon: add dotprod dispatch implementation

* read vector at runtime
     * add enum

cmake/OpenCVCompilerOptimizations.cmake

@@ -46,7 +46,7 @@
 
 set(CPU_ALL_OPTIMIZATIONS "SSE;SSE2;SSE3;SSSE3;SSE4_1;SSE4_2;POPCNT;AVX;FP16;AVX2;FMA3;AVX_512F")
 list(APPEND CPU_ALL_OPTIMIZATIONS "AVX512_COMMON;AVX512_KNL;AVX512_KNM;AVX512_SKX;AVX512_CNL;AVX512_CLX;AVX512_ICL")
-list(APPEND CPU_ALL_OPTIMIZATIONS NEON VFPV3 FP16)
+list(APPEND CPU_ALL_OPTIMIZATIONS NEON VFPV3 FP16 NEON_DOTPROD)
 list(APPEND CPU_ALL_OPTIMIZATIONS MSA)
 list(APPEND CPU_ALL_OPTIMIZATIONS VSX VSX3)
 list(REMOVE_DUPLICATES CPU_ALL_OPTIMIZATIONS)
@@ -326,6 +326,7 @@ if(X86 OR X86_64)
 elseif(ARM OR AARCH64)
   ocv_update(CPU_NEON_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_neon.cpp")
   ocv_update(CPU_FP16_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_fp16.cpp")
+  ocv_update(CPU_NEON_DOTPROD_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_dotprod.cpp")
   if(NOT AARCH64)
     ocv_update(CPU_KNOWN_OPTIMIZATIONS "VFPV3;NEON;FP16")
     if(NOT MSVC)
@@ -337,9 +338,11 @@ elseif(ARM OR AARCH64)
     endif()
     ocv_update(CPU_FP16_IMPLIES "NEON")
   else()
-    ocv_update(CPU_KNOWN_OPTIMIZATIONS "NEON;FP16")
+    ocv_update(CPU_KNOWN_OPTIMIZATIONS "NEON;FP16;NEON_DOTPROD")
     ocv_update(CPU_NEON_FLAGS_ON "")
     ocv_update(CPU_FP16_IMPLIES "NEON")
+    ocv_update(CPU_NEON_DOTPROD_FLAGS_ON "-march=armv8.2-a+dotprod")
+    ocv_update(CPU_NEON_DOTPROD_IMPLIES "NEON")
     set(CPU_BASELINE "NEON;FP16" CACHE STRING "${HELP_CPU_BASELINE}")
   endif()
 elseif(MIPS)

cmake/checks/cpu_dotprod.cpp

@@ -0,0 +1,24 @@
+#include <stdio.h>
+
+#if defined __GNUC__ && (defined __arm__ || defined __aarch64__)
+#include "arm_neon.h"
+int test()
+{
+    const unsigned int src[] = { 0, 0, 0, 0 };
+    unsigned int dst[4];
+    uint32x4_t v_src = *(uint32x4_t*)src;
+    uint8x16_t v_m0 = *(uint8x16_t*)src;
+    uint8x16_t v_m1 = *(uint8x16_t*)src;
+    uint32x4_t v_dst = vdotq_u32(v_src, v_m0, v_m1);
+    *(uint32x4_t*)dst = v_dst;
+    return (int)dst[0];
+}
+#else
+#error "DOTPROD is not supported"
+#endif
+
+int main()
+{
+  printf("%d\n", test());
+  return 0;
+}

modules/core/CMakeLists.txt

@@ -6,7 +6,7 @@ ocv_add_dispatched_file(arithm SSE2 SSE4_1 AVX2 VSX3)
 ocv_add_dispatched_file(convert SSE2 AVX2 VSX3)
 ocv_add_dispatched_file(convert_scale SSE2 AVX2)
 ocv_add_dispatched_file(count_non_zero SSE2 AVX2)
-ocv_add_dispatched_file(matmul SSE2 SSE4_1 AVX2 AVX512_SKX)
+ocv_add_dispatched_file(matmul SSE2 SSE4_1 AVX2 AVX512_SKX NEON_DOTPROD)
 ocv_add_dispatched_file(mean SSE2 AVX2)
 ocv_add_dispatched_file(merge SSE2 AVX2)
 ocv_add_dispatched_file(split SSE2 AVX2)

modules/core/include/opencv2/core/cv_cpu_dispatch.h

@@ -79,6 +79,10 @@
 #  endif
 #  define CV_FP16 1
 #endif
+#ifdef CV_CPU_COMPILE_NEON_DOTPROD
+#  include <arm_neon.h>
+#  define CV_NEON_DOT 1
+#endif
 #ifdef CV_CPU_COMPILE_AVX2
 #  include <immintrin.h>
 #  define CV_AVX2 1

modules/core/include/opencv2/core/cv_cpu_helper.h

@@ -420,6 +420,27 @@
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_NEON(fn, args, mode, ...)  CV_CPU_CALL_NEON(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
+#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_NEON_DOTPROD
+#  define CV_TRY_NEON_DOTPROD 1
+#  define CV_CPU_FORCE_NEON_DOTPROD 1
+#  define CV_CPU_HAS_SUPPORT_NEON_DOTPROD 1
+#  define CV_CPU_CALL_NEON_DOTPROD(fn, args) return (cpu_baseline::fn args)
+#  define CV_CPU_CALL_NEON_DOTPROD_(fn, args) return (opt_NEON_DOTPROD::fn args)
+#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_NEON_DOTPROD
+#  define CV_TRY_NEON_DOTPROD 1
+#  define CV_CPU_FORCE_NEON_DOTPROD 0
+#  define CV_CPU_HAS_SUPPORT_NEON_DOTPROD (cv::checkHardwareSupport(CV_CPU_NEON_DOTPROD))
+#  define CV_CPU_CALL_NEON_DOTPROD(fn, args) if (CV_CPU_HAS_SUPPORT_NEON_DOTPROD) return (opt_NEON_DOTPROD::fn args)
+#  define CV_CPU_CALL_NEON_DOTPROD_(fn, args) if (CV_CPU_HAS_SUPPORT_NEON_DOTPROD) return (opt_NEON_DOTPROD::fn args)
+#else
+#  define CV_TRY_NEON_DOTPROD 0
+#  define CV_CPU_FORCE_NEON_DOTPROD 0
+#  define CV_CPU_HAS_SUPPORT_NEON_DOTPROD 0
+#  define CV_CPU_CALL_NEON_DOTPROD(fn, args)
+#  define CV_CPU_CALL_NEON_DOTPROD_(fn, args)
+#endif
+#define __CV_CPU_DISPATCH_CHAIN_NEON_DOTPROD(fn, args, mode, ...)  CV_CPU_CALL_NEON_DOTPROD(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
+
 #if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_MSA
 #  define CV_TRY_MSA 1
 #  define CV_CPU_FORCE_MSA 1

modules/core/include/opencv2/core/cvdef.h

@@ -282,6 +282,7 @@ namespace cv {
 #define CV_CPU_AVX_5124FMAPS    27
 
 #define CV_CPU_NEON             100
+#define CV_CPU_NEON_DOTPROD     101
 
 #define CV_CPU_MSA              150
 
@@ -334,6 +335,7 @@ enum CpuFeatures {
     CPU_AVX_5124FMAPS   = 27,
 
     CPU_NEON            = 100,
+    CPU_NEON_DOTPROD    = 101,
 
     CPU_MSA             = 150,
 

modules/core/include/opencv2/core/hal/intrin_neon.hpp

@@ -78,8 +78,6 @@ CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
 #define CV_NEON_AARCH64 0
 #endif
 
-// TODO
-#define CV_NEON_DOT 0
 
 //////////// Utils ////////////
 
@@ -667,11 +665,22 @@ inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b, const v_int64
 }
 
 // 8 >> 32
+#ifdef CV_NEON_DOT
+#define OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_OP(_Tpvec1, _Tpvec2, suffix) \
+inline _Tpvec1 v_dotprod_expand(const _Tpvec2& a, const _Tpvec2& b)   \
+{ \
+    return _Tpvec1(vdotq_##suffix(vdupq_n_##suffix(0), a.val, b.val));\
+} \
+inline _Tpvec1 v_dotprod_expand(const _Tpvec2& a, const _Tpvec2& b, const _Tpvec1& c) \
+{ \
+    return _Tpvec1(vdotq_##suffix(c.val, a.val, b.val)); \
+}
+
+OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_OP(v_uint32x4, v_uint8x16, u32)
+OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_OP(v_int32x4,  v_int8x16,  s32)
+#else
 inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
 {
-#if CV_NEON_DOT
-    return v_uint32x4(vdotq_u32(vdupq_n_u32(0), a.val, b.val));
-#else
     const uint8x16_t zero   = vreinterpretq_u8_u32(vdupq_n_u32(0));
     const uint8x16_t mask   = vreinterpretq_u8_u32(vdupq_n_u32(0x00FF00FF));
     const uint16x8_t zero32 = vreinterpretq_u16_u32(vdupq_n_u32(0));
@@ -687,23 +696,15 @@ inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
     uint32x4_t s1 = vaddq_u32(vshrq_n_u32(vreinterpretq_u32_u16(even), 16),
                               vshrq_n_u32(vreinterpretq_u32_u16(odd),  16));
     return v_uint32x4(vaddq_u32(s0, s1));
-#endif
 }
 inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b,
                                    const v_uint32x4& c)
 {
-#if CV_NEON_DOT
-    return v_uint32x4(vdotq_u32(c.val, a.val, b.val));
-#else
     return v_dotprod_expand(a, b) + c;
-#endif
 }
 
 inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
 {
-#if CV_NEON_DOT
-    return v_int32x4(vdotq_s32(vdupq_n_s32(0), a.val, b.val));
-#else
     int16x8_t p0  = vmull_s8(vget_low_s8(a.val), vget_low_s8(b.val));
     int16x8_t p1  = vmull_s8(vget_high_s8(a.val), vget_high_s8(b.val));
     int16x8_t uzp1, uzp2;
@@ -712,18 +713,13 @@ inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
     int16x4_t uzpl1, uzpl2;
     _v128_unzip(vget_low_s16(sum), vget_high_s16(sum), uzpl1, uzpl2);
     return v_int32x4(vaddl_s16(uzpl1, uzpl2));
-#endif
 }
 inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b,
                                   const v_int32x4& c)
 {
-#if CV_NEON_DOT
-    return v_int32x4(vdotq_s32(c.val, a.val, b.val));
-#else
     return v_dotprod_expand(a, b) + c;
-#endif
 }
-
+#endif
 // 16 >> 64
 inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b)
 {
@@ -832,45 +828,44 @@ inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_
 }
 
 // 8 >> 32
+#ifdef CV_NEON_DOT
+#define OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_FAST_OP(_Tpvec1, _Tpvec2, suffix) \
+inline _Tpvec1 v_dotprod_expand_fast(const _Tpvec2& a, const _Tpvec2& b)   \
+{ \
+    return v_dotprod_expand(a, b); \
+} \
+inline _Tpvec1 v_dotprod_expand_fast(const _Tpvec2& a, const _Tpvec2& b, const _Tpvec1& c) \
+{ \
+    return v_dotprod_expand(a, b, c); \
+}
+
+OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_FAST_OP(v_uint32x4, v_uint8x16, u32)
+OPENCV_HAL_IMPL_NEON_DOT_PRODUCT_FAST_OP(v_int32x4,  v_int8x16,  s32)
+#else
 inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b)
 {
-#if CV_NEON_DOT
-    return v_uint32x4(vdotq_u32(vdupq_n_u32(0), a.val, b.val));
-#else
     uint16x8_t p0 = vmull_u8(vget_low_u8(a.val), vget_low_u8(b.val));
     uint16x8_t p1 = vmull_u8(vget_high_u8(a.val), vget_high_u8(b.val));
     uint32x4_t s0 = vaddl_u16(vget_low_u16(p0), vget_low_u16(p1));
     uint32x4_t s1 = vaddl_u16(vget_high_u16(p0), vget_high_u16(p1));
     return v_uint32x4(vaddq_u32(s0, s1));
-#endif
 }
 inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
 {
-#if CV_NEON_DOT
-    return v_uint32x4(vdotq_u32(c.val, a.val, b.val));
-#else
     return v_dotprod_expand_fast(a, b) + c;
-#endif
 }
 
 inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b)
 {
-#if CV_NEON_DOT
-    return v_int32x4(vdotq_s32(vdupq_n_s32(0), a.val, b.val));
-#else
     int16x8_t prod = vmull_s8(vget_low_s8(a.val), vget_low_s8(b.val));
     prod = vmlal_s8(prod, vget_high_s8(a.val), vget_high_s8(b.val));
     return v_int32x4(vaddl_s16(vget_low_s16(prod), vget_high_s16(prod)));
-#endif
 }
 inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
 {
-#if CV_NEON_DOT
-    return v_int32x4(vdotq_s32(c.val, a.val, b.val));
-#else
     return v_dotprod_expand_fast(a, b) + c;
-#endif
 }
+#endif
 
 // 16 >> 64
 inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b)

modules/core/src/system.cpp

@@ -411,6 +411,7 @@ struct HWFeatures
         g_hwFeatureNames[CPU_AVX_5124FMAPS] = "AVX5124FMAPS";
 
         g_hwFeatureNames[CPU_NEON] = "NEON";
+        g_hwFeatureNames[CPU_NEON_DOTPROD] = "NEON_DOTPROD";
 
         g_hwFeatureNames[CPU_VSX] = "VSX";
         g_hwFeatureNames[CPU_VSX3] = "VSX3";
@@ -555,6 +556,24 @@ struct HWFeatures
     #ifdef __aarch64__
         have[CV_CPU_NEON] = true;
         have[CV_CPU_FP16] = true;
+        int cpufile = open("/proc/self/auxv", O_RDONLY);
+
+        if (cpufile >= 0)
+        {
+            Elf64_auxv_t auxv;
+            const size_t size_auxv_t = sizeof(auxv);
+
+            while ((size_t)read(cpufile, &auxv, size_auxv_t) == size_auxv_t)
+            {
+                if (auxv.a_type == AT_HWCAP)
+                {
+                    have[CV_CPU_NEON_DOTPROD] = (auxv.a_un.a_val & (1 << 20)) != 0;
+                    break;
+                }
+            }
+
+            close(cpufile);
+        }
     #elif defined __arm__ && defined __ANDROID__
       #if defined HAVE_CPUFEATURES
         CV_LOG_INFO(NULL, "calling android_getCpuFeatures() ...");