added detection & dispatching of some modern NEON instructions (NEON_FP16, NEON_BF16) (#24420)

* added more or less cross-platform (based on POSIX signal() semantics) method to detect various NEON extensions, such as FP16 SIMD arithmetics, BF16 SIMD arithmetics, SIMD dotprod etc. It could be propagated to other instruction sets if necessary.

* hopefully fixed compile errors

* continue to fix CI

* another attempt to fix build on Linux aarch64

* * reverted to the original method to detect special arm neon instructions without signal()
* renamed FP16_SIMD & BF16_SIMD to NEON_FP16 and NEON_BF16, respectively

* removed extra whitespaces

cmake/OpenCVCompilerOptimizations.cmake

@@ -331,7 +331,9 @@ 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 "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_dotprod.cpp")
+  ocv_update(CPU_NEON_FP16_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_neon_fp16.cpp")
+  ocv_update(CPU_NEON_BF16_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_neon_bf16.cpp")
+  ocv_update(CPU_NEON_DOTPROD_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_neon_dotprod.cpp")
   if(NOT AARCH64)
     ocv_update(CPU_KNOWN_OPTIMIZATIONS "VFPV3;NEON;FP16")
     if(NOT MSVC)
@@ -343,12 +345,17 @@ elseif(ARM OR AARCH64)
     endif()
     ocv_update(CPU_FP16_IMPLIES "NEON")
   else()
-    ocv_update(CPU_KNOWN_OPTIMIZATIONS "NEON;FP16;NEON_DOTPROD")
+    ocv_update(CPU_KNOWN_OPTIMIZATIONS "NEON;FP16;NEON_DOTPROD;NEON_FP16;NEON_BF16")
     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")
+    ocv_update(CPU_NEON_FP16_FLAGS_ON "-march=armv8.2-a+fp16")
+    ocv_update(CPU_NEON_FP16_IMPLIES "NEON")
+    ocv_update(CPU_NEON_BF16_FLAGS_ON "-march=armv8.2-a+fp16+bf16")
+    ocv_update(CPU_NEON_BF16_IMPLIES "NEON")
     set(CPU_BASELINE "NEON;FP16" CACHE STRING "${HELP_CPU_BASELINE}")
+    set(CPU_DISPATCH "NEON_FP16;NEON_BF16;NEON_DOTPROD" CACHE STRING "${HELP_CPU_DISPATCH}")
   endif()
 elseif(MIPS)
   ocv_update(CPU_MSA_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_msa.cpp")

cmake/checks/cpu_neon_bf16.cpp

@@ -0,0 +1,46 @@
+#if defined __GNUC__ && (defined __arm__ || defined __aarch64__)
+#include <stdio.h>
+#include "arm_neon.h"
+
+/*#if defined __clang__
+#pragma clang attribute push (__attribute__((target("bf16"))), apply_to=function)
+#elif defined GCC
+#pragma GCC push_options
+#pragma GCC target("armv8.2-a", "bf16")
+#endif*/
+bfloat16x8_t vld1q_as_bf16(const float* src)
+{
+    float32x4_t s0 = vld1q_f32(src), s1 = vld1q_f32(src + 4);
+    return vcombine_bf16(vcvt_bf16_f32(s0), vcvt_bf16_f32(s1));
+}
+
+void vprintreg(const char* name, const float32x4_t& r)
+{
+    float data[4];
+    vst1q_f32(data, r);
+    printf("%s: (%.2f, %.2f, %.2f, %.2f)\n",
+        name, data[0], data[1], data[2], data[3]);
+}
+
+void test()
+{
+    const float src1[] = { 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f };
+    const float src2[] = { 1.f, 3.f, 6.f, 10.f, 15.f, 21.f, 28.f, 36.f };
+    bfloat16x8_t s1 = vld1q_as_bf16(src1), s2 = vld1q_as_bf16(src2);
+    float32x4_t d = vbfdotq_f32(vdupq_n_f32(0.f), s1, s2);
+    vprintreg("(s1[0]*s2[0] + s1[1]*s2[1], ... s1[6]*s2[6] + s1[7]*s2[7])", d);
+}
+/*#if defined __clang__
+#pragma clang attribute pop
+#elif defined GCC
+#pragma GCC pop_options
+#endif*/
+#else
+#error "BF16 is not supported"
+#endif
+
+int main()
+{
+    test();
+    return 0;
+}

cmake/checks/cpu_neon_fp16.cpp

@@ -0,0 +1,46 @@
+#include <stdio.h>
+
+#if defined __GNUC__ && (defined __arm__ || defined __aarch64__)
+#include "arm_neon.h"
+
+float16x8_t vld1q_as_f16(const float* src)
+{
+    float32x4_t s0 = vld1q_f32(src), s1 = vld1q_f32(src + 4);
+    return vcombine_f16(vcvt_f16_f32(s0), vcvt_f16_f32(s1));
+}
+
+void vprintreg(const char* name, const float16x8_t& r)
+{
+    float data[8];
+    vst1q_f32(data, vcvt_f32_f16(vget_low_f16(r)));
+    vst1q_f32(data + 4, vcvt_f32_f16(vget_high_f16(r)));
+    printf("%s: (%.2f, %.2f, %.2f, %.2f, %.2f, %.2f, %.2f, %.2f)\n",
+        name, data[0], data[1], data[2], data[3],
+        data[4], data[5], data[6], data[7]);
+}
+
+void test()
+{
+    const float src1[] = { 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f };
+    const float src2[] = { 1.f, 3.f, 6.f, 10.f, 15.f, 21.f, 28.f, 36.f };
+    float16x8_t s1 = vld1q_as_f16(src1), s2 = vld1q_as_f16(src2);
+    float16x8_t d = vsubq_f16(s1, s1);
+    d = vfmaq_laneq_f16(d, s1, s2, 0);
+    d = vfmaq_laneq_f16(d, s1, s2, 1);
+    d = vfmaq_laneq_f16(d, s1, s2, 2);
+    d = vfmaq_laneq_f16(d, s1, s2, 3);
+    d = vfmaq_laneq_f16(d, s1, s2, 4);
+    d = vfmaq_laneq_f16(d, s1, s2, 5);
+    d = vfmaq_laneq_f16(d, s1, s2, 6);
+    d = vfmaq_laneq_f16(d, s1, s2, 7);
+    vprintreg("s1*s2[0]+s1*s2[1] + ... + s1*s2[7]", d);
+}
+#else
+#error "FP16 is not supported"
+#endif
+
+int main()
+{
+    test();
+    return 0;
+}

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

@@ -269,6 +269,8 @@ namespace cv {
 
 #define CV_CPU_NEON             100
 #define CV_CPU_NEON_DOTPROD     101
+#define CV_CPU_NEON_FP16        102
+#define CV_CPU_NEON_BF16        103
 
 #define CV_CPU_MSA              150
 
@@ -328,6 +330,8 @@ enum CpuFeatures {
 
     CPU_NEON            = 100,
     CPU_NEON_DOTPROD    = 101,
+    CPU_NEON_FP16       = 102,
+    CPU_NEON_BF16       = 103,
 
     CPU_MSA             = 150,
 

modules/core/src/system.cpp

@@ -408,6 +408,8 @@ struct HWFeatures
 
         g_hwFeatureNames[CPU_NEON] = "NEON";
         g_hwFeatureNames[CPU_NEON_DOTPROD] = "NEON_DOTPROD";
+        g_hwFeatureNames[CPU_NEON_FP16] = "NEON_FP16";
+        g_hwFeatureNames[CPU_NEON_BF16] = "NEON_BF16";
 
         g_hwFeatureNames[CPU_VSX] = "VSX";
         g_hwFeatureNames[CPU_VSX3] = "VSX3";
@@ -566,10 +568,15 @@ struct HWFeatures
 
             while ((size_t)read(cpufile, &auxv, size_auxv_t) == size_auxv_t)
             {
+                // see https://elixir.bootlin.com/linux/latest/source/arch/arm64/include/uapi/asm/hwcap.h
                 if (auxv.a_type == AT_HWCAP)
                 {
-                    have[CV_CPU_NEON_DOTPROD] = (auxv.a_un.a_val & (1 << 20)) != 0;
-                    break;
+                    have[CV_CPU_NEON_DOTPROD] = (auxv.a_un.a_val & (1 << 20)) != 0; // HWCAP_ASIMDDP
+                    have[CV_CPU_NEON_FP16] = (auxv.a_un.a_val & (1 << 10)) != 0; // HWCAP_ASIMDHP
+                }
+                else if (auxv.a_type == AT_HWCAP2)
+                {
+                    have[CV_CPU_NEON_BF16] = (auxv.a_un.a_val & (1 << 14)) != 0; // HWCAP2_BF16
                 }
             }
 
@@ -623,16 +630,23 @@ struct HWFeatures
         have[CV_CPU_NEON] = true;
     #endif
     #if (defined __ARM_FP  && (((__ARM_FP & 0x2) != 0) && defined __ARM_NEON__))
-        have[CV_CPU_FP16] = true;
+        have[CV_CPU_FP16] = have[CV_CPU_NEON_FP16] = true;
     #endif
-    #if (defined __ARM_FEATURE_DOTPROD)
+    // system.cpp may be compiled w/o special -march=armv8...+dotprod, -march=armv8...+bf16 etc.,
+    // so we check for the features in any case, no mater what are the compile flags.
+    // We check the real hardware capabilities here.
     int has_feat_dotprod = 0;
     size_t has_feat_dotprod_size = sizeof(has_feat_dotprod);
     sysctlbyname("hw.optional.arm.FEAT_DotProd", &has_feat_dotprod, &has_feat_dotprod_size, NULL, 0);
     if (has_feat_dotprod) {
         have[CV_CPU_NEON_DOTPROD] = true;
     }
-    #endif
+    int has_feat_bf16 = 0;
+    size_t has_feat_bf16_size = sizeof(has_feat_bf16);
+    sysctlbyname("hw.optional.arm.FEAT_BF16", &has_feat_bf16, &has_feat_bf16_size, NULL, 0);
+    if (has_feat_bf16) {
+        have[CV_CPU_NEON_BF16] = true;
+    }
     #elif (defined __clang__)
     #if (defined __ARM_NEON__ || (defined __ARM_NEON && defined __aarch64__))
         have[CV_CPU_NEON] = true;