https://github.com/opencv/opencv/pull/24203

From 689fa6f372975d58e9f50fd17a0abd105b1815f1 Mon Sep 17 00:00:00 2001
From: Sam James <sam@gentoo.org>
Date: Mon, 28 Aug 2023 04:20:58 +0100
Subject: [PATCH] Fix compilation on arm64 with FP16 when disabled

If building with -mcpu=native or any other setting which implies the current
CPU has FP16 but with intrinsics disabled, we mistakenly try to use it even
though convolution.hpp conditionally defines it correctly based on whether
we should *use it*. convolution.cpp on the other hand was mismatched and
trying to use it if the CPU supported it, even if not enabled in the build
system.

Make the guards match.

Bug: https://bugs.gentoo.org/913031
Signed-off-by: Sam James <sam@gentoo.org>
--- a/modules/dnn/src/layers/cpu_kernels/convolution.cpp
+++ b/modules/dnn/src/layers/cpu_kernels/convolution.cpp
@@ -118,7 +118,7 @@ Ptr<FastConv> initFastConv(
     const size_t wstep = weightsMat.step1();
 
     conv->useFP16 = false;
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
     // TODO: add FP16 support for Winograd.
     if (_useFP16 && (conv->conv_type == CONV_TYPE_GENERIC || conv->conv_type == CONV_TYPE_DEPTHWISE_REMAIN))
         conv->useFP16 = true;
@@ -137,7 +137,7 @@ Ptr<FastConv> initFastConv(
         int padded_ksize = ((ksize + VEC_ALIGN-1) / VEC_ALIGN) * VEC_ALIGN;
         int nweights = C * padded_ksize;
 
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
         if (conv->useFP16)
         {
             conv->weightsBuf_FP16.resize(nweights + VEC_ALIGN);
@@ -190,7 +190,7 @@ Ptr<FastConv> initFastConv(
 #endif
         const int CONV_WINO_NATOMS_F32 = CONV_WINO_AREA / CONV_WINO_ATOM_F32; // for AVX2, it is 8, otherwise, it's 16.
 
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
         // FP 16
         const int CONV_WINO_ATOM_F16 = CONV_WINO_ATOM_F32 * 2;
         const int CONV_WINO_NATOMS_F16 = CONV_WINO_AREA / CONV_WINO_ATOM_F16;
@@ -208,7 +208,7 @@ Ptr<FastConv> initFastConv(
         size_t nweights = ngroups*Kg_nblocks*Cg*CONV_WINO_KBLOCK*CONV_WINO_AREA;
 
         float* wptrWino = nullptr;
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
         float16_t* wptrWino_FP16 = nullptr;
         if (conv->useFP16)
         {
@@ -264,7 +264,7 @@ Ptr<FastConv> initFastConv(
                 }
 
                 // repack the data.
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                 if (conv->useFP16)
                 {
                     float16_t* wptr = wptrWino_FP16 + (g*Kg_nblocks + ki) * Cg *CONV_WINO_KBLOCK*CONV_WINO_AREA +
@@ -308,7 +308,7 @@ Ptr<FastConv> initFastConv(
 
         float* weightsBufPtr = nullptr;
 
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
         int numStripsMR_FP16 = (Kg + CONV_MR_FP16 - 1) / CONV_MR_FP16;
         int Kg_aligned_FP16 = numStripsMR_FP16 * CONV_MR_FP16;
         size_t nweights_FP16 = ngroups * Kg_aligned_FP16 * DkHkWkCg;
@@ -331,7 +331,7 @@ Ptr<FastConv> initFastConv(
         }
 
         // Pack the weight.
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
         if (conv->useFP16)
         {
             parallel_for_(Range(0, ngroups * numStripsMR_FP16), [&](const Range& r0){
@@ -415,7 +415,7 @@ static inline void packData8(char*& inpbuf, float*& inptrIn, int& in_w, int& x0,
     char * inpbufC = inpbuf + s0 * esz;
     float* inptrInC = (float* )inptrIn;
 
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
     float16_t* inpbufC_FP16 = (float16_t *)inpbufC;
     if (esz == sizeof(float16_t))
     {
@@ -521,7 +521,7 @@ static inline void packData2(char *& inpbuf, float*& inptrIn, int& in_w, int& x0
     char* inpbufC = inpbuf + s0 * esz;
     float* inptrInC = inptrIn;
 
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
     float16_t* inpbufC_FP16 = (float16_t *)inpbufC;
     if (esz == sizeof(float16_t))
     {
@@ -553,7 +553,7 @@ static inline void packData2(char *& inpbuf, float*& inptrIn, int& in_w, int& x0
     in_w += stride_w;
 }
 
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
 // Fast convert float 32 to float16
 static inline void _cvt32f16f( const float* src, float16_t* dst, int len)
 {
@@ -623,7 +623,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
             {
                 // Make special branch where memcpy() is called with a constant buffer size.
                 // Compilers will likely unroll this loop properly.
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                 if (useFP16)
                 {
                     for (int c = 0; c < Cg; c++, inptr += inp_planesize, inpbuf += CONV_NR_esz)
@@ -636,7 +636,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
             }
             else
             {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                 if (useFP16)
                 {
                     for (int c = 0; c < Cg; c++, inptr += inp_planesize, inpbuf += CONV_NR_esz)
@@ -700,7 +700,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                             int w0 = std::max(0, (-in_w + dilation_w-1)/dilation_w);
                             int w1 = std::min(Wk, (Wi - in_w + dilation_w-1)/dilation_w);
                             const float* inptrInC = inptrIn;
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                             if (useFP16)
                             {
                                 float16_t* inpbufC = (float16_t *)inpbuf + s0;
@@ -761,7 +761,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                             int w1 = std::min(Wk, (Wi - in_w + dilation_w-1)/dilation_w);
 
                             const float* inptrInC = inptrIn;
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                             if (useFP16)
                             {
                                 float16_t* inpbufC = (float16_t *)inpbuf + s0;
@@ -834,7 +834,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                             int w0 = std::max(0, (-in_w + dilation_w-1)/dilation_w);
                             int w1 = std::min(Wk, (Wi - in_w + dilation_w-1)/dilation_w);
                             const float* inptrInC = inptrIn;
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                             if (useFP16)
                             {
                                 float16_t* inpbufC = (float16_t* )inpbuf + s0;
@@ -887,7 +887,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                 for (; i < CONV_NR;)
                 {
                     float* inpbuf_ki = (float* )inpbuf + k * CONV_NR * Cg + i;
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                     float16_t * inpbuf_ki_FP16 = (float16_t *)inpbuf + k * CONV_NR * Cg + i;
 #endif
 
@@ -903,7 +903,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                         {
                             if (stride_w == 1)
                             {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                                 if (useFP16)
                                 {
                                     for (int c = 0; c < Cg; c++, inpbuf_ki_FP16 += CONV_NR, inptr_ki += inp_planesize)
@@ -934,7 +934,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                             }
                             else if (stride_w == 2)
                             {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                                 if (useFP16)
                                 {
                                     for (int c = 0; c < Cg; c++, inpbuf_ki_FP16 += CONV_NR, inptr_ki += inp_planesize)
@@ -967,7 +967,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                             }
                             else
                             {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                                 if (useFP16)
                                 {
                                     for (int c = 0; c < Cg; c++, inpbuf_ki_FP16 += CONV_NR, inptr_ki += inp_planesize)
@@ -1006,7 +1006,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                         {
                             if (stride_w == 1)
                             {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                                 if (useFP16)
                                 {
                                     for (int c = 0; c < Cg; c++, inpbuf_ki_FP16 += CONV_NR, inptr_ki += inp_planesize)
@@ -1029,7 +1029,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                             }
                             else
                             {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                                 if (useFP16)
                                 {
                                     for (int c = 0; c < Cg; c++, inpbuf_ki_FP16 += CONV_NR, inptr_ki += inp_planesize)
@@ -1057,7 +1057,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                         }
                         else
                         {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                             if (useFP16)
                             {
                                 for (int c = 0; c < Cg; c++, inpbuf_ki_FP16 += CONV_NR, inptr_ki += inp_planesize)
@@ -1073,7 +1073,7 @@ static inline void packInputData(char* inpbuf_task, float* inp, const int* ofsta
                     }
                     else
                     {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                         if (useFP16)
                         {
                             for (int c = 0; c < Cg; c++, inpbuf_ki_FP16 += CONV_NR)
@@ -1260,7 +1260,7 @@ void runFastConv(InputArray _input, OutputArray _output, const Ptr<FastConv>& co
     int CONV_MR = CONV_MR_FP32;
     int esz = sizeof(float );
 
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
     if (useFP16)
     {
         // works at FP 16.
@@ -1433,7 +1433,7 @@ void runFastConv(InputArray _input, OutputArray _output, const Ptr<FastConv>& co
                 }
 
                 char *weights = nullptr;
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                 if (useFP16)
                 {
                     CV_Assert(!conv->weightsBuf_FP16.empty());
@@ -1474,7 +1474,7 @@ void runFastConv(InputArray _input, OutputArray _output, const Ptr<FastConv>& co
 #if CV_NEON && CV_NEON_AARCH64
                         if (conv->useNEON)
                         {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                             if (useFP16)
                             {
                                 opt_NEON::convBlockMR1_FP16(DkHkWkCg, weights, inptr, cptr, biasVal, fusedAdd, minval, maxval, ifMinMaxAct, outLen, CONV_NR);
@@ -1537,7 +1537,7 @@ void runFastConv(InputArray _input, OutputArray _output, const Ptr<FastConv>& co
 #if CV_NEON
                                 if (conv->useNEON)
                                 {
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                                     if (useFP16)
                                     {
                                         opt_NEON::convBlock_FP16(c1 - c0, wptr, inptr, (char *)cptr_f16, ldc, c0 == 0, outLen, CONV_MR, CONV_NR);
@@ -1567,7 +1567,7 @@ void runFastConv(InputArray _input, OutputArray _output, const Ptr<FastConv>& co
                         float biasval = biasptr[k];
                         int j = 0;
 
-#ifdef CONV_ARM_FP16
+#if defined(CONV_ARM_FP16) && CV_FP16
                         if (useFP16)
                         {
                             float32x4_t vbias = vdupq_n_f32(biasval);