From b1dddf02ccf8dcfaadee4e8a5ed8726725ec1b93 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Sun, 12 Aug 2018 23:43:45 +0000
Subject: [PATCH] Fixed AVX compiled bug

---
 src/gemm.c |  464 ++++++++++++++++++++++++++++++++++++++++++++++++++++-----
 1 files changed, 422 insertions(+), 42 deletions(-)

diff --git a/src/gemm.c b/src/gemm.c
index ff5edfa..b909814 100644
--- a/src/gemm.c
+++ b/src/gemm.c
@@ -318,6 +318,22 @@
 #include <immintrin.h>
 #include <smmintrin.h>
 
+static inline __int32 _mm256_extract_epi64(__m256i a, const int index) {
+    return a.m256i_i64[index];
+}
+
+static inline __int32 _mm256_extract_epi32(__m256i a, const int index) {
+    return a.m256i_i32[index];
+}
+
+static inline float _castu32_f32(uint32_t a) {
+    return *((float *)&a);
+}
+
+static inline float _mm256_extract_float32(__m256 a, const int index) {
+    return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), index));
+}
+
 #else    // Linux GCC/Clang
 #include <x86intrin.h>
 #include <ammintrin.h>
@@ -325,6 +341,14 @@
 #include <smmintrin.h>
 #include <cpuid.h>
 
+static inline float _castu32_f32(uint32_t a) {
+    return *((float *)&a);
+}
+
+static inline float _mm256_extract_float32(__m256 a, const int index) {
+    return _castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), index));
+}
+
 void asm_cpuid(uint32_t* abcd, uint32_t eax)
 {
     uint32_t ebx = 0, edx = 0, ecx = 0;
@@ -429,6 +453,180 @@
 }
 
 
+void convolution_2d_old(int w, int h, int ksize, int n, int c, int pad, int stride,
+    float *weights, float *input, float *output)
+{
+    int out_h = (h + 2 * pad - ksize) / stride + 1;    // output_height=input_height for stride=1 and pad=1
+    int out_w = (w + 2 * pad - ksize) / stride + 1;    // output_width=input_width for stride=1 and pad=1
+    int i, f, j;
+
+    int fil;
+    // filter index
+#pragma omp parallel for      // "omp parallel for" - automatic parallelization of loop by using OpenMP
+    for (fil = 0; fil < n; ++fil) {
+        int chan, y, x, f_y, f_x;
+        // channel index
+        for (chan = 0; chan < c; ++chan)
+            // input - y
+            for (y = 0; y < h; ++y)
+                // input - x
+                for (x = 0; x < w; ++x)
+                {
+                    int const output_index = fil*w*h + y*w + x;
+                    int const weights_pre_index = fil*c*ksize*ksize + chan*ksize*ksize;
+                    int const input_pre_index = chan*w*h;
+                    float sum = 0;
+
+                    // filter - y
+                    for (f_y = 0; f_y < ksize; ++f_y)
+                    {
+                        int input_y = y + f_y - pad;
+                        // filter - x
+                        for (f_x = 0; f_x < ksize; ++f_x)
+                        {
+                            int input_x = x + f_x - pad;
+                            if (input_y < 0 || input_x < 0 || input_y >= h || input_x >= w) continue;
+
+                            int input_index = input_pre_index + input_y*w + input_x;
+                            int weights_index = weights_pre_index + f_y*ksize + f_x;
+
+                            sum += input[input_index] * weights[weights_index];
+                        }
+                    }
+                    // l.output[filters][width][height] +=
+                    //        state.input[channels][width][height] *
+                    //        l.weights[filters][channels][filter_width][filter_height];
+                    output[output_index] += sum;
+                }
+    }
+}
+
+void convolution_2d(int w, int h, int ksize, int n, int c, int pad, int stride,
+    float *weights, float *input, float *output, float *mean)
+{
+    int out_h = (h + 2 * pad - ksize) / stride + 1;    // output_height=input_height for stride=1 and pad=1
+    int out_w = (w + 2 * pad - ksize) / stride + 1;    // output_width=input_width for stride=1 and pad=1
+    int i, f, j;
+
+#if defined(_OPENMP)
+    static int max_num_threads = 0;
+    if (max_num_threads == 0) {
+        max_num_threads = omp_get_max_threads();
+        omp_set_num_threads(4);// max_num_threads / 2);
+    }
+#endif
+
+    //convolution_2d_old(w, h, ksize, n, c, pad, stride, weights, input, output);
+
+    __m256i all256_sing1 = _mm256_set_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000);
+    for (i = 0; i < ksize*ksize*n*c; i+=8) {
+        *((__m256*)&weights[i]) = _mm256_and_ps(*((__m256*)&weights[i]), _mm256_castsi256_ps(all256_sing1));
+    }
+
+    for (i = 0; i < w*h*c; i += 8) {
+        //*((__m256*)&input[i]) = _mm256_and_ps(*((__m256*)&input[i]), _mm256_castsi256_ps(all256_sing1));
+    }
+
+
+    //__m256i all256_last_zero = _mm256_set1_epi32(0xFFFFFFFF);
+    //all256_last_zero.m256i_i32[7] = 0;
+    __m256i all256_last_zero =
+        _mm256_set_epi32(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0x0);
+
+    __m256i idx256 = _mm256_set_epi32(0, 7, 6, 5, 4, 3, 2, 1);
+    //__m256 all256_sing1 = _mm256_set1_ps(0x80000000);
+    __m256 all256_one = _mm256_set1_ps(1);
+    __m256i all256i_one = _mm256_set1_epi32(1);
+
+    ///__m256i src256 = _mm256_loadu_si256((__m256i *)(&src[i]));
+    ///__m256i result256 = _mm256_and_si256(src256, all256_sing1); // check sign in 8 x 32-bit floats
+
+    int fil;
+    // filter index
+#pragma omp parallel for      // "omp parallel for" - automatic parallelization of loop by using OpenMP
+    for (fil = 0; fil < n; ++fil) {
+        int chan, y, x, f_y, f_x;
+        float cur_mean = fabs(mean[fil]);
+        __m256 mean256 = _mm256_set1_ps(cur_mean);
+        // channel index
+        //for (chan = 0; chan < c; ++chan)
+            // input - y
+            for (y = 0; y < h; ++y)
+                // input - x
+                for (x = 0; x < w-8; x+=8)
+                {
+                    int const output_index = fil*w*h + y*w + x;
+                    float sum = 0;
+                    __m256 sum256 = _mm256_set1_ps(0);
+
+                    for (chan = 0; chan < c; ++chan) {
+                        int const weights_pre_index = fil*c*ksize*ksize + chan*ksize*ksize;
+                        int const input_pre_index = chan*w*h;
+
+
+                        // filter - y
+                        for (f_y = 0; f_y < ksize; ++f_y)
+                        {
+                            int input_y = y + f_y - pad;
+                            //__m256 in = *((__m256*)&input[input_pre_index + input_y*w]);
+                            if (input_y < 0 || input_y >= h) continue;
+                            //__m256 in = _mm256_loadu_ps(&input[input_pre_index + input_y*w + x - pad]);
+
+                            // filter - x
+                            for (f_x = 0; f_x < ksize; ++f_x)
+                            {
+                                int input_x = x + f_x - pad;
+                                //if (input_y < 0 || input_x < 0 || input_y >= h || input_x >= w) continue;
+
+                                int input_index = input_pre_index + input_y*w + input_x;
+                                int weights_index = weights_pre_index + f_y*ksize + f_x;
+                                //if (input_y < 0 || input_y >= h) continue;
+
+                                //sum += input[input_index] * weights[weights_index];
+
+                                __m256 in = *((__m256*)&input[input_index]);
+                                __m256 w = _mm256_set1_ps(weights[weights_index]);
+                                //__m256 w_sign = _mm256_and_ps(w, _mm256_castsi256_ps(all256_sing1)); // check sign in 8 x 32-bit floats
+                                __m256 xor256 = _mm256_xor_ps(w, in);
+                                //printf("\n xor256_1 = %f, xor256_2 = %f \n", xor256.m256_f32[0], xor256.m256_f32[1]);
+                                //printf("\n in = %f, w = %f, xor256 = %f \n", in.m256_f32[0], w_sign.m256_f32[0], xor256.m256_f32[0]);
+
+                                //__m256 pn1 = _mm256_and_ps(_mm256_castsi256_ps(all256i_one), xor256);
+
+
+                                //sum256 = xor256;
+                                sum256 = _mm256_add_ps(xor256, sum256);
+                                //printf("\n --- \n");
+                                //printf("\n 0 = %f, 1 = %f, 2 = %f, 3 = %f, 4 = %f, 5 = %f, 6 = %f, 7 = %f \n", in.m256_f32[0], in.m256_f32[1], in.m256_f32[2], in.m256_f32[3], in.m256_f32[4], in.m256_f32[5], in.m256_f32[6], in.m256_f32[7]);
+
+                                if (f_x < ksize-1) {
+                                    //in = _mm256_permutevar8x32_ps(in, idx256);
+                                    //in = _mm256_and_ps(in, _mm256_castsi256_ps(all256_last_zero));
+                                }
+                            }
+                        }
+                    }
+                    // l.output[filters][width][height] +=
+                    //        state.input[channels][width][height] *
+                    //        l.weights[filters][channels][filter_width][filter_height];
+                    //output[output_index] += sum;
+
+                    sum256 = _mm256_mul_ps(sum256, mean256);
+                    //printf("\n cur_mean = %f, sum256 = %f, sum256 = %f, in = %f \n",
+                    //    cur_mean, sum256.m256_f32[0], sum256.m256_f32[1], input[input_pre_index]);
+
+                    //__m256 out = *((__m256*)&output[output_index]);
+                    //out = _mm256_add_ps(out, sum256);
+                    //*((__m256*)&output[output_index]) = out;
+                    *((__m256*)&output[output_index]) = sum256;
+
+                    //_mm256_storeu_ps(&C[i*ldc + j], result256);
+                }
+    }
+}
+
+
+
 // http://graphics.stanford.edu/~seander/bithacks.html
 // https://stackoverflow.com/questions/17354971/fast-counting-the-number-of-set-bits-in-m128i-register
 // https://arxiv.org/pdf/1611.07612.pdf
@@ -466,10 +664,14 @@
 static inline int popcnt256_custom(__m256i n) {
     __m256i val = count256(n);
 
-    return val.m256i_i64[0] +
-    val.m256i_i64[1] +
-    val.m256i_i64[2] +
-    val.m256i_i64[3];
+    //return val.m256i_i64[0] +
+    //val.m256i_i64[1] +
+    //val.m256i_i64[2] +
+    //val.m256i_i64[3];
+    return _mm256_extract_epi64(val, 0)
+        + _mm256_extract_epi64(val, 1)
+        + _mm256_extract_epi64(val, 2)
+        + _mm256_extract_epi64(val, 3);
 }
 
 void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
@@ -483,7 +685,7 @@
     static int max_num_threads = 0;
     if (max_num_threads == 0) {
         max_num_threads = omp_get_max_threads();
-        omp_set_num_threads(max_num_threads / 2);
+        //omp_set_num_threads(max_num_threads / 2);
     }
 #endif
 
@@ -514,10 +716,14 @@
             }
 
             // count of 1 bits
-            count = count_sum.m256i_i64[0] +
-                count_sum.m256i_i64[1] +
-                count_sum.m256i_i64[2] +
-                count_sum.m256i_i64[3];
+            //count = count_sum.m256i_i64[0] +
+            //    count_sum.m256i_i64[1] +
+            //    count_sum.m256i_i64[2] +
+             //   count_sum.m256i_i64[3];
+            count = _mm256_extract_epi64(count_sum, 0)
+                + _mm256_extract_epi64(count_sum, 1)
+                + _mm256_extract_epi64(count_sum, 2)
+                + _mm256_extract_epi64(count_sum, 3);
 
             int f1 = (K % bit_step == 0) ? 0 : (bit_step - (K % bit_step));
             count = count - f1;    // remove extra bits (from empty space for align only)
@@ -541,6 +747,121 @@
 
 //From Berkeley Vision's Caffe!
 //https://github.com/BVLC/caffe/blob/master/LICENSE
+void im2col_cpu_custom_transpose(float* data_im,
+    int channels, int height, int width,
+    int ksize, int stride, int pad, float* data_col, int ldb_align)
+{
+    int c, h, w;
+    int height_col = (height + 2 * pad - ksize) / stride + 1;
+    int width_col = (width + 2 * pad - ksize) / stride + 1;
+    int channels_col = channels * ksize * ksize;
+
+    // optimized version
+    if (height_col == height && width_col == width && stride == 1 && pad == 1)
+    {
+#pragma omp parallel for
+        for (c = 0; c < channels_col; ++c) {
+            int w_offset = c % ksize;
+            int h_offset = (c / ksize) % ksize;
+            int c_im = c / ksize / ksize;
+            for (h = pad; h < height_col - pad; ++h) {
+                for (w = pad; w < width_col - pad - 4; w+=8) {
+                    int im_row = h_offset + h - pad;
+                    int im_col = w_offset + w - pad;
+                    //int col_index = (c * height_col + h) * width_col + w;
+                    int col_index = (h * width_col + w)*ldb_align + c;   // transposed & aligned
+
+                    //data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
+                    __m256 src256 = _mm256_loadu_ps((float *)(&data_im[im_col + width*(im_row + height*c_im)]));
+                    data_col[col_index + ldb_align * 0] = _mm256_extract_float32(src256, 0);// src256.m256_f32[0];
+                    data_col[col_index + ldb_align * 1] = _mm256_extract_float32(src256, 1);// src256.m256_f32[1];
+                    data_col[col_index + ldb_align * 2] = _mm256_extract_float32(src256, 2);// src256.m256_f32[2];
+                    data_col[col_index + ldb_align * 3] = _mm256_extract_float32(src256, 3);// src256.m256_f32[3];
+                    data_col[col_index + ldb_align * 4] = _mm256_extract_float32(src256, 4);// src256.m256_f32[4];
+                    data_col[col_index + ldb_align * 5] = _mm256_extract_float32(src256, 5);// src256.m256_f32[5];
+                    data_col[col_index + ldb_align * 6] = _mm256_extract_float32(src256, 6);// src256.m256_f32[6];
+                    data_col[col_index + ldb_align * 7] = _mm256_extract_float32(src256, 7);// src256.m256_f32[7];
+
+                    //_mm256_storeu_ps(&data_col[col_index], src256);
+                }
+
+                for (; w < width_col - pad; ++w) {
+                    int im_row = h_offset + h - pad;
+                    int im_col = w_offset + w - pad;
+                    int col_index = (h * width_col + w)*ldb_align + c;   // transposed & aligned
+                    data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
+                }
+            }
+
+            {
+                w = 0;
+                for (h = 0; h < height_col; ++h) {
+                    int im_row = h_offset + h;
+                    int im_col = w_offset + w;
+                    int col_index = (h * width_col + w)*ldb_align + c;   // transposed & aligned
+                    data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
+                        im_row, im_col, c_im, pad);
+                }
+            }
+
+            {
+                w = width_col - 1;
+                for (h = 0; h < height_col; ++h) {
+                    int im_row = h_offset + h;
+                    int im_col = w_offset + w;
+                    int col_index = (h * width_col + w)*ldb_align + c;   // transposed & aligned
+                    data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
+                        im_row, im_col, c_im, pad);
+                }
+            }
+
+            {
+                h = 0;
+                for (w = 0; w < width_col; ++w) {
+                    int im_row = h_offset + h;
+                    int im_col = w_offset + w;
+                    int col_index = (h * width_col + w)*ldb_align + c;   // transposed & aligned
+                    data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
+                        im_row, im_col, c_im, pad);
+                }
+            }
+
+            {
+                h = height_col - 1;
+                for (w = 0; w < width_col; ++w) {
+                    int im_row = h_offset + h;
+                    int im_col = w_offset + w;
+                    int col_index = (h * width_col + w)*ldb_align + c;   // transposed & aligned
+                    data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
+                        im_row, im_col, c_im, pad);
+                }
+            }
+        }
+
+    }
+    else {
+        #pragma omp parallel for
+        for (c = 0; c < channels_col; ++c) {
+            int w_offset = c % ksize;
+            int h_offset = (c / ksize) % ksize;
+            int c_im = c / ksize / ksize;
+            for (h = 0; h < height_col; ++h) {
+                for (w = 0; w < width_col; ++w) {
+                    int im_row = h_offset + h * stride;
+                    int im_col = w_offset + w * stride;
+
+                    int col_index = (h * width_col + w)*ldb_align + c;   // transposed & aligned
+                    data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
+                        im_row, im_col, c_im, pad);
+                }
+            }
+        }
+    }
+}
+
+
+//From Berkeley Vision's Caffe!
+//https://github.com/BVLC/caffe/blob/master/LICENSE
 void im2col_cpu_custom(float* data_im,
     int channels, int height, int width,
     int ksize, int stride, int pad, float* data_col)
@@ -566,7 +887,7 @@
                     int col_index = (c * height_col + h) * width_col + w;
 
                     //data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
-                    __m256 src256 = _mm256_loadu_ps((__m256i *)(&data_im[im_col + width*(im_row + height*c_im)]));
+                    __m256 src256 = _mm256_loadu_ps((float *)(&data_im[im_col + width*(im_row + height*c_im)]));
                     _mm256_storeu_ps(&data_col[col_index], src256);
                 }
 
@@ -641,16 +962,16 @@
         __m256i all256_sing1 = _mm256_set_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000);
         __m256 all256_01 = _mm256_set1_ps(0.1F);
 
-        for (i = 0; i < n; i += 8) {
+        for (i = 0; i < n-8; i += 8) {
             //x[i] = (x[i]>0) ? x[i] : .1*x[i];
 
-            __m256 src256 = _mm256_loadu_ps((__m256 *)(&x[i]));
+            __m256 src256 = _mm256_loadu_ps(&x[i]);
             __m256 mult256 = _mm256_mul_ps((src256), all256_01); // mult * 0.1
 
             __m256i sign256 = _mm256_and_si256(_mm256_castps_si256(src256), all256_sing1); // check sign in 8 x 32-bit floats
 
             __m256 result256 = _mm256_blendv_ps(src256, mult256, _mm256_castsi256_ps(sign256)); // (sign>0) ? src : mult;
-            _mm256_storeu_ps((__m256 *)(&x[i]), result256);
+            _mm256_storeu_ps(&x[i], result256);
         }
 
         for (; i < n; ++i) {
@@ -686,47 +1007,50 @@
 
 static inline void transpose4x4_SSE(float *A, float *B, const int lda, const int ldb)
 {
-    __m128 row1 = _mm_load_ps(&A[0 * lda]);
-    __m128 row2 = _mm_load_ps(&A[1 * lda]);
-    __m128 row3 = _mm_load_ps(&A[2 * lda]);
-    __m128 row4 = _mm_load_ps(&A[3 * lda]);
+    __m128 row1 = _mm_loadu_ps(&A[0 * lda]);
+    __m128 row2 = _mm_loadu_ps(&A[1 * lda]);
+    __m128 row3 = _mm_loadu_ps(&A[2 * lda]);
+    __m128 row4 = _mm_loadu_ps(&A[3 * lda]);
     _MM_TRANSPOSE4_PS(row1, row2, row3, row4);
-    _mm_store_ps(&B[0 * ldb], row1);
-    _mm_store_ps(&B[1 * ldb], row2);
-    _mm_store_ps(&B[2 * ldb], row3);
-    _mm_store_ps(&B[3 * ldb], row4);
+    _mm_storeu_ps(&B[0 * ldb], row1);
+    _mm_storeu_ps(&B[1 * ldb], row2);
+    _mm_storeu_ps(&B[2 * ldb], row3);
+    _mm_storeu_ps(&B[3 * ldb], row4);
 }
 
 void transpose_block_SSE4x4(float *A, float *B, const int n, const int m,
     const int lda, const int ldb, const int block_size)
 {
     int i;
-    if (block_size % 4 == 0) {
-        #pragma omp parallel for
-        for (i = 0; i < n; i += block_size) {
-            int j, i2, j2;
+    #pragma omp parallel for
+    for (i = 0; i < n; i += block_size) {
+        int j, i2, j2;
+        //int max_i2 = (i + block_size < n) ? (i + block_size) : n;
+        if (i + block_size < n) {
+            int max_i2 = i + block_size;
             for (j = 0; j < m; j += block_size) {
-                int max_i2 = i + block_size < n ? i + block_size : n;
-                int max_j2 = j + block_size < m ? j + block_size : m;
-                for (i2 = i; i2 < max_i2; i2 += 4) {
-                    for (j2 = j; j2 < max_j2; j2 += 4) {
-                        transpose4x4_SSE(&A[i2*lda + j2], &B[j2*ldb + i2], lda, ldb);
+                //int max_j2 = (j + block_size < m) ? (j + block_size) : m;
+                if (j + block_size < m) {
+                    int max_j2 = j + block_size;
+                    for (i2 = i; i2 < max_i2; i2 += 4) {
+                        for (j2 = j; j2 < max_j2; j2 += 4) {
+                            transpose4x4_SSE(&A[i2*lda + j2], &B[j2*ldb + i2], lda, ldb);
+                        }
+                    }
+                }
+                else {
+                    for (i2 = i; i2 < max_i2; ++i2) {
+                        for (j2 = j; j2 < m; ++j2) {
+                            B[j2*ldb + i2] = A[i2*lda + j2];
+                        }
                     }
                 }
             }
         }
-    }
-    else {
-        #pragma omp parallel for
-        for (i = 0; i < n; i += block_size) {
-            int j, i2, j2;
-            for (j = 0; j < m; j += block_size) {
-                int max_i2 = i + block_size < n ? i + block_size : n;
-                int max_j2 = j + block_size < m ? j + block_size : m;
-                for (i2 = i; i2 < max_i2; ++i2) {
-                    for (j2 = j; j2 < max_j2; ++j2) {
-                        B[j2*ldb + i2] = A[i2*lda + j2];
-                    }
+        else {
+            for (i2 = i; i2 < n; ++i2) {
+                for (j2 = 0; j2 < m; ++j2) {
+                    B[j2*ldb + i2] = A[i2*lda + j2];
                 }
             }
         }
@@ -752,6 +1076,55 @@
     }
 }
 
+
+void convolution_2d(int w, int h, int ksize, int n, int c, int pad, int stride,
+    float *weights, float *input, float *output, float *mean)
+{
+    int out_h = (h + 2 * pad - ksize) / stride + 1;    // output_height=input_height for stride=1 and pad=1
+    int out_w = (w + 2 * pad - ksize) / stride + 1;    // output_width=input_width for stride=1 and pad=1
+    int i, f, j;
+
+    int fil;
+    // filter index
+#pragma omp parallel for      // "omp parallel for" - automatic parallelization of loop by using OpenMP
+    for (fil = 0; fil < n; ++fil) {
+        int chan, y, x, f_y, f_x;
+        // channel index
+        for (chan = 0; chan < c; ++chan)
+            // input - y
+            for (y = 0; y < h; ++y)
+                // input - x
+                for (x = 0; x < w; ++x)
+                {
+                    int const output_index = fil*w*h + y*w + x;
+                    int const weights_pre_index = fil*c*ksize*ksize + chan*ksize*ksize;
+                    int const input_pre_index = chan*w*h;
+                    float sum = 0;
+
+                    // filter - y
+                    for (f_y = 0; f_y < ksize; ++f_y)
+                    {
+                        int input_y = y + f_y - pad;
+                        // filter - x
+                        for (f_x = 0; f_x < ksize; ++f_x)
+                        {
+                            int input_x = x + f_x - pad;
+                            if (input_y < 0 || input_x < 0 || input_y >= h || input_x >= w) continue;
+
+                            int input_index = input_pre_index + input_y*w + input_x;
+                            int weights_index = weights_pre_index + f_y*ksize + f_x;
+
+                            sum += input[input_index] * weights[weights_index];
+                        }
+                    }
+                    // l.output[filters][width][height] +=
+                    //        state.input[channels][width][height] *
+                    //        l.weights[filters][channels][filter_width][filter_height];
+                    output[output_index] += sum;
+                }
+    }
+}
+
 void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
     unsigned char *A, int lda,
     unsigned char *B, int ldb,
@@ -788,6 +1161,13 @@
     }
 }
 
+void im2col_cpu_custom_transpose(float* data_im,
+    int channels, int height, int width,
+    int ksize, int stride, int pad, float* data_col, int ldb_align)
+{
+    printf("\n im2col_cpu_custom_transpose() isn't implemented without AVX \n");
+}
+
 //From Berkeley Vision's Caffe!
 //https://github.com/BVLC/caffe/blob/master/LICENSE
 void im2col_cpu_custom(float* data_im,

--
Gitblit v1.10.0