From a723e1c62a27aeb39aaf7fcdeb3beb4e89fba32d Mon Sep 17 00:00:00 2001
From: Alexey <AlexeyAB@users.noreply.github.com>
Date: Wed, 15 Aug 2018 20:52:09 +0000
Subject: [PATCH] Merge pull request #766 from HotChick91/AlexeyAB-mask

---
 src/gemm.c |  834 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++--
 1 files changed, 802 insertions(+), 32 deletions(-)

diff --git a/src/gemm.c b/src/gemm.c
index ee7fa15..2fd7f50 100644
--- a/src/gemm.c
+++ b/src/gemm.c
@@ -1,10 +1,15 @@
 #include "gemm.h"
 #include "utils.h"
+#include "im2col.h"
 #include "cuda.h"
 #include <stdlib.h>
 #include <stdio.h>
 #include <math.h>
 
+#if defined(_OPENMP)
+#include <omp.h>
+#endif
+
 void gemm_bin(int M, int N, int K, float ALPHA,
         char  *A, int lda,
         float *B, int ldb,
@@ -313,6 +318,25 @@
 #include <immintrin.h>
 #include <smmintrin.h>
 
+#if defined(_MSC_VER) && _MSC_VER <= 1900
+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];
+}
+
+#endif
+
+static inline float _castu32_f32(uint32_t a) {
+    return *((float *)&a);
+}
+
+static inline float _mm256_extract_float32(__m256 a, const int index) {
+    return a.m256_f32[index];
+}
+
 #else    // Linux GCC/Clang
 #include <x86intrin.h>
 #include <ammintrin.h>
@@ -320,6 +344,14 @@
 #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;
@@ -424,10 +456,183 @@
 }
 
 
+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
-
-// 2 x faster than popcnt: https://arxiv.org/pdf/1611.07612.pdf
+// https://arxiv.org/pdf/1611.07612.pdf
 
 static inline int popcnt128(__m128i n) {
     const __m128i n_hi = _mm_unpackhi_epi64(n, n);
@@ -442,77 +647,447 @@
     return popcnt128(_mm256_extractf128_si256(n, 0)) + popcnt128(_mm256_extractf128_si256(n, 1));
 }
 
+static inline __m256i count256(__m256i v) {
+    __m256i lookup =
+        _mm256_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2,
+            2, 3, 2, 3, 3, 4, 0, 1, 1, 2, 1, 2, 2, 3,
+            1, 2, 2, 3, 2, 3, 3, 4);
+
+    __m256i low_mask = _mm256_set1_epi8(0x0f);
+
+    __m256i lo = _mm256_and_si256(v, low_mask);
+    __m256i hi = _mm256_and_si256(_mm256_srli_epi32(v, 4), low_mask);
+    __m256i popcnt1 = _mm256_shuffle_epi8(lookup, lo);
+    __m256i popcnt2 = _mm256_shuffle_epi8(lookup, hi);
+    __m256i total = _mm256_add_epi8(popcnt1, popcnt2);
+
+    return _mm256_sad_epu8(total, _mm256_setzero_si256());
+}
+
+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 _mm256_extract_epi64(val, 0)
+        + _mm256_extract_epi64(val, 1)
+        + _mm256_extract_epi64(val, 2)
+        + _mm256_extract_epi64(val, 3);
+}
+
+// 5x times faster than gemm()-float32
+// further optimizations: do mean-mult only for the last layer
 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,
     float *C, int ldc, float *mean_arr)
 {
-    __m256i all_1 = _mm256_set1_epi8(255);
-    int i, j, k, h;
+    int i;
+
+#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(max_num_threads / 2);
+    }
+#endif
 
     #pragma omp parallel for
-    for (i = 0; i < M; ++i) {   // l.n - filters [16 - 55 - 1024]
+    for (i = 0; i < M; ++i)
+    {   // l.n - filters [16 - 55 - 1024]
         float mean_val = mean_arr[i];
+        int j, k;
+        __m256i all_1 = _mm256_set1_epi8(255);
 
         for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
             int count = 0;
             const int bit_step = 256;
+            __m256i count_sum = _mm256_set1_epi8(0);
 
             for (k = 0; k < K; k += bit_step) {   // l.size*l.size*l.c - one filter size [27 - 9216]
-
-                //__m128i a_bit128 = _mm_loadu_si128((__m128i *)(A + (i*lda + k) / 8));
-                //__m128i b_bit128 = _mm_loadu_si128((__m128i *)(B + (j*ldb + k) / 8));
-                //__m128i xor128 = _mm_xor_si128(a_bit128, b_bit128);
-                //__m128i c_bit128 = _mm_andnot_si128(xor128, all_1);
-                //int tmp_count = popcnt128(c_bit128);
-
                 __m256i a_bit256 = _mm256_loadu_si256((__m256i *)(A + (i*lda + k) / 8));
                 __m256i b_bit256 = _mm256_loadu_si256((__m256i *)(B + (j*ldb + k) / 8));
-                __m256i xor256 = _mm256_xor_si256(a_bit256, b_bit256);
-                __m256i c_bit256 = _mm256_andnot_si256(xor256, all_1); //we can do NOT for wegihts once and do not do this NOT
-                int tmp_count = popcnt256(c_bit256);
+                __m256i xor256 = _mm256_xor_si256(a_bit256, b_bit256);  // xnor = not(xor(a,b))
+                __m256i c_bit256 = _mm256_andnot_si256(xor256, all_1);  // can be optimized - we can do other NOT for wegihts once and do not do this NOT
 
-                if (K - k < bit_step)  tmp_count = tmp_count - (bit_step - (K - k));    // remove extra bits
-                count += tmp_count;
+                count_sum = _mm256_add_epi64(count256(c_bit256), count_sum);    //  Mula�s algorithm
+
+                //count += popcnt256(c_bit256);
+
                 //binary_int64_printf(c_bit64);
                 //printf(", count = %d \n\n", tmp_count);
             }
 
+            // 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 = _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)
+
             C[i*ldc + j] = (2 * count - K) * mean_val;
         }
     }
 }
 
 
+static inline float im2col_get_pixel(float *im, int height, int width, int channels,
+    int row, int col, int channel, int pad)
+{
+    row -= pad;
+    col -= pad;
+
+    if (row < 0 || col < 0 ||
+        row >= height || col >= width) return 0;
+    return im[col + width*(row + height*channel)];
+}
+
+//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)
+{
+
+    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 && is_fma_avx())
+    {
+        #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-8; 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;
+
+                    //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)]));
+                    _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 = (c * height_col + h) * width_col + w;
+
+                    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 = (c * height_col + h) * width_col + w;
+                    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 = (c * height_col + h) * width_col + w;
+                    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 = (c * height_col + h) * width_col + w;
+                    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 = (c * height_col + h) * width_col + w;
+                    data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
+                        im_row, im_col, c_im, pad);
+                }
+            }
+        }
+
+    }
+    else {
+        //printf("\n Error: is no non-optimized version \n");
+        im2col_cpu(data_im, channels, height, width, ksize, stride, pad, data_col);
+    }
+}
+
+void transpose_8x8_bits(unsigned char A[8], unsigned char B[8], int m, int n)
+{
+    unsigned x, y, t;
+
+    // Load the array and pack it into x and y.
+
+    x = (A[0] << 24) | (A[m] << 16) | (A[2 * m] << 8) | A[3 * m];
+    y = (A[4 * m] << 24) | (A[5 * m] << 16) | (A[6 * m] << 8) | A[7 * m];
+
+    t = (x ^ (x >> 7)) & 0x00AA00AA; x = x ^ t ^ (t << 7);
+    t = (y ^ (y >> 7)) & 0x00AA00AA; y = y ^ t ^ (t << 7);
+
+    t = (x ^ (x >> 14)) & 0x0000CCCC; x = x ^ t ^ (t << 14);
+    t = (y ^ (y >> 14)) & 0x0000CCCC; y = y ^ t ^ (t << 14);
+
+    t = (x & 0xF0F0F0F0) | ((y >> 4) & 0x0F0F0F0F);
+    y = ((x << 4) & 0xF0F0F0F0) | (y & 0x0F0F0F0F);
+    x = t;
+
+    B[0] = x >> 24; B[n] = x >> 16; B[2 * n] = x >> 8; B[3 * n] = x;
+    B[4 * n] = y >> 24; B[5 * n] = y >> 16; B[6 * n] = y >> 8; B[7 * n] = y;
+}
+
+void activate_array_cpu_custom(float *x, const int n, const ACTIVATION a)
+{
+    int i = 0;
+    if (a == LINEAR)
+    {}
+    else if (a == LEAKY)
+    {
+        if (is_fma_avx()) {
+            __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 - 8; i += 8) {
+                //x[i] = (x[i]>0) ? x[i] : .1*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(&x[i], result256);
+            }
+        }
+
+        for (; i < n; ++i) {
+            x[i] = (x[i]>0) ? x[i] : .1*x[i];
+        }
+    }
+    else {
+        for (i = 0; i < n; ++i) {
+            x[i] = activate(x[i], a);
+        }
+    }
+}
+
 void float_to_bit(float *src, unsigned char *dst, size_t size)
 {
     size_t dst_size = size / 8 + 1;
     memset(dst, 0, dst_size);
 
     size_t i;
-    __m128i all128_0 = _mm_set_epi32(0, 0, 0, 0);
-    __m256 all256_0 = _mm256_set1_ps(0);
-    __m256i bits_asc = _mm256_set_epi32(1, 2, 4, 8, 16, 32, 64, 128);
-    //for(i = 0; i < 8; ++i) bits_asc.m256i_i32[i] = 1 << i;
+    __m256i all256_sing1 = _mm256_set_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000);
 
     for (i = 0; i < size; i+=8)
     {
-        __m256 src256 = _mm256_loadu_ps((__m256i *)(&src[i]));  // load 256 bits
-        __m256 result256 = _mm256_cmp_ps(src256, all256_0, _CMP_GT_OS); // compare dst[i] = (float[i] > 0)
+        __m256i src256 = _mm256_loadu_si256((__m256i *)(&src[i]));
+        __m256i result256 = _mm256_and_si256(src256, all256_sing1); // check sign in 8 x 32-bit floats
 
-        __m256i bits256 = _mm256_castps_si256(result256);       // floats to ints32
-        __m256i and256 = _mm256_and_si256(bits256, bits_asc);   // bitwise and
+        uint32_t mask = _mm256_movemask_ps(_mm256_castsi256_ps(result256)); // (val >= 0) ? 0 : 1
+        mask = ~mask;   // inverse mask,  (val >= 0) ? 1 : 0
 
-        // sum all elements from single and256
-        __m128i tmp128 = _mm_hadd_epi32(_mm256_extractf128_si256(and256, 0), _mm256_extractf128_si256(and256, 1));
-        tmp128 = _mm_hadd_epi32(tmp128, all128_0);
-        tmp128 = _mm_hadd_epi32(tmp128, all128_0);
-
-        dst[i / 8] = tmp128.m128i_i32[0];
+        dst[i / 8] = mask;
     }
-    // int _mm256_movemask_epi8 (__m256i a)
 }
 
+static inline void transpose4x4_SSE(float *A, float *B, const int lda, const int ldb)
+{
+    __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_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;
+    #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_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 {
+            for (i2 = i; i2 < n; ++i2) {
+                for (j2 = 0; j2 < m; ++j2) {
+                    B[j2*ldb + i2] = A[i2*lda + j2];
+                }
+            }
+        }
+    }
+}
+
+
 #else
 
 void gemm_nn(int M, int N, int K, float ALPHA,
@@ -531,6 +1106,55 @@
     }
 }
 
+
+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,
@@ -567,6 +1191,122 @@
     }
 }
 
+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,
+    int channels, int height, int width,
+    int ksize, int stride, int pad, float* data_col)
+{
+
+    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; ++w) {
+                    int im_row = h_offset + h - pad;
+                    int im_col = w_offset + w - pad;
+                    int col_index = (c * height_col + h) * width_col + w;
+
+                    data_col[col_index] = data_im[im_col + width*(im_row + height*c_im)];
+    }
+
+                for (; w < width_col - pad; ++w) {
+                    int im_row = h_offset + h - pad;
+                    int im_col = w_offset + w - pad;
+                    int col_index = (c * height_col + h) * width_col + w;
+
+                    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 = (c * height_col + h) * width_col + w;
+                    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 = (c * height_col + h) * width_col + w;
+                    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 = (c * height_col + h) * width_col + w;
+                    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 = (c * height_col + h) * width_col + w;
+                    data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
+                        im_row, im_col, c_im, pad);
+                }
+            }
+        }
+
+    }
+    else {
+        //printf("\n Error: is no non-optimized version \n");
+        im2col_cpu(data_im, channels, height, width, ksize, stride, pad, data_col);
+    }
+}
+
+void activate_array_cpu_custom(float *x, const int n, const ACTIVATION a)
+{
+    int i;
+    if (a == LINEAR)
+    {
+    }
+    else if (a == LEAKY)
+    {
+        for (i = 0; i < n; ++i) {
+            x[i] = (x[i]>0) ? x[i] : .1*x[i];
+        }
+    }
+    else {
+        for (i = 0; i < n; ++i) {
+            x[i] = activate(x[i], a);
+        }
+    }
+}
+
 void float_to_bit(float *src, unsigned char *dst, size_t size)
 {
     size_t dst_size = size / 8 + 1;
@@ -596,6 +1336,36 @@
     }
     free(byte_arr);
 }
+
+static inline void transpose_scalar_block(float *A, float *B, const int lda, const int ldb, const int block_size)
+{
+    int i, j;
+    //#pragma omp parallel for
+    for (i = 0; i<block_size; i++) {
+        for (j = 0; j<block_size; j++) {
+            B[j*ldb + i] = A[i*lda + j];
+        }
+    }
+}
+
+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;
+    #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];
+                }
+                }
+            }
+        }
+}
 #endif    // __x86_64
 
 void gemm_nt(int M, int N, int K, float ALPHA,

--
Gitblit v1.10.0