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 |  250 +++++++++++++++++++++++++++++++++++++++++++------
 1 files changed, 218 insertions(+), 32 deletions(-)

diff --git a/src/gemm.c b/src/gemm.c
index 75ce59c..2fd7f50 100644
--- a/src/gemm.c
+++ b/src/gemm.c
@@ -318,6 +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>
@@ -325,6 +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;
@@ -429,7 +456,7 @@
 }
 
 
-void convolution_2d(int w, int h, int ksize, int n, int c, int pad, int stride,
+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
@@ -477,6 +504,130 @@
     }
 }
 
+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
@@ -516,12 +667,18 @@
 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);
 }
 
+// 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,
@@ -533,7 +690,7 @@
     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
 
@@ -564,10 +721,14 @@
             }
 
             // 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)
@@ -616,15 +777,15 @@
                     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((__m256i *)(&data_im[im_col + width*(im_row + height*c_im)]));
-                    data_col[col_index + ldb_align * 0] = src256.m256_f32[0];
-                    data_col[col_index + ldb_align * 1] = src256.m256_f32[1];
-                    data_col[col_index + ldb_align * 2] = src256.m256_f32[2];
-                    data_col[col_index + ldb_align * 3] = src256.m256_f32[3];
-                    data_col[col_index + ldb_align * 4] = src256.m256_f32[4];
-                    data_col[col_index + ldb_align * 5] = src256.m256_f32[5];
-                    data_col[col_index + ldb_align * 6] = src256.m256_f32[6];
-                    data_col[col_index + ldb_align * 7] = src256.m256_f32[7];
+                    __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);
                 }
@@ -717,7 +878,7 @@
     int channels_col = channels * ksize * ksize;
 
     // optimized version
-    if (height_col == height && width_col == width && stride == 1 && pad == 1)
+    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) {
@@ -731,7 +892,7 @@
                     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);
                 }
 
@@ -796,26 +957,51 @@
     }
 }
 
+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;
+    int i = 0;
     if (a == LINEAR)
     {}
     else if (a == LEAKY)
     {
-        __m256i all256_sing1 = _mm256_set_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000);
-        __m256 all256_01 = _mm256_set1_ps(0.1F);
+        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];
+            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 mult256 = _mm256_mul_ps((src256), all256_01); // mult * 0.1
+                __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
+                __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);
+                __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) {
@@ -922,7 +1108,7 @@
 
 
 void convolution_2d(int w, int h, int ksize, int n, int c, int pad, int stride,
-    float *weights, float *input, float *output)
+    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

--
Gitblit v1.10.0