From 1f2155b8864031d1f434aa910ceaf53477217c75 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Fri, 10 Aug 2018 23:49:55 +0000
Subject: [PATCH] Experiments
---
src/gemm.c | 347 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 345 insertions(+), 2 deletions(-)
diff --git a/src/gemm.c b/src/gemm.c
index 4a7dad7..580f4da 100644
--- a/src/gemm.c
+++ b/src/gemm.c
@@ -429,6 +429,178 @@
}
+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 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 xor = _mm256_xor_ps(w, in);
+ //printf("\n xor1 = %f, xor2 = %f \n", xor.m256_f32[0], xor.m256_f32[1]);
+ //printf("\n in = %f, w = %f, xor = %f \n", in.m256_f32[0], w_sign.m256_f32[0], xor.m256_f32[0]);
+
+ //__m256 pn1 = _mm256_and_ps(_mm256_castsi256_ps(all256i_one), xor);
+
+
+ //sum256 = xor;
+ sum256 = _mm256_add_ps(xor, 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
@@ -483,7 +655,7 @@
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
@@ -541,6 +713,121 @@
//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((__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];
+
+ //_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)
@@ -641,7 +928,7 @@
__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]));
@@ -755,6 +1042,55 @@
}
}
+
+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,
@@ -791,6 +1127,13 @@
}
}
+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