Implemented SIMD optimizations in tricorn_simd including negative height handling which needs to be fixed in normal tricorn.c

Implementierung/Makefile

@@ -1,5 +1,5 @@
 CC = gcc
-CFLAGS = -Wall -Wextra -O3 -std=c17
+CFLAGS = -Wall -Wextra -O3 -std=c17 -msse4.1
 DEPFLAGS = -MMD -MP
 LDFLAGS = -lm
 

Implementierung/main.c

@@ -6,6 +6,7 @@
 #include <getopt.h> // for getopt_long
 #include <string.h> // for strcpy
 #include "tricorn.h"
+#include "tricorn_simd.h"
 
 void save_bmp(const char *filename, unsigned char *img, size_t width, ssize_t height,_Bool color);
 
@@ -17,7 +18,7 @@ unsigned tests(void) {
 void benchmarking(void) {
     // do the benchmarking
 }
-// REMOVE LATER: interesting test command  ./tricorn -s -2.0,0.6 -r 0.0010  -n 200 -d 1920 1080 -o image.bmp  -C
+// REMOVE LATER: interesting test command  ./tricorn -s -2.0,0.6 -r 0.0010  -n 200 -d 1920,-1080 -o image.bmp  -C
 const char* help_messsage = "Help Message\n\n"
                 "Usage:\n"
                 "   ./tricorn [OPTIONS]\n\n"
@@ -200,7 +201,7 @@ int main(int argc, char** argv)
     
     unsigned char tempArr[width*height*bytesPerColour]; // if not colored, rem
     printf("yashar\n");
-    tricorn(complexA,width,height,resolution,n,colored,tempArr);
+    tricorn_simd(complexA,width,height,resolution,n,colored,tempArr);
     printf("yashar\n");
     save_bmp(output_data,tempArr,width,height,colored);
     
@@ -218,14 +219,16 @@ void save_bmp(const char *filename, unsigned char *img, size_t width, ssize_t he
         exit(EXIT_FAILURE);
     }
 
+    size_t bytesPerColor = color ? 3 : 1;
+
+
     int actual_height = (height < 0) ? -height : height; // Calculate absolute height for iteration and calculation purposes
-    size_t row_padding = (4 - (width) * 3 % 4) % 4; // Calculate padding for each row as .bmp is 4 bytes aligned
+    size_t row_padding = (4 - (width * bytesPerColor) % 4) % 4; // Calculate padding for each row as .bmp is 4 bytes aligned
-    size_t row_width = width * 3 + row_padding; // Calculate row width with padding
+    size_t row_width = width * bytesPerColor + row_padding; // Calculate row width with padding
 
     size_t image_size = actual_height * row_width;
     size_t file_size = image_size + 54; // File size = Image Size + Headers
 
-    size_t bytesPerColor = color ? 3 : 1;
 
     // File header - contains information about the file
     unsigned char file_header[14] = { 

Implementierung/tricorn_simd.c

@@ -1,51 +1,115 @@
-#include "tricorn.h"
+#include "tricorn_simd.h"
+#include <immintrin.h>
 
-static const float convergenceLimit = 2.0f; // Mandelbrot set limit
+static unsigned converge(float c_rea, float c_imag, unsigned n);
-static unsigned converge(float complex position,unsigned n);
 
 void tricorn_simd(float complex start, size_t width, ssize_t height, float res, unsigned n, _Bool color, unsigned char* img){
+    int bytesPerColor = color ? 3 : 1;
     
     // Absolute height for iteration purposes
-    height = height < 0 ?  height*-1 : height;
+    ssize_t actual_height = height < 0 ?  (ssize_t)(-height) : height;
+    ssize_t imag_step = height < 0 ? -1:1;
+    size_t simd_width = (width / 4) * 4;
 
-    // Simple iteration over every pixel in the picture
+    float c_start_real = creal(start);
-    for(size_t i = 0; i < width; i++){
+    float c_start_imag = cimag(start);
-        for(ssize_t j = 0; j < height; j++){
 
-            unsigned value = converge(start + (i*res - I*j*res),n);
+    __m128 four = _mm_set1_ps(4.0f);
+    __m128 minus_two = _mm_set1_ps(-2.0f);
 
-            // Pixel value position in the output array
+    for(ssize_t y = 0; y < actual_height;y++) {
-            size_t position = (j*width +i) *3;
+        float c_imag = c_start_imag - (y*imag_step) * res;
+        __m128 c_imag_vector = _mm_set1_ps(c_imag);
 
+        size_t x = 0;
+        for(; x < simd_width;x += 4) { //Simd
+            float c_real = c_start_real + (x*res);
+            __m128 c_real_vector = _mm_setr_ps(c_real,c_real+1.0f*res,c_real+2.0f*res,c_real+3.0f*res);
 
-            if(!color){ // Grayscale colors as specified by the assignment r = g = b
+            __m128 real = c_real_vector;
-                img[position] =  value;
+            __m128 imag = c_imag_vector;
-                img[position + 1] = value;
+            __m128i iterations = _mm_setzero_si128();
-                img[position + 2] = value;
+            __m128i active_mask = _mm_set1_epi32(-1);
+
+            for(unsigned i = 0; i < n;i++) {
+                __m128 real2 = _mm_mul_ps(real,real);
+                __m128 imag2 = _mm_mul_ps(imag,imag);
+
+                __m128 unbounded = _mm_cmplt_ps(_mm_add_ps(real2,imag2),four);
+                __m128i unbounded_i = _mm_castps_si128(unbounded);
+
+                active_mask = _mm_and_si128(active_mask,unbounded_i);
+
+                if(_mm_test_all_zeros(active_mask,active_mask)) {
+                    break;
+                }
+
+                iterations = _mm_sub_epi32(iterations,active_mask);
+
+                imag = _mm_add_ps(c_imag_vector,_mm_mul_ps(minus_two, _mm_mul_ps(real,imag)));
+                real = _mm_add_ps(_mm_sub_ps(real2,imag2),c_real_vector);
             }
-            else{ // Temporary colors with random multipliers
+            
+
+            __m128i mod64 = _mm_and_si128(iterations, _mm_set1_epi32(63));
+            __m128i times4 = _mm_slli_epi32(mod64, 2);
+            __m128i values = _mm_sub_epi32(_mm_set1_epi32(255), times4);
+
+            __m128i max_iter = _mm_set1_epi32(n);
+            __m128i is_max = _mm_cmpeq_epi32(iterations, max_iter);
+            values = _mm_andnot_si128(is_max, values);
+
+            unsigned int value_array[4];
+            _mm_storeu_si128((__m128i*)value_array,values);
+
+            for(unsigned i = 0 ; i < 4;i++) {
+                unsigned int value = value_array[i];
+                size_t position = (y*width + (x+i)) * bytesPerColor;
+
+                if(!color) {
+                    img[position] = value;
+                }
+                else {
+                    img[position] = ((value*3) % 255);
+                    img[position + 1] = ((value*2) % 255);
+                    img[position + 2] = ((value*5) % 255);
+                }
+            }
+        }
+
+        for(;x < width; x++) {
+            float c_real = c_start_real + (x*res);
+            unsigned int value = converge(c_real,c_imag,n);
+
+            size_t position = (y*width + x) * bytesPerColor;
+
+            if(!color) {
+                img[position] = value;
+            }
+            else {
                 img[position] = ((value*3) % 255);
-                img[position + 1] = ((value*2) % 255) ;
+                img[position + 1] = ((value*2) % 255);
                 img[position + 2] = ((value*5) % 255);
             }
-
         }
+
     }
 }
 
-// Iterative application of the equation given by the assignment
+static unsigned converge(float c_real, float c_imag, unsigned n) {
-static unsigned converge(float complex position,unsigned n){
+    float real = 0.0f;
+    float imag = 0.0f;
 
-    complex float result  = 0;
+    for (unsigned i = 0; i < n; i++) {
-    for(unsigned i = 0; i < n; i++){
+        float real2 = real * real;
-        // Application of the given complex equation with m = 2
+        float imag2 = imag * imag;
-        result = cpowf(conjf(result),2) + position;
 
-        // Detection of unbounded behaviour
+        if (real2 + imag2 > 4.0f) {
-        // If the magnitude of the result vector is greater than the convergence limit, we know that the function will grow unbounded
+            return 255 - 4 * (i % 64);
-        // Current convergence limit taken from mandelbrot set
+        }
-        if(sqrtf(powf (crealf(result),2)+powf(cimagf(result),2)) > convergenceLimit)
+
-            return 255-4*(i % 64);
+        imag = -2.0f * real * imag + c_imag;
+        real = real2 - imag2 + c_real;
     }
-    return 0;
+    return 0; 
 }