AVX draftHEADmaster

4 files changed, 60 insertions, 24 deletions

diff --git a/Makefile b/Makefile
index 5ee5d60..27c551e 100644
--- a/Makefile
+++ b/Makefile
@@ -8,7 +8,7 @@ OBJ_DIR = obj
 
 CC = gcc
 OFLAG = -Ofast -funroll-loops -ffast-math
-CCFLAGS = -I$(INC_DIR) -Wall -Wextra $(OFLAG)
+CCFLAGS = -I$(INC_DIR) -Wall -Wextra $(OFLAG) -mavx
 LDFLAGS = -lm -lpthread \
 		  $(shell pkg-config --libs --cflags sdl2 SDL2_ttf)
 
diff --git a/inc/mandel.h b/inc/mandel.h
index 80cb5f4..d5c8d9b 100644
--- a/inc/mandel.h
+++ b/inc/mandel.h
@@ -1,6 +1,10 @@
 #ifndef MANDEL_H
 # define MANDEL_H
 
+# ifndef __AVX__
+#  error "AVX not supported"
+# endif
+
 # include <stdlib.h>
 # include <stdbool.h>
 # include <math.h>
@@ -78,6 +82,7 @@ typedef struct
 
 // mandelbrot.c
 int					mandelbrot(double ca, double cb, int iterations);
+void				mandelbrot_avx(State *state, Color *pixels, int width, int height);
 
 // state.c
 bool				state_init(State *state);
diff --git a/src/mandelbrot.c b/src/mandelbrot.c
index 3c8ab98..3f47d1b 100644
--- a/src/mandelbrot.c
+++ b/src/mandelbrot.c
@@ -21,3 +21,54 @@ int mandelbrot(double ca, double cb, int iterations)
     }
     return n;
 }
+
+void mandelbrot_avx(State *state, Color *pixels, int width, int height)
+{
+	__m256d real_step = _mm256_set1_pd((state->real_end - state->real_end) / (double)width);
+	__m256d imag_step = _mm256_set1_pd((state->imag_end - state->imag_end) / (double)height);
+
+	__m256i	ones       = _mm256_set1_epi64x(1);
+	__m256i iterations = _mm256_set1_epi64x(state->iterations);
+
+	__m256d	twos  = _mm256_set1_pd(2.0);
+	__m256d	fours = _mm256_set1_pd(4.0);
+
+	__m256d	offset = _mm256_setr_pd(0.0, 1.0, 2.0, 3.0);
+
+	__m256d ci = _mm256_set1_pd(0.0);
+	for (int y = 0; y < height; y++)
+	{
+		__m256d cr = _mm256_set1_pd(0.0);
+		for (int x = 0; x < width; x += 4)
+		{
+			__m256d	zr = cr;
+			__m256d	zi = ci;
+			__m256d	zr_square;
+			__m256d	zi_square;
+			__m256i	n;
+
+			/* while { */
+
+				zi_square = _mm256_mul_pd(zi, zi);
+				zr_square = _mm256_mul_pd(zr, zr);
+
+				__m256d dist = _mm256_add_pd(zi_square, zr_square);
+				__m256d escaped = _mm256_cmp_pd(dist, fours, _CMP_LT_OQ);
+
+
+				// zi = 2.0 * zi * zr
+				zi = _mm256_mul_pd(zi, zr);
+				zi = _mm256_mul_pd(zi, twos);
+
+				zr = _mm256_sub_pd(zr_square, zi_square);
+				zi = _mm256_add_pd(zi, ci);
+				zr = _mm256_add_pd(zr, cr);
+
+			/* } */
+
+
+			cr = _mm256_add_pd(cr, real_step);
+		}
+		ci = _mm256_add_pd(ci, imag_step);
+	}
+}
diff --git a/src/state.c b/src/state.c
index c1bdacb..b0bc38b 100644
--- a/src/state.c
+++ b/src/state.c
@@ -47,36 +47,16 @@ void	state_run(State *state)
     {
         event_handle(state);
 
-		double	real_step;
-		double	imag_step;
-		double	real;
-		double	imag;
+		int		_;
 		void	*pixels;
-		int		pitch;
 		int		width;
 		int		height;
 
-		SDL_CALL(SDL_LockTexture(state->texture, NULL, &pixels, &pitch));
+		SDL_CALL(SDL_LockTexture(state->texture, NULL, &pixels, &_));
 		SDL_CALL(SDL_QueryTexture(state->texture, NULL, NULL, &width, &height));
 
-		real_step = (state->real_end - state->real_start) / width;
-		imag_step = (state->imag_end - state->imag_start) / height;
-		imag = state->imag_start;
-
 		uint32_t render_start_time = SDL_GetTicks();
-
-		for (int i = 0; i < height; i++)
-		{
-			real = state->real_start;
-			for (int j = 0; j < width; j++)
-			{
-				int n = mandelbrot(real, imag, state->iterations);
-				((Color*)pixels)[i * width + j] = state->palette[n];
-				real += real_step;
-			}
-			imag += imag_step;
-		}
-
+		mandelbrot_avx(state, pixels, width, height);
 		uint32_t render_end_time = SDL_GetTicks();
 
 		SDL_UnlockTexture(state->texture);