Enumerate physical devices to check for any supporting shaderFloat64

Use double precision
2025-01-17 15:14:24 +00:00 · 2025-01-17 15:14:24 +00:00
5 changed files with 81 additions and 384 deletions
--- a/7
+++ b/7
@ -10,10 +10,3 @@ double precision shaders.
 To build and run:

   ./build.sh && ./gpu-frac
-
-Since both GLSL and C use CPP directives, I parameterised the program
-a bunch and stuck a load of cool presets in config.h, which can be
-simply be uncommented and commented as is required.  All the #ifdefs
-make the code a bit harder to read but it's nice being able to twiddle
-numbers and change the visualisation without messing around editing
-and commenting out code.
--- a/build.sh
+++ b/build.sh
@ -2,4 +2,4 @@
 set -e
 glslc -fshader-stage=vert vert_shader.glsl -o vert_shader.spv
 glslc -fshader-stage=frag frag_shader.glsl -o frag_shader.spv
-clang -std=c11 -pedantic -Wall -Wextra -lglfw -lm -lvulkan main.c -o gpu-fractal
+clang -std=c11 -pedantic -Wall -Wextra -lglfw -lvulkan main.c -o gpu-fractal
--- a/config.h
+++ b/config.h
@ -1,77 +0,0 @@
-/*
- * Copyright (c) Camden Dixie O'Brien
- * SPDX-License-Identifier: AGPL-3.0-only
- */
-
-#define MAXITER 1000
-
-// Pretty Julia set zoom
-//#define SS 9
-//#define JULIA
-//#define JULIA_C_RE 0.285
-//#define JULIA_C_IM 0.010
-//#define CENTRE_RE -0.47353731
-//#define CENTRE_IM -0.18894516
-//#define ZOOMRATE 0.01
-
-// Cool zoomy swirly animated Julia set
-//#define SS 9
-//#define JULIA
-//#define JULIA_C_RE 0.300
-//#define JULIA_DTHETA 0.00001
-//#define ZOOMRATE 0.0004
-
-// Nice Mandelbrot zoom
-//#define SS 9
-//#define CENTRE_RE -0.743643887
-//#define CENTRE_IM  0.131825904
-//#define ZOOMRATE 0.004
-
-// Cool warping Mandelbrot zoom
-//#define SS 4
-//#define CENTRE_RE -0.900000000
-//#define POWRATE 0.0001
-//#define ZOOMRATE 0.001
-
-// Funky swirly, warping Julia zoom
-//#define SS 9
-//#define JULIA
-//#define JULIA_C_RE 0.5
-//#define JULIA_C_IM 0.5
-//#define CENTRE_RE 0.5
-//#define CENTRE_IM -0.5
-//#define JULIA_DTHETA -0.00002
-//#define POWRATE 0.0001
-//#define ZOOMRATE 0.0003
-
-// Cool tendril stuff (complex power)
-//#define SS 4
-//#define CPOW_DTHETA 0.000
-//#define MAXMAG 100.0
-//#define CENTRE_RE -0.84880415
-//#define CENTRE_IM  0.03921157
-//#define CPOW_START_RE 1.96550429
-//#define CPOW_START_IM 0.36985874
-//#define ZOOMRATE 0.01
-
-// ??
-//#define SS 4
-//#define THING
-//#define JULIA_C_RE -0.9999999506519784
-//#define JULIA_C_IM 3.141592601914583e-4
-//#define JULIA_DTHETA 0.00001
-//#define MAXMAG 100.0
-//#define STARTSCALE 4
-//#define ZOOMRATE 0.01
-
-// Funky complex power julia set
-#define SS 4
-#define JULIA
-#define JULIA_C_RE -0.010
-#define JULIA_C_IM 0.285
-#define JULIA_DTHETA 0.0001
-#define CPOW_START_RE 0.0
-#define CPOW_START_IM 2.0
-#define CPOW_DTHETA 0.0001
-#define MAXMAG 100.0
-#define STARTSCALE 1
--- a/frag_shader.glsl
+++ b/frag_shader.glsl
@ -4,142 +4,54 @@
 */

 #version 450
+#extension GL_ARB_gpu_shader_fp64 : enable

-#include "config.h"
-
-#define NCOLS 8
-
-#if !defined(CPOW_DTHETA) && !defined(THING)
-#define MAXMAG 2.0
-#endif
+#define MAXITER 1000

 layout(location = 0) out vec4 out_colour;

 layout(push_constant) uniform Constants {
-	vec2 img_size;
-#if defined(JULIA) || defined(THING)
-	vec2 julia;
-#endif
-	vec2 centre;
-#ifdef CPOW_DTHETA
-	vec2 cpow;
-#endif
-#ifdef POWRATE
-	float zpow;
-#endif
-	float scale;
+	dvec2 img_size;
+	dvec2 shift;
+	double zoom;
 } params;

-vec3 incol = vec3(0.000000, 0.014444, 0.027321);
-
-vec3 palette[NCOLS] = vec3[](
-	vec3(0.000000, 0.049707, 0.107023),
-	vec3(0.025187, 0.064803, 0.177888),
-	vec3(0.254152, 0.080220, 0.274677),
-	vec3(0.502886, 0.080220, 0.278894),
-	vec3(1.000000, 0.124772, 0.119538),
-	vec3(1.000000, 0.234551, 0.030713),
-	vec3(1.000000, 0.381326, 0.000000),
-	vec3(1.000000, 0.651406, 0.215861)
-);
-
-#if SS == 4
-vec2 ss_offsets[SS] = vec2[](
+vec2 ss_offsets[4] = vec2[](
 	vec2(-0.25, -0.25),
 	vec2( 0.25, -0.25),
 	vec2(-0.25,  0.25),
 	vec2( 0.25,  0.25)
 );
-#elif SS == 9
-vec2 ss_offsets[SS] = vec2[](
-    vec2(-0.33, -0.33), vec2(0.0, -0.33), vec2(0.33, -0.33),
-    vec2(-0.33,  0.0),  vec2(0.0,  0.0),  vec2(0.33,  0.0),
-    vec2(-0.33,  0.33), vec2(0.0,  0.33), vec2(0.33,  0.33)
-);
-#else
-#error "Unsupported supersampling count"
-#endif

-vec2 maptoz(vec2 xy)
+dvec2 maptoz(vec2 xy)
 {
-	float aspect = params.img_size.x / params.img_size.y;
-	vec2 z = 4.0 * (xy / params.img_size.xy - 0.5);
-	z.x *= aspect;
-	return params.scale * z + params.centre;
+	dvec2 dxy = dvec2(xy);
+	dvec2 uv = ((dxy / params.img_size.xy) * 4.0LF - 2.0LF);
+	return params.zoom * uv + params.shift;
 }

-vec2 cmul(vec2 a, vec2 b)
-{
-	return vec2(a.x * b.x - a.y * b.y, a.x * b.y + a.y * b.x);
-}
-
-vec2 cexp(vec2 z)
-{
-	return exp(z.x) * vec2(cos(z.y), sin(z.y));
-}
-
-vec2 cln(vec2 z)
-{
-	return vec2(log(length(z)), atan(z.y, z.x));
-}
-
-vec2 cpow(vec2 z, vec2 p)
-{
-	if (z.x == 0.0 && z.y == 0.0) return vec2(0.0, 0.0);
-	return cexp(cmul(p, cln(z)));
-}
-
-vec2 zpow(vec2 z, float p)
-{
-    float mag = pow(length(z), p);
-    float arg = p * atan(z.y, z.x);
-    return mag * vec2(cos(arg), sin(arg));
-}
-
-int fractal(vec2 c) {
-#ifdef JULIA
-	vec2 z = c;
-	c = params.julia;
-#else
-	vec2 z = vec2(0.0, 0.0);
-#endif
+double mandelbrot(dvec2 c) {
+	dvec2 z = dvec2(0.0LF, 0.0LF);
 	for (int i = 0; i < MAXITER; ++i) {
-#ifdef CPOW_DTHETA
-		z = cpow(z, params.cpow) + c;
-#elif defined(POWRATE)
-		z = zpow(z, params.zpow) + c;
-#elif defined(THING)
-		z = cpow(z, c) + params.julia;
-#else
-		z = vec2(z.x * z.x - z.y * z.y, 2.0 * z.x * z.y) + c;
-#endif
-		if (length(z) > MAXMAG)
-			return i;
+		z = dvec2(z.x * z.x - z.y * z.y, 2.0LF * z.x * z.y) + c;
+		if (length(z) > 2.0LF)
+			return double(i) / double(MAXITER);
 	}
-	return -1;
+	return 0.0;
 }

-vec3 colour(int i) {
-	if (i == -1) {
-		return incol;
-	} else {
-		float c = float(NCOLS) * sqrt(float(i) / float(MAXITER));
-		int fc = int(floor(c));
-		int cc = int(ceil(c));
-		float p = c - float(fc);
-		p = smoothstep(0.0, 1.0, p);
-		return mix(palette[fc], palette[cc], p);
-	}
+vec3 colour(double i) {
+	float fi = float(i);
+	return vec3(fi, fi * fi, fi * fi * fi);
 }

 void main()
 {
-	vec2 c;
+	dvec2 c;
 	vec3 col = vec3(0.0, 0.0, 0.0);
-	for (int i = 0; i < SS; ++i) {
+	for (int i = 0; i < 4; ++i) {
 		c = maptoz(gl_FragCoord.xy + ss_offsets[i]);
-		col += colour(fractal(c));
+		col += colour(mandelbrot(c));
 	}
-	col /= SS;
 	out_colour = vec4(col, 1.0);
 }
--- a/main.c
+++ b/main.c
@ -5,12 +5,8 @@

 #define GLFW_INCLUDE_VULKAN

-#include "config.h"
-
 #include <GLFW/glfw3.h>
 #include <assert.h>
-#include <complex.h>
-#include <math.h>
 #include <stdbool.h>
 #include <stdio.h>
 #include <string.h>
@ -19,130 +15,25 @@
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define CLAMP(x, min, max) MAX((min), MIN(x, (max)))

+#define PHYSICAL_DEV_BUFFER_SIZE 8
 #define QUEUE_FAMILY_PROP_BUFFER_SIZE 16
 #define EXT_PROP_BUFFER_SIZE 256
 #define SURFACE_FMT_BUFFER_SIZE 64
 #define PRESENT_MODE_BUFFER_SIZE 8
 #define SWAPCHAIN_IMG_BUFFER_SIZE 4

-#define MAX_SHADER_SIZE (8 * 1024)
-
-#define FASTFORWARD_RATE 100
-
-#if defined(CPOW_DTHETA) && defined(POWRATE)
-#error "Cannot define CPOW_DTHETA and POWRATE simulataneously"
-#endif
-
-#ifndef STARTSCALE
-#define STARTSCALE 1.0
-#endif
+#define MAX_SHADER_SIZE (4 * 1024)

 typedef struct {
-	float width, height;
-#if defined(JULIA) || defined(THING)
+	double width, height;
 	struct {
-		float re, im;
-	} julia;
-#endif
-	struct {
-		float re, im;
-	} centre;
-#ifdef CPOW_DTHETA
-	struct {
-		float re, im;
-	} cpow;
-#endif
-#ifdef POWRATE
-	float zpow;
-#endif
-	float scale;
+		double x, y;
+	} shift;
+	double zoom;
 } params_t;

 static const char *layers[] = { "VK_LAYER_KHRONOS_validation" };
-static const char *extensions[] = {
-	VK_KHR_SWAPCHAIN_EXTENSION_NAME,
-};
-
-#define NEXTENSIONS (sizeof(extensions) / sizeof(extensions[0]))
-
-// Initial parameters
-static params_t params = {
-#if defined(JULIA) || defined(THING)
-	.julia = {
-#ifdef JULIA_C_RE
-		.re = JULIA_C_RE,
-#endif
-#ifdef JULIA_C_IM
-		.im = JULIA_C_IM,
-#endif
-	},
-#endif
-#if defined(CENTRE_RE) || defined(CENTRE_IM)
-	.centre = {
-#ifdef CENTRE_RE
-		.re = CENTRE_RE,
-#endif
-#ifdef CENTRE_IM
-		.im = CENTRE_IM,
-#endif
-	},
-#endif
-#ifdef CPOW_DTHETA
-	.cpow = { .re = CPOW_START_RE, .im = CPOW_START_IM },
-#endif
-#ifdef POWRATE
-	.zpow = 2.0,
-#endif
-	.scale = STARTSCALE,
-};
-
-static bool paused = false, fastforward = false;
-
-static void
-mouse_button_callback(GLFWwindow *window, int button, int action, int mods)
-{
-	(void)window;
-	(void)button;
-	(void)mods;
-
-	if (action == GLFW_RELEASE)
-		return;
-
-	double x, y;
-	glfwGetCursorPos(window, &x, &y);
-
-	double zx = 4.0 * (x / params.width - 0.5);
-	double zy = 4.0 * (y / params.height - 0.5);
-	zx *= params.width / params.height;
-
-	zx = params.scale * zx + params.centre.re;
-	zy = params.scale * zy + params.centre.im;
-
-	printf("CLICKED: %.8f + %.8fi\n", zx, zy);
-	params.centre.re = zx;
-	params.centre.im = zy;
-}
-
-static void
-key_callback(GLFWwindow *window, int key, int scancode, int action, int mods)
-{
-	(void)window;
-	(void)scancode;
-	(void)mods;
-
-	switch (key) {
-	case GLFW_KEY_SPACE:
-		if (action == GLFW_PRESS)
-			paused = !paused;
-		break;
-	case GLFW_KEY_F:
-		if (action == GLFW_PRESS)
-			fastforward = true;
-		else if (action == GLFW_RELEASE)
-			fastforward = false;
-		break;
-	}
-}
+static const char *swapchain_ext_name = VK_KHR_SWAPCHAIN_EXTENSION_NAME;

 static VkShaderModule load_shader_module(VkDevice dev, const char *path)
 {
@ -168,24 +59,18 @@ static VkShaderModule load_shader_module(VkDevice dev, const char *path)

 int main(void)
 {
-	// Set up window
+	// Set up window and surface
 	glfwInit();
 	glfwWindowHint(GLFW_CLIENT_API, GLFW_NO_API);
 	glfwWindowHint(GLFW_RESIZABLE, GLFW_FALSE);
-	GLFWmonitor *monitor = glfwGetPrimaryMonitor();
-	const GLFWvidmode *mode = glfwGetVideoMode(monitor);
-	GLFWwindow *window = glfwCreateWindow(
-	    mode->width, mode->height, "GPU Fractal", monitor, NULL);
+	GLFWwindow *window
+	    = glfwCreateWindow(900, 900, "GPU Fractal", NULL, NULL);
 	assert(window);

-	// Register input callbacks
-	glfwSetMouseButtonCallback(window, mouse_button_callback);
-	glfwSetKeyCallback(window, key_callback);
-
 	// Initialise Vulkan instance.
 	const VkApplicationInfo app_info = {
 		.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO,
-		.pApplicationName = "GPU Fractal",
+		.pApplicationName = "Vulkan Test",
 		.applicationVersion = VK_MAKE_VERSION(1, 0, 0),
 		.pEngineName = "No Engine",
 		.engineVersion = VK_MAKE_VERSION(1, 0, 0),
@ -212,10 +97,22 @@ int main(void)
 	assert(result == VK_SUCCESS);

 	// Select a physical device
+	VkPhysicalDevice physical_devs[PHYSICAL_DEV_BUFFER_SIZE];
+	uint32_t dev_count = PHYSICAL_DEV_BUFFER_SIZE;
+	result = vkEnumeratePhysicalDevices(inst, &dev_count, physical_devs);
+	assert(result == VK_SUCCESS);
+	bool found_physical_dev = false;
 	VkPhysicalDevice physical_dev;
-	uint32_t dev_count = 1;
-	result = vkEnumeratePhysicalDevices(inst, &dev_count, &physical_dev);
-	assert(result == VK_SUCCESS || result == VK_INCOMPLETE);
+	for (unsigned i = 0; i < dev_count; ++i) {
+		// Check that the device supports double precision
+		VkPhysicalDeviceFeatures features;
+		vkGetPhysicalDeviceFeatures(physical_devs[i], &features);
+		if (features.shaderFloat64) {
+			physical_dev = physical_devs[i];
+			found_physical_dev = true;
+		}
+	}
+	assert(found_physical_dev);

 	// Select queue family
 	uint32_t queue_family_count = QUEUE_FAMILY_PROP_BUFFER_SIZE;
@ -239,28 +136,6 @@ int main(void)
 	}
 	assert(found_queue_family);

-	// Check that the physical device supports all extensions
-	uint32_t ext_count = EXT_PROP_BUFFER_SIZE;
-	VkExtensionProperties ext_props[EXT_PROP_BUFFER_SIZE];
-	result = vkEnumerateDeviceExtensionProperties(
-	    physical_dev, NULL, &ext_count, ext_props);
-	assert(result == VK_SUCCESS);
-	bool supported[NEXTENSIONS] = { 0 };
-	for (unsigned i = 0; i < ext_count; ++i) {
-		for (unsigned j = 0; j < NEXTENSIONS; ++j) {
-			if (strcmp(extensions[j], ext_props[i].extensionName) == 0) {
-				supported[j] = true;
-				break;
-			}
-		}
-	}
-	for (unsigned j = 0; j < NEXTENSIONS; ++j) {
-		if (!supported[j]) {
-			fprintf(stderr, "Unsupported: %s\n", extensions[j]);
-			assert(false);
-		}
-	}
-
 	// Create the logical device and get the queue handle.
 	float queue_priority = 1.0;
 	const VkDeviceQueueCreateInfo queue_create_info = {
@ -269,12 +144,16 @@ int main(void)
 		.queueCount = 1,
 		.pQueuePriorities = &queue_priority,
 	};
+	const VkPhysicalDeviceFeatures enabled_features = {
+		.shaderFloat64 = VK_TRUE,
+	};
 	const VkDeviceCreateInfo dev_create_info = {
 		.sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
 		.pQueueCreateInfos = &queue_create_info,
 		.queueCreateInfoCount = 1,
-		.enabledExtensionCount = NEXTENSIONS,
-		.ppEnabledExtensionNames = extensions,
+		.enabledExtensionCount = 1,
+		.ppEnabledExtensionNames = &swapchain_ext_name,
+		.pEnabledFeatures = &enabled_features,
 	};
 	VkDevice dev;
 	result = vkCreateDevice(physical_dev, &dev_create_info, NULL, &dev);
@ -282,6 +161,21 @@ int main(void)
 	VkQueue queue;
 	vkGetDeviceQueue(dev, queue_family, 0, &queue);

+	// Check that the physical device has swap chain support
+	uint32_t ext_count = EXT_PROP_BUFFER_SIZE;
+	VkExtensionProperties ext_props[EXT_PROP_BUFFER_SIZE];
+	result = vkEnumerateDeviceExtensionProperties(
+	    physical_dev, NULL, &ext_count, ext_props);
+	assert(result == VK_SUCCESS);
+	bool swapchain_support = false;
+	for (unsigned i = 0; i < ext_count; ++i) {
+		if (strcmp(swapchain_ext_name, ext_props[i].extensionName) == 0) {
+			swapchain_support = true;
+			break;
+		}
+	}
+	assert(swapchain_support);
+
 	// Select a surface format, preferring R8G8B8A8_SRGB with
 	// SRGB_NONLINEAR colour space.
 	uint32_t surface_fmt_count = SURFACE_FMT_BUFFER_SIZE;
@ -584,15 +478,15 @@ int main(void)
 	result = vkCreateFence(dev, &fence_config, NULL, &in_flight);
 	assert(result == VK_SUCCESS);

-	// Initialise shader parameters
-	params.width = (float)swapchain_extent.width;
-	params.height = (float)swapchain_extent.height;
-#ifdef JULIA_DTHETA
-	const double complex julia_rotz = cexp(I * JULIA_DTHETA);
-#endif
-#ifdef CPOW_DTHETA
-	const double complex cpow_rotz = cexp(I * CPOW_DTHETA);
-#endif
+	params_t params = {
+		.width = (double)swapchain_extent.width,
+		.height = (double)swapchain_extent.height,
+		.shift = {
+			.x = -0.743643887037158704752191506114774,
+			.y = 0.131825904205311970493132056385139,
+		},
+		.zoom = 1.0,
+	};

 	while (!glfwWindowShouldClose(window)) {
 		glfwPollEvents();
@ -629,7 +523,6 @@ int main(void)
 		vkCmdBeginRenderPass(
 		    cmd_buf, &render_pass_begin, VK_SUBPASS_CONTENTS_INLINE);

-		// Send parameters
 		vkCmdPushConstants(
 		    cmd_buf, pipeline_layout, VK_SHADER_STAGE_FRAGMENT_BIT, 0,
 		    sizeof(params), &params);
@ -672,32 +565,8 @@ int main(void)
 		result = vkQueuePresentKHR(queue, &present_info);
 		assert(result == VK_SUCCESS);

-		if (paused)
-			continue;
-
-		for (int i = 0; i < (fastforward ? FASTFORWARD_RATE : 1); ++i) {
-#ifdef JULIA_DTHETA
-			double complex julia = params.julia.re + I * params.julia.im;
-			julia *= julia_rotz;
-			params.julia.re = (float)creal(julia);
-			params.julia.im = (float)cimag(julia);
-#endif
-
-#ifdef ZOOMRATE
-			params.scale *= (1 - ZOOMRATE);
-#endif
-
-#ifdef CPOW_DTHETA
-			double complex cpow = params.cpow.re + I * params.cpow.im;
-			cpow *= cpow_rotz;
-			params.cpow.re = (float)creal(cpow);
-			params.cpow.im = (float)cimag(cpow);
-#endif
-
-#ifdef POWRATE
-			params.zpow += POWRATE;
-#endif
-		}
+		// Increment zoom
+		params.zoom *= 0.99;
 	}

 	vkDeviceWaitIdle(dev);
Author	SHA1	Message	Date
Camden Dixie O'Brien	a8acdc4b01	Enumerate physical devices to check for any supporting shaderFloat64	2025-01-17 15:14:24 +00:00
Camden Dixie O'Brien	03364ad387	Use double precision	2025-01-17 15:14:24 +00:00