-
Notifications
You must be signed in to change notification settings - Fork 3
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Merge pull request #8 from ryosuke-yamada/egl_multigpu_devel
[WIP] Multi gpu dynamic loading #6
- Loading branch information
Showing
2 changed files
with
355 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,305 @@ | ||
| # depends: pegl, PyGLM (not glm), PyOpenGL, plyfile, OpenEXR, PIL, matplotlib | ||
|
|
||
| import argparse | ||
|
|
||
| import pegl | ||
| from OpenGL.GL import * | ||
| from OpenGL.GLU import * | ||
| import glm | ||
| import ctypes | ||
| from plyfile import PlyData, PlyElement | ||
| import numpy as np | ||
| import OpenEXR, array, Imath | ||
| from PIL import Image | ||
| import math | ||
| import time | ||
| from matplotlib import pyplot as plt | ||
| import os | ||
| import glob | ||
| import random | ||
| from PIL import Image | ||
| from typing import List,Tuple | ||
| from tqdm import tqdm | ||
| from joblib import Parallel, delayed | ||
|
|
||
| # def myGLDebugCallback(source, mtype, id, severity, length, message, userParam): | ||
| # print("[GLDBG]") | ||
| # if mtype == GL_DEBUG_TYPE_ERROR: | ||
| # raise SystemError("[GLDBG]") | ||
|
|
||
| def createShader(vertFile:str, fragFile:str) -> GLuint: | ||
| with open(vertFile,'r',encoding='utf8') as fp: | ||
| vertShader_code = fp.read() | ||
| vertShader = glCreateShader(GL_VERTEX_SHADER) | ||
| glShaderSource(vertShader, [vertShader_code]) | ||
| glCompileShader(vertShader) | ||
| compiledStatus = glGetShaderiv(vertShader, GL_COMPILE_STATUS) | ||
| infoLog = glGetShaderInfoLog(vertShader) | ||
| if infoLog != '': | ||
| print(infoLog.decode('ascii')) | ||
| if compiledStatus == GL_FALSE: | ||
| raise Exception("Compile error in vertex shader.") | ||
|
|
||
| with open(fragFile,'r',encoding='utf8') as fp: | ||
| fragShader_code = fp.read() | ||
| fragShader = glCreateShader(GL_FRAGMENT_SHADER) | ||
| glShaderSource(fragShader, [fragShader_code]) | ||
| glCompileShader(fragShader) | ||
| compiledStatus = glGetShaderiv(fragShader, GL_COMPILE_STATUS) | ||
| infoLog = glGetShaderInfoLog(fragShader) | ||
| if infoLog != '': | ||
| print(infoLog.decode('ascii')) | ||
| if compiledStatus == GL_FALSE: | ||
| raise Exception("Compile error in fragment shader.") | ||
|
|
||
| shader_prog = glCreateProgram() | ||
| glAttachShader(shader_prog, vertShader) | ||
| glAttachShader(shader_prog, fragShader) | ||
| glDeleteShader(vertShader) | ||
| glDeleteShader(fragShader) | ||
|
|
||
| glLinkProgram(shader_prog) | ||
| shader_linked = ctypes.c_uint(0) | ||
| glGetProgramiv(shader_prog, GL_LINK_STATUS, ctypes.pointer(shader_linked)) | ||
| infoLog = glGetProgramInfoLog(shader_prog) | ||
| if infoLog != '': | ||
| print(infoLog.decode('ascii')) | ||
| if shader_linked == GL_FALSE: | ||
| raise Exception("Link error.") | ||
|
|
||
| return shader_prog | ||
|
|
||
| import torch, torchvision | ||
|
|
||
| class DynamicRenderer(torch.utils.data.Dataset): | ||
| def __init__(self, use_gpus : List[int], data_dir : str, base_seed: int, transform=None): | ||
| self.use_gpus = use_gpus | ||
| self.worker_id = -1 | ||
| self.base_seed = base_seed | ||
| self.transform = transform | ||
| self.width=480 | ||
| self.height=480 | ||
| self.npts=4000 | ||
|
|
||
| print("Constructing ply file list from %s..."%data_dir) | ||
| self.cat_list = sorted(os.listdir(data_dir)) | ||
| self.ply_lists = [] | ||
| self.ptss = [] | ||
| for cat_id, cat_ in tqdm(enumerate(self.cat_list)): | ||
| cat_path = os.path.join(data_dir, cat_) | ||
| ply_list = sorted(glob.glob(cat_path+"/*.ply")) | ||
| self.ply_lists += [(cat_id, ply) for ply in ply_list] | ||
| print("Done.") | ||
| self.ply_lists=self.ply_lists[0:512] | ||
| print("Loading ply files...") | ||
| def load_ply(cat_ply) -> Tuple[int,np.ndarray]: | ||
| cat_id, ply = cat_ply | ||
| with open(ply, 'rb') as f: | ||
| plydata = PlyData.read(f) | ||
| pts = np.array([ | ||
| np.asarray(plydata.elements[0].data['x']), | ||
| np.asarray(plydata.elements[0].data['y']), | ||
| np.asarray(plydata.elements[0].data['z']) | ||
| ], dtype=np.float32).T | ||
| return (cat_id, pts) | ||
| self.ptss=Parallel(n_jobs=-1,verbose=10)(delayed(load_ply)(cat_ply) for cat_ply in self.ply_lists) | ||
| # for cat_id, ply in self.ply_lists: | ||
| # pts=load_ply(ply) | ||
| # self.ptss.append((cat_id, pts)) | ||
| print("Done.") | ||
|
|
||
| def open(self, gpu_id:int): | ||
| '''open | ||
| Initialize OpenGL. | ||
| An OpenGL context requires a process. So, it shoud be executed on worker_init_fn. | ||
| ''' | ||
| if gpu_id<0 or gpu_id>=len(self.use_gpus): | ||
| raise ValueError("Invalid gpu id : %d" % gpu_id) | ||
|
|
||
| self.dpy = pegl.Display(self.use_gpus[gpu_id]) | ||
| print("Device %d: "%gpu_id,self.dpy.vendor,self.dpy.version_string) | ||
| pegl.bind_api(pegl.ClientAPI.OPENGL_API) | ||
| conf:pegl.Config = self.dpy.choose_config({ | ||
| pegl.ConfigAttrib.SURFACE_TYPE: pegl.SurfaceTypeFlag.PBUFFER_BIT, | ||
| pegl.ConfigAttrib.BLUE_SIZE: 8, | ||
| pegl.ConfigAttrib.GREEN_SIZE: 8, | ||
| pegl.ConfigAttrib.RED_SIZE: 8, | ||
| pegl.ConfigAttrib.ALPHA_SIZE: 8, | ||
| pegl.ConfigAttrib.DEPTH_SIZE: 24, | ||
| pegl.ConfigAttrib.RENDERABLE_TYPE: pegl.ClientAPIFlag.OPENGL, | ||
| #pegl.ConfigAttrib.CONTEXT_OPENGL_DEBUG: pegl.EGL_TRUE, | ||
| })[0] | ||
| self.ctx = conf.create_context() | ||
| self.surf = conf.create_pbuffer_surface({pegl.SurfaceAttrib.WIDTH: self.width, | ||
| pegl.SurfaceAttrib.HEIGHT: self.height}) | ||
| self.ctx.make_current(draw=self.surf) | ||
| print("config_id: ",self.ctx.config_id, self.ctx.config) | ||
|
|
||
| # print(glGetInteger(GL_VERSION)) | ||
| print("GL version", glGetInteger(GL_MAJOR_VERSION), glGetInteger(GL_MINOR_VERSION)) | ||
|
|
||
| # generate gpu memories | ||
| #self.npts = 10000 | ||
| col_buf=np.ones((self.npts,3,3),np.float32) | ||
| self.pts_vao = glGenVertexArrays(1) | ||
| self.pts_pos_vbo = glGenBuffers(1) | ||
| self.pts_col_vbo = glGenBuffers(1) | ||
| glBindVertexArray(self.pts_vao) | ||
| glBindBuffer(GL_ARRAY_BUFFER, self.pts_pos_vbo) | ||
| glBufferData(GL_ARRAY_BUFFER, sizeof(ctypes.c_float)*3*self.npts, None, GL_DYNAMIC_DRAW) | ||
| glEnableVertexAttribArray(0) | ||
| glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE, 0, ctypes.c_void_p(0)) | ||
| glBindBuffer(GL_ARRAY_BUFFER, self.pts_col_vbo) | ||
| glBufferData(GL_ARRAY_BUFFER, sizeof(ctypes.c_float)*3*self.npts, col_buf, GL_DYNAMIC_DRAW) | ||
| glEnableVertexAttribArray(1) | ||
| glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, 0, ctypes.c_void_p(0)) | ||
| glBindVertexArray(0) | ||
|
|
||
| self.pts_fbo = glGenFramebuffers(1) | ||
| self.pts_rbo_color = glGenRenderbuffers(1) | ||
| self.pts_rbo_depth = glGenRenderbuffers(1) | ||
| glBindRenderbuffer(GL_RENDERBUFFER, self.pts_rbo_color) | ||
| glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, self.width, self.height) | ||
| glBindRenderbuffer(GL_RENDERBUFFER, self.pts_rbo_depth) | ||
| glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT24, self.width, self.height) | ||
| glBindRenderbuffer(GL_RENDERBUFFER, 0) | ||
| glBindFramebuffer(GL_FRAMEBUFFER, self.pts_fbo) | ||
| glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, self.pts_rbo_color) | ||
| glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, self.pts_rbo_depth) | ||
| glBindFramebuffer(GL_FRAMEBUFFER, 0) | ||
|
|
||
| self.pts_pbo_color = glGenBuffers(1) | ||
| self.pts_pbo_depth = glGenBuffers(1) | ||
| glBindBuffer(GL_PIXEL_PACK_BUFFER, self.pts_pbo_color) | ||
| glBufferData(GL_PIXEL_PACK_BUFFER, sizeof(ctypes.c_uint8)*4*self.width*self.height, None, GL_STREAM_READ) | ||
| glBindBuffer(GL_PIXEL_PACK_BUFFER, self.pts_pbo_depth) | ||
| glBufferData(GL_PIXEL_PACK_BUFFER, 24//8*self.width*self.height, None, GL_STREAM_READ) | ||
| glBindBuffer(GL_PIXEL_PACK_BUFFER, 0) | ||
|
|
||
| # prepare shaders | ||
| self.shader = createShader( | ||
| os.path.join(os.path.dirname(os.path.abspath(__file__)),"unlit_shader.vert"), | ||
| os.path.join(os.path.dirname(os.path.abspath(__file__)),"unlit_shader.frag")) | ||
|
|
||
| mat_proj_ul = glGetUniformLocation(self.shader,"mat_proj") | ||
| self.mat_view_ul = glGetUniformLocation(self.shader,"mat_view") | ||
| self.mat_model_ul = glGetUniformLocation(self.shader,"mat_model") | ||
| fovy=math.pi/4 | ||
| aspect=self.width/self.height | ||
| zNear=0.1 | ||
| zFar=10.0 | ||
| mat_proj=glm.perspective(fovy,aspect,zNear,zFar) | ||
| camPos=glm.vec3(1,1,1) | ||
| gazePos=glm.vec3(0,0,0) | ||
| upDir=glm.vec3(0,1,0) | ||
| mat_view = glm.lookAt(camPos, gazePos, upDir) | ||
| mat_model = glm.mat4(1.0) | ||
| glUseProgram(self.shader) | ||
| glUniformMatrix4fv(mat_proj_ul, 1, GL_FALSE, glm.value_ptr(mat_proj)) | ||
| glUniformMatrix4fv(self.mat_view_ul, 1, GL_FALSE, glm.value_ptr(mat_view)) | ||
| glUniformMatrix4fv(self.mat_model_ul, 1, GL_FALSE, glm.value_ptr(mat_model)) | ||
| glUseProgram(0) | ||
|
|
||
| return | ||
|
|
||
| def __len__(self): | ||
| return len(self.ply_lists) | ||
|
|
||
| def __getitem__(self, id): | ||
| cat_id, pts = self.ptss[id] | ||
| pts = pts.astype(np.float32) | ||
| # send pts to gpu | ||
| glBindBuffer(GL_ARRAY_BUFFER, self.pts_pos_vbo) | ||
| glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(ctypes.c_float)*3*self.npts, pts) | ||
| glBindBuffer(GL_ARRAY_BUFFER, 0) | ||
|
|
||
|
|
||
| # render | ||
| glBindFramebuffer(GL_FRAMEBUFFER, self.pts_fbo) | ||
| glClearColor(0,0,0.5,1) | ||
| glViewport(0,0,self.width,self.height) | ||
| glClear(GL_COLOR_BUFFER_BIT|GL_DEPTH_BUFFER_BIT) | ||
| glUseProgram(self.shader) | ||
| glBindBuffer(GL_ARRAY_BUFFER, self.pts_vao) | ||
| glDrawArrays(GL_POINTS, 0, self.npts) | ||
| glBindBuffer(GL_ARRAY_BUFFER, 0) | ||
| glUseProgram(0) | ||
| glBindFramebuffer(GL_FRAMEBUFFER, 0) | ||
|
|
||
| try: | ||
| glBindFramebuffer(GL_READ_FRAMEBUFFER, self.pts_fbo) | ||
| glBindBuffer(GL_PIXEL_PACK_BUFFER, self.pts_pbo_color) | ||
| glReadPixels(0, 0, self.width, self.height, GL_RGBA, GL_UNSIGNED_BYTE, ctypes.c_void_p(0)) | ||
| # glBindBuffer(GL_PIXEL_PACK_BUFFER, self.pts_pbo_depth) | ||
| # glReadPixels(0, 0, self.width, self.height, GL_DEPTH_COMPONENT, GL_UNSIGNED_SHORT) | ||
| glBindBuffer(GL_PIXEL_PACK_BUFFER, 0) | ||
| glBindFramebuffer(GL_READ_FRAMEBUFFER, 0) | ||
| except OpenGL.error.GLError as e: | ||
| print(e.err, e.baseOperation,e.description) | ||
| print(e) | ||
| raise e | ||
|
|
||
|
|
||
| glFlush() | ||
| glFinish() | ||
|
|
||
| glBindBuffer(GL_PIXEL_PACK_BUFFER, self.pts_pbo_color) | ||
| p = glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY) | ||
| color_img = np.ctypeslib.as_array(ctypes.cast(p, ctypes.POINTER(ctypes.c_uint8)), shape=(self.height,self.width,4)) | ||
| glUnmapBuffer(GL_PIXEL_PACK_BUFFER) | ||
| # glBindBuffer(GL_PIXEL_PACK_BUFFER, self.pts_pbo_depth) | ||
| # p = glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY) | ||
| # zbuf_img = np.ctypeslib.as_array(ctypes.cast(p, ctypes.POINTER(ctypes.c_uint16)), shape=(self.height,self.width)).astype(np.float)/65535.0 | ||
| # glUnmapBuffer(GL_PIXEL_PACK_BUFFER) | ||
| glBindBuffer(GL_PIXEL_PACK_BUFFER, 0) | ||
|
|
||
| color_img = torch.Tensor(color_img[:, :, 0:3]).permute(2,0,1) | ||
| if self.transform is not None: | ||
| color_img = self.transform(color_img) | ||
|
|
||
| return cat_id, color_img | ||
|
|
||
|
|
||
|
|
||
| def worker_init_fn(worker_id:int) -> None: | ||
| info = torch.utils.data.get_worker_info() | ||
| _seed = info.dataset.base_seed+worker_id | ||
| random.seed(_seed) | ||
| np.random.seed(_seed) | ||
| torch.manual_seed(_seed) | ||
| torch.cuda.manual_seed(_seed) | ||
| info.dataset.open(worker_id) | ||
| return | ||
|
|
||
| if __name__=="__main__": | ||
| base_seed = 7743 | ||
|
|
||
| use_gpus=[0,1,2,3] | ||
| data_dir='./' | ||
| dataset = DynamicRenderer(use_gpus=use_gpus, data_dir=data_dir, base_seed = base_seed, transform = None) | ||
| dataloader = torch.utils.data.DataLoader(dataset, num_workers=len(use_gpus), batch_size=256, shuffle=True, | ||
| drop_last=True, worker_init_fn=worker_init_fn, persistent_workers=True) | ||
|
|
||
|
|
||
| random.seed(base_seed) | ||
| np.random.seed(base_seed) | ||
| torch.manual_seed(base_seed) | ||
| torch.cuda.manual_seed(base_seed) | ||
|
|
||
| # test loading | ||
| NUM_STEPS = 100 | ||
| LOG_STEP=1 | ||
| bt = time.time() | ||
| _imgs = None | ||
| for i, (labels, imgs) in enumerate(dataloader): | ||
| _imgs=imgs.clone() | ||
| if i%LOG_STEP==0: | ||
| print("Iter %d done."%i) | ||
| if i >= NUM_STEPS-1: | ||
| break | ||
| ct = time.time() | ||
| print("elapsed time: %.3f s (avg. %.3f s)"%(ct-bt,(ct-bt)/NUM_STEPS)) | ||
|
|
||
| pimg=Image.fromarray(np.array(torchvision.utils.make_grid(_imgs,nrow=16).permute(1,2,0)).astype(np.uint8)) | ||
| pimg.resize((1024,1024)) | ||
| pimg.save("test.png") |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,50 @@ | ||
| '''eglinfo - display all available graphic deviecs | ||
| ''' | ||
|
|
||
| import ctypes, pegl | ||
| from typing import List | ||
|
|
||
| EGLDeviceEXT = ctypes.c_void_p | ||
|
|
||
| def enumerateGraphicDevices(_maxDevices : int) -> List[EGLDeviceEXT]: | ||
| '''enumerateGraphicDevices | ||
| Enumerate available graphic devices. It includes software renderer if it is available. | ||
| Args: | ||
| _maxDevices : int - num of devices to enumerate | ||
| Return: | ||
| List[EGLDeviceEXT] - list of devices | ||
| Comment: | ||
| EGLDeviceEXT = ctypes.c_void_p | ||
| ''' | ||
| maxDevices = ctypes.c_int32(_maxDevices) | ||
| PEGLDevEXT = EGLDeviceEXT * _maxDevices | ||
| _eglDevs = [None] * _maxDevices | ||
| eglDevs = PEGLDevEXT(*_eglDevs) | ||
| numDevices = ctypes.c_int32(_maxDevices) | ||
| eglQueryDevicesEXT=pegl.egl._common._load_function("eglQueryDevicesEXT", None, [ctypes.c_int32, ctypes.POINTER(EGLDeviceEXT), ctypes.POINTER(ctypes.c_int32)]) | ||
| eglQueryDevicesEXT(maxDevices,eglDevs,ctypes.byref(numDevices)) | ||
| return [eglDevs[i] for i in range(numDevices.value)] | ||
|
|
||
|
|
||
| def printGraphicDevices(eglDevs : List[EGLDeviceEXT]) -> None: | ||
| '''printGraphicDevices | ||
| Print device name of graphic devices. | ||
| Args: | ||
| eglDevs : List[EGLDeviceEXT] - list of devices | ||
| Comment: | ||
| EGLDeviceEXT = ctypes.c_void_p | ||
| ''' | ||
| eglQueryDeviceStringEXT=pegl.egl._common._load_function("eglQueryDeviceStringEXT", ctypes.c_char_p, [EGLDeviceEXT, ctypes.c_int32]) | ||
| for i,d in enumerate(eglDevs): | ||
| s = eglQueryDeviceStringEXT(d, pegl.egl.EGL_EXTENSIONS) | ||
| print(i,s) | ||
| # EGL_NV_device_cuda | ||
| return | ||
|
|
||
| if __name__=="__main__": | ||
| printGraphicDevices(enumerateGraphicDevices(10)) |