aboutsummaryrefslogtreecommitdiff
path: root/Gem/manual/Texture.html
diff options
context:
space:
mode:
authorTravis CI <zmoelnig@travis-ci.umlaeute.mur.at>2019-05-09 20:35:45 +0000
committerTravis CI <zmoelnig@travis-ci.umlaeute.mur.at>2019-05-09 20:35:45 +0000
commit9b873693bdcbfe439a4157564ece781f67e58239 (patch)
treec9b187b8fbfb7ce45034f9ac38e4981f0c6f7902 /Gem/manual/Texture.html
parentf1c1528406b219db176d913890d86532c1b9aaa0 (diff)
Gem 01b1861db023a225d01a4d5a519cf54f09f631f5 linux/amd64
built 'master:01b1861db023a225d01a4d5a519cf54f09f631f5' for linux/amd64
Diffstat (limited to 'Gem/manual/Texture.html')
-rw-r--r--Gem/manual/Texture.html252
1 files changed, 126 insertions, 126 deletions
diff --git a/Gem/manual/Texture.html b/Gem/manual/Texture.html
index 1de889c..0e35a1d 100644
--- a/Gem/manual/Texture.html
+++ b/Gem/manual/Texture.html
@@ -1,126 +1,126 @@
-<!doctype html public "-//w3c//dtd html 4.0 transitional//en">
-<html>
-<head>
- <meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1">
- <meta name="Author" content="Mark Danks">
- <meta name="Author" content="IOhannes m zmölnig">
- <title>Texture mapping</title>
-</head>
-<body>
-
-<center>
-<h2>
-<u>Texture Mapping</u></h2></center>
-<a href="Gloss.html#Texture">Texture mapping</a> is the act of applying
-pixel data to a geometric object. In GEM, this is achieved with the
-<i>[pix_texture]</i>
-object. It is important to understand that the
-<i>[pix_texture]</i>
-object merely sets the pix as the current texture. It does not do
-any rendering! You need to use a geo object which does texture mapping.
-All of the basic geo objects can texture map, such as <i>[square]</i> or
-<i>[sphere]</i>.
-<p><img SRC="tribar.gif" height=13 width=561>
-<p>A simple example of texture mapping is the following patch:
-<center>
-<p><img SRC="texture.jpg" BORDER=1 height=182 width=160></center>
-
-<p>This patch can be found at 07.texture/01.texture.pd. Change
-the number box connected to the rotate object to see what a texture map
-on a cube looks like.
-<p>The <i>[pix_image]</i> object loads in the fractal image file. The
-<i>[pix_texture]</i>
-object says that the pix data should be used as a texture map. Notice
-that this is different than the previous manual section when we used the
-<i>[pix_draw]</i> object. The final object in the chain is the <i>[cube]</i>
-object. Because we have enabled texture mapping with the <i>[pix_texture]</i>
-object, the cube takes the pix data and applies it to the geometry.
-<p><img SRC="tribar.gif" height=13 width=561>
-<p>Texture mapping can be used with any GEM object. In the previous
-manual section, you saw how to load in pix data with a variety of objects,
-including <i>[pix_multiimage]</i> and <i>[pix_video]</i>. All of these
-objects can be used with the <i>[pix_texture]</i> object.
-<p>Because the pix data is applied to geometry, you can move, rotate, and
-scale the image. This is extremely useful on the <i>[square]</i> object.
-Instead of doing a one-to-one pixel mapping as occurs with the <i>[pix_draw]</i>
-object, you can resize and reshape the image.
-<p>OpenGL originally required that images must have dimensions that are power-of-2, such as 64, 128, or 256. This restriction has been released with recent gfx-cards
-(like some radeon/nvidia products).
-However, if the width or height of an image is not a power of two,
-then the <i>[pix_texture]</i> object will take care of this,
-and still render it (depending on you hardware with some tricks).
-You can thus texture images of any size, but since this is based on tricking
-the texture-coordinates, <i>[pix_coordinate]</i> might not give the wanted result any more.
-<p><img SRC="tribar.gif" height=13 width=561>
-<p>The example patch 07.texture/02.moveImages.pd is a much more complex
-patch which uses alpha blending to create a transparent object, in this
-case, the dancer. Make sure to turn on the rotation with the <i>[metro]</i>
-object.
-<p><img SRC="tribar.gif" height=13 width=561><a href="index.html"></a>
-<p>People have been asking how textures are handled in GEM. Here
-is a long explanation from an email which I wrote.
-<p><tt> Here is how textures are dealt with under OpenGL and hardware
-accelerators. This can obviously change in the future, but right
-now, I am fairly certain that the info is correct (I make games in my day
-job, so I have vested interest in this :-)</tt><tt></tt>
-<p><tt> The amount of memory (VRAM) on the card (12mb for Voodoo2,
-16mb for TNT, 64mb for GeForce2, etc) is used for both textures (TRAM)
-and frame buffer space. If you have a large rendering window, like
-1600x1200, it will take up 1600x1200x4x3 in 32-bit mode with double buffering
-and a Z buffer (or 23mb). Most people run at TV resolution, like
-NTSC, so it takes 640x480x4x3 = 3.7mb All of the space left
-is for textures onboard the card (FYI, if you have heard that people are
-having problems with the PlayStation2, notice that it only has 4mb of VRAM...not
-much onboard texture space, huh? :-) Thankfully it has an <i>extremely</i>
-fast DMA bus)</tt><tt></tt>
-<p><tt> Sooo, when GEM "creates" a texture, it immediately tries
-to send the texture to the card, which uses some of the left over space
-in the VRAM. If you had a 640x480 window on a Voodoo2, you have ~8mb
-of texture space left over. On a GeForce2, ~60mb. The problem
-is what happens if you want more textures than can fit into TRAM.
-OpenGL requires that the video drivers deal with the problem, so GEM doesn't
-care too much (more about this later).</tt><tt></tt>
-<p><tt> In most cases, the drivers cache the textures in main memory
-and if a texture is requested for rendering and it isn't resident on the
-card, it will download it. If you have AGP, then this is pretty quick,
-although none of 3dfx cards really take advantage of this (ie, those cards
-are about the same speed as the PCI bus). So depending on the number
-of textures, and how complex the scene is, you might be able to display
-more textures than you have TRAM.</tt><tt></tt>
-<p><tt> One slowdown that can happen with GEM is that it makes a
-copy of the image before sending it down the chain of objects. If
-you are constantly changing images with a pix_multiimage, this can be a
-performance hit, but you can modify the actual pixel data with the pix
-objects. The pixels aren't sent to the graphics card until the pix_texture
-object is reached.</tt><tt></tt>
-<p><tt> GEM tries to help with this with a few objects. pix_imageInPlace
-acts much the same as pix_multiimage, but it downloads _every_ image in
-the sequence to the card when a download message is recieved. It
-also immediately turns on texturing, instead of making a copy (ie, you
-don't need a pix_texture object). Much faster, but not as flexible.
-pix_movie does much the same thing. It sends the pixel data without
-copying it if there is a new frame to display.</tt><tt></tt>
-<p><tt> The entire pix system uses a caching system so that the copying
-and processing only occurs if something actually changes. For example,
-if you had a pix_threshold object, it would only process when rendering
-started...and every time that the values actually changed. You can
-use pix_buf to isolate parts which don't change from those that do, but
-it involves another copy.</tt><tt></tt>
-<p><tt> On the Voodoo2, the hardware itself limits textures to 256x256...this
-will never change. The newest Voodoo5 boards have a higher texture
-size.</tt><tt></tt>
-<p><tt> If you load the _exact_ same image (this means the exact
-same file/path name), then the pix_image has a cache system which means
-that it is only loaded into the</tt>
-<br><tt>computers memory once. However, each pix_image still sends
-its own copy down to the gfx card.</tt><tt></tt>
-<p><tt> You could use a single [pix_image]/[pix_texture] with [separator]
-to do this...I have done it a lot in the past.</tt><tt></tt>
-<p><tt> The reason that [pix_image] doesn't share the actual texture
-data is that you can modify the pixel data with other pix objects...[pix_image]
-doesn't actually send the texture data to the gfx card, [pix_texture] does.</tt>
-<p><img SRC="tribar.gif" height=13 width=561><a href="index.html"></a>
-<p><a href="index.html">[return]</a>
-<br>
-</body>
-</html>
+<!doctype html public "-//w3c//dtd html 4.0 transitional//en">
+<html>
+<head>
+ <meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1">
+ <meta name="Author" content="Mark Danks">
+ <meta name="Author" content="IOhannes m zmölnig">
+ <title>Texture mapping</title>
+</head>
+<body>
+
+<center>
+<h2>
+<u>Texture Mapping</u></h2></center>
+<a href="Gloss.html#Texture">Texture mapping</a> is the act of applying
+pixel data to a geometric object. In GEM, this is achieved with the
+<i>[pix_texture]</i>
+object. It is important to understand that the
+<i>[pix_texture]</i>
+object merely sets the pix as the current texture. It does not do
+any rendering! You need to use a geo object which does texture mapping.
+All of the basic geo objects can texture map, such as <i>[square]</i> or
+<i>[sphere]</i>.
+<p><img SRC="tribar.gif" height=13 width=561>
+<p>A simple example of texture mapping is the following patch:
+<center>
+<p><img SRC="texture.jpg" BORDER=1 height=182 width=160></center>
+
+<p>This patch can be found at 07.texture/01.texture.pd. Change
+the number box connected to the rotate object to see what a texture map
+on a cube looks like.
+<p>The <i>[pix_image]</i> object loads in the fractal image file. The
+<i>[pix_texture]</i>
+object says that the pix data should be used as a texture map. Notice
+that this is different than the previous manual section when we used the
+<i>[pix_draw]</i> object. The final object in the chain is the <i>[cube]</i>
+object. Because we have enabled texture mapping with the <i>[pix_texture]</i>
+object, the cube takes the pix data and applies it to the geometry.
+<p><img SRC="tribar.gif" height=13 width=561>
+<p>Texture mapping can be used with any GEM object. In the previous
+manual section, you saw how to load in pix data with a variety of objects,
+including <i>[pix_multiimage]</i> and <i>[pix_video]</i>. All of these
+objects can be used with the <i>[pix_texture]</i> object.
+<p>Because the pix data is applied to geometry, you can move, rotate, and
+scale the image. This is extremely useful on the <i>[square]</i> object.
+Instead of doing a one-to-one pixel mapping as occurs with the <i>[pix_draw]</i>
+object, you can resize and reshape the image.
+<p>OpenGL originally required that images must have dimensions that are power-of-2, such as 64, 128, or 256. This restriction has been released with recent gfx-cards
+(like some radeon/nvidia products).
+However, if the width or height of an image is not a power of two,
+then the <i>[pix_texture]</i> object will take care of this,
+and still render it (depending on you hardware with some tricks).
+You can thus texture images of any size, but since this is based on tricking
+the texture-coordinates, <i>[pix_coordinate]</i> might not give the wanted result any more.
+<p><img SRC="tribar.gif" height=13 width=561>
+<p>The example patch 07.texture/02.moveImages.pd is a much more complex
+patch which uses alpha blending to create a transparent object, in this
+case, the dancer. Make sure to turn on the rotation with the <i>[metro]</i>
+object.
+<p><img SRC="tribar.gif" height=13 width=561><a href="index.html"></a>
+<p>People have been asking how textures are handled in GEM. Here
+is a long explanation from an email which I wrote.
+<p><tt> Here is how textures are dealt with under OpenGL and hardware
+accelerators. This can obviously change in the future, but right
+now, I am fairly certain that the info is correct (I make games in my day
+job, so I have vested interest in this :-)</tt><tt></tt>
+<p><tt> The amount of memory (VRAM) on the card (12mb for Voodoo2,
+16mb for TNT, 64mb for GeForce2, etc) is used for both textures (TRAM)
+and frame buffer space. If you have a large rendering window, like
+1600x1200, it will take up 1600x1200x4x3 in 32-bit mode with double buffering
+and a Z buffer (or 23mb). Most people run at TV resolution, like
+NTSC, so it takes 640x480x4x3 = 3.7mb All of the space left
+is for textures onboard the card (FYI, if you have heard that people are
+having problems with the PlayStation2, notice that it only has 4mb of VRAM...not
+much onboard texture space, huh? :-) Thankfully it has an <i>extremely</i>
+fast DMA bus)</tt><tt></tt>
+<p><tt> Sooo, when GEM "creates" a texture, it immediately tries
+to send the texture to the card, which uses some of the left over space
+in the VRAM. If you had a 640x480 window on a Voodoo2, you have ~8mb
+of texture space left over. On a GeForce2, ~60mb. The problem
+is what happens if you want more textures than can fit into TRAM.
+OpenGL requires that the video drivers deal with the problem, so GEM doesn't
+care too much (more about this later).</tt><tt></tt>
+<p><tt> In most cases, the drivers cache the textures in main memory
+and if a texture is requested for rendering and it isn't resident on the
+card, it will download it. If you have AGP, then this is pretty quick,
+although none of 3dfx cards really take advantage of this (ie, those cards
+are about the same speed as the PCI bus). So depending on the number
+of textures, and how complex the scene is, you might be able to display
+more textures than you have TRAM.</tt><tt></tt>
+<p><tt> One slowdown that can happen with GEM is that it makes a
+copy of the image before sending it down the chain of objects. If
+you are constantly changing images with a pix_multiimage, this can be a
+performance hit, but you can modify the actual pixel data with the pix
+objects. The pixels aren't sent to the graphics card until the pix_texture
+object is reached.</tt><tt></tt>
+<p><tt> GEM tries to help with this with a few objects. pix_imageInPlace
+acts much the same as pix_multiimage, but it downloads _every_ image in
+the sequence to the card when a download message is recieved. It
+also immediately turns on texturing, instead of making a copy (ie, you
+don't need a pix_texture object). Much faster, but not as flexible.
+pix_movie does much the same thing. It sends the pixel data without
+copying it if there is a new frame to display.</tt><tt></tt>
+<p><tt> The entire pix system uses a caching system so that the copying
+and processing only occurs if something actually changes. For example,
+if you had a pix_threshold object, it would only process when rendering
+started...and every time that the values actually changed. You can
+use pix_buf to isolate parts which don't change from those that do, but
+it involves another copy.</tt><tt></tt>
+<p><tt> On the Voodoo2, the hardware itself limits textures to 256x256...this
+will never change. The newest Voodoo5 boards have a higher texture
+size.</tt><tt></tt>
+<p><tt> If you load the _exact_ same image (this means the exact
+same file/path name), then the pix_image has a cache system which means
+that it is only loaded into the</tt>
+<br><tt>computers memory once. However, each pix_image still sends
+its own copy down to the gfx card.</tt><tt></tt>
+<p><tt> You could use a single [pix_image]/[pix_texture] with [separator]
+to do this...I have done it a lot in the past.</tt><tt></tt>
+<p><tt> The reason that [pix_image] doesn't share the actual texture
+data is that you can modify the pixel data with other pix objects...[pix_image]
+doesn't actually send the texture data to the gfx card, [pix_texture] does.</tt>
+<p><img SRC="tribar.gif" height=13 width=561><a href="index.html"></a>
+<p><a href="index.html">[return]</a>
+<br>
+</body>
+</html>