Shader Pipeline Details¶
The heart of Shady is a customizable, specialized “fragment shader” program written in GLSL. It is designed to ensure that the GPU takes on nearly all of the burden of frame-by-frame pixel processing, including signal generation, animation, spatial windowing, contrast modulation in time and space, gamma correction, quantization, and dynamic-range enhancement tricks.
Specifically, the steps in the shader pipeline are as follows:
Compute a carrier pattern. Depending on the settings of the
Stimulus, this may be (a) a patch of solid color, (b) a pre-determined texture (i.e. an array of discrete pixel values), (c) a signal function that generates patterns procedurally at run-time, or (d) a combination of these options.
The rules for combining different carrier types are as follows:
- Texture and
- If you omit the texture, solid
.backgroundColoris used as the signal baseline instead.
- If you omit the
.signalFunction, it looks the same as a signal function that outputs 0 everywhere.
.color, if supplied, is a multiplier for both texture and
.signalFunction(though often with qualitatively different effects, because texture values are always in the range 0 to 1 whereas signal functions can go negative).
- An exception to all the above rules is if you omit both texture and
.signalFunction, but specify a
.color: then it is assumed you want a solid patch of the specified
.color, independent of the
The table below summarizes the possible outcomes:
.color+ signal *
texture + signal texture no texture,
.backgroundColor+ signal *
no texture, no
- Texture and
Translate, rotate or scale the carrier pattern if requested. Carriers are treated as infinitely, cyclically repeating patterns for this purpose. In the case of procedurally-generated signals, the transformations are actually applied to the coordinate system before the
Apply contrast effects. This may include (a) a
.windowingFunctionthat attenuates the edges of the stimulus, (b) another more arbitrary procedural
.modulationFunction(for example, sinusoidal contrast modulation), (c) an overall
.contrastscaling factor, or (d) a multiplicative combination of these options.
For linearized, psychophysically-accurate stimuli, the
backgroundAlphaproperty should be 1.0, in which case contrast effects cause the carrier to be blended with the specified
backgroundColor. On the other hand, if the
backgroundAlpha < 1.0, contrast effects are mediated through alpha blending with other stimuli (and you should not expect the composite result to be accurately linearized).
Translate, rotate or scale the complete stimulus if requested. (Note that scaling, and any rotation except a multiple of 90 degrees, will compromise the pixel-perfect accuracy of the stimulus content due to interpolation artifacts. As for translations: under normal circumstances Shady automatically rounds to the nearest pixel, to avoid such interpolation artifacts in ordinary stimulus repositioning.)
Add noise, if requested. A two-dimensional uniform or Gaussian pixel noise pattern can be added to the stimulus at this stage. It is useful at very low amplitudes if we intend to apply a bit-stealing lookup-table in the next step, or at higher amplitudes if we actually want a visible noise effect.
Apply gamma-correction and dynamic-range enhancement effects. This is done by one of the following procedures:
- DEFAULT OPTION:
transform each channel’s pixel values through the inverse of the screen non-linearity, then dither each color-channel independently between the DAC values immediately above and immediately below the transformed target level, according to the noisy-bit algorithm of Allard & Faubert (2008);
transform each channel’s pixel values through the inverse of the screen non-linearity, then re-express the resulting values as 16-bit integers, distributing the more- and less-significant bytes either between color channels or between adjacent pixels in video memory: specialized hardware such as the ViewPixx or Bits# can then reinterpret the video content as a high-dynamic-range pattern at the expense of either color or resolution;
quantize pixel values according to the size of a large (say, 16-bit) pre-generated lookup table, then use the table to look up a triplet of (red,green,blue) DAC values—for monochromatic stimuli this can simultaneously accomplish linearization, bit-stealing (after Tyler, 1997) if desired, and further quantization down to the native precision of the video hardware;
If not already accomplished in the previous step, quantize down to 8 bits per color channel (or however many bits are supported natively in video memory).