Added deferred shading workload

examples/README.txt

@@ -13,6 +13,7 @@ against regular serial C++ implementations, printing out a comparison of
 the runtimes and the speedup delivered by ispc.  It may be instructive to
 do a side-by-side diff of the C++ and ispc implementations of these
 algorithms to learn more about wirting ispc code.
+
  
 AOBench
 =======
@@ -27,6 +28,7 @@ It executes the program for the given number of iterations, rendering an
 (xres x yres) image each time and measuring the computation time with both
 serial and ispc implementations.
 
+
 AOBench_Instrumented
 ====================
 
@@ -40,12 +42,47 @@ is provided in the instrument.cpp file.
 *** Note: on Linux, this example currently hits an assertion in LLVM during
 *** compilation
 
+
+Deferred
+========
+
+This example shows an extensive example of using ispc for efficient
+deferred shading of scenes with thousands of lights; it's an implementation
+of the algorithm that Johan Andersson described at SIGGRAPH 2009,
+implemented by Andrew Lauritzen and Jefferson Montgomery.  The basic idea
+is that a pre-rendered G-buffer is partitioned into tiles, and in each
+tile, the set of lights that contribute to the tile is first computed.
+Then, the pixels in the tile are then shaded using just those light
+sources. (See slides 19-29 of
+http://s09.idav.ucdavis.edu/talks/04-JAndersson-ParallelFrostbite-Siggraph09.pdf
+for more details on the algorithm.)
+
+This directory includes three implementations of the algorithm:
+
+- An ispc implementation that first does a static partitioning of the
+  screen into tiles to parallelize across the CPU cores.  Within each tile
+  ispc kernels provide highly efficient implementations of the light
+  culling and shading calculations.
+- A "best practices" serial C++ implementation.  This implementation does a
+  dynamic partitioning of the screen, refining tiles with significant Z
+  depth complexity (these tiles often have a large number of lights that
+  affect them).  Within each final tile, the pixels are shaded using
+  regular C++ code.
+- If the Cilk extensions are available in your compiler, an ispc
+  implementation that uses Cilk will also be built.
+  (See http://software.intel.com/en-us/articles/intel-cilk-plus/).  Like 
+  the "best practices" serial implementation, this version does dynamic
+  tile partitioning for better load balancing and then uses ispc for the
+  light culling and shading.
+
+
 Mandelbrot
 ==========
 
 Mandelbrot set generation.  This example is extensively documented at the
 http://ispc.github.com/example.html page.
 
+
 Mandelbrot_tasks
 ================
 
@@ -58,6 +95,7 @@ using tasks with ispc, no task system is mandated; the user is free to plug
 in any task system they want, for ease of interoperating with existing task
 systems.
 
+
 Noise
 =====
 
@@ -71,6 +109,7 @@ Options
 This program implements both the Black-Scholes and Binomial options pricing
 models in both ispc and regular serial C++ code.
 
+
 RT
 ==
 
@@ -87,6 +126,7 @@ and triangle intersection code from pbrt; see the pbrt source code and/or
 "Physically Based Rendering" book for more about the basic algorithmic
 details.
 
+
 Simple
 ======
 
@@ -94,6 +134,7 @@ This is a simple "hello world" type program that shows a ~10 line
 application program calling out to a ~5 line ispc program to do a simple
 computation.
 
+
 Volume
 ======