Effectively Utilizing 3DNow! in Linux

XYGRAD is passed as a pointer to the original image, called img_ptr, along with the size of the row to process. The function steps through each image row four pixels at a time. For a set of four pixels, the change in intensity in the x direction is calculated first. The quadword containing P0 and P1 is moved with movq into MMX register MM0. P2 and P3 are moved into the MM1 register. These registers are then subtracted using the 3DNow! packed subtraction (i.e., PFSUB MM0, MM1). The result is stored in the first operand, MM0, which is later squared using the packed multiplication operation (PFMUL MM0, MM0). Similar steps are followed to calculate the gradient in the y direction, with the squared difference being stored in the MM2 register. MM0 and MM2 are then added to get delta(x)<+>2<+> +delta(y)<+>2<+>, which is stored in the output array referenced by result_img_ptr. After the whole image has been processed, a second 3DNow!-optimized module is called by the C program to calculate the square root of each pixel in the resulting image. This completes the gradient calculation. The complete source code for both the C and assembly modules used in the range image gradient calculation program can be downloaded from http://merlin.cs.uah.edu/visgig/threednow/gradient/.

One thing we kept in mind during the coding process is the ability of the K6-2 and K6-3 CPUs to pipeline instructions in two execution pipelines. Due to the architecture of these processors, certain restrictions apply to the pipelining of 3DNow! instructions. Namely, each 3DNow! instruction belongs to one of two subsets, and two instructions from the same subset cannot be issued in parallel. For instance, the packed floating-point subtraction (PFSUB) and packed floating-point addition (PFADD) both belong to the same subset, and therefore cannot be issued in the same clock cycle. On the other hand, the packed floating-point multiplication (PFMUL) belongs to the other subset of 3DNow! instructions; therefore, PFMUL and PFADD instructions can execute simultaneously, provided there is no operand dependency between them. It's beneficial to interweave instructions from each subset as much as possible to increase the likelihood of parallel computation. Most integer MMX operations do not have such a restriction on the K6-2 and K6-3; achieving optimal floating-point performance does require a little more consideration by the programmer.

Our 3DNow!-optimized gradient-calculation programs delivered excellent performance. We conducted tests of the programs on a dual-bootable PC with a 300MHz AMD K6-2 CPU to determine 3DNow! performance in both Windows and Linux environments. On Linux, we used the Netwide Assembler (NASM) version 0.98 to assemble the assembly modules. We used GNU C (gcc) version 2.7 to compile a C driver and link the driver to the assembly code. On Windows, we used a popular commercial C compiler to assemble, compile and link the assembly and C codes. The Windows assembler/compiler did not recognize 3DNow! instructions, so we had to include the AMD header file that defines 3DNow! using the pseudo-instruction macros. We tested the 3DNow!-optimized gradient calculations on several range images and 3-D volumes.

Tables 1 and 2 show performance results for two representative data sets. In Table 1, the average CPU times for execution of the 3DNow!-optimized range image gradient calculation in Linux and Windows environments are shown. (Times are averaged over 2000 trials for a 240x240 range image.) By comparison, the 3DNow!-optimized code ran about nine times faster under Linux than unoptimized standard C code which computes the data according to the same pattern of pixel access, i.e., the unoptimized code was implemented purely in C without use of 3DNow! and was compiled without any compiler optimizations. When the standard C version of the gradient was compiled with maximal optimizations by gcc in Linux (using optimization switches of -O2 -ffast-math), it was still more than three times slower than the 3DNow!-optimized code. This time, improvement is significant; the 3DNow!-optimized version of the range image gradient calculation can be done in real time at frame rates under Linux. The performance improvement was nearly as impressive for the volume gradient computations—the times shown in Table 2 are for a 72x256x256 volume. The 3DNow!-optimized volume gradient executed more than 2.5 times faster than fully optimized standard C implementation, and better than 4.5 times faster than an unoptimized standard C implementation.

Table 1

We've also tested the performance of the standard C range image gradient implementation (i.e., the implementation without 3DNow! code) on a Pentium II/333 running Linux. Using full compiler optimizations, the 240x240 range image's gradient was computed in 17% more time on the PII than on the K6-2. Thus, we estimate that the K6-2 can calculate range image gradient under Linux about 30% faster than a comparably clocked PII. Of course, we should add that it's generally easier to develop modules in C than in assembly language.

Table 2