A comparison of pixel throughput when executing a simple shader program that fetches once from a floatN texture, performs a few ADD operations, and outputs the result to a floatN buffer.

fpfilltest -p -n -c 1
fpfilltest -p -n -c 2
fpfilltest -p -n -c 3
fpfilltest -p -n -c 4

fpfilltest -b -n -c 1
fpfilltest -b -n -c 2
fpfilltest -b -n -c 3
fpfilltest -b -n -c 4

Note: Both graphs represent the same test. One shows MPix/sec and one reports GB/sec.

A comparison of performance obtained when issuing a shader using either a screen covering triangle or a large quad. The shader reads from a (1 or 4-component) float texture, performs a few ADD instructions, and outputs the result to a (1 or 4-component) pbuffer.

fpfilltest.exe -n -c 1 -r triangle
fpfilltest.exe -n -c 1 -r quad
fpfilltest.exe -n -c 4 -r triangle
fpfilltest.exe -n -c 4 -r quad

Floating point input bandwidth test (4-component texels). Single access test (SGL) measures bandwidth when repeatedly accessing texel (0,0). Sequential access (SEQ) is 1-to-1 copy of input texels to output texels. Random access (DEP-RAND) uses dependent texturing to randomly fetch from the dependent texture. The shaders perform a total of 4 fetches, each from a unique input texture (dependent texturing case performs an additional fetch from an index texture, which is factored into the bandwidth computation).

floatbandwidth -n -c 4 -f 4 -a single
floatbandwidth -n -c 4 -f 4 -a seq
floatbandwidth -n -c 4 -f 4 -a random -d

Measures output bandwidth when storing results into multiple 4-component floating point render targets. Output into 1 to 4 render targets is tested.

outputbandwidth -n -o 1
outputbandwidth -n -o 2
outputbandwidth -n -o 3
outputbandwidth -n -o 4

Measures time taken to execute a shader containing a fixed number texture fetches followed by various numbers of MAD instructions. The number of instructions following the fetches is increased (x-axis) until the shader becomes compute bound. Above this threshold, running time is a linear function of the length of the program.

fetchcosts -n -m 1 -x 60 -f 1 -a single -i 2
fetchcosts -n -m 1 -x 60 -f 2 -a single -i 2
fetchcosts -n -m 1 -x 60 -f 3 -a single -i 2
fetchcosts -n -m 1 -x 60 -f 4 -a single -i 2
fetchcosts -n -m 1 -x 60 -f 5 -a single -i 2
fetchcosts -n -m 1 -x 60 -f 6 -a single -i 2

Measures time taken to execute a shader containing a fixed number texture fetches followed by various numbers of MAD instructions. The number of instructions following the fetches is increased (x-axis) until the shader becomes compute bound. Above this threshold, running time is a linear function of the length of the program.

fetchcosts -n -m 1 -x 60 -f 1 -a seq -i 2
fetchcosts -n -m 1 -x 60 -f 2 -a seq -i 2
fetchcosts -n -m 1 -x 60 -f 3 -a seq -i 2
fetchcosts -n -m 1 -x 60 -f 4 -a seq -i 2
fetchcosts -n -m 1 -x 60 -f 5 -a seq -i 2
fetchcosts -n -m 1 -x 60 -f 6 -a seq -i 2

Measures readback performance of glReadPixels(0, 0, 512, 512, FORMAT, TYPE, ptr) from a single buffered window and 4-component float pbuffer.

readback.exe -x -r
readback.exe -f -a

Measures rate at which texture data can be loaded onto the card using multiple calls to glTexSubImage2D(). The texture is a 512x512 n-component float texture.

download.exe -n -c 1
download.exe -n -c 2
download.exe -n -c 3
download.exe -n -c 4

Measures rates at which various shader instructions can be executed. Vector instructions are executed with 4-component vector operands.

instrissue -n -a -l 64 -m

Compares the rate a which the card can perform ADD, SUB, MUL, and MAD instructions when operating on both scalar and 4-component vector operands. Any dual issue capability in the hardware should be apparent when performing the single component version of the instructions.

instrissue -n -c 1 -i add -l 40 -m
instrissue -n -c 4 -i add -l 40 -m
instrissue -n -c 1 -i sub -l 40 -m
instrissue -n -c 4 -i sub -l 40 -m
instrissue -n -c 1 -i mul -l 40 -m
instrissue -n -c 4 -i mul -l 40 -m
instrissue -n -c 1 -i mad -l 40 -m
instrissue -n -c 4 -i mad -l 40 -m

Measures the effectiveness of the early-z test at saving work. The test generates a computation mask which is loaded into the depth buffer and used to kill all fragments beneath a given threshold while a big shader is run. The test measures a uniformly distributed set of thresholds, a uniformly distributed set of 4x4 blocks, and a wavefront pattern.

branching --step 5 --block --earlyz
branching --step 5 --blockN 4 --earlyz
branching --step 5 --rand --earlyz

Measures the effectiveness of pixel-shader 3.0 branching at saving work. The test generates a computation mask and then runs a pixel-shader 3.0 shader which fetches the mask value and then conditionally executes a big shader. The test measures a uniformly distributed set of thresholds, a uniformly distributed set of 4x4 blocks, and a wavefront pattern.

branching --step 5 --block --ps30
branching --step 5 --blockN 2 --ps30
branching --step 5 --blockN 4 --ps30
branching --step 5 --blockN 8 --ps30
branching --step 5 --blockN 16 --ps30
branching --step 5 --blockN 32 --ps30
branching --step 5 --rand --ps30

Plots average and maximum error of complex arithmetic instructions. Comparison is performed against double precision CPU calculations. The first graph samples inputs over the range of 0.001 to PI/2. The second graph samples over 10^-12 to 0.001 (only tests functions which behave nicely at the origin).

precision.exe -m 0.001 -x 1.5707963267949 -b
precision.exe -m 1.0e-12 -x 0.001 -bsce