|
|
|
|
Orc TutorialOrc Tutorial — Getting started writing Orc code. |
This section walks you through several examples of increasing complexity to get you started working with Orc. Each of these examples are available in the Orc source code, in the examples directory. The first three examples use static Orc code that is in a source file, and is compiled into intermediate C code by the orcc tool.
The first example demonstrates how to add two arrays of 16-bit signed integers together. A possible use case for this is combining two stereo audio streams together.
The second example builds from the first, replacing one of the stereo input streams with a mono stream, converting it to stereo in the process, and also adjusting the volume of the stream.
The third example shows how to convert a planar 4:2:0 video image into a packed 4:4:4 video image with an alpha channel.
This example demonstrates combining two stereo audio streams by adding. Uncompressed audio streams (i.e., PCM format) can be in a variety of formats, but one of the most common is interleaved signed 16-bit integers, and we will choose that for the purposes of this example. Extending to other formats is left as an exercise for the reader. Interleaved means that left and right channel samples are consecutive: in memory, the data look like LRLRLR... The sampling rate is unimportant, as long as both streams are the same.
One important feature/limitation of signed 16-bit audio samples is that adding two together could cause an overflow. For example, adding the value 25000 to 10000 gives 35000, but this overflows 16 bits, so a standard addition would instead give the value -30536 (35000-65536). Overflows handled this way sound like crackling or worse, so we would like a better solution. One solution is to use saturating addition: in this case, the addition of 25000 and 10000 would be limited by the upper end of signed 16-bit values to give 32767. Although this still causes distortion in the output signal, it is much less audible and annoying.
In normal C code, 16-bit saturating addition is difficult to express without using 32-bit intermediates. In Orc, saturating addition is a basic operation with opcodes for each size, both signed and unsigned. In this case, we want "addssw", for "add signed saturated word".
Also, we're going to make a one simplification: Adding two interleaved stereo streams is the same as adding two mono streams with twice as many samples. So we'll use 2*n_samples in the calling code.
To the code:
.function audio_add_s16 .dest 2 d1 .source 2 s1 .source 2 s2 addssw d1, s1, s2
Line by line:
.function audio_add_s16
This starts a function. A function (represented internally by the object OrcProgram) is equivalent to a C function. When you generate C code from this Orc exmaple using the orcc tool, it generates a C stub function called "audio_add_s16()", which at runtime will generate an OrcProgram object corresponding to the above code, compile it, and then run it.
.dest 2 d1 short
This specifies that you want a destination (output) array named "d1", with the element size being 2. Orc does not differentiate between signed and unsigned arrays (or even floating point), however, you may optionally specify a type afterwards that will be used in any autogenerated C code.
.source 2 s1 short .source 2 s2 short
This specifies that you want two source (input) arrays, "s1" and "s2", similar to the destination array.
addssw d1, s1, s2
This specifies the (only) opcode that we want for this program: signed saturated addition of each member of the two source arrays, and store the result in the destination array.
A few notes about the above program: The loop over the array members is implied. Everything that Orc does is based on looping over each array element and executing the opcodes in a program.
When you generate C code from the above Orc code using 'orcc --implementation example1.orc', you get a bunch of boilerplate code, plus three C functions:
/* audio_add_s16 */
#ifdef DISABLE_ORC
void
audio_add_s16 (int16 * d1, const int16 * s1, const int16 * s2, int n)
{
...
}
This function is used if DISABLE_ORC is defined. As one might guess, if you define DISABLE_ORC, no runtime Orc features are used, and all calls to audio_add_s16() use this function. The interior of the function is a for() loop that implements the Orc function. The generated code may not necessarily be easy to read, but it is straightforward: all the verbosity and use of unions is to avoid compiler warnings without making the compiler too complex. But this is the place to go if you are trying to understand what Orc is doing.
#else
static void
_backup_audio_add_s16 (OrcExecutor * ORC_RESTRICT ex)
{
...
}
This function is used when runtime Orc is enabled, but Orc was unable to generate code for the function at runtime. There are various reasons why that might happen -- unimplemented rules for a target, or more temporary variables used than available registers.
void
audio_add_s16 (short * d1, const short * s1, const short * s2, int n)
{
...
}
The third generated function is the important part: It is used when Orc is enabled at runtime, and creates the OrcProgram corresponding to the function you defined. Then it compiles the function and calls it.
After generating the C code, you should generate the header file, using: 'orcc --header example1orc.orc -o example1orc.h'. After similar boilerplate code, there is the expected declaration of audio_add_s16():
void audio_add_s16 (short * d1, const short * s1, const short * s2, int n);
Some C code to generate sample data, call the generated code, and print out the results:
#include <stdio.h>
#include "example1orc.h"
#define N 10
short a[N];
short b[N];
short c[N];
int
main (int argc, char *argv[])
{
int i;
/* Create some data in the source arrays */
for(i=0;i < N;i++){
a[i] = 100*i;
b[i] = 32000;
}
/* Call a function that uses Orc */
audio_add_s16 (c, a, b, N);
/* Print the results */
for(i=0;i < N;i++){
printf("%d: %d %d -> %d\n", i, a[i], b[i], c[i]);
}
return 0;
}
The output of the program:
0: 0 32000 -> 32000 1: 100 32000 -> 32100 2: 200 32000 -> 32200 3: 300 32000 -> 32300 4: 400 32000 -> 32400 5: 500 32000 -> 32500 6: 600 32000 -> 32600 7: 700 32000 -> 32700 8: 800 32000 -> 32767 9: 900 32000 -> 32767
In this example, we will expand on the previous example by making one of the input arrays a mono stream, and also scale the mono input stream by a volume. Rather than iterating over each signed 16-bit value, in this example we will iterate over samples, meaning the member size for the stereo arrays is 4, since each array member contains a left and right 16 bit value.
.function audio_add_mono_to_stereo_scaled_s16 .dest 4 d1 short .source 4 s1 short .source 2 s2 short .param 2 volume .temp 4 s2_scaled .temp 2 t .temp 4 s2_stereo mulswl s2_scaled, s2, volume shrsl s2_scaled, s2_scaled, 12 convssslw t, s2_scaled mergewl s2_stereo, t, t x2 addssw d1, s1, s2_stereo
Piece by piece:
.function audio_add_mono_to_stereo_scaled_s16 .dest 4 d1 short .source 4 s1 short .source 2 s2 short
This is the same as the previous example, except that the stereo arrays are increased in size to 4. However, we'll use the short type, since Orc does not care what type we use, and short is the type of the array we want to use in the C code.
.param 2 volume
This specifies a parameter, which is an integer that is passed to an Orc function. In the generated C code, parameters are always of type int. There are also float parameters for the floating point equivalent.
.temp 4 s2_scaled .temp 2 t .temp 4 s2_stereo
This specifies a few temporary variables that are used later in the code. These definitions are similar to defining local variables in C code. Note that the size is important: each opcode has specific sizes for source and destination operands, and it is important to match these correctly with temporary variables.
mulswl s2_scaled, s2, volume shrsl s2_scaled, s2_scaled, 12
This scales the mono input: signed multiply of s2 and volume, giving a 32-bit value, and then a signed right shift by 12. Since the second operand of mulswl is 16-bit, only the lower 16 bits of volume will be used in the multiply. The right shift is effectively the same as dividing by 4096. Thus, a neutral scaling that does not increase or decrease the mono input would correspond to calling the function with a parameter value of 4096.
convssslw t, s2_scaled mergewl s2_stereo, t, t
The first instruction is "convert saturated signed 32-bit to signed 16-bit", and the second merges the two values of (16 bit) t into the high and low halves of s2_stereo. This duplicates the mono signal into the right and left channels. It is important to use the saturated conversion, since the effective scaling value may have been greater than 1.0, thus the larger values may need to be clipped.
x2 addssw d1, s1, s2_stereo
The "x2" prefix indicates that we want the operation specified to be done twice, first to the upper half of all operands, and again separately to the lower half of all operands. Since addssw is normally a 16-bit operation, the x2 prefix causes it to be a 32-bit operation. And so, it adds the newly created right and left values of the scaled mono signal into the s1 signal.
There are several variations of the above program that might be more suitable for a particular application. This function only handles a limited dynamic range of volume scaling factors, however, by changing the shift constant, or turning the shift into a parameter, the dynamic range can be increased significantly.
The third example shows how to convert a planar 4:2:0 video image into a packed 4:4:4 video image with an alpha channel. The first format is often referred to as I420 and the second as AYUV.
For simplicity in the following discussion, we'll assume that the image dimensions are 640x480. The 4:2:0 subsampling means the input chroma planes are 320x240 (subsampled by 2 in each direction). These need to be upsampled to 640x480, then repacked with the input Y plane, with an added dummy alpha value. There are many ways to perform upsampling; the simplest is to duplicate each value horizontally and vertically. The result is low quality, but adequate for demonstration purposes.
There are several choices for the Orc array size and dimensionality. Iterating vertically can be done in the C code or in the Orc code. If done in the Orc code, we would need to use an array size of 240 and have two separate arrays for the even and odd Y rows. If done in the C code, there is no such limitation. Horizontally, the story is different: we can use the loadupsdb opcode to duplicate each byte in the U and V arrays, so we can iterate over 640 array elements. It is also possible to iterate over 320 elements and duplicate the U and V elements using mergebw. There is a very slight speed advantage to iterating vertically in Orc, and for demonstration purposes, we will choose to use the loadupsdb opcode, thus we will be iterating over 320x240 elements.
The code:
.function convert_I420_AYUV .flags 2d .dest 4 d1 .dest 4 d2 .source 1 y1 .source 1 y2 .source 1 u .source 1 v .const 1 c255 255 .temp 2 uv .temp 2 ay .temp 1 tu .temp 1 tv loadupdb tu, u loadupdb tv, v mergebw uv, tu, tv mergebw ay, c255, y1 mergewl d1, ay, uv mergebw ay, c255, y2 mergewl d2, ay, uv
A few things of note: The ".flags 2d" line is used to indicate that Orc should iterate over two dimensions, and generate a prototype that includes row strides for each array and a size parameter for the second dimension.
Since we are working on two input Y lines and two output AYUV lines at a time, we need two source and destination arrays corresponding to the even and odd lines. The row strides for these are doubled compared to the normal 2-D array.
The mergebw and mergewl opcodes join two 8-bit values into one 16-bit value (or 16-bit values into a 32-bit value) by concatinating them in memory order. Thus, to get AYUV in memory order, we merge AY and UV, and to get UV, we merge U and V. Since we're duplicating each U and V line, we use the same UV value for the even and odd output lines.
The prototype that is generated is:
void convert_I420_AYUV (orc_uint32 * d1, int d1_stride, orc_uint32 * d2, int d2_stride, const orc_uint8 * s1, int s1_stride, const orc_uint8 * s2, int s2_stride, const orc_uint8 * s3, int s3_stride, const orc_uint8 * s4, int s4_stride, int n, int m);
The orcc tool unhelpfully changed the names of the parameters, however, the order is standard: first destinations, then sources, then parameters, then array sizes. Think of it like memcpy() or memset().
Calling the function:
convert_I420_AYUV (output, 1280*4, output + 640, 1280 * 4,
input_y, 1280, input_y + 640, 1280,
input_u, 320, input_v, 320,
320, 240);