NEON HelloWorld - Revision history

JulienB: /* How to compile ? */

2018-06-01T16:42:32Z

‎How to compile ?

JulienPeeters: /* And in the add3 function ? */

2009-11-02T12:49:50Z

‎And in the add3 function ?

JulienPeeters: Add the pretty print result

2009-11-02T12:49:22Z

Add the pretty print result

JulienPeeters: /* Links */

2009-11-02T12:31:20Z

‎Links

JulienPeeters: Add categories

2009-11-02T12:29:03Z

Add categories

JulienPeeters: /* Let's go programming */

2009-11-02T12:23:39Z

‎Let's go programming

JulienPeeters: /* Let's go programming */

2009-11-02T12:21:40Z

‎Let's go programming

JulienPeeters: /* Links */

2009-11-02T12:20:54Z

‎Links

JulienPeeters: Add links

2009-11-02T12:20:33Z

Add links

JulienPeeters: Add a "Hello World" example with SIMD intrinsics

2009-11-02T12:18:15Z

Add a "Hello World" example with SIMD intrinsics

New page

Developing with SIMD instructions or intrinsics need a bit of experience. That is a complete different way of thinking and programming.

== Prerequisites ==

* Install a emulation environment
* Build a GCC toolchain which support NEON intrinsics

== Let's go programming ==

Here is a brief example of what is possible with SIMD programming. This piece of code only add the value "3" to each value of the SIMD vector.

On the Cortex-A platform there is both 64 bits and 128 bits vector registers. Here we use 128 bits ones then we can code sixteen 8 bits values (unsigned integers).

{{Warning | Not all ARM Cortex processors have a NEON unit and old ARM processors may have a SIMD unit not compliant with NEON (cf. ARM reference manuals).}}

<source lang="c">
#include <stdio.h>

#include "arm_neon.h"

void add3 (uint8x16_t *data) {
/* Set each sixteen values of the vector to 3.
*
* Remark: a 'q' suffix to intrinsics indicates
* the instruction run for 128 bits registers.
*/
uint8x16_t three = vmovq_n_u8 (3);

/* Add 3 to the value given in argument. */
*data = vaddq_u8 (*data, three);
}

void print_uint8 (uint8x16_t data, char* name) {
int i;
static uint8_t p[16];

vst1q_u8 (p, data);

printf ("%s = ", name);
for (i = 0; i < 16; i++) {
printf ("%02d ", p[i]);
}
printf ("\n");
}

int main () {
int i;

/* Create custom arbitrary data. */
const uint8_t uint8_data[] = { 1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16 };
/* Create the vector with our data. */
uint8x16_t data;

/* Load our custom data into the vector register. */
data = vld1q_u8 (uint8_data);

print_uint8 (data, "data");

/* Call of the add3 function. */
add3(&data);

print_uint8 (data, "data (new)");

return 0;
}
</source>

What is done ?
* We first create custom data: an array of 16 uint8_t values.
* We load the custom values in a vector register through the '''vld1q_u8''' intrinsic (note the 'q' suffix which means we use 128 registers)
* We call our '''add3''' function

And in the '''add3''' function ?
* Because we use SIMD vectors and want to add 3 to each value we need to set the same value (3) with the '''vmovq_n_u8''' intrinsic
* We call the '''vaddq_u8''' intrinsic to do the addition

{{Note | all sixteen values of each vector are compute in real parallelism}}

=== How to compile ? ===

The gcc options to give in order to compile NEON intrinsics in case of an ARM Cortex-A8 is :
<pre class="host">
$ gcc -mcpu=cortex-a8 -mfloat-abi=softfp -mfpu=neon <sources>
</pre>

{{Note | Do not forget to include the "arm_neon.h" header file.}}

← Older revision		Revision as of 12:31, 2 November 2009
Line 97:		Line 97:
	[[Category:Software]]		[[Category:Software]]
	[[Category:Architecture dependent]]		[[Category:Architecture dependent]]
−	~~[[Category:SIMD]]~~

← Older revision		Revision as of 12:29, 2 November 2009
Line 93:		Line 93:
	* [http://gcc.gnu.org/onlinedocs/gcc/ARM-NEON-Intrinsics.html GCC NEON intrinsics]		* [http://gcc.gnu.org/onlinedocs/gcc/ARM-NEON-Intrinsics.html GCC NEON intrinsics]
	* [http://code.google.com/p/math-neon/source/browse/#svn/trunk Example of a mathematic library using NEON]		* [http://code.google.com/p/math-neon/source/browse/#svn/trunk Example of a mathematic library using NEON]
		+
		+
		+	[[Category:Software]]
		+	[[Category:Architecture dependent]]
		+	[[Category:SIMD]]

@@ Line 88: / Line 88: @@
 {{Note | Do not forget to include the "arm_neon.h" header file.}}
-=== Links ===
+== Links ==
 * [http://gcc.gnu.org/onlinedocs/gcc/ARM-NEON-Intrinsics.html GCC NEON intrinsics]
 * [http://code.google.com/p/math-neon/source/browse/#svn/trunk Example of a mathematic library using NEON]

← Older revision		Revision as of 12:20, 2 November 2009
Line 87:		Line 87:

	{{Note \| Do not forget to include the "arm_neon.h" header file.}}		{{Note \| Do not forget to include the "arm_neon.h" header file.}}
		+
		+	=== Links ===
		+
		+	* [http://gcc.gnu.org/onlinedocs/gcc/ARM-NEON-Intrinsics.html GCC NEON intrinsics]
		+	* [http://code.google.com/p/math-neon/source/browse/#svn/trunk Example of a mathematic library using NEON]

@@ Line 90: / Line 90: @@
 === How to compile ? ===
-The gcc options to give in order to compile NEON intrinsics in case of an ARM Cortex-A8 is :
+The gcc options to give in order to compile NEON intrinsics in case of an ARM Cortex-A8/9 is :
 <pre class="host">
-$ gcc -mcpu=cortex-a8 -mfloat-abi=softfp -mfpu=neon <sources>
+$ ./buildroot/output/host/usr/bin/arm-linux-gcc -mfpu=neon -mcpu=cortex-a9 neon.c    (APF6)
 </pre>

@@ Line 76: / Line 76: @@
 * Because we use SIMD vectors and want to add 3 to each value we need to set the same value (3) with the '''vmovq_n_u8''' intrinsic
 * We call the '''vaddq_u8''' intrinsic to do the addition
 {{Note | all sixteen values of each vector are compute in real parallelism}}

@@ Line 13: / Line 13: @@
 {{Warning | Not all ARM Cortex processors have a NEON unit and old ARM processors may have a SIMD unit not compliant with NEON (cf. ARM reference manuals).}}
 <source lang="c">
@@ Line 66: / Line 68: @@
 </source>
-What is done ?
+=== What is done ? ===
 * We first create custom data: an array of 16 uint8_t values.
 * We load the custom values in a vector register through the '''vld1q_u8''' intrinsic (note the 'q' suffix which means we use 128 registers)
 * We call our '''add3''' function
-And in the '''add3''' function ?
+=== And in the '''add3''' function ? ===
 * Because we use SIMD vectors and want to add 3 to each value we need to set the same value (3) with the '''vmovq_n_u8''' intrinsic
 * We call the '''vaddq_u8''' intrinsic to do the addition

@@ Line 45: / Line 45: @@
 int main () {
      /* Create custom arbitrary data. */
      const uint8_t uint8_data[] = { 1, 2, 3, 4, 5, 6, 7, 8,
 , 10, 11, 12, 13, 14, 15, 16 };
      /* Create the vector with our data. */
      uint8x16_t data;