Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Optimize float32 IQ code path using ARM NEON #89

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

Optimize float32 IQ code path using ARM NEON #89

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
8916439
Optimize float32 IQ code path using ARM NEON
sergeyvfx Oct 15, 2022
2dc5463
Merge branch 'master' into simd_neon
sergeyvfx Jun 18, 2023
9467f39
Fix vaddvq_f32() used on 32bit ARM platform
sergeyvfx Jun 18, 2023
204228a
Use multiple accumulators to further improve performance
sergeyvfx Jun 27, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 25 additions & 0 deletions libairspy/src/airspy.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,6 +39,11 @@ ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSI
#include "iqconverter_int16.h"
#include "filters.h"

#if defined(__ARM_NEON) && __ARM_NEON
#include <arm_neon.h>
#define USE_NEON
#endif

#ifndef bool
typedef int bool;
#define true 1
Expand Down Expand Up @@ -306,13 +311,33 @@ static void convert_samples_int16(uint16_t *src, int16_t *dest, int count)

static void convert_samples_float(uint16_t *src, float *dest, int count)
{
#if defined(USE_NEON)
const float32x4_t offset_f32 = vmovq_n_f32(2048);
const float32x4_t sample_scale_f32 = vmovq_n_f32(SAMPLE_SCALE);
#endif

int i;
for (i = 0; i < count; i += 4)
{

#if defined(USE_NEON)

const uint16x4_t src_u16 = vld1_u16(src + i);
const uint32x4_t src_u32 = vmovl_u16(src_u16);
const float32x4_t src_f32 = vcvtq_f32_u32(src_u32);
const float32x4_t src_offset_f32 = vsubq_f32(src_f32, offset_f32);
const float32x4_t dest_f32 = vmulq_f32(src_offset_f32, sample_scale_f32);

vst1q_f32(dest + i, dest_f32);

#else

dest[i + 0] = (src[i + 0] - 2048) * SAMPLE_SCALE;
dest[i + 1] = (src[i + 1] - 2048) * SAMPLE_SCALE;
dest[i + 2] = (src[i + 2] - 2048) * SAMPLE_SCALE;
dest[i + 3] = (src[i + 3] - 2048) * SAMPLE_SCALE;

#endif
}
}

Expand Down
113 changes: 107 additions & 6 deletions libairspy/src/iqconverter_float.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,11 @@ void *_aligned_malloc(size_t size, size_t alignment)
#endif
#endif

#if defined(__ARM_NEON) && __ARM_NEON
#include <arm_neon.h>
#define USE_NEON
#endif

#define SIZE_FACTOR 32
#define DEFAULT_ALIGNMENT 16
#define HPF_COEFF 0.01f
Expand All @@ -72,6 +77,18 @@ void *_aligned_malloc(size_t size, size_t alignment)
#define ALIGNED
#endif

#if defined(USE_NEON)
static _inline float horizontal_sum_neon_f32(const float32x4_t v) {
# if defined(__aarch64__) || defined(_M_ARM64)
return vaddvq_f32(v);
# else
float32x2_t r = vadd_f32(vget_low_f32(v), vget_high_f32(v));
r = vpadd_f32(r, r);
return vget_lane_f32(r, 0);
# endif
}
#endif

iqconverter_float_t *iqconverter_float_create(const float *hb_kernel, int len)
{
int i, j;
Expand Down Expand Up @@ -122,6 +139,10 @@ static _inline float process_fir_taps(const float *kernel, const float *queue, i

__m128 acc = _mm_set_ps(0, 0, 0, 0);

#elif defined(USE_NEON)

float32x4_t acc = vmovq_n_f32(0);

#else

float sum = 0.0f;
Expand Down Expand Up @@ -150,8 +171,26 @@ static _inline float process_fir_taps(const float *kernel, const float *queue, i

queue += 8;
kernel += 8;

#elif defined(USE_NEON)
float32x4_t acc2 = vmovq_n_f32(0);

for (i = 0; i < it; i++)
{
const float32x4_t head1 = vld1q_f32(queue);
const float32x4_t kern1 = vld1q_f32(kernel);
const float32x4_t head2 = vld1q_f32(queue + 4);
const float32x4_t kern2 = vld1q_f32(kernel + 4);

acc = vmlaq_f32(acc, kern1, head1);
acc2 = vmlaq_f32(acc2, kern2, head2);

queue += 8;
kernel += 8;
}

acc = vaddq_f32(acc, acc2);

#else

for (i = 0; i < it; i++)
Expand Down Expand Up @@ -183,6 +222,12 @@ static _inline float process_fir_taps(const float *kernel, const float *queue, i
__m128 mul = _mm_mul_ps(kern, head);
acc = _mm_add_ps(acc, mul);

#elif defined(USE_NEON)

const float32x4_t head = vld1q_f32(queue);
const float32x4_t kern = vld1q_f32(kernel);
acc = vmlaq_f32(acc, kern, head);

#else

sum += kernel[0] * queue[0]
Expand All @@ -208,6 +253,8 @@ static _inline float process_fir_taps(const float *kernel, const float *queue, i
float sum = acc.m128_f32[0];
#endif

#elif defined(USE_NEON)
float sum = horizontal_sum_neon_f32(acc);
#endif

if (len >= 2)
Expand Down Expand Up @@ -327,6 +374,14 @@ static void fir_interleaved_12(iqconverter_float_t *cnv, float *samples, int len
cnv->fir_index = fir_index;
}

#if defined(USE_NEON)
static inline float32x4_t vld1q_f32_reversed(const float* values) {
float32x4_t v = vld1q_f32(values); /* 0 1 2 3 */
v = vrev64q_f32(v); /* 1 0 3 2 */
return vextq_f32(v, v, 2); /* 3 2 1 0 */
}
#endif

static void fir_interleaved_24(iqconverter_float_t *cnv, float *samples, int len)
{
int i;
Expand All @@ -335,14 +390,39 @@ static void fir_interleaved_24(iqconverter_float_t *cnv, float *samples, int len
float *fir_kernel = cnv->fir_kernel;
float *fir_queue = cnv->fir_queue;
float *queue;

#if defined(USE_NEON)
const float32x4_t kernel1 = vld1q_f32(fir_kernel + 0);
const float32x4_t kernel2 = vld1q_f32(fir_kernel + 4);
const float32x4_t kernel3 = vld1q_f32(fir_kernel + 8);

float32x4_t acc;
#else
float acc = 0;
#endif

for (i = 0; i < len; i += 2)
{
queue = fir_queue + fir_index;

queue[0] = samples[i];

#if defined(USE_NEON)

const float32x4_t queue1_1 = vld1q_f32(queue + 0);
const float32x4_t queue2_1 = vld1q_f32(queue + 4);
const float32x4_t queue3_1 = vld1q_f32(queue + 8);

const float32x4_t queue3_2 = vld1q_f32_reversed(queue + 12);
const float32x4_t queue2_2 = vld1q_f32_reversed(queue + 16);
const float32x4_t queue1_2 = vld1q_f32_reversed(queue + 20);

acc = vmulq_f32(kernel1, vaddq_f32(queue1_1, queue1_2));
acc = vmlaq_f32(acc, kernel2, vaddq_f32(queue2_1, queue2_2));
acc = vmlaq_f32(acc, kernel3, vaddq_f32(queue3_1, queue3_2));

samples[i] = horizontal_sum_neon_f32(acc);
#else
acc = fir_kernel[0] * (queue[0] + queue[24 - 1])
+ fir_kernel[1] * (queue[1] + queue[24 - 2])
+ fir_kernel[2] * (queue[2] + queue[24 - 3])
Expand All @@ -357,6 +437,7 @@ static void fir_interleaved_24(iqconverter_float_t *cnv, float *samples, int len
+ fir_kernel[11] * (queue[11] + queue[24 - 12]);

samples[i] = acc;
#endif

if (--fir_index < 0)
{
Expand Down Expand Up @@ -446,21 +527,23 @@ static void delay_interleaved(iqconverter_float_t *cnv, float *samples, int len)

static void remove_dc(iqconverter_float_t *cnv, float *samples, int len)
{
int i;
float *sample = samples;
const float *samples_end = sample + len;

ALIGNED float avg = cnv->avg;

for (i = 0; i < len; i++)
while (sample < samples_end)
{
samples[i] -= avg;
avg += SCALE * samples[i];
*sample -= avg;
avg += SCALE * (*sample);
++sample;
}

cnv->avg = avg;
}

static void translate_fs_4(iqconverter_float_t *cnv, float *samples, int len)
{
int i;
ALIGNED float hbc = cnv->hbc;

#ifdef USE_SSE2
Expand All @@ -476,9 +559,27 @@ static void translate_fs_4(iqconverter_float_t *cnv, float *samples, int len)
_mm_storeu_ps(buf, vec);
}

#elif defined(USE_NEON)

float *buf = samples;
const float *buf_end = buf + len;
float32x4_t vec;
float rot_data[4] = {-1.0f, -hbc, 1.0f, hbc};
const float32x4_t rot = vld1q_f32(rot_data);

while (buf < buf_end)
{
vec = vld1q_f32(buf);

vec = vmulq_f32(vec, rot);
vst1q_f32(buf, vec);

buf += 4;
}

#else

int j;
int i, j;

for (i = 0; i < len / 4; i++)
{
Expand Down