Change Details

Test with D47639 for `AFMT_FLOAT`, aka `AFMT_F32_NE`. On a little-endian machine, to test `AFMT_F32_BE` explicitly, you can apply the following patch to beep(1) instead: ``` diff --git a/usr.bin/beep/beep.c b/usr.bin/beep/beep.c index 151236b4825b..96e13cc86e9d 100644 --- a/usr.bin/beep/beep.c +++ b/usr.bin/beep/beep.c @@ -152,7 +152,7 @@ usage(void) int main(int argc, char **argv) { - int32_t *buffer; + float *buffer; size_t slope; size_t size; size_t off; @@ -208,9 +208,9 @@ main(int argc, char **argv) if (ioctl(f, SOUND_PCM_WRITE_CHANNELS, &c) != 0) errx(1, "ioctl SOUND_PCM_WRITE_CHANNELS(1) failed"); - c = AFMT_S32_NE; + c = AFMT_F32_BE; if (ioctl(f, SNDCTL_DSP_SETFMT, &c) != 0) - errx(1, "ioctl SNDCTL_DSP_SETFMT(AFMT_S32_NE) failed"); + errx(1, "ioctl SNDCTL_DSP_SETFMT(AFMT_F32_BE) failed"); if (ioctl(f, SNDCTL_DSP_SPEED, &sample_rate) != 0) errx(1, "ioctl SNDCTL_DSP_SPEED(%d) failed", sample_rate); @@ -251,7 +251,15 @@ main(int argc, char **argv) else if (off > (size - slope)) sample = sample * (size - off - 1) / (float)slope; - buffer[off] = sample * 0x7fffff00; + int32_t tmp; + memcpy(&tmp, &sample, sizeof(tmp)); + tmp = ((uint32_t)tmp & 0x000000ff) << 24 | + (tmp & 0x0000ff00) << 8 | + (tmp & 0x00ff0000) >> 8 | + (tmp & 0xff000000) >> 24; + memcpy(&sample, &tmp, sizeof(sample)); + + buffer[off] = sample; } if (write(f, buffer, size * sizeof(buffer[0])) != ```

Apply D47639 to test sound. --- Below is a benchmark for `AFMT_FLOAT` operations in `pcm_sample_read()` and `pcm_sample_write()`. The results are in cycles, and are measured by adding 2 `rdtscp()` calls in `pcm_sample_read()` and `pcm_sample_write()` respectively - one at the start and one at the end, and calculating the difference. The test program is a simple loopback program that reads and writes zeroes 512 times: ``` #include <sys/soundcard.h> #include <err.h> #include <fcntl.h> #include <unistd.h> int main(int argc, char *argv[]) { uint32_t sample; int fd, fmt, chans, rate, n; if ((fd = open(argv[1], O_RDWR)) < 0) err(1, "open(%s)", argv[1]); chans = 1; if (ioctl(fd, SNDCTL_DSP_CHANNELS, &chans) < 0) err(1, "ioctl(SNDCTL_DSP_CHANNELS)"); fmt = AFMT_S16_LE; if (ioctl(fd, SNDCTL_DSP_SETFMT, &fmt) < 0) err(1, "ioctl(SNDCTL_DSP_SETFMT)"); rate = 48000; if (ioctl(fd, SNDCTL_DSP_SPEED, &rate) < 0) err(1, "ioctl(SNDCTL_DSP_SPEED)"); for (n = 512; n--;) { sample = 0; if (read(fd, &sample, sizeof(sample)) < 0) err(1, "read"); if (write(fd, &sample, sizeof(sample)) < 0) err(1, "write"); } close(fd); } ``` The sound card used is a Focusrite Scarlett Solo (`snd_uaudio(4)`), with VCHANs disabled to simplify the test. | function | cycles | | `pcm_sample_read(AFMT_S32_LE)` | 330.806 | | `pcm_sample_write(AFMT_S32_LE)` | 183.298 | | `pcm_sample_read(AFMT_F32_LE)` | 3246.6 | | `pcm_sample_write(AFMT_F32_LE)` | 1893.8 | The `AFMT_FLOAT` operations seem to be ~10x more expensive.

Test with D47639 for `AFMT_FLOAT`,Apply D47639 to test sound. --- Below is a benchmark for `AFMT_FLOAT` operations in `pcm_sample_read()` and `pcm_sample_write()`. aka `AFMT_F32_NE`.The results are in cycles, On a little-endian machineand are measured by adding 2 `rdtscp()` calls in `pcm_sample_read()` and `pcm_sample_write()` respectively - one at the start and one at the end, to test `AFMT_F32_BE` explicitly, you can apply the following patch to beep(1) instead:and calculating the difference. The test program is a simple loopback program that reads and writes zeroes 512 times: ``` diff --git a/usr.bin/beep/beep.c b/usr.bin/beep/beep.c#include <sys/soundcard.h> #include <err.h> index 151236b4825b..96e13cc86e9d 100644#include <fcntl.h> --- a/usr.bin/beep/beep.c +++ b/usr.bin/beep/beep.c @@ -152,7 +152,7 @@ usage(void) #include <unistd.h> int main(int argc, char **argv[]) { - uint32_t *buffer;sample; + float *buffer;int fd, fmt, chans, rate, n; if ((fd = open(argv[1], O_RDWR)) < 0) err(1, "open(%s)", argv[1]); size_t slope;chans = 1; size_t size;if (ioctl(fd, SNDCTL_DSP_CHANNELS, &chans) < 0) err(1, "ioctl(SNDCTL_DSP_CHANNELS)"); size_t off;fmt = AFMT_S16_LE; @@ -208,9 +208,9 @@ main(int argc, char **argv) if (ioctl(fd, SOUND_PCM_WRITE_CHANNELSSNDCTL_DSP_SETFMT, &c) != 0&fmt) < 0) errx(1, "ioctl SOUND_PCM_WRITE_CHANNELS(1) failed");(SNDCTL_DSP_SETFMT)"); - c = AFMT_S32_NE; rate = 48000; + c = AFMT_F32_BE; if (ioctl(fd, SNDCTL_DSP_SETFMTPEED, &c) != 0&rate) < 0) - errx(1, "ioctl SNDCTL_DSP_SETFMT(AFMT_S32_NE) failed"); + errx(1(1, "ioctl (SNDCTL_DSP_SETFMT(AFMT_F32_BE) failed");PEED)"); if (ioctl(f, SNDCTL_DSP_SPEED,for (n = 512; &sample_rate) != 0)n--;) { errx(1, "ioctl SNDCTL_DSP_SPEED(%d) failed", sample_rate); = 0; @@ -251,7 +251,15 @@ main(int argc, char **argv) else if (off > (if (read(fd, &sample, size - slope)of(sample)) < 0) sample = sample * (size - off - 1) / (float)slope; - buffer[off] = sample * 0x7fffff00;err(1, "read"); + int32_t tmp;f (write(fd, &sample, sizeof(sample)) < 0) + memcpy(&tmp, &sample err(1, sizeof(tmp));"write"); + } tmp = ((uint32_t)tmp & 0x000000ff) << 24 | close(fd); + (tmp & 0x0000ff00) << 8 |} + (tmp & 0x00ff0000) >> 8``` The sound card used is a Focusrite Scarlett Solo (`snd_uaudio(4)`), with VCHANs disabled to simplify the test. | function | cycles | + (tmp & 0xff000000) >> 24;| `pcm_sample_read(AFMT_S32_LE)` | 330.806 | + memcpy(&sample, &tmp, sizeof(sample));| `pcm_sample_write(AFMT_S32_LE)` | 183.298 | +| `pcm_sample_read(AFMT_F32_LE)` | 3246.6 | + buffer[off] = sample; }| `pcm_sample_write(AFMT_F32_LE)` | 1893.8 | if (write(f, buffer, size * sizeof(buffer[0])) != ```The `AFMT_FLOAT` operations seem to be ~10x more expensive.