On Sat, 2013-01-05 at 20:25 +0100, Peter Meerwald wrote:
> From: Peter Meerwald <p.meerwald at bct-electronic.com>
>
> use (1<<15) instead of 0x7fff as a factor when converting from s16 to float32
> use (1<<31) instead of 0x7fffffff as a factor when converting from s32 to float32
>
> the change is motivated by the following desireable properties:
> * s16_from_f32(f32_from_s16(x)) == x for all possible s16 values
> * x / (1.0f << 15) == x * (1.0f / (1 << 15)) for all x in s16
>
> above changes enable easier optimization while guaranteeing bit-exact results
>
> further, other audio sample conversion code (libavresample) does it the same way
>
> Signed-off-by: Peter Meerwald <p.meerwald at bct-electronic.com>
> ---
> src/pulsecore/sconv-s16le.c | 84 +++++++++++++++++++------------------------
> src/pulsecore/sconv_neon.c | 17 ++++-----
> src/pulsecore/sconv_sse.c | 66 +++++++++++++++++-----------------
> 3 files changed, 75 insertions(+), 92 deletions(-)
>
> diff --git a/src/pulsecore/sconv-s16le.c b/src/pulsecore/sconv-s16le.c
> index 138e418..e0d7975 100644
> --- a/src/pulsecore/sconv-s16le.c
> +++ b/src/pulsecore/sconv-s16le.c
> @@ -85,11 +85,11 @@ void pa_sconv_s16le_to_float32ne(unsigned n, const int16_t *a, float *b) {
> #if SWAP_WORDS == 1
> for (; n > 0; n--) {
> int16_t s = *(a++);
> - *(b++) = ((float) INT16_FROM(s))/(float) 0x7FFF;
> + *(b++) = INT16_FROM(s) * (1.0f / (1 << 15));
> }
> #else
> for (; n > 0; n--)
> - *(b++) = ((float) (*(a++)))/(float) 0x7FFF;
> + *(b++) = *(a++) * (1.0f / (1 << 15));
> #endif
> }
>
> @@ -100,11 +100,11 @@ void pa_sconv_s32le_to_float32ne(unsigned n, const int32_t *a, float *b) {
> #if SWAP_WORDS == 1
> for (; n > 0; n--) {
> int32_t s = *(a++);
> - *(b++) = (float) (((double) INT32_FROM(s))/0x7FFFFFFF);
> + *(b++) = INT32_FROM(s) * (1.0f / (1U << 31));
> }
> #else
> for (; n > 0; n--)
> - *(b++) = (float) (((double) (*(a++)))/0x7FFFFFFF);
> + *(b++) = *(a++) * (1.0f / (1U << 31));
> #endif
> }
>
> @@ -115,18 +115,16 @@ void pa_sconv_s16le_from_float32ne(unsigned n, const float *a, int16_t *b) {
> #if SWAP_WORDS == 1
> for (; n > 0; n--) {
> int16_t s;
> - float v = *(a++);
> + float v = *(a++) * (1 << 15);
>
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.f);
> - s = (int16_t) lrintf(v * 0x7FFF);
> + s = (int16_t) PA_CLAMP_UNLIKELY(lrintf(v), -0x8000, 0x7FFF);
If you call lrintf() before clamping the input, you get undefined
results in case the input is very large. I guess this would work:
s = (int16_t) lrintf(PA_CLAMP_UNLIKELY(v, (float) -0x8000, (float) 0x7FFF));
The same comment applies to all "from float" conversions.
> *(b++) = INT16_TO(s);
> }
> #else
> for (; n > 0; n--) {
> - float v = *(a++);
> + float v = *(a++) * (1 << 15);
>
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.f);
> - *(b++) = (int16_t) lrintf(v * 0x7FFF);
> + *(b++) = (int16_t) PA_CLAMP_UNLIKELY(lrintf(v), -0x8000, 0x7FFF);
> }
> #endif
> }
> @@ -138,18 +136,16 @@ void pa_sconv_s32le_from_float32ne(unsigned n, const float *a, int32_t *b) {
> #if SWAP_WORDS == 1
> for (; n > 0; n--) {
> int32_t s;
> - float v = *(a++);
> + float v = *(a++) * (1U << 31);
>
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.0f);
> - s = (int32_t) lrint((double) v * (double) 0x7FFFFFFF);
> + s = (int32_t) PA_CLAMP_UNLIKELY(lrintf(v), -0x80000000LL, 0x7FFFFFFFLL);
You probably meant to use llrintf() rather than lrintf(), but if you do
the clamping before the rounding, I guess lrintf() will be enough here
and in the other "from float" functions.
> *(b++) = INT32_TO(s);
> }
> #else
> for (; n > 0; n--) {
> - float v = *(a++);
> + float v = *(a++) * (1U << 31);
>
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.0f);
> - *(b++) = (int32_t) lrint((double) v * (double) 0x7FFFFFFF);
> + *(b++) = (int32_t) PA_CLAMP_UNLIKELY(lrintf(v), -0x80000000LL, 0x7FFFFFFFLL);
> }
> #endif
> }
> @@ -160,7 +156,7 @@ void pa_sconv_s16le_to_float32re(unsigned n, const int16_t *a, float *b) {
>
> for (; n > 0; n--) {
> int16_t s = *(a++);
> - float k = ((float) INT16_FROM(s))/0x7FFF;
> + float k = INT16_FROM(s) * (1.0f / (1 << 15));
> k = PA_FLOAT32_SWAP(k);
> *(b++) = k;
> }
> @@ -172,7 +168,7 @@ void pa_sconv_s32le_to_float32re(unsigned n, const int32_t *a, float *b) {
>
> for (; n > 0; n--) {
> int32_t s = *(a++);
> - float k = (float) (((double) INT32_FROM(s))/0x7FFFFFFF);
> + float k = INT32_FROM(s) * (1.0f / (1U << 31));
> k = PA_FLOAT32_SWAP(k);
> *(b++) = k;
> }
> @@ -185,9 +181,8 @@ void pa_sconv_s16le_from_float32re(unsigned n, const float *a, int16_t *b) {
> for (; n > 0; n--) {
> int16_t s;
> float v = *(a++);
> - v = PA_FLOAT32_SWAP(v);
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.0f);
> - s = (int16_t) lrintf(v * 0x7FFF);
> + v = PA_FLOAT32_SWAP(v) * (1 << 15);
> + s = (int16_t) PA_CLAMP_UNLIKELY(lrintf(v), -0x8000, 0x7FFF);
> *(b++) = INT16_TO(s);
> }
> }
> @@ -199,9 +194,8 @@ void pa_sconv_s32le_from_float32re(unsigned n, const float *a, int32_t *b) {
> for (; n > 0; n--) {
> int32_t s;
> float v = *(a++);
> - v = PA_FLOAT32_SWAP(v);
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.0f);
> - s = (int32_t) lrint((double) v * 0x7FFFFFFF);
> + v = PA_FLOAT32_SWAP(v) * (1U << 31);
> + s = (int32_t) PA_CLAMP_UNLIKELY(llrintf(v), -0x80000000LL, 0x7FFFFFFFLL);
> *(b++) = INT32_TO(s);
> }
> }
> @@ -304,9 +298,9 @@ void pa_sconv_s24le_to_float32ne(unsigned n, const uint8_t *a, float *b) {
>
> for (; n > 0; n--) {
> int32_t s = READ24(a) << 8;
> - *b = ((float) s) / 0x7FFFFFFF;
> + *b = s * (1.0f / (1U << 31));
> a += 3;
> - b ++;
> + b++;
> }
> }
>
> @@ -316,12 +310,11 @@ void pa_sconv_s24le_from_float32ne(unsigned n, const float *a, uint8_t *b) {
>
> for (; n > 0; n--) {
> int32_t s;
> - float v = *a;
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.0f);
> - s = (int32_t) lrint((double) v * (double) 0x7FFFFFFF);
> + float v = *a * (1U << 31);
> + s = (int32_t) PA_CLAMP_UNLIKELY(llrint(v), -0x80000000LL, 0x7FFFFFFFLL);
> WRITE24(b, ((uint32_t) s) >> 8);
> a++;
> - b+=3;
> + b += 3;
> }
> }
>
> @@ -331,10 +324,10 @@ void pa_sconv_s24le_to_float32re(unsigned n, const uint8_t *a, float *b) {
>
> for (; n > 0; n--) {
> int32_t s = READ24(a) << 8;
> - float k = ((float) s) / 0x7FFFFFFF;
> + float k = s * (1.0f / (1U << 31));
> *b = PA_FLOAT32_SWAP(k);
> a += 3;
> - b ++;
> + b++;
> }
> }
>
> @@ -345,9 +338,8 @@ void pa_sconv_s24le_from_float32re(unsigned n, const float *a, uint8_t *b) {
> for (; n > 0; n--) {
> int32_t s;
> float v = *a;
> - v = PA_FLOAT32_SWAP(v);
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.0f);
> - s = (int32_t) lrint((double) v * (double) 0x7FFFFFFF);
> + v = PA_FLOAT32_SWAP(v) * (1U << 31);
> + s = (int32_t) PA_CLAMP_UNLIKELY(llrint(v), -0x80000000LL, 0x7FFFFFFFLL);
> WRITE24(b, ((uint32_t) s) >> 8);
> a++;
> b+=3;
> @@ -406,9 +398,9 @@ void pa_sconv_s24_32le_to_float32ne(unsigned n, const uint32_t *a, float *b) {
>
> for (; n > 0; n--) {
> int32_t s = (int32_t) (UINT32_FROM(*a) << 8);
> - *b = (float) s / (float) 0x7FFFFFFF;
> - a ++;
> - b ++;
> + *b = s * (1.0f / (1U << 31));
> + a++;
> + b++;
> }
> }
>
> @@ -418,10 +410,10 @@ void pa_sconv_s24_32le_to_float32re(unsigned n, const uint32_t *a, float *b) {
>
> for (; n > 0; n--) {
> int32_t s = (int32_t) (UINT32_FROM(*a) << 8);
> - float k = (float) s / (float) 0x7FFFFFFF;
> + float k = s * (1.0f / (1U << 31));
> *b = PA_FLOAT32_SWAP(k);
> - a ++;
> - b ++;
> + a++;
> + b++;
> }
> }
>
> @@ -431,9 +423,8 @@ void pa_sconv_s24_32le_from_float32ne(unsigned n, const float *a, uint32_t *b) {
>
> for (; n > 0; n--) {
> int32_t s;
> - float v = *a;
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.0f);
> - s = (int32_t) lrint((double) v * (double) 0x7FFFFFFF);
> + float v = *a * (1U << 31);
> + s = (int32_t) PA_CLAMP_UNLIKELY(llrint(v), -0x80000000LL, 0x7FFFFFFFLL);
> *b = UINT32_TO(((uint32_t) s) >> 8);
> a++;
> b++;
> @@ -447,9 +438,8 @@ void pa_sconv_s24_32le_from_float32re(unsigned n, const float *a, uint32_t *b) {
> for (; n > 0; n--) {
> int32_t s;
> float v = *a;
> - v = PA_FLOAT32_SWAP(v);
> - v = PA_CLAMP_UNLIKELY(v, -1.0f, 1.0f);
> - s = (int32_t) lrint((double) v * (double) 0x7FFFFFFF);
> + v = PA_FLOAT32_SWAP(v) * (1U << 31);
> + s = (int32_t) PA_CLAMP_UNLIKELY(llrint(v), -0x80000000LL, 0x7FFFFFFFLL);
> *b = UINT32_TO(((uint32_t) s) >> 8);
> a++;
> b++;
> diff --git a/src/pulsecore/sconv_neon.c b/src/pulsecore/sconv_neon.c
> index 6fd966d..111b56f 100644
> --- a/src/pulsecore/sconv_neon.c
> +++ b/src/pulsecore/sconv_neon.c
> @@ -36,16 +36,11 @@ static void pa_sconv_s16le_from_f32ne_neon(unsigned n, const float *src, int16_t
> "movs %[n], %[n], lsr #2 \n\t"
> "beq 2f \n\t"
>
> - "vdup.f32 q2, %[plusone] \n\t"
> - "vneg.f32 q3, q2 \n\t"
> - "vdup.f32 q4, %[scale] \n\t"
> - "vdup.u32 q5, %[mask] \n\t"
> + "vdup.f32 q1, %[scale] \n\t"
>
> "1: \n\t"
> "vld1.32 {q0}, [%[src]]! \n\t"
> - "vmin.f32 q0, q0, q2 \n\t" /* clamp */
> - "vmax.f32 q0, q0, q3 \n\t"
> - "vmul.f32 q0, q0, q4 \n\t" /* scale */
> + "vmul.f32 q0, q0, q1 \n\t" /* scale */
> "vcvt.s32.f32 q0, q0, #16 \n\t" /* narrow */
You removed clamping - what happens if there's need for clamping? (I'm
not very good at reading assembly.)
> "vrshrn.s32 d0, q0, #16 \n\t"
> "subs %[n], %[n], #1 \n\t"
> @@ -55,13 +50,13 @@ static void pa_sconv_s16le_from_f32ne_neon(unsigned n, const float *src, int16_t
> "2: \n\t"
>
> : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n) /* output operands (or input operands that get modified) */
> - : [plusone] "r" (1.0f), [scale] "r" (32767.0f), [mask] "r" (0x80000000) /* input operands */
> - : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "q6" /* clobber list */
> + : [scale] "r" (32768.0f) /* input operands */
> + : "memory", "cc", "q0", "q1" /* clobber list */
> );
>
> /* leftovers */
> while (i--) {
> - *dst++ = (int16_t) lrintf(PA_CLAMP_UNLIKELY(*src, -1.0f, 1.0f) * 0x7FFF);
> + *dst++ = (int16_t) PA_CLAMP_UNLIKELY(lrintf(*src * (1 << 15)), -0x8000, 0x7FFF);
> src++;
> }
> }
> @@ -69,7 +64,7 @@ static void pa_sconv_s16le_from_f32ne_neon(unsigned n, const float *src, int16_t
> static void pa_sconv_s16le_to_f32ne_neon(unsigned n, const int16_t *src, float *dst) {
> unsigned i = n & 3;
>
> - const float invscale = 1.0f / 0x7FFF;
> + const float invscale = 1.0f / (1 << 15);
>
> __asm__ __volatile__ (
> "movs %[n], %[n], lsr #2 \n\t"
> diff --git a/src/pulsecore/sconv_sse.c b/src/pulsecore/sconv_sse.c
> index 3b56a9c..072d329 100644
> --- a/src/pulsecore/sconv_sse.c
> +++ b/src/pulsecore/sconv_sse.c
> @@ -35,17 +35,13 @@
>
> #if !defined(__APPLE__) && defined (__i386__) || defined (__amd64__)
>
> -static const PA_DECLARE_ALIGNED (16, float, one[4]) = { 1.0, 1.0, 1.0, 1.0 };
> -static const PA_DECLARE_ALIGNED (16, float, mone[4]) = { -1.0, -1.0, -1.0, -1.0 };
> -static const PA_DECLARE_ALIGNED (16, float, scale[4]) = { 0x7fff, 0x7fff, 0x7fff, 0x7fff };
> +static const PA_DECLARE_ALIGNED (16, float, scale[4]) = { 0x8000, 0x8000, 0x8000, 0x8000 };
>
> static void pa_sconv_s16le_from_f32ne_sse(unsigned n, const float *a, int16_t *b) {
> pa_reg_x86 temp, i;
>
> __asm__ __volatile__ (
> " movaps %5, %%xmm5 \n\t"
> - " movaps %6, %%xmm6 \n\t"
> - " movaps %7, %%xmm7 \n\t"
> " xor %0, %0 \n\t"
>
> " mov %4, %1 \n\t"
> @@ -56,12 +52,8 @@ static void pa_sconv_s16le_from_f32ne_sse(unsigned n, const float *a, int16_t *b
> "1: \n\t"
> " movups (%q2, %0, 2), %%xmm0 \n\t" /* read 8 floats */
> " movups 16(%q2, %0, 2), %%xmm2 \n\t"
> - " minps %%xmm5, %%xmm0 \n\t" /* clamp to 1.0 */
> - " minps %%xmm5, %%xmm2 \n\t"
> - " maxps %%xmm6, %%xmm0 \n\t" /* clamp to -1.0 */
> - " maxps %%xmm6, %%xmm2 \n\t"
> - " mulps %%xmm7, %%xmm0 \n\t" /* *= 0x7fff */
> - " mulps %%xmm7, %%xmm2 \n\t"
> + " mulps %%xmm5, %%xmm0 \n\t" /* *= 0x8000 */
> + " mulps %%xmm5, %%xmm2 \n\t"
>
> " cvtps2pi %%xmm0, %%mm0 \n\t" /* low part to int */
> " cvtps2pi %%xmm2, %%mm2 \n\t"
> @@ -73,7 +65,7 @@ static void pa_sconv_s16le_from_f32ne_sse(unsigned n, const float *a, int16_t *b
> " packssdw %%mm1, %%mm0 \n\t" /* pack parts */
> " packssdw %%mm3, %%mm2 \n\t"
> " movq %%mm0, (%q3, %0) \n\t"
> - " movq %%mm2, 8(%q3, %0) \n\t"
> + " movq %%mm2, 8(%q3, %0) \n\t"
>
> " add $16, %0 \n\t"
> " dec %1 \n\t"
> @@ -82,24 +74,30 @@ static void pa_sconv_s16le_from_f32ne_sse(unsigned n, const float *a, int16_t *b
> "2: \n\t"
> " mov %4, %1 \n\t" /* prepare for leftovers */
> " and $7, %1 \n\t"
> - " je 4f \n\t"
> + " je 5f \n\t"
>
> "3: \n\t"
> " movss (%q2, %0, 2), %%xmm0 \n\t"
> - " minss %%xmm5, %%xmm0 \n\t"
> - " maxss %%xmm6, %%xmm0 \n\t"
> - " mulss %%xmm7, %%xmm0 \n\t"
> + " mulss %%xmm5, %%xmm0 \n\t"
> + " cvtss2si %%xmm0, %4 \n\t"
> + " add $0x8000, %4 \n\t" /* check for saturation */
> + " and $~0xffff, %4 \n\t"
> " cvtss2si %%xmm0, %4 \n\t"
> - " movw %w4, (%q3, %0) \n\t"
> + " je 4f \n\t"
> + " sar $31, %4 \n\t"
> + " xor $0x7fff, %4 \n\t"
> +
> + "4: \n\t"
> + " movw %w4, (%q3, %0) \n\t" /* store leftover */
> " add $2, %0 \n\t"
> " dec %1 \n\t"
> " jne 3b \n\t"
>
> - "4: \n\t"
> + "5: \n\t"
> " emms \n\t"
>
> : "=&r" (i), "=&r" (temp)
> - : "r" (a), "r" (b), "r" ((pa_reg_x86)n), "m" (*one), "m" (*mone), "m" (*scale)
> + : "r" (a), "r" (b), "r" ((pa_reg_x86)n), "m" (*scale)
> : "cc", "memory"
> );
> }
> @@ -109,8 +107,6 @@ static void pa_sconv_s16le_from_f32ne_sse2(unsigned n, const float *a, int16_t *
>
> __asm__ __volatile__ (
> " movaps %5, %%xmm5 \n\t"
> - " movaps %6, %%xmm6 \n\t"
> - " movaps %7, %%xmm7 \n\t"
> " xor %0, %0 \n\t"
>
> " mov %4, %1 \n\t"
> @@ -121,12 +117,8 @@ static void pa_sconv_s16le_from_f32ne_sse2(unsigned n, const float *a, int16_t *
> "1: \n\t"
> " movups (%q2, %0, 2), %%xmm0 \n\t" /* read 8 floats */
> " movups 16(%q2, %0, 2), %%xmm2 \n\t"
> - " minps %%xmm5, %%xmm0 \n\t" /* clamp to 1.0 */
> - " minps %%xmm5, %%xmm2 \n\t"
> - " maxps %%xmm6, %%xmm0 \n\t" /* clamp to -1.0 */
> - " maxps %%xmm6, %%xmm2 \n\t"
> - " mulps %%xmm7, %%xmm0 \n\t" /* *= 0x7fff */
> - " mulps %%xmm7, %%xmm2 \n\t"
> + " mulps %%xmm5, %%xmm0 \n\t" /* *= 0x8000 */
> + " mulps %%xmm5, %%xmm2 \n\t"
>
> " cvtps2dq %%xmm0, %%xmm0 \n\t"
> " cvtps2dq %%xmm2, %%xmm2 \n\t"
> @@ -141,23 +133,29 @@ static void pa_sconv_s16le_from_f32ne_sse2(unsigned n, const float *a, int16_t *
> "2: \n\t"
> " mov %4, %1 \n\t" /* prepare for leftovers */
> " and $7, %1 \n\t"
> - " je 4f \n\t"
> + " je 5f \n\t"
>
> "3: \n\t"
> " movss (%q2, %0, 2), %%xmm0 \n\t"
> - " minss %%xmm5, %%xmm0 \n\t"
> - " maxss %%xmm6, %%xmm0 \n\t"
> - " mulss %%xmm7, %%xmm0 \n\t"
> + " mulss %%xmm5, %%xmm0 \n\t"
> + " cvtss2si %%xmm0, %4 \n\t"
> + " add $0x8000, %4 \n\t"
> + " and $~0xffff, %4 \n\t" /* check for saturation */
> " cvtss2si %%xmm0, %4 \n\t"
> - " movw %w4, (%q3, %0) \n\t"
> + " je 4f \n\t"
> + " sar $31, %4 \n\t"
> + " xor $0x7fff, %4 \n\t"
> +
> + "4: \n\t"
> + " movw %w4, (%q3, %0) \n\t" /* store leftover */
> " add $2, %0 \n\t"
> " dec %1 \n\t"
> " jne 3b \n\t"
>
> - "4: \n\t"
> + "5: \n\t"
>
> : "=&r" (i), "=&r" (temp)
> - : "r" (a), "r" (b), "r" ((pa_reg_x86)n), "m" (*one), "m" (*mone), "m" (*scale)
> + : "r" (a), "r" (b), "r" ((pa_reg_x86)n), "m" (*scale)
> : "cc", "memory"
> );
> }
The x86 code remains unreviewed. I was able to guess to some extent
what's happening in the arm code, but can't do that for the x86 code.
Maybe Arun will check this, or maybe I should learn this stuff properly.
--
Tanu
