audioconvert: refactor peaks resampler

Use common code in macro and generate arch specific version.
Compile with -Ofast to optimize some fmaxf calls.

spa/plugins/audioconvert/meson.build

@@ -27,7 +27,7 @@ if have_sse
       'resample-peaks-sse.c',
       'volume-ops-sse.c',
       'channelmix-ops-sse.c' ],
-    c_args : [sse_args, '-O3', '-DHAVE_SSE'],
+    c_args : [sse_args, '-Ofast', '-DHAVE_SSE'],
     dependencies : [ spa_dep ],
     install : false
     )

spa/plugins/audioconvert/resample-native-avx.c

@@ -27,7 +27,7 @@
 #include <assert.h>
 #include <immintrin.h>
 
-static void inner_product_avx(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_avx(float *d, const float * SPA_RESTRICT s,
 		const float * SPA_RESTRICT taps, uint32_t n_taps)
 {
 	__m256 sy[2] = { _mm256_setzero_ps(), _mm256_setzero_ps() }, ty;
@@ -56,7 +56,7 @@ static void inner_product_avx(float *d, const float * SPA_RESTRICT s,
 	_mm_store_ss(d, sx[0]);
 }
 
-static void inner_product_ip_avx(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_ip_avx(float *d, const float * SPA_RESTRICT s,
 	const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x,
 	uint32_t n_taps)
 {

spa/plugins/audioconvert/resample-native-c.c

@@ -24,7 +24,7 @@
 
 #include "resample-native-impl.h"
 
-static void inner_product_c(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_c(float *d, const float * SPA_RESTRICT s,
 		const float * SPA_RESTRICT taps, uint32_t n_taps)
 {
 	float sum = 0.0f;
@@ -40,7 +40,7 @@ static void inner_product_c(float *d, const float * SPA_RESTRICT s,
 	*d = sum;
 }
 
-static void inner_product_ip_c(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_ip_c(float *d, const float * SPA_RESTRICT s,
 	const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x,
 	uint32_t n_taps)
 {

spa/plugins/audioconvert/resample-native-neon.c

@@ -26,7 +26,7 @@
 
 #include <arm_neon.h>
 
-static void inner_product_neon(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_neon(float *d, const float * SPA_RESTRICT s,
 		const float * SPA_RESTRICT taps, uint32_t n_taps)
 {
 	unsigned int remainder = n_taps % 16;
@@ -137,7 +137,7 @@ static void inner_product_neon(float *d, const float * SPA_RESTRICT s,
 #endif
 }
 
-static void inner_product_ip_neon(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_ip_neon(float *d, const float * SPA_RESTRICT s,
 	const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x,
 	uint32_t n_taps)
 {

spa/plugins/audioconvert/resample-native-sse.c

@@ -26,7 +26,7 @@
 
 #include <xmmintrin.h>
 
-static void inner_product_sse(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_sse(float *d, const float * SPA_RESTRICT s,
 		const float * SPA_RESTRICT taps, uint32_t n_taps)
 {
 	__m128 sum = _mm_setzero_ps();
@@ -68,7 +68,7 @@ static void inner_product_sse(float *d, const float * SPA_RESTRICT s,
 	_mm_store_ss(d, sum);
 }
 
-static void inner_product_ip_sse(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_ip_sse(float *d, const float * SPA_RESTRICT s,
 	const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x,
 	uint32_t n_taps)
 {

spa/plugins/audioconvert/resample-native-ssse3.c

@@ -26,7 +26,7 @@
 
 #include <tmmintrin.h>
 
-static void inner_product_ssse3(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_ssse3(float *d, const float * SPA_RESTRICT s,
 		const float * SPA_RESTRICT taps, uint32_t n_taps)
 {
 	__m128 sum = _mm_setzero_ps();
@@ -97,7 +97,7 @@ static void inner_product_ssse3(float *d, const float * SPA_RESTRICT s,
 	_mm_store_ss(d, sum);
 }
 
-static void inner_product_ip_ssse3(float *d, const float * SPA_RESTRICT s,
+static inline void inner_product_ip_ssse3(float *d, const float * SPA_RESTRICT s,
 	const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x,
 	uint32_t n_taps)
 {

spa/plugins/audioconvert/resample-peaks-c.c

@@ -26,48 +26,12 @@
 
 #include "resample-peaks-impl.h"
 
-void resample_peaks_process_c(struct resample *r,
+static inline float find_abs_max_c(const float *s, uint32_t n_samples, float m)
-			const void * SPA_RESTRICT src[], uint32_t *in_len,
-			void * SPA_RESTRICT dst[], uint32_t *out_len)
 {
-	struct peaks_data *pd = r->data;
+	uint32_t n;
-	uint32_t c, i, o, end, chunk, o_count, i_count;
+	for (n = 0; n < n_samples; n++)
-
+		m = fmaxf(fabsf(s[n]), m);
-	if (SPA_UNLIKELY(r->channels == 0))
+	return m;
-		return;
-
-	for (c = 0; c < r->channels; c++) {
-		const float *s = src[c];
-		float *d = dst[c], m = pd->max_f[c];
-
-		o_count = pd->o_count;
-		i_count = pd->i_count;
-		o = i = 0;
-
-		while (i < *in_len && o < *out_len) {
-			end = ((uint64_t) (o_count + 1) * r->i_rate) / r->o_rate;
-			end = end > i_count ? end - i_count : 0;
-			chunk = SPA_MIN(end, *in_len);
-
-			for (; i < chunk; i++)
-				m = SPA_MAX(fabsf(s[i]), m);
-
-			if (i == end) {
-				d[o++] = m;
-				m = 0.0f;
-				o_count++;
-			}
-		}
-		pd->max_f[c] = m;
-	}
-
-	*out_len = o;
-	*in_len = i;
-	pd->o_count = o_count;
-	pd->i_count = i_count + i;
-
-	while (pd->i_count >= r->i_rate) {
-		pd->i_count -= r->i_rate;
-		pd->o_count -= r->o_rate;
-	}
 }
+
+MAKE_PEAKS(c);

spa/plugins/audioconvert/resample-peaks-impl.h

@@ -34,11 +34,59 @@ struct peaks_data {
 	float max_f[];
 };
 
-void resample_peaks_process_c(struct resample *r,
+#define DEFINE_PEAKS(arch)						\
-	const void * SPA_RESTRICT src[], uint32_t *in_len,
+void resample_peaks_process_##arch(struct resample *r,			\
-	void * SPA_RESTRICT dst[], uint32_t *out_len);
+	const void * SPA_RESTRICT src[], uint32_t *in_len,		\
+	void * SPA_RESTRICT dst[], uint32_t *out_len)
+
+#define MAKE_PEAKS(arch)						\
+DEFINE_PEAKS(arch)							\
+{									\
+	struct peaks_data *pd = r->data;				\
+	uint32_t c, i, o, end, chunk, i_count, o_count;			\
+									\
+	if (SPA_UNLIKELY(r->channels == 0))				\
+		return;							\
+									\
+	for (c = 0; c < r->channels; c++) {				\
+		const float *s = src[c];				\
+		float *d = dst[c], m = pd->max_f[c];			\
+									\
+		o_count = pd->o_count;					\
+		i_count = pd->i_count;					\
+		o = i = 0;						\
+									\
+		while (i < *in_len && o < *out_len) {			\
+			end = ((uint64_t) (o_count + 1) 		\
+				* r->i_rate) / r->o_rate;		\
+			end = end > i_count ? end - i_count : 0;	\
+			chunk = SPA_MIN(end, *in_len);			\
+									\
+			m = find_abs_max_##arch(&s[i], chunk - i, m);	\
+									\
+			i += chunk;					\
+									\
+			if (i == end) {					\
+				d[o++] = m;				\
+				m = 0.0f;				\
+				o_count++;				\
+			}						\
+		}							\
+		pd->max_f[c] = m;					\
+	}								\
+	*out_len = o;							\
+	*in_len = i;							\
+	pd->o_count = o_count;						\
+	pd->i_count = i_count + i;					\
+									\
+	while (pd->i_count >= r->i_rate) {				\
+		pd->i_count -= r->i_rate;				\
+		pd->o_count -= r->o_rate;				\
+	}								\
+}
+
+
+DEFINE_PEAKS(c);
 #if defined (HAVE_SSE)
-void resample_peaks_process_sse(struct resample *r,
+DEFINE_PEAKS(sse);
-	const void * SPA_RESTRICT src[], uint32_t *in_len,
-	void * SPA_RESTRICT dst[], uint32_t *out_len);
 #endif

spa/plugins/audioconvert/resample-peaks-sse.c

@@ -37,60 +37,27 @@ static inline float hmax_ps(__m128 val)
 	return _mm_cvtss_f32(val);
 }
 
-void resample_peaks_process_sse(struct resample *r,
+static inline float find_abs_max_sse(const float *s, uint32_t n_samples, float m)
-			const void * SPA_RESTRICT src[], uint32_t *in_len,
-			void * SPA_RESTRICT dst[], uint32_t *out_len)
 {
-	struct peaks_data *pd = r->data;
+	__m128 in, max;
-	uint32_t c, i, o, end, chunk, unrolled, i_count, o_count;
+	uint32_t n, unrolled;
-	__m128 in, max, mask = _mm_andnot_ps(_mm_set_ps1(-0.0f),
+	const __m128 mask = _mm_andnot_ps(
-			_mm_cmpeq_ps(_mm_setzero_ps(), _mm_setzero_ps()));
+				_mm_set_ps1(-0.0f),
+				_mm_cmpeq_ps(_mm_setzero_ps(), _mm_setzero_ps()));
 
-	if (r->channels == 0)
+	max = _mm_set1_ps(m);
-		return;
 
-	for (c = 0; c < r->channels; c++) {
+	unrolled = n_samples & ~3;
-		const float *s = src[c];
-		float *d = dst[c], m = pd->max_f[c];
 
-		o_count = pd->o_count;
+	for (n = 0; n < unrolled; n+=4) {
-		i_count = pd->i_count;
+		in = _mm_loadu_ps(&s[n]);
-		o = i = 0;
+		in = _mm_and_ps(mask, in);
-
+		max = _mm_max_ps(in, max);
-		max = _mm_set1_ps(m);
-
-		while (i < *in_len && o < *out_len) {
-			end = ((uint64_t) (o_count + 1) * r->i_rate) / r->o_rate;
-			end = end > i_count ? end - i_count : 0;
-			chunk = SPA_MIN(end, *in_len);
-
-			unrolled = chunk - ((chunk - i) & 3);
-
-			for (; i < unrolled; i+=4) {
-				in = _mm_loadu_ps(&s[i]);
-				in = _mm_and_ps(mask, in);
-				max = _mm_max_ps(in, max);
-			}
-			for (; i < chunk; i++)
-				m = SPA_MAX(fabsf(s[i]), m);
-
-			if (i == end) {
-				d[o++] = SPA_MAX(hmax_ps(max), m);
-				m = 0.0f;
-				max = _mm_set1_ps(m);
-				o_count++;
-			}
-		}
-		pd->max_f[c] = SPA_MAX(hmax_ps(max), m);
 	}
+	for (; n < n_samples; n++)
+		m = fmaxf(fabsf(s[n]), m);
 
-	*out_len = o;
+	return fmaxf(hmax_ps(max), m);
-	*in_len = i;
-	pd->o_count = o_count;
-	pd->i_count = i_count + i;
-
-	while (pd->i_count >= r->i_rate) {
-		pd->i_count -= r->i_rate;
-		pd->o_count -= r->o_rate;
-	}
 }
+
+MAKE_PEAKS(sse);