filter-chain: add avx mix function

src/modules/module-filter-chain/dsp-ops-avx.c

@@ -8,10 +8,70 @@
 
 #include <spa/utils/defs.h>
 
+#include "config.h"
 #include "dsp-ops.h"
 
 #include <immintrin.h>
 
+void dsp_mix_gain_avx(struct dsp_ops *ops,
+		void * SPA_RESTRICT dst,
+		const void * SPA_RESTRICT src[],
+		float gain[], uint32_t n_src, uint32_t n_samples)
+{
+	if (n_src == 0) {
+		memset(dst, 0, n_samples * sizeof(float));
+	} else if (n_src == 1 && gain[0] == 1.0f) {
+		if (dst != src[0])
+			spa_memcpy(dst, src[0], n_samples * sizeof(float));
+	} else {
+		uint32_t n, i, unrolled;
+		__m256 in[4], g;
+		const float **s = (const float **)src;
+		float *d = dst;
+
+		if (SPA_LIKELY(SPA_IS_ALIGNED(dst, 32))) {
+			unrolled = n_samples & ~31;
+			for (i = 0; i < n_src; i++) {
+				if (SPA_UNLIKELY(!SPA_IS_ALIGNED(src[i], 32))) {
+					unrolled = 0;
+					break;
+				}
+			}
+		} else
+			unrolled = 0;
+
+		for (n = 0; n < unrolled; n += 32) {
+			g = _mm256_set1_ps(gain[0]);
+			in[0] = _mm256_mul_ps(g, _mm256_load_ps(&s[0][n+ 0]));
+			in[1] = _mm256_mul_ps(g, _mm256_load_ps(&s[0][n+ 8]));
+			in[2] = _mm256_mul_ps(g, _mm256_load_ps(&s[0][n+16]));
+			in[3] = _mm256_mul_ps(g, _mm256_load_ps(&s[0][n+24]));
+
+			for (i = 1; i < n_src; i++) {
+				g = _mm256_set1_ps(gain[i]);
+				in[0] = _mm256_add_ps(in[0], _mm256_mul_ps(g, _mm256_load_ps(&s[i][n+ 0])));
+				in[1] = _mm256_add_ps(in[1], _mm256_mul_ps(g, _mm256_load_ps(&s[i][n+ 8])));
+				in[2] = _mm256_add_ps(in[2], _mm256_mul_ps(g, _mm256_load_ps(&s[i][n+16])));
+				in[3] = _mm256_add_ps(in[3], _mm256_mul_ps(g, _mm256_load_ps(&s[i][n+24])));
+			}
+			_mm256_store_ps(&d[n+ 0], in[0]);
+			_mm256_store_ps(&d[n+ 8], in[1]);
+			_mm256_store_ps(&d[n+16], in[2]);
+			_mm256_store_ps(&d[n+24], in[3]);
+		}
+		for (; n < n_samples; n++) {
+			__m128 in[1], g;
+			g = _mm_set_ss(gain[0]);
+			in[0] = _mm_mul_ss(g, _mm_load_ss(&s[0][n]));
+			for (i = 1; i < n_src; i++) {
+				g = _mm_set_ss(gain[i]);
+				in[0] = _mm_add_ss(in[0], _mm_mul_ss(g, _mm_load_ss(&s[i][n])));
+			}
+			_mm_store_ss(&d[n], in[0]);
+		}
+	}
+}
+
 void dsp_sum_avx(struct dsp_ops *ops, float *r, const float *a, const float *b, uint32_t n_samples)
 {
 	uint32_t n, unrolled;

src/modules/module-filter-chain/dsp-ops.c

@@ -25,7 +25,7 @@ static struct dsp_info dsp_table[] =
 	{ SPA_CPU_FLAG_AVX,
 		.funcs.clear = dsp_clear_c,
 		.funcs.copy = dsp_copy_c,
-		.funcs.mix_gain = dsp_mix_gain_sse,
+		.funcs.mix_gain = dsp_mix_gain_avx,
 		.funcs.biquad_run = dsp_biquad_run_sse,
 		.funcs.sum = dsp_sum_avx,
 		.funcs.linear = dsp_linear_c,

src/modules/module-filter-chain/dsp-ops.h

@@ -150,6 +150,7 @@ MAKE_BIQUADN_RUN_FUNC(sse);
 MAKE_DELAY_FUNC(sse);
 #endif
 #if defined (HAVE_AVX)
+MAKE_MIX_GAIN_FUNC(avx);
 MAKE_SUM_FUNC(avx);
 #endif