From 2c132be6262494517bad0249ef6af9ba10e46e12 Mon Sep 17 00:00:00 2001
From: Wim Taymans <wtaymans@redhat.com>
Date: Wed, 23 Oct 2024 12:54:23 +0200
Subject: [PATCH] audioconvert: align some buffers

so that we can use aligned read and writes in SSE.
---
 spa/plugins/audioconvert/channelmix-ops-sse.c | 10 +++++-----
 spa/plugins/audioconvert/channelmix-ops.c     |  6 +++++-
 spa/plugins/audioconvert/channelmix-ops.h     | 10 ++++++----
 3 files changed, 16 insertions(+), 10 deletions(-)

diff --git a/spa/plugins/audioconvert/channelmix-ops-sse.c b/spa/plugins/audioconvert/channelmix-ops-sse.c
index 2b04ace0a..9a128177f 100644
--- a/spa/plugins/audioconvert/channelmix-ops-sse.c
+++ b/spa/plugins/audioconvert/channelmix-ops-sse.c
@@ -266,7 +266,7 @@ static inline void convolver_run(const float *src, float *dst,
 	sum[0] = _mm_setzero_ps();
 	for(i = 0; i < n_taps; i+=4) {
 		t[0] = _mm_loadu_ps(&src[i]);
-		sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(_mm_loadu_ps(&taps[i]), t[0]));
+		sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(_mm_load_ps(&taps[i]), t[0]));
 	}
 	sum[0] = _mm_add_ps(sum[0], _mm_movehl_ps(sum[0], sum[0]));
 	sum[0] = _mm_add_ss(sum[0], _mm_shuffle_ps(sum[0], sum[0], 0x55));
@@ -303,8 +303,8 @@ static inline void delay_convolve_run_sse(float *buffer, uint32_t *pos,
 			w += 4;
 			if (w >= n_buffer) {
 				w -= n_buffer;
-				t[0] = _mm_loadu_ps(&buffer[n_buffer]);
-				_mm_storeu_ps(&buffer[0], t[0]);
+				t[0] = _mm_load_ps(&buffer[n_buffer]);
+				_mm_store_ps(&buffer[0], t[0]);
 			}
 		}
 		for(; n < n_samples; n++) {
@@ -326,8 +326,8 @@ static inline void delay_convolve_run_sse(float *buffer, uint32_t *pos,
 			w += 4;
 			if (w >= n_buffer) {
 				w -= n_buffer;
-				t[0] = _mm_loadu_ps(&buffer[n_buffer]);
-				_mm_storeu_ps(&buffer[0], t[0]);
+				t[0] = _mm_load_ps(&buffer[n_buffer]);
+				_mm_store_ps(&buffer[0], t[0]);
 			}
 		}
 		for(; n < n_samples; n++) {
diff --git a/spa/plugins/audioconvert/channelmix-ops.c b/spa/plugins/audioconvert/channelmix-ops.c
index f7db4c61c..d774112bc 100644
--- a/spa/plugins/audioconvert/channelmix-ops.c
+++ b/spa/plugins/audioconvert/channelmix-ops.c
@@ -775,7 +775,11 @@ int channelmix_init(struct channelmix *mix)
 	mix->delay = (uint32_t)(mix->rear_delay * mix->freq / 1000.0f);
 	mix->func_name = info->name;
 
-	spa_zero(mix->taps);
+	spa_zero(mix->taps_mem);
+	mix->taps = SPA_PTR_ALIGN(mix->taps_mem, CHANNELMIX_OPS_MAX_ALIGN, float);
+	mix->buffer[0] = SPA_PTR_ALIGN(&mix->buffer_mem[0], CHANNELMIX_OPS_MAX_ALIGN, float);
+	mix->buffer[1] = SPA_PTR_ALIGN(&mix->buffer_mem[2*BUFFER_SIZE], CHANNELMIX_OPS_MAX_ALIGN, float);
+
 	if (mix->hilbert_taps > 0) {
 		mix->n_taps = SPA_CLAMP(mix->hilbert_taps, 15u, MAX_TAPS) | 1;
 		blackman_window(mix->taps, mix->n_taps);
diff --git a/spa/plugins/audioconvert/channelmix-ops.h b/spa/plugins/audioconvert/channelmix-ops.h
index e1102c5ec..26e2efc3a 100644
--- a/spa/plugins/audioconvert/channelmix-ops.h
+++ b/spa/plugins/audioconvert/channelmix-ops.h
@@ -25,6 +25,8 @@
 #define BUFFER_SIZE 4096
 #define MAX_TAPS 255u
 
+#define CHANNELMIX_OPS_MAX_ALIGN 16
+
 struct channelmix {
 	uint32_t src_chan;
 	uint32_t dst_chan;
@@ -59,10 +61,12 @@ struct channelmix {
 	uint32_t hilbert_taps;				/* to phase shift, 0 disabled */
 	struct lr4 lr4[SPA_AUDIO_MAX_CHANNELS];
 
-	float buffer[2][BUFFER_SIZE*2 + 16];
+	float buffer_mem[2 * BUFFER_SIZE*2 + CHANNELMIX_OPS_MAX_ALIGN/4];
+	float *buffer[2];
 	uint32_t pos[2];
 	uint32_t delay;
-	float taps[MAX_TAPS];
+	float taps_mem[MAX_TAPS + CHANNELMIX_OPS_MAX_ALIGN/4];
+	float *taps;
 	uint32_t n_taps;
 
 	void (*process) (struct channelmix *mix, void * SPA_RESTRICT dst[],
@@ -104,8 +108,6 @@ void channelmix_##name##_##arch(struct channelmix *mix,				\
 		void * SPA_RESTRICT dst[], const void * SPA_RESTRICT src[],	\
 		uint32_t n_samples);
 
-#define CHANNELMIX_OPS_MAX_ALIGN 16
-
 DEFINE_FUNCTION(copy, c);
 DEFINE_FUNCTION(f32_n_m, c);
 DEFINE_FUNCTION(f32_1_2, c);