From cd15b1775e9c81283f998a874bda55821e79f3fc Mon Sep 17 00:00:00 2001
From: Chris Robinson <chris.kcat@gmail.com>
Date: Wed, 30 Aug 2017 11:30:19 -0700
Subject: Avoid some extraneous load calls

This likely doesn't change anything given a working optimizer, but it cleans up
the code some.
---
 Alc/mixer_neon.c | 23 +++++++++++------------
 Alc/mixer_sse.c  | 26 ++++++++++++--------------
 2 files changed, 23 insertions(+), 26 deletions(-)

(limited to 'Alc')

diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 58929e1a..b99dcf69 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -144,8 +144,8 @@ const ALfloat *Resample_bsinc_Neon(const InterpState *state,
     const ALfloat *const filter = state->bsinc.filter;
     const float32x4_t sf4 = vdupq_n_f32(state->bsinc.sf);
     const ALsizei m = state->bsinc.m;
-    const ALfloat *fil, *scd, *phd, *spd;
-    ALsizei pi, i, j;
+    const float32x4_t *fil, *scd, *phd, *spd;
+    ALsizei pi, i, j, offset;
     float32x4_t r4;
     ALfloat pf;
 
@@ -158,23 +158,22 @@ const ALfloat *Resample_bsinc_Neon(const InterpState *state,
         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
 #undef FRAC_PHASE_BITDIFF
 
-        fil = ASSUME_ALIGNED(filter + m*pi*4, 16);
-        scd = ASSUME_ALIGNED(fil + m, 16);
-        phd = ASSUME_ALIGNED(scd + m, 16);
-        spd = ASSUME_ALIGNED(phd + m, 16);
+        offset = m*pi*4;
+        fil = ASSUME_ALIGNED(filter + offset, 16); offset += m;
+        scd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
+        phd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
+        spd = ASSUME_ALIGNED(filter + offset, 16);
 
         // Apply the scale and phase interpolated filter.
         r4 = vdupq_n_f32(0.0f);
         {
             const float32x4_t pf4 = vdupq_n_f32(pf);
-            for(j = 0;j < m;j+=4)
+            for(j = 0;j < m;j+=4,fil++,scd++,phd++,spd++)
             {
                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
-                const float32x4_t f4 = vmlaq_f32(vmlaq_f32(vld1q_f32(&fil[j]),
-                                                           sf4, vld1q_f32(&scd[j])),
-                    pf4, vmlaq_f32(vld1q_f32(&phd[j]),
-                        sf4, vld1q_f32(&spd[j])
-                    )
+                const float32x4_t f4 = vmlaq_f32(
+                    vmlaq_f32(*fil, sf4, *scd),
+                    pf4, vmlaq_f32(*phd, sf4, *spd)
                 );
                 /* r += f*src */
                 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c
index 5273a088..773a1c2f 100644
--- a/Alc/mixer_sse.c
+++ b/Alc/mixer_sse.c
@@ -19,8 +19,8 @@ const ALfloat *Resample_bsinc_SSE(const InterpState *state, const ALfloat *restr
     const ALfloat *const filter = state->bsinc.filter;
     const __m128 sf4 = _mm_set1_ps(state->bsinc.sf);
     const ALsizei m = state->bsinc.m;
-    const ALfloat *fil, *scd, *phd, *spd;
-    ALsizei pi, i, j;
+    const __m128 *fil, *scd, *phd, *spd;
+    ALsizei pi, i, j, offset;
     ALfloat pf;
     __m128 r4;
 
@@ -33,30 +33,28 @@ const ALfloat *Resample_bsinc_SSE(const InterpState *state, const ALfloat *restr
         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
 #undef FRAC_PHASE_BITDIFF
 
-        fil = ASSUME_ALIGNED(filter + m*pi*4, 16);
-        scd = ASSUME_ALIGNED(fil + m, 16);
-        phd = ASSUME_ALIGNED(scd + m, 16);
-        spd = ASSUME_ALIGNED(phd + m, 16);
+        offset = m*pi*4;
+        fil = ASSUME_ALIGNED(filter + offset, 16); offset += m;
+        scd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
+        phd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
+        spd = ASSUME_ALIGNED(filter + offset, 16);
 
         // Apply the scale and phase interpolated filter.
         r4 = _mm_setzero_ps();
         {
             const __m128 pf4 = _mm_set1_ps(pf);
-#define LD4(x) _mm_load_ps(x)
-#define ULD4(x) _mm_loadu_ps(x)
 #define MLA4(x, y, z) _mm_add_ps(x, _mm_mul_ps(y, z))
-            for(j = 0;j < m;j+=4)
+            for(j = 0;j < m;j+=4,fil++,scd++,phd++,spd++)
             {
                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
-                const __m128 f4 = MLA4(MLA4(LD4(&fil[j]), sf4, LD4(&scd[j])),
-                    pf4, MLA4(LD4(&phd[j]), sf4, LD4(&spd[j]))
+                const __m128 f4 = MLA4(
+                    MLA4(*fil, sf4, *scd),
+                    pf4, MLA4(*phd, sf4, *spd)
                 );
                 /* r += f*src */
-                r4 = MLA4(r4, f4, ULD4(&src[j]));
+                r4 = MLA4(r4, f4, _mm_loadu_ps(&src[j]));
             }
 #undef MLA4
-#undef ULD4
-#undef LD4
         }
         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
-- 
cgit v1.2.3