From db26e381a45aadcd82851075b85e2466e7de77d2 Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <jmvalin@amazon.com>
Date: Tue, 28 Nov 2023 14:16:57 -0500
Subject: Trying to use fma instructions when possible

Compilers sometimes replace vmlaq*() with fmul+fadd instead of fmla.
Trying to use vfmaq*() instead when possible.
---
 celt/arm/celt_neon_intr.c  | 8 ++++++++
 celt/arm/pitch_neon_intr.c | 7 +++++++
 dnn/vec_neon.h             | 6 ++++++
 3 files changed, 21 insertions(+)

diff --git a/celt/arm/celt_neon_intr.c b/celt/arm/celt_neon_intr.c
index effda769..c8301101 100644
--- a/celt/arm/celt_neon_intr.c
+++ b/celt/arm/celt_neon_intr.c
@@ -97,6 +97,14 @@ void xcorr_kernel_neon_fixed(const opus_val16 * x, const opus_val16 * y, opus_va
 }
 
 #else
+
+#if defined(__ARM_FEATURE_FMA) && defined(__ARM_ARCH_ISA_A64)
+/* If we can, force the compiler to use an FMA instruction rather than break
+ *    vmlaq_f32() into fmul/fadd. */
+#define vmlaq_lane_f32(a,b,c,lane) vfmaq_lane_f32(a,b,c,lane)
+#endif
+
+
 /*
  * Function: xcorr_kernel_neon_float
  * ---------------------------------
diff --git a/celt/arm/pitch_neon_intr.c b/celt/arm/pitch_neon_intr.c
index 35cc46e2..43885f52 100644
--- a/celt/arm/pitch_neon_intr.c
+++ b/celt/arm/pitch_neon_intr.c
@@ -130,6 +130,13 @@ void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, const opus
 
 /* ========================================================================== */
 
+#ifdef __ARM_FEATURE_FMA
+/* If we can, force the compiler to use an FMA instruction rather than break
+   vmlaq_f32() into fmul/fadd. */
+#define vmlaq_f32(a,b,c) vfmaq_f32(a,b,c)
+#endif
+
+
 #ifdef OPUS_CHECK_ASM
 
 /* This part of code simulates floating-point NEON operations. */
diff --git a/dnn/vec_neon.h b/dnn/vec_neon.h
index 785fc056..acf49f47 100644
--- a/dnn/vec_neon.h
+++ b/dnn/vec_neon.h
@@ -49,6 +49,12 @@ static OPUS_INLINE int16x8_t vmull_high_s8(int8x16_t a, int8x16_t b) {
 }
 #endif
 
+#ifdef __ARM_FEATURE_FMA
+/* If we can, force the compiler to use an FMA instruction rather than break
+   vmlaq_f32() into fmul/fadd. */
+#define vmlaq_f32(a,b,c) vfmaq_f32(a,b,c)
+#endif
+
 #ifndef LPCNET_TEST
 static inline float32x4_t exp4_approx(float32x4_t x) {
   int32x4_t i;
-- 
cgit v1.2.3