Add NEON intrinsics version for WebRtcSpl_MinValueW16Neon
WebRtcSpl_MinValueW16Neon is added. SplTest in common_audio_unittests
is passed on ARM32/ARM64 platforms.
BUG=4002
R=andrew@webrtc.org, jridges@masque.com
Change-Id: I33c3853766d7594ed121166288e5325a03b3d6fe
Review URL: https://webrtc-codereview.appspot.com/32639004
Patch from Yang Zhang <yang.zhang@arm.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk/webrtc@7867 4adac7df-926f-26a2-2b94-8c16560cd09d
diff --git a/common_audio/signal_processing/min_max_operations_neon.c b/common_audio/signal_processing/min_max_operations_neon.c
index aa6cf1f..704911e 100644
--- a/common_audio/signal_processing/min_max_operations_neon.c
+++ b/common_audio/signal_processing/min_max_operations_neon.c
@@ -67,3 +67,43 @@
return (int16_t)maximum;
}
+// Minimum value of word16 vector. NEON intrinsics version for
+// ARM 32-bit/64-bit platforms.
+int16_t WebRtcSpl_MinValueW16Neon(const int16_t* vector, int length) {
+ int16_t minimum = WEBRTC_SPL_WORD16_MAX;
+ int i = 0;
+ int residual = length & 0x7;
+
+ if (vector == NULL || length <= 0) {
+ return minimum;
+ }
+
+ const int16_t* p_start = vector;
+ int16x8_t min16x8 = vdupq_n_s16(WEBRTC_SPL_WORD16_MAX);
+
+ // First part, unroll the loop 8 times.
+ for (i = length - residual; i >0; i -= 8) {
+ int16x8_t in16x8 = vld1q_s16(p_start);
+ min16x8 = vminq_s16(min16x8, in16x8);
+ p_start += 8;
+ }
+
+#if defined(WEBRTC_ARCH_ARM64)
+ minimum = vminvq_s16(min16x8);
+#else
+ int16x4_t min16x4 = vmin_s16(vget_low_s16(min16x8), vget_high_s16(min16x8));
+ min16x4 = vpmin_s16(min16x4, min16x4);
+ min16x4 = vpmin_s16(min16x4, min16x4);
+
+ minimum = vget_lane_s16(min16x4, 0);
+#endif
+
+ // Second part, do the remaining iterations (if any).
+ for (i = residual; i > 0; i--) {
+ if (*p_start < minimum)
+ minimum = *p_start;
+ p_start++;
+ }
+ return minimum;
+}
+
