From: Scott LaVarnway <slavarnway@google.com>
Date: Fri, 20 Jul 2018 10:50:49 +0000 (-0700)
Subject: VPX: Improve HBD vpx_hadamard_16x16_sse2()
X-Git-Tag: v1.8.0~514^2
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=17ffb22d4daf4c4645b4c79086d8c27a2768775e;p=libvpx

VPX: Improve HBD vpx_hadamard_16x16_sse2()

~12% improvement.

Change-Id: Ieca4d870a4c1c5ea2c689e27fc4550fcbab9f867
---

diff --git a/vpx_dsp/x86/avg_intrin_sse2.c b/vpx_dsp/x86/avg_intrin_sse2.c
index 0362c63c5..bb744a981 100644
--- a/vpx_dsp/x86/avg_intrin_sse2.c
+++ b/vpx_dsp/x86/avg_intrin_sse2.c
@@ -264,8 +264,9 @@ static void hadamard_col8_sse2(__m128i *in, int iter) {
   }
 }
 
-void vpx_hadamard_8x8_sse2(int16_t const *src_diff, ptrdiff_t src_stride,
-                           tran_low_t *coeff) {
+static INLINE void hadamard_8x8_sse2(int16_t const *src_diff,
+                                     ptrdiff_t src_stride, tran_low_t *_coeff,
+                                     int is_final) {
   __m128i src[8];
   src[0] = _mm_load_si128((const __m128i *)src_diff);
   src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
@@ -279,37 +280,72 @@ void vpx_hadamard_8x8_sse2(int16_t const *src_diff, ptrdiff_t src_stride,
   hadamard_col8_sse2(src, 0);
   hadamard_col8_sse2(src, 1);
 
-  store_tran_low(src[0], coeff);
-  coeff += 8;
-  store_tran_low(src[1], coeff);
-  coeff += 8;
-  store_tran_low(src[2], coeff);
-  coeff += 8;
-  store_tran_low(src[3], coeff);
-  coeff += 8;
-  store_tran_low(src[4], coeff);
-  coeff += 8;
-  store_tran_low(src[5], coeff);
-  coeff += 8;
-  store_tran_low(src[6], coeff);
-  coeff += 8;
-  store_tran_low(src[7], coeff);
+  if (is_final) {
+    store_tran_low(src[0], _coeff);
+    _coeff += 8;
+    store_tran_low(src[1], _coeff);
+    _coeff += 8;
+    store_tran_low(src[2], _coeff);
+    _coeff += 8;
+    store_tran_low(src[3], _coeff);
+    _coeff += 8;
+    store_tran_low(src[4], _coeff);
+    _coeff += 8;
+    store_tran_low(src[5], _coeff);
+    _coeff += 8;
+    store_tran_low(src[6], _coeff);
+    _coeff += 8;
+    store_tran_low(src[7], _coeff);
+  } else {
+    int16_t *coeff = (int16_t *)_coeff;
+    _mm_store_si128((__m128i *)coeff, src[0]);
+    coeff += 8;
+    _mm_store_si128((__m128i *)coeff, src[1]);
+    coeff += 8;
+    _mm_store_si128((__m128i *)coeff, src[2]);
+    coeff += 8;
+    _mm_store_si128((__m128i *)coeff, src[3]);
+    coeff += 8;
+    _mm_store_si128((__m128i *)coeff, src[4]);
+    coeff += 8;
+    _mm_store_si128((__m128i *)coeff, src[5]);
+    coeff += 8;
+    _mm_store_si128((__m128i *)coeff, src[6]);
+    coeff += 8;
+    _mm_store_si128((__m128i *)coeff, src[7]);
+  }
+}
+
+void vpx_hadamard_8x8_sse2(int16_t const *src_diff, ptrdiff_t src_stride,
+                           tran_low_t *coeff) {
+  hadamard_8x8_sse2(src_diff, src_stride, coeff, 1);
 }
 
 void vpx_hadamard_16x16_sse2(int16_t const *src_diff, ptrdiff_t src_stride,
                              tran_low_t *coeff) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  // For high bitdepths, it is unnecessary to store_tran_low
+  // (mult/unpack/store), then load_tran_low (load/pack) the same memory in the
+  // next stage.  Output to an intermediate buffer first, then store_tran_low()
+  // in the final stage.
+  DECLARE_ALIGNED(32, int16_t, temp_coeff[16 * 16]);
+  int16_t *t_coeff = temp_coeff;
+#else
+  int16_t *t_coeff = coeff;
+#endif
   int idx;
   for (idx = 0; idx < 4; ++idx) {
     int16_t const *src_ptr =
         src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8;
-    vpx_hadamard_8x8_sse2(src_ptr, src_stride, coeff + idx * 64);
+    hadamard_8x8_sse2(src_ptr, src_stride, (tran_low_t *)(t_coeff + idx * 64),
+                      0);
   }
 
   for (idx = 0; idx < 64; idx += 8) {
-    __m128i coeff0 = load_tran_low(coeff);
-    __m128i coeff1 = load_tran_low(coeff + 64);
-    __m128i coeff2 = load_tran_low(coeff + 128);
-    __m128i coeff3 = load_tran_low(coeff + 192);
+    __m128i coeff0 = _mm_load_si128((const __m128i *)t_coeff);
+    __m128i coeff1 = _mm_load_si128((const __m128i *)(t_coeff + 64));
+    __m128i coeff2 = _mm_load_si128((const __m128i *)(t_coeff + 128));
+    __m128i coeff3 = _mm_load_si128((const __m128i *)(t_coeff + 192));
 
     __m128i b0 = _mm_add_epi16(coeff0, coeff1);
     __m128i b1 = _mm_sub_epi16(coeff0, coeff1);
@@ -332,6 +368,7 @@ void vpx_hadamard_16x16_sse2(int16_t const *src_diff, ptrdiff_t src_stride,
     store_tran_low(coeff3, coeff + 192);
 
     coeff += 8;
+    t_coeff += 8;
   }
 }