Add 1D version of vpx_sum_squares_i16

Change-Id: I1829f931749a26aec38c896b609c5a2640d6dfaf

vpx_dsp/sum_squares.c

@@ -27,3 +27,13 @@ uint64_t vpx_sum_squares_2d_i16_c(const int16_t *src, int src_stride,
 
   return ss;
 }
+
+uint64_t vpx_sum_squares_i16_c(const int16_t *src, uint32_t n) {
+  uint64_t ss = 0;
+  do {
+    const int16_t v = *src++;
+    ss += v*v;
+  } while (--n);
+
+  return ss;
+}

vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -972,6 +972,9 @@ if (vpx_config("CONFIG_VP10_ENCODER") eq "yes") {
   #
   add_proto qw/uint64_t vpx_sum_squares_2d_i16/, "const int16_t *src, int stride, int size";
   specialize qw/vpx_sum_squares_2d_i16 sse2/;
+
+  add_proto qw/uint64_t vpx_sum_squares_i16/, "const int16_t *src, uint32_t N";
+  specialize qw/vpx_sum_squares_i16 sse2/;
 }
 
 if ((vpx_config("CONFIG_VP9_ENCODER") eq "yes") || (vpx_config("CONFIG_VP10_ENCODER") eq "yes")) {

vpx_dsp/x86/sum_squares_sse2.c

@@ -12,6 +12,8 @@
 #include <emmintrin.h>
 #include <stdio.h>
 
+#include "vpx_dsp/x86/synonyms.h"
+
 #include "./vpx_dsp_rtcd.h"
 
 static uint64_t vpx_sum_squares_2d_i16_4x4_sse2(const int16_t *src,
@@ -117,3 +119,77 @@ uint64_t vpx_sum_squares_2d_i16_sse2(const int16_t *src, int stride,
     return vpx_sum_squares_2d_i16_nxn_sse2(src, stride, size);
   }
 }
+
+//////////////////////////////////////////////////////////////////////////////
+// 1D version
+//////////////////////////////////////////////////////////////////////////////
+
+static uint64_t vpx_sum_squares_i16_64n_sse2(const int16_t *src, uint32_t n) {
+  const __m128i v_zext_mask_q = _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff);
+  __m128i v_acc0_q = _mm_setzero_si128();
+  __m128i v_acc1_q = _mm_setzero_si128();
+
+  const int16_t *const end = src + n;
+
+  assert(n % 64 == 0);
+
+  while (src < end) {
+    const __m128i v_val_0_w = xx_load_128(src);
+    const __m128i v_val_1_w = xx_load_128(src + 8);
+    const __m128i v_val_2_w = xx_load_128(src + 16);
+    const __m128i v_val_3_w = xx_load_128(src + 24);
+    const __m128i v_val_4_w = xx_load_128(src + 32);
+    const __m128i v_val_5_w = xx_load_128(src + 40);
+    const __m128i v_val_6_w = xx_load_128(src + 48);
+    const __m128i v_val_7_w = xx_load_128(src + 56);
+
+    const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
+    const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
+    const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
+    const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
+    const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w);
+    const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w);
+    const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w);
+    const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w);
+
+    const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
+    const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
+    const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d);
+    const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d);
+
+    const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
+    const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d);
+
+    const __m128i v_sum_d = _mm_add_epi32(v_sum_0123_d, v_sum_4567_d);
+
+    v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_and_si128(v_sum_d, v_zext_mask_q));
+    v_acc1_q = _mm_add_epi64(v_acc1_q, _mm_srli_epi64(v_sum_d, 32));
+
+    src += 64;
+  }
+
+  v_acc0_q = _mm_add_epi64(v_acc0_q, v_acc1_q);
+  v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8));
+
+#if ARCH_X86_64
+  return (uint64_t)_mm_cvtsi128_si64(v_acc0_q);
+#else
+  {
+    uint64_t tmp;
+    _mm_storel_epi64((__m128i*)&tmp, v_acc0_q);
+    return tmp;
+  }
+#endif
+}
+
+uint64_t vpx_sum_squares_i16_sse2(const int16_t *src, uint32_t n) {
+  if (n % 64 == 0) {
+    return vpx_sum_squares_i16_64n_sse2(src, n);
+  } else if (n > 64) {
+    int k = n & ~(64-1);
+    return vpx_sum_squares_i16_64n_sse2(src, k) +
+           vpx_sum_squares_i16_c(src + k, n - k);
+  } else {
+    return vpx_sum_squares_i16_c(src, n);
+  }
+}