Update vp9_scale_and_extend_frame_ssse3()

Change-Id: I22622faebfcc36f7a4d1f37e3800ae8ab87c8cd4

diff --git a/test/vp9_scale_test.cc b/test/vp9_scale_test.cc
index f16cd48..5d7d38e 100644 (file)
--- a/test/vp9_scale_test.cc
+++ b/test/vp9_scale_test.cc
@@ -48,17 +48,19 @@ class ScaleTest : public VpxScaleBase,
   }
 
   void RunTest() {
-    static const int kNumSizesToTest = 4;
+    static const int kNumSizesToTest = 20;
     static const int kNumScaleFactorsToTest = 4;
-    static const int kWidthsToTest[] = { 16, 32, 48, 64 };
-    static const int kHeightsToTest[] = { 16, 20, 24, 28 };
+    static const int kSizesToTest[] = {
+      2,  4,  6,  8,  10, 12, 14, 16, 18,  20,
+      22, 24, 26, 28, 30, 32, 34, 68, 128, 134
+    };
     static const int kScaleFactors[] = { 1, 2, 3, 4 };
     for (INTERP_FILTER filter_type = 0; filter_type < 4; ++filter_type) {
       for (int phase_scaler = 0; phase_scaler < 16; ++phase_scaler) {
         for (int h = 0; h < kNumSizesToTest; ++h) {
-          const int src_height = kHeightsToTest[h];
+          const int src_height = kSizesToTest[h];
           for (int w = 0; w < kNumSizesToTest; ++w) {
-            const int src_width = kWidthsToTest[w];
+            const int src_width = kSizesToTest[w];
             for (int sf_up_idx = 0; sf_up_idx < kNumScaleFactorsToTest;
                  ++sf_up_idx) {
               const int sf_up = kScaleFactors[sf_up_idx];
@@ -71,7 +73,13 @@ class ScaleTest : public VpxScaleBase,
                   continue;
                 }
                 // I420 frame width and height must be even.
-                if (dst_width & 1 || dst_height & 1) {
+                if (!dst_width || !dst_height || dst_width & 1 ||
+                    dst_height & 1) {
+                  continue;
+                }
+                // vpx_convolve8_c() has restriction on the step which cannot
+                // exceed 64 (ratio 1 to 4).
+                if (src_width > 4 * dst_width || src_height > 4 * dst_height) {
                   continue;
                 }
                 ASSERT_NO_FATAL_FAILURE(ResetScaleImages(

diff --git a/vp9/encoder/x86/vp9_frame_scale_ssse3.c b/vp9/encoder/x86/vp9_frame_scale_ssse3.c
index c0c92a9..0d7bd1d 100644 (file)
--- a/vp9/encoder/x86/vp9_frame_scale_ssse3.c
+++ b/vp9/encoder/x86/vp9_frame_scale_ssse3.c
@@ -13,158 +13,147 @@
 #include "./vp9_rtcd.h"
 #include "./vpx_dsp_rtcd.h"
 #include "./vpx_scale_rtcd.h"
+#include "vpx_dsp/x86/convolve_ssse3.h"
+#include "vpx_dsp/x86/mem_sse2.h"
 #include "vpx_scale/yv12config.h"
 
-static void downsample_2_to_1_ssse3(const uint8_t *src, ptrdiff_t src_stride,
-                                    uint8_t *dst, ptrdiff_t dst_stride, int w,
-                                    int h) {
+static void scale_plane_2_to_1_phase_0(const uint8_t *src,
+                                       const ptrdiff_t src_stride, uint8_t *dst,
+                                       const ptrdiff_t dst_stride,
+                                       const int dst_w, const int dst_h) {
   const __m128i mask = _mm_set1_epi16(0x00FF);
-  const int max_width = w & ~15;
-  int y;
-  for (y = 0; y < h; ++y) {
-    int x;
-    for (x = 0; x < max_width; x += 16) {
-      const __m128i a = _mm_loadu_si128((const __m128i *)(src + x * 2 + 0));
-      const __m128i b = _mm_loadu_si128((const __m128i *)(src + x * 2 + 16));
+  const int max_width = (dst_w + 15) & ~15;
+  int y = dst_h;
+
+  do {
+    int x = max_width;
+    do {
+      const __m128i a = _mm_loadu_si128((const __m128i *)(src + 0));
+      const __m128i b = _mm_loadu_si128((const __m128i *)(src + 16));
       const __m128i a_and = _mm_and_si128(a, mask);
       const __m128i b_and = _mm_and_si128(b, mask);
       const __m128i c = _mm_packus_epi16(a_and, b_and);
-      _mm_storeu_si128((__m128i *)(dst + x), c);
-    }
-    for (; x < w; ++x) dst[x] = src[x * 2];
-    src += src_stride * 2;
-    dst += dst_stride;
-  }
+      _mm_storeu_si128((__m128i *)dst, c);
+      src += 32;
+      dst += 16;
+      x -= 16;
+    } while (x);
+    src += 2 * (src_stride - max_width);
+    dst += dst_stride - max_width;
+  } while (--y);
 }
 
-static INLINE __m128i filter(const __m128i *const a, const __m128i *const b,
-                             const __m128i *const c, const __m128i *const d,
-                             const __m128i *const e, const __m128i *const f,
-                             const __m128i *const g, const __m128i *const h) {
-  // TODO(linfengz): hard coded coefficients should be replaced with general
-  // coefficients
-  // reading.
-  const __m128i coeffs_ab =
-      _mm_set_epi8(6, -1, 6, -1, 6, -1, 6, -1, 6, -1, 6, -1, 6, -1, 6, -1);
-  const __m128i coeffs_cd = _mm_set_epi8(78, -19, 78, -19, 78, -19, 78, -19, 78,
-                                         -19, 78, -19, 78, -19, 78, -19);
-  const __m128i const64_x16 = _mm_set1_epi16(64);
-  const __m128i ab = _mm_unpacklo_epi8(*a, *b);
-  const __m128i cd = _mm_unpacklo_epi8(*c, *d);
-  const __m128i fe = _mm_unpacklo_epi8(*f, *e);
-  const __m128i hg = _mm_unpacklo_epi8(*h, *g);
-  const __m128i ab_terms = _mm_maddubs_epi16(ab, coeffs_ab);
-  const __m128i cd_terms = _mm_maddubs_epi16(cd, coeffs_cd);
-  const __m128i fe_terms = _mm_maddubs_epi16(fe, coeffs_cd);
-  const __m128i hg_terms = _mm_maddubs_epi16(hg, coeffs_ab);
-  // can not overflow
-  const __m128i abcd_terms = _mm_add_epi16(ab_terms, cd_terms);
-  // can not overflow
-  const __m128i fehg_terms = _mm_add_epi16(fe_terms, hg_terms);
-  // can overflow, use saturating add
-  const __m128i terms = _mm_adds_epi16(abcd_terms, fehg_terms);
-  const __m128i round = _mm_adds_epi16(terms, const64_x16);
-  const __m128i shift = _mm_srai_epi16(round, 7);
-  return _mm_packus_epi16(shift, shift);
+static INLINE __m128i scale_1_to_2_phase_0_kernel(const __m128i *const s,
+                                                  const __m128i *const f) {
+  __m128i ss[4], temp;
+
+  ss[0] = _mm_unpacklo_epi8(s[0], s[1]);
+  ss[1] = _mm_unpacklo_epi8(s[2], s[3]);
+  ss[2] = _mm_unpacklo_epi8(s[4], s[5]);
+  ss[3] = _mm_unpacklo_epi8(s[6], s[7]);
+  temp = convolve8_8_ssse3(ss, f);
+  return _mm_packus_epi16(temp, temp);
 }
 
-static void eight_tap_row_ssse3(const uint8_t *src, uint8_t *dst, int w) {
-  const int max_width = w & ~7;
-  int x = 0;
-  for (; x < max_width; x += 8) {
-    const __m128i a = _mm_loadl_epi64((const __m128i *)(src + x + 0));
-    const __m128i b = _mm_loadl_epi64((const __m128i *)(src + x + 1));
-    const __m128i c = _mm_loadl_epi64((const __m128i *)(src + x + 2));
-    const __m128i d = _mm_loadl_epi64((const __m128i *)(src + x + 3));
-    const __m128i e = _mm_loadl_epi64((const __m128i *)(src + x + 4));
-    const __m128i f = _mm_loadl_epi64((const __m128i *)(src + x + 5));
-    const __m128i g = _mm_loadl_epi64((const __m128i *)(src + x + 6));
-    const __m128i h = _mm_loadl_epi64((const __m128i *)(src + x + 7));
-    const __m128i pack = filter(&a, &b, &c, &d, &e, &f, &g, &h);
-    _mm_storel_epi64((__m128i *)(dst + x), pack);
-  }
+// Only calculate odd columns since even columns are just src pixels' copies.
+static void scale_1_to_2_phase_0_row(const uint8_t *src, uint8_t *dst,
+                                     const int w, const __m128i *const f) {
+  int x = w;
+
+  do {
+    __m128i s[8], temp;
+    s[0] = _mm_loadl_epi64((const __m128i *)(src + 0));
+    s[1] = _mm_loadl_epi64((const __m128i *)(src + 1));
+    s[2] = _mm_loadl_epi64((const __m128i *)(src + 2));
+    s[3] = _mm_loadl_epi64((const __m128i *)(src + 3));
+    s[4] = _mm_loadl_epi64((const __m128i *)(src + 4));
+    s[5] = _mm_loadl_epi64((const __m128i *)(src + 5));
+    s[6] = _mm_loadl_epi64((const __m128i *)(src + 6));
+    s[7] = _mm_loadl_epi64((const __m128i *)(src + 7));
+    temp = scale_1_to_2_phase_0_kernel(s, f);
+    _mm_storel_epi64((__m128i *)dst, temp);
+    src += 8;
+    dst += 8;
+    x -= 8;
+  } while (x);
 }
 
-static void upsample_1_to_2_ssse3(const uint8_t *src, ptrdiff_t src_stride,
-                                  uint8_t *dst, ptrdiff_t dst_stride, int dst_w,
-                                  int dst_h) {
-  dst_w /= 2;
-  dst_h /= 2;
-  {
-    DECLARE_ALIGNED(16, uint8_t, tmp[1920 * 8]);
-    uint8_t *tmp0 = tmp + dst_w * 0;
-    uint8_t *tmp1 = tmp + dst_w * 1;
-    uint8_t *tmp2 = tmp + dst_w * 2;
-    uint8_t *tmp3 = tmp + dst_w * 3;
-    uint8_t *tmp4 = tmp + dst_w * 4;
-    uint8_t *tmp5 = tmp + dst_w * 5;
-    uint8_t *tmp6 = tmp + dst_w * 6;
-    uint8_t *tmp7 = tmp + dst_w * 7;
-    uint8_t *tmp8 = NULL;
-    const int max_width = dst_w & ~7;
-    int y;
-    eight_tap_row_ssse3(src - src_stride * 3 - 3, tmp0, dst_w);
-    eight_tap_row_ssse3(src - src_stride * 2 - 3, tmp1, dst_w);
-    eight_tap_row_ssse3(src - src_stride * 1 - 3, tmp2, dst_w);
-    eight_tap_row_ssse3(src + src_stride * 0 - 3, tmp3, dst_w);
-    eight_tap_row_ssse3(src + src_stride * 1 - 3, tmp4, dst_w);
-    eight_tap_row_ssse3(src + src_stride * 2 - 3, tmp5, dst_w);
-    eight_tap_row_ssse3(src + src_stride * 3 - 3, tmp6, dst_w);
-    for (y = 0; y < dst_h; y++) {
-      int x;
-      eight_tap_row_ssse3(src + src_stride * 4 - 3, tmp7, dst_w);
-      for (x = 0; x < max_width; x += 8) {
-        const __m128i A = _mm_loadl_epi64((const __m128i *)(src + x));
-        const __m128i B = _mm_loadl_epi64((const __m128i *)(tmp3 + x));
-        const __m128i AB = _mm_unpacklo_epi8(A, B);
-        __m128i C, D, CD;
-        _mm_storeu_si128((__m128i *)(dst + x * 2), AB);
-        {
-          const __m128i a =
-              _mm_loadl_epi64((const __m128i *)(src + x - src_stride * 3));
-          const __m128i b =
-              _mm_loadl_epi64((const __m128i *)(src + x - src_stride * 2));
-          const __m128i c =
-              _mm_loadl_epi64((const __m128i *)(src + x - src_stride * 1));
-          const __m128i d =
-              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 0));
-          const __m128i e =
-              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 1));
-          const __m128i f =
-              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 2));
-          const __m128i g =
-              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 3));
-          const __m128i h =
-              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 4));
-          C = filter(&a, &b, &c, &d, &e, &f, &g, &h);
-        }
-        {
-          const __m128i a = _mm_loadl_epi64((const __m128i *)(tmp0 + x));
-          const __m128i b = _mm_loadl_epi64((const __m128i *)(tmp1 + x));
-          const __m128i c = _mm_loadl_epi64((const __m128i *)(tmp2 + x));
-          const __m128i d = _mm_loadl_epi64((const __m128i *)(tmp3 + x));
-          const __m128i e = _mm_loadl_epi64((const __m128i *)(tmp4 + x));
-          const __m128i f = _mm_loadl_epi64((const __m128i *)(tmp5 + x));
-          const __m128i g = _mm_loadl_epi64((const __m128i *)(tmp6 + x));
-          const __m128i h = _mm_loadl_epi64((const __m128i *)(tmp7 + x));
-          D = filter(&a, &b, &c, &d, &e, &f, &g, &h);
-        }
-        CD = _mm_unpacklo_epi8(C, D);
-        _mm_storeu_si128((__m128i *)(dst + x * 2 + dst_stride), CD);
-      }
-      src += src_stride;
-      dst += dst_stride * 2;
-      tmp8 = tmp0;
-      tmp0 = tmp1;
-      tmp1 = tmp2;
-      tmp2 = tmp3;
-      tmp3 = tmp4;
-      tmp4 = tmp5;
-      tmp5 = tmp6;
-      tmp6 = tmp7;
-      tmp7 = tmp8;
+static void scale_plane_1_to_2_phase_0(const uint8_t *src,
+                                       const ptrdiff_t src_stride, uint8_t *dst,
+                                       const ptrdiff_t dst_stride,
+                                       const int src_w, const int src_h,
+                                       const int16_t *const coef,
+                                       uint8_t *const temp_buffer) {
+  int max_width;
+  int y;
+  uint8_t *tmp[9];
+  __m128i f[4];
+
+  max_width = (src_w + 7) & ~7;
+  tmp[0] = temp_buffer + 0 * max_width;
+  tmp[1] = temp_buffer + 1 * max_width;
+  tmp[2] = temp_buffer + 2 * max_width;
+  tmp[3] = temp_buffer + 3 * max_width;
+  tmp[4] = temp_buffer + 4 * max_width;
+  tmp[5] = temp_buffer + 5 * max_width;
+  tmp[6] = temp_buffer + 6 * max_width;
+  tmp[7] = temp_buffer + 7 * max_width;
+
+  shuffle_filter_ssse3(coef, f);
+
+  scale_1_to_2_phase_0_row(src - 3 * src_stride - 3, tmp[0], max_width, f);
+  scale_1_to_2_phase_0_row(src - 2 * src_stride - 3, tmp[1], max_width, f);
+  scale_1_to_2_phase_0_row(src - 1 * src_stride - 3, tmp[2], max_width, f);
+  scale_1_to_2_phase_0_row(src + 0 * src_stride - 3, tmp[3], max_width, f);
+  scale_1_to_2_phase_0_row(src + 1 * src_stride - 3, tmp[4], max_width, f);
+  scale_1_to_2_phase_0_row(src + 2 * src_stride - 3, tmp[5], max_width, f);
+  scale_1_to_2_phase_0_row(src + 3 * src_stride - 3, tmp[6], max_width, f);
+
+  y = src_h;
+  do {
+    int x;
+    scale_1_to_2_phase_0_row(src + 4 * src_stride - 3, tmp[7], max_width, f);
+    for (x = 0; x < max_width; x += 8) {
+      __m128i s[8], C, D, CD;
+
+      // Even rows
+      const __m128i a = _mm_loadl_epi64((const __m128i *)(src + x));
+      const __m128i b = _mm_loadl_epi64((const __m128i *)(tmp[3] + x));
+      const __m128i ab = _mm_unpacklo_epi8(a, b);
+      _mm_storeu_si128((__m128i *)(dst + 2 * x), ab);
+
+      // Odd rows
+      // Even columns
+      load_8bit_8x8(src + x - 3 * src_stride, src_stride, s);
+      C = scale_1_to_2_phase_0_kernel(s, f);
+
+      // Odd columns
+      s[0] = _mm_loadl_epi64((const __m128i *)(tmp[0] + x));
+      s[1] = _mm_loadl_epi64((const __m128i *)(tmp[1] + x));
+      s[2] = _mm_loadl_epi64((const __m128i *)(tmp[2] + x));
+      s[3] = _mm_loadl_epi64((const __m128i *)(tmp[3] + x));
+      s[4] = _mm_loadl_epi64((const __m128i *)(tmp[4] + x));
+      s[5] = _mm_loadl_epi64((const __m128i *)(tmp[5] + x));
+      s[6] = _mm_loadl_epi64((const __m128i *)(tmp[6] + x));
+      s[7] = _mm_loadl_epi64((const __m128i *)(tmp[7] + x));
+      D = scale_1_to_2_phase_0_kernel(s, f);
+
+      CD = _mm_unpacklo_epi8(C, D);
+      _mm_storeu_si128((__m128i *)(dst + dst_stride + 2 * x), CD);
     }
-  }
+
+    src += src_stride;
+    dst += 2 * dst_stride;
+    tmp[8] = tmp[0];
+    tmp[0] = tmp[1];
+    tmp[1] = tmp[2];
+    tmp[2] = tmp[3];
+    tmp[3] = tmp[4];
+    tmp[4] = tmp[5];
+    tmp[5] = tmp[6];
+    tmp[6] = tmp[7];
+    tmp[7] = tmp[8];
+  } while (--y);
 }
 
 void vp9_scale_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src,
@@ -174,33 +163,43 @@ void vp9_scale_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src,
   const int src_h = src->y_crop_height;
   const int dst_w = dst->y_crop_width;
   const int dst_h = dst->y_crop_height;
-  const int dst_uv_w = dst_w / 2;
-  const int dst_uv_h = dst_h / 2;
+  int scaled = 0;
 
   if (dst_w * 2 == src_w && dst_h * 2 == src_h && phase_scaler == 0) {
-    downsample_2_to_1_ssse3(src->y_buffer, src->y_stride, dst->y_buffer,
-                            dst->y_stride, dst_w, dst_h);
-    downsample_2_to_1_ssse3(src->u_buffer, src->uv_stride, dst->u_buffer,
-                            dst->uv_stride, dst_uv_w, dst_uv_h);
-    downsample_2_to_1_ssse3(src->v_buffer, src->uv_stride, dst->v_buffer,
-                            dst->uv_stride, dst_uv_w, dst_uv_h);
-    vpx_extend_frame_borders(dst);
-  } else if (dst_w == src_w * 2 && dst_h == src_h * 2 && filter_type == 0 &&
-             phase_scaler == 0) {
-    // The upsample() supports widths up to 1920 * 2.  If greater, fall back
-    // to vp9_scale_and_extend_frame_c().
-    if (dst_w / 2 <= 1920) {
-      upsample_1_to_2_ssse3(src->y_buffer, src->y_stride, dst->y_buffer,
-                            dst->y_stride, dst_w, dst_h);
-      upsample_1_to_2_ssse3(src->u_buffer, src->uv_stride, dst->u_buffer,
-                            dst->uv_stride, dst_uv_w, dst_uv_h);
-      upsample_1_to_2_ssse3(src->v_buffer, src->uv_stride, dst->v_buffer,
-                            dst->uv_stride, dst_uv_w, dst_uv_h);
-      vpx_extend_frame_borders(dst);
-    } else {
-      vp9_scale_and_extend_frame_c(src, dst, filter_type, phase_scaler);
+    // 2 to 1
+    const int dst_uv_w = dst_w / 2;
+    const int dst_uv_h = dst_h / 2;
+    scaled = 1;
+    scale_plane_2_to_1_phase_0(src->y_buffer, src->y_stride, dst->y_buffer,
+                               dst->y_stride, dst_w, dst_h);
+    scale_plane_2_to_1_phase_0(src->u_buffer, src->uv_stride, dst->u_buffer,
+                               dst->uv_stride, dst_uv_w, dst_uv_h);
+    scale_plane_2_to_1_phase_0(src->v_buffer, src->uv_stride, dst->v_buffer,
+                               dst->uv_stride, dst_uv_w, dst_uv_h);
+  } else if (dst_w == src_w * 2 && dst_h == src_h * 2 && phase_scaler == 0) {
+    // 1 to 2
+    uint8_t *const temp_buffer = (uint8_t *)malloc(8 * ((src_w + 7) & ~7));
+    if (temp_buffer) {
+      scaled = 1;
+      scale_plane_1_to_2_phase_0(
+          src->y_buffer, src->y_stride, dst->y_buffer, dst->y_stride, src_w,
+          src_h, vp9_filter_kernels[filter_type][8], temp_buffer);
+      scale_plane_1_to_2_phase_0(src->u_buffer, src->uv_stride, dst->u_buffer,
+                                 dst->uv_stride, src_w / 2, src_h / 2,
+                                 vp9_filter_kernels[filter_type][8],
+                                 temp_buffer);
+      scale_plane_1_to_2_phase_0(src->v_buffer, src->uv_stride, dst->v_buffer,
+                                 dst->uv_stride, src_w / 2, src_h / 2,
+                                 vp9_filter_kernels[filter_type][8],
+                                 temp_buffer);
+      free(temp_buffer);
     }
+  }
+
+  if (scaled) {
+    vpx_extend_frame_borders(dst);
   } else {
+    // Call c version for all other scaling ratios.
     vp9_scale_and_extend_frame_c(src, dst, filter_type, phase_scaler);
   }
 }