Add AVX2 version of vpx_convolve8_avg.

vpx_convolve8_avg works by first running a normal horizontal filter then a
vertical filter averages at the end.

The added vpx_convolve8_avg_avx2 calls pre-existing AVX2 code for the
horizontal step.

vpx_convolve8_avg_vert_avx2 is also added, but only uses ssse3 code.

Change-Id: If5160c0c8e778e10de61ee9bf42ee4be5975c983

test/convolve_test.cc

@@ -603,6 +603,29 @@ TEST_P(ConvolveTest, DISABLED_Scale_Speed) {
          UUT_->use_highbd_ ? UUT_->use_highbd_ : 8, elapsed_time);
 }
 
+TEST_P(ConvolveTest, DISABLED_8Tap_Avg_Speed) {
+  const uint8_t *const in = input();
+  uint8_t *const out = output();
+  const InterpKernel *const eighttap = vp9_filter_kernels[EIGHTTAP_SHARP];
+  const int kNumTests = 5000000;
+  const int width = Width();
+  const int height = Height();
+  vpx_usec_timer timer;
+
+  SetConstantInput(127);
+
+  vpx_usec_timer_start(&timer);
+  for (int n = 0; n < kNumTests; ++n) {
+    UUT_->hv8_[1](in, kInputStride, out, kOutputStride, eighttap, 8, 16, 8, 16,
+                  width, height);
+  }
+  vpx_usec_timer_mark(&timer);
+
+  const int elapsed_time = static_cast<int>(vpx_usec_timer_elapsed(&timer));
+  printf("convolve8_avg_%dx%d_%d: %d us\n", width, height,
+         UUT_->use_highbd_ ? UUT_->use_highbd_ : 8, elapsed_time);
+}
+
 TEST_P(ConvolveTest, Copy) {
   uint8_t *const in = input();
   uint8_t *const out = output();
@@ -1178,7 +1201,7 @@ INSTANTIATE_TEST_CASE_P(AVX2, ConvolveTest,
 const ConvolveFunctions convolve8_avx2(
     vpx_convolve_copy_c, vpx_convolve_avg_c, vpx_convolve8_horiz_avx2,
     vpx_convolve8_avg_horiz_ssse3, vpx_convolve8_vert_avx2,
-    vpx_convolve8_avg_vert_ssse3, vpx_convolve8_avx2, vpx_convolve8_avg_ssse3,
+    vpx_convolve8_avg_vert_avx2, vpx_convolve8_avx2, vpx_convolve8_avg_avx2,
     vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c, vpx_scaled_vert_c,
     vpx_scaled_avg_vert_c, vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
 const ConvolveParam kArrayConvolve8_avx2[] = { ALL_SIZES(convolve8_avx2) };

vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -364,13 +364,13 @@ add_proto qw/void vpx_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride
 specialize qw/vpx_convolve8_vert sse2 ssse3 avx2 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8_avg sse2 ssse3 neon dspr2 msa vsx/;
+specialize qw/vpx_convolve8_avg sse2 ssse3 avx2 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h";
 specialize qw/vpx_convolve8_avg_horiz sse2 ssse3 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8_avg_vert sse2 ssse3 neon dspr2 msa vsx/;
+specialize qw/vpx_convolve8_avg_vert sse2 ssse3 avx2 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_scaled_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h";
 specialize qw/vpx_scaled_2d ssse3 neon/;

vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c

@@ -539,6 +539,12 @@ filter8_1dfunction vpx_filter_block1d4_h8_ssse3;
 #define vpx_filter_block1d8_h8_avx2 vpx_filter_block1d8_h8_ssse3
 #define vpx_filter_block1d4_h8_avx2 vpx_filter_block1d4_h8_ssse3
 #endif  // ARCH_X86_64
+filter8_1dfunction vpx_filter_block1d16_v8_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d8_v8_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d4_v8_avg_ssse3;
+#define vpx_filter_block1d16_v8_avg_avx2 vpx_filter_block1d16_v8_avg_ssse3
+#define vpx_filter_block1d8_v8_avg_avx2 vpx_filter_block1d8_v8_avg_ssse3
+#define vpx_filter_block1d4_v8_avg_avx2 vpx_filter_block1d4_v8_avg_ssse3
 filter8_1dfunction vpx_filter_block1d16_v2_ssse3;
 filter8_1dfunction vpx_filter_block1d16_h2_ssse3;
 filter8_1dfunction vpx_filter_block1d8_v2_ssse3;
@@ -552,6 +558,12 @@ filter8_1dfunction vpx_filter_block1d4_h2_ssse3;
 #define vpx_filter_block1d8_h2_avx2 vpx_filter_block1d8_h2_ssse3
 #define vpx_filter_block1d4_v2_avx2 vpx_filter_block1d4_v2_ssse3
 #define vpx_filter_block1d4_h2_avx2 vpx_filter_block1d4_h2_ssse3
+filter8_1dfunction vpx_filter_block1d16_v2_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d8_v2_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d4_v2_avg_ssse3;
+#define vpx_filter_block1d16_v2_avg_avx2 vpx_filter_block1d16_v2_avg_ssse3
+#define vpx_filter_block1d8_v2_avg_avx2 vpx_filter_block1d8_v2_avg_ssse3
+#define vpx_filter_block1d4_v2_avg_avx2 vpx_filter_block1d4_v2_avg_ssse3
 // void vpx_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride,
 //                                uint8_t *dst, ptrdiff_t dst_stride,
 //                                const InterpKernel *filter, int x0_q4,
@@ -562,13 +574,25 @@ filter8_1dfunction vpx_filter_block1d4_h2_ssse3;
 //                               const InterpKernel *filter, int x0_q4,
 //                               int32_t x_step_q4, int y0_q4, int y_step_q4,
 //                               int w, int h);
+// void vpx_convolve8_avg_vert_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                                   uint8_t *dst, ptrdiff_t dst_stride,
+//                                   const InterpKernel *filter, int x0_q4,
+//                                   int32_t x_step_q4, int y0_q4,
+//                                   int y_step_q4, int w, int h);
 FUN_CONV_1D(horiz, x0_q4, x_step_q4, h, src, , avx2);
 FUN_CONV_1D(vert, y0_q4, y_step_q4, v, src - src_stride * 3, , avx2);
+FUN_CONV_1D(avg_vert, y0_q4, y_step_q4, v, src - src_stride * 3, avg_, avx2);
 
 // void vpx_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride,
 //                          uint8_t *dst, ptrdiff_t dst_stride,
 //                          const InterpKernel *filter, int x0_q4,
 //                          int32_t x_step_q4, int y0_q4, int y_step_q4,
 //                          int w, int h);
+// void vpx_convolve8_avg_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                              uint8_t *dst, ptrdiff_t dst_stride,
+//                              const InterpKernel *filter, int x0_q4,
+//                              int32_t x_step_q4, int y0_q4, int y_step_q4,
+//                              int w, int h);
 FUN_CONV_2D(, avx2);
+FUN_CONV_2D(avg_, avx2);
 #endif  // HAVE_AX2 && HAVE_SSSE3