Add AVX2 implementation for motion compensation function

AVX2 Code for av1_convolve_2d_sr_c()

Change-Id: Id8a2192b78bbb2c6ac22da3134a7c256941985c8

av1/common/av1_rtcd_defs.pl

@@ -521,7 +521,7 @@ if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
 add_proto qw/void av1_convolve_2d/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params";
 specialize qw/av1_convolve_2d sse2 avx2/;
 add_proto qw/void av1_convolve_2d_sr/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params";
-specialize qw/av1_convolve_2d_sr c sse2/;
+specialize qw/av1_convolve_2d_sr c sse2 avx2/;
 add_proto qw/void av1_convolve_rounding/, "const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits";
 specialize qw/av1_convolve_rounding avx2/;
 

av1/common/x86/convolve_2d_avx2.c

@@ -225,3 +225,200 @@ void av1_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst0,
     }
   }
 }
+
+void av1_convolve_2d_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst,
+                             int dst_stride, int w, int h,
+                             InterpFilterParams *filter_params_x,
+                             InterpFilterParams *filter_params_y,
+                             const int subpel_x_q4, const int subpel_y_q4,
+                             ConvolveParams *conv_params) {
+  const int bd = 8;
+
+  DECLARE_ALIGNED(32, int16_t,
+                  im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]);
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = MAX_SB_SIZE;
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+
+  const __m256i zero = _mm256_setzero_si256();
+  /* Horizontal filter */
+  {
+    const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+        *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+
+    const __m128i coeffs_x8 = _mm_loadu_si128((__m128i *)x_filter);
+    // since not all compilers yet support _mm256_set_m128i()
+    const __m256i coeffs_x = _mm256_insertf128_si256(
+        _mm256_castsi128_si256(coeffs_x8), coeffs_x8, 1);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m256i tmp_0 = _mm256_unpacklo_epi32(coeffs_x, coeffs_x);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m256i tmp_1 = _mm256_unpackhi_epi32(coeffs_x, coeffs_x);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m256i coeff_01 = _mm256_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m256i coeff_23 = _mm256_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m256i coeff_45 = _mm256_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m256i coeff_67 = _mm256_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m256i round_const = _mm256_set1_epi32(
+        ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
+    const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0);
+
+    for (i = 0; i < im_h; ++i) {
+      for (j = 0; j < w; j += 16) {
+        const __m256i data = _mm256_permute4x64_epi64(
+            _mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]),
+            _MM_SHUFFLE(2, 1, 1, 0));
+
+        const __m256i src_lo = _mm256_unpacklo_epi8(data, zero);
+        const __m256i src_hi = _mm256_unpackhi_epi8(data, zero);
+
+        // Filter even-index pixels
+        const __m256i res_0 = _mm256_madd_epi16(src_lo, coeff_01);
+        const __m256i src_2 = _mm256_alignr_epi8(src_hi, src_lo, 4);
+        const __m256i res_2 = _mm256_madd_epi16(src_2, coeff_23);
+        const __m256i src_4 = _mm256_alignr_epi8(src_hi, src_lo, 8);
+        const __m256i res_4 = _mm256_madd_epi16(src_4, coeff_45);
+        const __m256i src_6 = _mm256_alignr_epi8(src_hi, src_lo, 12);
+        const __m256i res_6 = _mm256_madd_epi16(src_6, coeff_67);
+
+        __m256i res_even = _mm256_add_epi32(_mm256_add_epi32(res_0, res_4),
+                                            _mm256_add_epi32(res_2, res_6));
+        res_even = _mm256_sra_epi32(_mm256_add_epi32(res_even, round_const),
+                                    round_shift);
+
+        // Filter odd-index pixels
+        const __m256i src_1 = _mm256_alignr_epi8(src_hi, src_lo, 2);
+        const __m256i res_1 = _mm256_madd_epi16(src_1, coeff_01);
+        const __m256i src_3 = _mm256_alignr_epi8(src_hi, src_lo, 6);
+        const __m256i res_3 = _mm256_madd_epi16(src_3, coeff_23);
+        const __m256i src_5 = _mm256_alignr_epi8(src_hi, src_lo, 10);
+        const __m256i res_5 = _mm256_madd_epi16(src_5, coeff_45);
+        const __m256i src_7 = _mm256_alignr_epi8(src_hi, src_lo, 14);
+        const __m256i res_7 = _mm256_madd_epi16(src_7, coeff_67);
+
+        __m256i res_odd = _mm256_add_epi32(_mm256_add_epi32(res_1, res_5),
+                                           _mm256_add_epi32(res_3, res_7));
+        res_odd = _mm256_sra_epi32(_mm256_add_epi32(res_odd, round_const),
+                                   round_shift);
+
+        __m256i res = _mm256_packs_epi32(res_even, res_odd);
+        _mm256_storeu_si256((__m256i *)&im_block[i * im_stride + j], res);
+      }
+    }
+  }
+
+  /* Vertical filter */
+  {
+    const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+        *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+
+    const __m128i coeffs_y8 = _mm_loadu_si128((__m128i *)y_filter);
+    const __m256i coeffs_y = _mm256_insertf128_si256(
+        _mm256_castsi128_si256(coeffs_y8), coeffs_y8, 1);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m256i tmp_0 = _mm256_unpacklo_epi32(coeffs_y, coeffs_y);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m256i tmp_1 = _mm256_unpackhi_epi32(coeffs_y, coeffs_y);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m256i coeff_01 = _mm256_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m256i coeff_23 = _mm256_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m256i coeff_45 = _mm256_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m256i coeff_67 = _mm256_unpackhi_epi64(tmp_1, tmp_1);
+
+    const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+    const int bits =
+        FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+    const __m256i round_const = _mm256_set1_epi32(
+        ((((1 << bits) + 1) << conv_params->round_1) - (1 << offset_bits)) >>
+        1);
+    const __m128i round_shift = _mm_cvtsi32_si128(bits + conv_params->round_1);
+
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 16) {
+        // Filter even-index pixels
+        const int16_t *data = &im_block[i * im_stride + j];
+        const __m256i src_0 =
+            _mm256_unpacklo_epi16(*(__m256i *)(data + 0 * im_stride),
+                                  *(__m256i *)(data + 1 * im_stride));
+        const __m256i src_2 =
+            _mm256_unpacklo_epi16(*(__m256i *)(data + 2 * im_stride),
+                                  *(__m256i *)(data + 3 * im_stride));
+        const __m256i src_4 =
+            _mm256_unpacklo_epi16(*(__m256i *)(data + 4 * im_stride),
+                                  *(__m256i *)(data + 5 * im_stride));
+        const __m256i src_6 =
+            _mm256_unpacklo_epi16(*(__m256i *)(data + 6 * im_stride),
+                                  *(__m256i *)(data + 7 * im_stride));
+
+        const __m256i res_0 = _mm256_madd_epi16(src_0, coeff_01);
+        const __m256i res_2 = _mm256_madd_epi16(src_2, coeff_23);
+        const __m256i res_4 = _mm256_madd_epi16(src_4, coeff_45);
+        const __m256i res_6 = _mm256_madd_epi16(src_6, coeff_67);
+
+        const __m256i res_even = _mm256_add_epi32(
+            _mm256_add_epi32(res_0, res_2), _mm256_add_epi32(res_4, res_6));
+
+        // Filter odd-index pixels
+        const __m256i src_1 =
+            _mm256_unpackhi_epi16(*(__m256i *)(data + 0 * im_stride),
+                                  *(__m256i *)(data + 1 * im_stride));
+        const __m256i src_3 =
+            _mm256_unpackhi_epi16(*(__m256i *)(data + 2 * im_stride),
+                                  *(__m256i *)(data + 3 * im_stride));
+        const __m256i src_5 =
+            _mm256_unpackhi_epi16(*(__m256i *)(data + 4 * im_stride),
+                                  *(__m256i *)(data + 5 * im_stride));
+        const __m256i src_7 =
+            _mm256_unpackhi_epi16(*(__m256i *)(data + 6 * im_stride),
+                                  *(__m256i *)(data + 7 * im_stride));
+
+        const __m256i res_1 = _mm256_madd_epi16(src_1, coeff_01);
+        const __m256i res_3 = _mm256_madd_epi16(src_3, coeff_23);
+        const __m256i res_5 = _mm256_madd_epi16(src_5, coeff_45);
+        const __m256i res_7 = _mm256_madd_epi16(src_7, coeff_67);
+
+        const __m256i res_odd = _mm256_add_epi32(
+            _mm256_add_epi32(res_1, res_3), _mm256_add_epi32(res_5, res_7));
+
+        // Rearrange pixels back into the order 0 ... 7
+        const __m256i res_lo = _mm256_unpacklo_epi32(res_even, res_odd);
+        const __m256i res_hi = _mm256_unpackhi_epi32(res_even, res_odd);
+
+        const __m256i res_lo_round = _mm256_sra_epi32(
+            _mm256_add_epi32(res_lo, round_const), round_shift);
+        const __m256i res_hi_round = _mm256_sra_epi32(
+            _mm256_add_epi32(res_hi, round_const), round_shift);
+
+        const __m256i res16 = _mm256_packs_epi32(res_lo_round, res_hi_round);
+        const __m256i res = _mm256_packus_epi16(res16, res16);
+
+        __m128i *const p = (__m128i *)&dst[i * dst_stride + j];
+
+        if (w == 2) {
+          *(uint16_t *)p = _mm_cvtsi128_si32(_mm256_castsi256_si128(res));
+        } else if (w == 4) {
+          *(uint32_t *)p = _mm_cvtsi128_si32(_mm256_castsi256_si128(res));
+        } else if (w == 8) {
+          _mm_storel_epi64(p, _mm256_castsi256_si128(res));
+        } else {
+          *p = _mm_unpacklo_epi64(_mm256_castsi256_si128(res),
+                                  _mm256_extracti128_si256(res, 1));
+        }
+      }
+    }
+  }
+}

test/av1_convolve_2d_test.cc

@@ -93,6 +93,12 @@ INSTANTIATE_TEST_CASE_P(
     SSE2, AV1Convolve2DSrTest,
     libaom_test::AV1Convolve2D::BuildParams(av1_convolve_2d_sr_sse2, 1, 1, 0));
 
+#if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(
+    AVX2, AV1Convolve2DSrTest,
+    libaom_test::AV1Convolve2D::BuildParams(av1_convolve_2d_sr_avx2, 1, 1, 0));
+#endif
+
 #if CONFIG_JNT_COMP && HAVE_SSE4_1
 TEST_P(AV1JntConvolve2DTest, CheckOutput) { RunCheckOutput(GET_PARAM(0)); }