AVX2 implementation of av1_highbd_convolve_y_sr_c
AVX2 implementation of av1_highbd_convolve_y_sr_c has been added.
Code cleanup.
Change-Id: I1ed05ba539b0abe3ca0c73657572b63af963abc9
diff --git a/aom_dsp/x86/highbd_convolve_avx2.c b/aom_dsp/x86/highbd_convolve_avx2.c
index 2f6a820..a19af49 100644
--- a/aom_dsp/x86/highbd_convolve_avx2.c
+++ b/aom_dsp/x86/highbd_convolve_avx2.c
@@ -194,6 +194,155 @@
}
}
+void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride,
+ uint16_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, int bd) {
+ int i, j;
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const uint16_t *const src_ptr = src - fo_vert * src_stride;
+ (void)filter_params_x;
+ (void)subpel_x_q4;
+ (void)conv_params;
+
+ assert(conv_params->round_0 <= FILTER_BITS);
+ assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) ||
+ ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS)));
+
+ __m256i s[8], coeffs_y[4];
+
+ const int bits = FILTER_BITS;
+
+ const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+ const __m256i round_const_bits = _mm256_set1_epi32((1 << bits) >> 1);
+ const __m256i clip_pixel =
+ _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+ const __m256i zero = _mm256_setzero_si256();
+
+ prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y);
+
+ for (j = 0; j < w; j += 8) {
+ const uint16_t *data = &src_ptr[j];
+ /* Vertical filter */
+ {
+ __m256i src6;
+ __m256i s01 = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 0 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
+ 0x20);
+ __m256i s12 = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
+ 0x20);
+ __m256i s23 = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
+ 0x20);
+ __m256i s34 = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
+ 0x20);
+ __m256i s45 = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
+ 0x20);
+ src6 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 6 * src_stride)));
+ __m256i s56 = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
+ src6, 0x20);
+
+ s[0] = _mm256_unpacklo_epi16(s01, s12);
+ s[1] = _mm256_unpacklo_epi16(s23, s34);
+ s[2] = _mm256_unpacklo_epi16(s45, s56);
+
+ s[4] = _mm256_unpackhi_epi16(s01, s12);
+ s[5] = _mm256_unpackhi_epi16(s23, s34);
+ s[6] = _mm256_unpackhi_epi16(s45, s56);
+
+ for (i = 0; i < h; i += 2) {
+ data = &src_ptr[i * src_stride + j];
+
+ const __m256i s67 = _mm256_permute2x128_si256(
+ src6,
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
+ 0x20);
+
+ src6 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 8 * src_stride)));
+
+ const __m256i s78 = _mm256_permute2x128_si256(
+ _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
+ src6, 0x20);
+
+ s[3] = _mm256_unpacklo_epi16(s67, s78);
+ s[7] = _mm256_unpackhi_epi16(s67, s78);
+
+ const __m256i res_a = convolve(s, coeffs_y);
+
+ __m256i res_a_round = _mm256_sra_epi32(
+ _mm256_add_epi32(res_a, round_const_bits), round_shift_bits);
+
+ if (w - j > 4) {
+ const __m256i res_b = convolve(s + 4, coeffs_y);
+ __m256i res_b_round = _mm256_sra_epi32(
+ _mm256_add_epi32(res_b, round_const_bits), round_shift_bits);
+
+ __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
+ res_16bit = _mm256_min_epi16(res_16bit, clip_pixel);
+ res_16bit = _mm256_max_epi16(res_16bit, zero);
+
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j],
+ _mm256_castsi256_si128(res_16bit));
+ _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
+ _mm256_extracti128_si256(res_16bit, 1));
+ } else if (w == 4) {
+ res_a_round = _mm256_packs_epi32(res_a_round, res_a_round);
+ res_a_round = _mm256_min_epi16(res_a_round, clip_pixel);
+ res_a_round = _mm256_max_epi16(res_a_round, zero);
+
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j],
+ _mm256_castsi256_si128(res_a_round));
+ _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+ _mm256_extracti128_si256(res_a_round, 1));
+ } else {
+ res_a_round = _mm256_packs_epi32(res_a_round, res_a_round);
+ res_a_round = _mm256_min_epi16(res_a_round, clip_pixel);
+ res_a_round = _mm256_max_epi16(res_a_round, zero);
+
+ xx_storel_32((__m128i *)&dst[i * dst_stride + j],
+ _mm256_castsi256_si128(res_a_round));
+ xx_storel_32((__m128i *)&dst[i * dst_stride + j + dst_stride],
+ _mm256_extracti128_si256(res_a_round, 1));
+ }
+
+ s[0] = s[1];
+ s[1] = s[2];
+ s[2] = s[3];
+
+ s[4] = s[5];
+ s[5] = s[6];
+ s[6] = s[7];
+ }
+ }
+ }
+}
+
void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride,
uint16_t *dst, int dst_stride, int w, int h,
InterpFilterParams *filter_params_x,
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl
index 9e8102c..508f1e4 100755
--- a/av1/common/av1_rtcd_defs.pl
+++ b/av1/common/av1_rtcd_defs.pl
@@ -519,6 +519,7 @@
add_proto qw/void av1_highbd_convolve_y/, "const uint16_t *src, int src_stride, uint16_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, int bd";
add_proto qw/void av1_highbd_convolve_y_sr/, "const uint16_t *src, int src_stride, uint16_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, int bd";
+specialize qw/av1_highbd_convolve_y_sr c avx2/;
add_proto qw/void av1_highbd_convolve_rounding/, "const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits, int bd";
specialize qw/av1_highbd_convolve_rounding avx2/;
diff --git a/av1/common/convolve.c b/av1/common/convolve.c
index 27462ac..e96aa44 100644
--- a/av1/common/convolve.c
+++ b/av1/common/convolve.c
@@ -1230,7 +1230,6 @@
const int bits = FILTER_BITS - conv_params->round_0;
(void)filter_params_y;
(void)subpel_y_q4;
- (void)conv_params;
assert(bits >= 0);
assert((FILTER_BITS - conv_params->round_1) >= 0 ||
@@ -1262,7 +1261,6 @@
(void)filter_params_x;
(void)subpel_x_q4;
(void)conv_params;
- (void)bd;
assert(conv_params->round_0 <= FILTER_BITS);
assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) ||
diff --git a/test/av1_convolve_2d_test.cc b/test/av1_convolve_2d_test.cc
index fc03c16..ed59a6a 100644
--- a/test/av1_convolve_2d_test.cc
+++ b/test/av1_convolve_2d_test.cc
@@ -176,6 +176,9 @@
INSTANTIATE_TEST_CASE_P(AVX2_X, AV1HighbdConvolve2DSrTest,
libaom_test::AV1HighbdConvolve2D::BuildParams(
av1_highbd_convolve_x_sr_avx2, 1, 0, 0));
+INSTANTIATE_TEST_CASE_P(AVX2_Y, AV1HighbdConvolve2DSrTest,
+ libaom_test::AV1HighbdConvolve2D::BuildParams(
+ av1_highbd_convolve_y_sr_avx2, 0, 1, 0));
#endif
#if CONFIG_JNT_COMP && HAVE_SSE4_1
TEST_P(AV1HighbdJntConvolve2DTest, CheckOutput) {
