AVX2 implementation of av1_highbd_convolve_2d_sr_c AVX2 implementation of av1_highbd_convolve_2d_sr_c have been added. Code cleanup of AVX2 convolve functions. BUG FIX: Intermediate buffer of av1_highbd_convolve_2d_sr_c has been changed from int32_t to int16_t. Change-Id: I71667faf19443b2c4748fc7079810371d5534698
diff --git a/aom_dsp/x86/convolve_avx2.h b/aom_dsp/x86/convolve_avx2.h index bcc28a7..ec5868e 100644 --- a/aom_dsp/x86/convolve_avx2.h +++ b/aom_dsp/x86/convolve_avx2.h
@@ -33,9 +33,9 @@ 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14 }; -static INLINE void prepare_coeffs(const InterpFilterParams *const filter_params, - const int subpel_q4, - __m256i *const coeffs /* [4] */) { +static INLINE void prepare_coeffs_lowbd( + const InterpFilterParams *const filter_params, const int subpel_q4, + __m256i *const coeffs /* [4] */) { const int16_t *const filter = av1_get_interp_filter_subpel_kernel( *filter_params, subpel_q4 & SUBPEL_MASK); const __m128i coeffs_8 = _mm_loadu_si128((__m128i *)filter); @@ -61,27 +61,27 @@ coeffs[3] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0e0cu)); } -static INLINE void prepare_coeffs_y_2d( - const InterpFilterParams *const filter_params_y, const int subpel_y_q4, - __m256i *const coeffs /* [4] */) { - const int16_t *y_filter = av1_get_interp_filter_subpel_kernel( - *filter_params_y, subpel_y_q4 & SUBPEL_MASK); +static INLINE void prepare_coeffs(const InterpFilterParams *const filter_params, + const int subpel_q4, + __m256i *const coeffs /* [4] */) { + const int16_t *filter = av1_get_interp_filter_subpel_kernel( + *filter_params, subpel_q4 & SUBPEL_MASK); - const __m128i coeffs_y8 = _mm_loadu_si128((__m128i *)y_filter); - const __m256i coeffs_y = _mm256_broadcastsi128_si256(coeffs_y8); + const __m128i coeff_8 = _mm_loadu_si128((__m128i *)filter); + const __m256i coeff = _mm256_broadcastsi128_si256(coeff_8); // coeffs 0 1 0 1 0 1 0 1 - coeffs[0] = _mm256_shuffle_epi32(coeffs_y, 0x00); + coeffs[0] = _mm256_shuffle_epi32(coeff, 0x00); // coeffs 2 3 2 3 2 3 2 3 - coeffs[1] = _mm256_shuffle_epi32(coeffs_y, 0x55); + coeffs[1] = _mm256_shuffle_epi32(coeff, 0x55); // coeffs 4 5 4 5 4 5 4 5 - coeffs[2] = _mm256_shuffle_epi32(coeffs_y, 0xaa); + coeffs[2] = _mm256_shuffle_epi32(coeff, 0xaa); // coeffs 6 7 6 7 6 7 6 7 - coeffs[3] = _mm256_shuffle_epi32(coeffs_y, 0xff); + coeffs[3] = _mm256_shuffle_epi32(coeff, 0xff); } -static INLINE __m256i convolve(const __m256i *const s, - const __m256i *const coeffs) { +static INLINE __m256i convolve_lowbd(const __m256i *const s, + const __m256i *const coeffs) { const __m256i res_01 = _mm256_maddubs_epi16(s[0], coeffs[0]); const __m256i res_23 = _mm256_maddubs_epi16(s[1], coeffs[1]); const __m256i res_45 = _mm256_maddubs_epi16(s[2], coeffs[2]); @@ -94,8 +94,8 @@ return res; } -static INLINE __m256i convolve_y_2d(const __m256i *const s, - const __m256i *const coeffs) { +static INLINE __m256i convolve(const __m256i *const s, + const __m256i *const coeffs) { const __m256i res_0 = _mm256_madd_epi16(s[0], coeffs[0]); const __m256i res_1 = _mm256_madd_epi16(s[1], coeffs[1]); const __m256i res_2 = _mm256_madd_epi16(s[2], coeffs[2]); @@ -107,9 +107,9 @@ return res; } -static INLINE __m256i convolve_x(const __m256i data, - const __m256i *const coeffs, - const __m256i *const filt) { +static INLINE __m256i convolve_lowbd_x(const __m256i data, + const __m256i *const coeffs, + const __m256i *const filt) { __m256i s[4]; s[0] = _mm256_shuffle_epi8(data, filt[0]); @@ -117,7 +117,7 @@ s[2] = _mm256_shuffle_epi8(data, filt[2]); s[3] = _mm256_shuffle_epi8(data, filt[3]); - return convolve(s, coeffs); + return convolve_lowbd(s, coeffs); } static INLINE void add_store_aligned(CONV_BUF_TYPE *const dst,
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl index 139b184..94990ea 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_2d_copy_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"; add_proto qw/void av1_highbd_convolve_2d_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_2d_sr c avx2/; add_proto qw/void av1_highbd_convolve_x/, "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";
diff --git a/av1/common/convolve.c b/av1/common/convolve.c index 2a96da5..fef179c 100644 --- a/av1/common/convolve.c +++ b/av1/common/convolve.c
@@ -1242,7 +1242,7 @@ InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd) { - int32_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]; + 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 = w; const int fo_vert = filter_params_y->taps / 2 - 1; @@ -1267,7 +1267,7 @@ } // vertical filter - int32_t *src_vert = im_block + fo_vert * im_stride; + int16_t *src_vert = im_block + fo_vert * im_stride; const int16_t *y_filter = av1_get_interp_filter_subpel_kernel( *filter_params_y, subpel_y_q4 & SUBPEL_MASK); const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
diff --git a/av1/common/x86/convolve_2d_avx2.c b/av1/common/x86/convolve_2d_avx2.c index 9c1a32b..fafe344 100644 --- a/av1/common/x86/convolve_2d_avx2.c +++ b/av1/common/x86/convolve_2d_avx2.c
@@ -50,8 +50,8 @@ filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2); filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2); - prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_x); - prepare_coeffs_y_2d(filter_params_y, subpel_y_q4, coeffs_y); + prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs_x); + prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y); for (j = 0; j < w; j += 8) { /* Horizontal filter */ @@ -70,7 +70,7 @@ _mm_loadu_si128( (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]), 1); - __m256i res = convolve_x(data, coeffs_x, filt); + __m256i res = convolve_lowbd_x(data, coeffs_x, filt); res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const), round_shift); @@ -112,8 +112,8 @@ s[3] = _mm256_unpacklo_epi16(s6, s7); s[7] = _mm256_unpackhi_epi16(s6, s7); - const __m256i res_a = convolve_y_2d(s, coeffs_y); - const __m256i res_b = convolve_y_2d(s + 4, coeffs_y); + const __m256i res_a = convolve(s, coeffs_y); + const __m256i res_b = convolve(s + 4, coeffs_y); const __m256i res_a_round = _mm256_sra_epi32(_mm256_add_epi32(res_a, round_const), round_shift); @@ -185,8 +185,8 @@ filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2); filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2); - prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_h); - prepare_coeffs_y_2d(filter_params_y, subpel_y_q4, coeffs_v); + prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs_h); + prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_v); const __m256i round_const_h = _mm256_set1_epi16( ((1 << (conv_params->round_0 - 1)) >> 1) + (1 << (bd + FILTER_BITS - 2))); @@ -214,7 +214,7 @@ (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]), 1); - __m256i res = convolve_x(data, coeffs_h, filt); + __m256i res = convolve_lowbd_x(data, coeffs_h, filt); res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); @@ -251,8 +251,8 @@ s[3] = _mm256_unpacklo_epi16(s6, s7); s[7] = _mm256_unpackhi_epi16(s6, s7); - __m256i res_a = convolve_y_2d(s, coeffs_v); - __m256i res_b = convolve_y_2d(s + 4, coeffs_v); + __m256i res_a = convolve(s, coeffs_v); + __m256i res_b = convolve(s + 4, coeffs_v); // Combine V round and 2F-H-V round into a single rounding res_a =
diff --git a/av1/common/x86/convolve_avx2.c b/av1/common/x86/convolve_avx2.c index c4d7447..63f6138 100644 --- a/av1/common/x86/convolve_avx2.c +++ b/av1/common/x86/convolve_avx2.c
@@ -361,7 +361,7 @@ assert((FILTER_BITS - conv_params->round_0) >= 0); - prepare_coeffs(filter_params_y, subpel_y_q4, coeffs); + prepare_coeffs_lowbd(filter_params_y, subpel_y_q4, coeffs); (void)conv_params; (void)filter_params_x; @@ -442,7 +442,7 @@ s[3] = _mm256_unpacklo_epi8(src_67a, src_78a); s[7] = _mm256_unpackhi_epi8(src_67a, src_78a); - const __m256i res_lo = convolve(s, coeffs); + const __m256i res_lo = convolve_lowbd(s, coeffs); const __m256i res_lo_0_32b = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(res_lo)); @@ -465,7 +465,7 @@ &avg_mask); if (w - j > 8) { - const __m256i res_hi = convolve(s + 4, coeffs); + const __m256i res_hi = convolve_lowbd(s + 4, coeffs); const __m256i res_hi_0_32b = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(res_hi)); @@ -520,7 +520,7 @@ assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) || ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS))); - prepare_coeffs(filter_params_y, subpel_y_q4, coeffs); + prepare_coeffs_lowbd(filter_params_y, subpel_y_q4, coeffs); (void)filter_params_x; (void)subpel_x_q4; @@ -599,7 +599,7 @@ s[3] = _mm256_unpacklo_epi8(src_67a, src_78a); s[7] = _mm256_unpackhi_epi8(src_67a, src_78a); - const __m256i res_lo = convolve(s, coeffs); + const __m256i res_lo = convolve_lowbd(s, coeffs); /* rounding code */ // shift by F - 1 @@ -609,7 +609,7 @@ __m256i res_8b_lo = _mm256_packus_epi16(res_16b_lo, res_16b_lo); if (w - j > 8) { - const __m256i res_hi = convolve(s + 4, coeffs); + const __m256i res_hi = convolve_lowbd(s + 4, coeffs); /* rounding code */ // shift by F - 1 @@ -679,7 +679,7 @@ filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2); filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2); - prepare_coeffs(filter_params_x, subpel_x_q4, coeffs); + prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs); const __m256i round_const = _mm256_set1_epi16((1 << (conv_params->round_0 - 1)) >> 1); @@ -698,7 +698,7 @@ _mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]), _MM_SHUFFLE(2, 1, 1, 0)); - __m256i res = convolve_x(data, coeffs, filt); + __m256i res = convolve_lowbd_x(data, coeffs, filt); res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const), round_shift); @@ -738,7 +738,7 @@ filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2); filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2); - prepare_coeffs(filter_params_x, subpel_x_q4, coeffs); + prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs); const __m256i round_0_const = _mm256_set1_epi16((1 << (conv_params->round_0 - 1)) >> 1); @@ -762,7 +762,7 @@ _mm256_loadu_si256((__m256i *)&src_ptr[(i * src_stride) + j]), _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + (j + 8)]), 1); - __m256i res_16b = convolve_x(data, coeffs, filt); + __m256i res_16b = convolve_lowbd_x(data, coeffs, filt); res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_0_const), round_0_shift);
diff --git a/av1/common/x86/highbd_convolve_2d_avx2.c b/av1/common/x86/highbd_convolve_2d_avx2.c index 9ac8e9e..cb13e9b 100644 --- a/av1/common/x86/highbd_convolve_2d_avx2.c +++ b/av1/common/x86/highbd_convolve_2d_avx2.c
@@ -14,10 +14,169 @@ #include "./aom_dsp_rtcd.h" #include "aom_dsp/aom_convolve.h" +#include "aom_dsp/x86/convolve_avx2.h" +#include "aom_dsp/x86/synonyms.h" #include "aom_dsp/aom_dsp_common.h" #include "aom_dsp/aom_filter.h" #include "av1/common/convolve.h" +void av1_highbd_convolve_2d_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) { + DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]); + int im_h = h + filter_params_y->taps - 1; + int im_stride = 8; + int i, j; + const int fo_vert = filter_params_y->taps / 2 - 1; + const int fo_horiz = filter_params_x->taps / 2 - 1; + const uint16_t *const src_ptr = src - fo_vert * src_stride - fo_horiz; + + // Check that, even with 12-bit input, the intermediate values will fit + // into an unsigned 16-bit intermediate array. + assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16); + + __m256i s[8], coeffs_y[4], coeffs_x[4]; + + const __m256i round_const_x = _mm256_set1_epi32( + ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1))); + const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0); + + const __m256i round_const_y = _mm256_set1_epi32( + ((1 << conv_params->round_1) >> 1) - + (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1))); + const __m128i round_shift_y = _mm_cvtsi32_si128(conv_params->round_1); + + const __m128i bits = _mm_cvtsi32_si128( + FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_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_x, subpel_x_q4, coeffs_x); + prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y); + + for (j = 0; j < w; j += 8) { + /* Horizontal filter */ + { + for (i = 0; i < im_h; i += 2) { + const __m256i row0 = + _mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]); + __m256i row1 = _mm256_set1_epi16(0); + if (i + 1 < im_h) + row1 = + _mm256_loadu_si256((__m256i *)&src_ptr[(i + 1) * src_stride + j]); + + const __m256i r0 = _mm256_permute2x128_si256(row0, row1, 0x20); + const __m256i r1 = _mm256_permute2x128_si256(row0, row1, 0x31); + + // even pixels + s[0] = _mm256_alignr_epi8(r1, r0, 0); + s[1] = _mm256_alignr_epi8(r1, r0, 4); + s[2] = _mm256_alignr_epi8(r1, r0, 8); + s[3] = _mm256_alignr_epi8(r1, r0, 12); + + __m256i res_even = convolve(s, coeffs_x); + res_even = _mm256_sra_epi32(_mm256_add_epi32(res_even, round_const_x), + round_shift_x); + + // odd pixels + s[0] = _mm256_alignr_epi8(r1, r0, 2); + s[1] = _mm256_alignr_epi8(r1, r0, 6); + s[2] = _mm256_alignr_epi8(r1, r0, 10); + s[3] = _mm256_alignr_epi8(r1, r0, 14); + + __m256i res_odd = convolve(s, coeffs_x); + res_odd = _mm256_sra_epi32(_mm256_add_epi32(res_odd, round_const_x), + round_shift_x); + + __m256i res_even1 = _mm256_packs_epi32(res_even, res_even); + __m256i res_odd1 = _mm256_packs_epi32(res_odd, res_odd); + __m256i res = _mm256_unpacklo_epi16(res_even1, res_odd1); + + _mm256_store_si256((__m256i *)&im_block[i * im_stride], res); + } + } + + /* Vertical filter */ + { + __m256i s0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); + __m256i s1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); + __m256i s2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride)); + __m256i s3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride)); + __m256i s4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride)); + __m256i s5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride)); + + s[0] = _mm256_unpacklo_epi16(s0, s1); + s[1] = _mm256_unpacklo_epi16(s2, s3); + s[2] = _mm256_unpacklo_epi16(s4, s5); + + s[4] = _mm256_unpackhi_epi16(s0, s1); + s[5] = _mm256_unpackhi_epi16(s2, s3); + s[6] = _mm256_unpackhi_epi16(s4, s5); + + for (i = 0; i < h; i += 2) { + const int16_t *data = &im_block[i * im_stride]; + + const __m256i s6 = + _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); + const __m256i s7 = + _mm256_loadu_si256((__m256i *)(data + 7 * im_stride)); + + s[3] = _mm256_unpacklo_epi16(s6, s7); + s[7] = _mm256_unpackhi_epi16(s6, s7); + + const __m256i res_a = convolve(s, coeffs_y); + __m256i res_a_round = _mm256_sra_epi32( + _mm256_add_epi32(res_a, round_const_y), round_shift_y); + res_a_round = _mm256_sra_epi32(res_a_round, bits); + res_a_round = _mm256_min_epi16(res_a_round, clip_pixel); + res_a_round = _mm256_max_epi16(res_a_round, zero); + + 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_y), round_shift_y); + res_b_round = _mm256_sra_epi32(res_b_round, bits); + res_b_round = _mm256_min_epi16(res_b_round, clip_pixel); + res_b_round = _mm256_max_epi16(res_b_round, zero); + + const __m256i res_16bit = + _mm256_packs_epi32(res_a_round, res_b_round); + + _mm_store_si128((__m128i *)&dst[i * dst_stride + j], + _mm256_castsi256_si128(res_16bit)); + _mm_store_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); + _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); + 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_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, int h, InterpFilterParams *filter_params_x,
diff --git a/test/av1_convolve_2d_test.cc b/test/av1_convolve_2d_test.cc index b8b5971..9d04418 100644 --- a/test/av1_convolve_2d_test.cc +++ b/test/av1_convolve_2d_test.cc
@@ -169,7 +169,11 @@ INSTANTIATE_TEST_CASE_P(C_COPY, AV1HighbdConvolve2DSrTest, libaom_test::AV1HighbdConvolve2D::BuildParams( av1_highbd_convolve_2d_copy_sr_c, 0, 0, 0)); - +#if HAVE_AVX2 +INSTANTIATE_TEST_CASE_P(AVX2, AV1HighbdConvolve2DSrTest, + libaom_test::AV1HighbdConvolve2D::BuildParams( + av1_highbd_convolve_2d_sr_avx2, 1, 1, 0)); +#endif #if CONFIG_JNT_COMP && HAVE_SSE4_1 TEST_P(AV1HighbdJntConvolve2DTest, CheckOutput) { RunCheckOutput(GET_PARAM(1));