From ec910d81405c736a4490383a250299a7837c2e64 Mon Sep 17 00:00:00 2001 From: trav90 Date: Thu, 18 Oct 2018 21:53:44 -0500 Subject: Update aom to commit id e87fb2378f01103d5d6e477a4ef6892dc714e614 --- .../aom/av1/encoder/x86/av1_quantize_avx2.c | 49 +-- .../aom/av1/encoder/x86/av1_quantize_sse2.c | 74 ++-- third_party/aom/av1/encoder/x86/dct_intrin_sse2.c | 461 +------------------- third_party/aom/av1/encoder/x86/dct_ssse3.c | 469 --------------------- .../aom/av1/encoder/x86/error_intrin_avx2.c | 17 +- .../aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c | 6 +- .../aom/av1/encoder/x86/hybrid_fwd_txfm_avx2.c | 4 +- 7 files changed, 84 insertions(+), 996 deletions(-) delete mode 100644 third_party/aom/av1/encoder/x86/dct_ssse3.c (limited to 'third_party/aom/av1/encoder/x86') diff --git a/third_party/aom/av1/encoder/x86/av1_quantize_avx2.c b/third_party/aom/av1/encoder/x86/av1_quantize_avx2.c index 1c0a120ca..078a67510 100644 --- a/third_party/aom/av1/encoder/x86/av1_quantize_avx2.c +++ b/third_party/aom/av1/encoder/x86/av1_quantize_avx2.c @@ -16,24 +16,24 @@ #include "aom_dsp/aom_dsp_common.h" static INLINE void read_coeff(const tran_low_t *coeff, __m256i *c) { -#if CONFIG_HIGHBITDEPTH - const __m256i x0 = _mm256_loadu_si256((const __m256i *)coeff); - const __m256i x1 = _mm256_loadu_si256((const __m256i *)coeff + 1); - *c = _mm256_packs_epi32(x0, x1); - *c = _mm256_permute4x64_epi64(*c, 0xD8); -#else - *c = _mm256_loadu_si256((const __m256i *)coeff); -#endif + if (sizeof(tran_low_t) == 4) { + const __m256i x0 = _mm256_loadu_si256((const __m256i *)coeff); + const __m256i x1 = _mm256_loadu_si256((const __m256i *)coeff + 1); + *c = _mm256_packs_epi32(x0, x1); + *c = _mm256_permute4x64_epi64(*c, 0xD8); + } else { + *c = _mm256_loadu_si256((const __m256i *)coeff); + } } static INLINE void write_zero(tran_low_t *qcoeff) { const __m256i zero = _mm256_setzero_si256(); -#if CONFIG_HIGHBITDEPTH - _mm256_storeu_si256((__m256i *)qcoeff, zero); - _mm256_storeu_si256((__m256i *)qcoeff + 1, zero); -#else - _mm256_storeu_si256((__m256i *)qcoeff, zero); -#endif + if (sizeof(tran_low_t) == 4) { + _mm256_storeu_si256((__m256i *)qcoeff, zero); + _mm256_storeu_si256((__m256i *)qcoeff + 1, zero); + } else { + _mm256_storeu_si256((__m256i *)qcoeff, zero); + } } static INLINE void init_one_qp(const __m128i *p, __m256i *qp) { @@ -83,19 +83,16 @@ static INLINE void update_qp(int log_scale, __m256i *thr, __m256i *qp) { _mm256_storeu_si256((__m256i *)addr + 1, x1); \ } while (0) -#if CONFIG_HIGHBITDEPTH -#define store_two_quan(q, addr1, dq, addr2) \ - do { \ - store_quan(q, addr1); \ - store_quan(dq, addr2); \ - } while (0) -#else -#define store_two_quan(q, addr1, dq, addr2) \ - do { \ - _mm256_storeu_si256((__m256i *)addr1, q); \ - _mm256_storeu_si256((__m256i *)addr2, dq); \ +#define store_two_quan(q, addr1, dq, addr2) \ + do { \ + if (sizeof(tran_low_t) == 4) { \ + store_quan(q, addr1); \ + store_quan(dq, addr2); \ + } else { \ + _mm256_storeu_si256((__m256i *)addr1, q); \ + _mm256_storeu_si256((__m256i *)addr2, dq); \ + } \ } while (0) -#endif static INLINE void quantize(const __m256i *thr, const __m256i *qp, __m256i *c, const int16_t *iscan_ptr, tran_low_t *qcoeff, diff --git a/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c b/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c index 190317389..4f7c09546 100644 --- a/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c +++ b/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c @@ -18,53 +18,53 @@ static INLINE void read_coeff(const tran_low_t *coeff, intptr_t offset, __m128i *c0, __m128i *c1) { const tran_low_t *addr = coeff + offset; -#if CONFIG_HIGHBITDEPTH - const __m128i x0 = _mm_load_si128((const __m128i *)addr); - const __m128i x1 = _mm_load_si128((const __m128i *)addr + 1); - const __m128i x2 = _mm_load_si128((const __m128i *)addr + 2); - const __m128i x3 = _mm_load_si128((const __m128i *)addr + 3); - *c0 = _mm_packs_epi32(x0, x1); - *c1 = _mm_packs_epi32(x2, x3); -#else - *c0 = _mm_load_si128((const __m128i *)addr); - *c1 = _mm_load_si128((const __m128i *)addr + 1); -#endif + if (sizeof(tran_low_t) == 4) { + const __m128i x0 = _mm_load_si128((const __m128i *)addr); + const __m128i x1 = _mm_load_si128((const __m128i *)addr + 1); + const __m128i x2 = _mm_load_si128((const __m128i *)addr + 2); + const __m128i x3 = _mm_load_si128((const __m128i *)addr + 3); + *c0 = _mm_packs_epi32(x0, x1); + *c1 = _mm_packs_epi32(x2, x3); + } else { + *c0 = _mm_load_si128((const __m128i *)addr); + *c1 = _mm_load_si128((const __m128i *)addr + 1); + } } static INLINE void write_qcoeff(const __m128i *qc0, const __m128i *qc1, tran_low_t *qcoeff, intptr_t offset) { tran_low_t *addr = qcoeff + offset; -#if CONFIG_HIGHBITDEPTH - const __m128i zero = _mm_setzero_si128(); - __m128i sign_bits = _mm_cmplt_epi16(*qc0, zero); - __m128i y0 = _mm_unpacklo_epi16(*qc0, sign_bits); - __m128i y1 = _mm_unpackhi_epi16(*qc0, sign_bits); - _mm_store_si128((__m128i *)addr, y0); - _mm_store_si128((__m128i *)addr + 1, y1); - - sign_bits = _mm_cmplt_epi16(*qc1, zero); - y0 = _mm_unpacklo_epi16(*qc1, sign_bits); - y1 = _mm_unpackhi_epi16(*qc1, sign_bits); - _mm_store_si128((__m128i *)addr + 2, y0); - _mm_store_si128((__m128i *)addr + 3, y1); -#else - _mm_store_si128((__m128i *)addr, *qc0); - _mm_store_si128((__m128i *)addr + 1, *qc1); -#endif + if (sizeof(tran_low_t) == 4) { + const __m128i zero = _mm_setzero_si128(); + __m128i sign_bits = _mm_cmplt_epi16(*qc0, zero); + __m128i y0 = _mm_unpacklo_epi16(*qc0, sign_bits); + __m128i y1 = _mm_unpackhi_epi16(*qc0, sign_bits); + _mm_store_si128((__m128i *)addr, y0); + _mm_store_si128((__m128i *)addr + 1, y1); + + sign_bits = _mm_cmplt_epi16(*qc1, zero); + y0 = _mm_unpacklo_epi16(*qc1, sign_bits); + y1 = _mm_unpackhi_epi16(*qc1, sign_bits); + _mm_store_si128((__m128i *)addr + 2, y0); + _mm_store_si128((__m128i *)addr + 3, y1); + } else { + _mm_store_si128((__m128i *)addr, *qc0); + _mm_store_si128((__m128i *)addr + 1, *qc1); + } } static INLINE void write_zero(tran_low_t *qcoeff, intptr_t offset) { const __m128i zero = _mm_setzero_si128(); tran_low_t *addr = qcoeff + offset; -#if CONFIG_HIGHBITDEPTH - _mm_store_si128((__m128i *)addr, zero); - _mm_store_si128((__m128i *)addr + 1, zero); - _mm_store_si128((__m128i *)addr + 2, zero); - _mm_store_si128((__m128i *)addr + 3, zero); -#else - _mm_store_si128((__m128i *)addr, zero); - _mm_store_si128((__m128i *)addr + 1, zero); -#endif + if (sizeof(tran_low_t) == 4) { + _mm_store_si128((__m128i *)addr, zero); + _mm_store_si128((__m128i *)addr + 1, zero); + _mm_store_si128((__m128i *)addr + 2, zero); + _mm_store_si128((__m128i *)addr + 3, zero); + } else { + _mm_store_si128((__m128i *)addr, zero); + _mm_store_si128((__m128i *)addr + 1, zero); + } } void av1_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, diff --git a/third_party/aom/av1/encoder/x86/dct_intrin_sse2.c b/third_party/aom/av1/encoder/x86/dct_intrin_sse2.c index 496c33395..e5b19a44c 100644 --- a/third_party/aom/av1/encoder/x86/dct_intrin_sse2.c +++ b/third_party/aom/av1/encoder/x86/dct_intrin_sse2.c @@ -205,7 +205,7 @@ static void fidtx4_sse2(__m128i *in) { void av1_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i in[4]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -308,447 +308,6 @@ void av1_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, } } -void av1_fdct8x8_quant_sse2(const int16_t *input, int stride, - int16_t *coeff_ptr, intptr_t n_coeffs, - int skip_block, const int16_t *zbin_ptr, - const int16_t *round_ptr, const int16_t *quant_ptr, - const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, - int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, - uint16_t *eob_ptr, const int16_t *scan_ptr, - const int16_t *iscan_ptr) { - __m128i zero; - int pass; - // Constants - // When we use them, in one case, they are all the same. In all others - // it's a pair of them that we need to repeat four times. This is done - // by constructing the 32 bit constant corresponding to that pair. - const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); - const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); - const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); - const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); - const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); - const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); - const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); - const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); - const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); - // Load input - __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); - __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); - __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); - __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride)); - __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride)); - __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride)); - __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride)); - __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); - __m128i *in[8]; - int index = 0; - - (void)scan_ptr; - (void)zbin_ptr; - (void)quant_shift_ptr; - (void)coeff_ptr; - - // Pre-condition input (shift by two) - in0 = _mm_slli_epi16(in0, 2); - in1 = _mm_slli_epi16(in1, 2); - in2 = _mm_slli_epi16(in2, 2); - in3 = _mm_slli_epi16(in3, 2); - in4 = _mm_slli_epi16(in4, 2); - in5 = _mm_slli_epi16(in5, 2); - in6 = _mm_slli_epi16(in6, 2); - in7 = _mm_slli_epi16(in7, 2); - - in[0] = &in0; - in[1] = &in1; - in[2] = &in2; - in[3] = &in3; - in[4] = &in4; - in[5] = &in5; - in[6] = &in6; - in[7] = &in7; - - // We do two passes, first the columns, then the rows. The results of the - // first pass are transposed so that the same column code can be reused. The - // results of the second pass are also transposed so that the rows (processed - // as columns) are put back in row positions. - for (pass = 0; pass < 2; pass++) { - // To store results of each pass before the transpose. - __m128i res0, res1, res2, res3, res4, res5, res6, res7; - // Add/subtract - const __m128i q0 = _mm_add_epi16(in0, in7); - const __m128i q1 = _mm_add_epi16(in1, in6); - const __m128i q2 = _mm_add_epi16(in2, in5); - const __m128i q3 = _mm_add_epi16(in3, in4); - const __m128i q4 = _mm_sub_epi16(in3, in4); - const __m128i q5 = _mm_sub_epi16(in2, in5); - const __m128i q6 = _mm_sub_epi16(in1, in6); - const __m128i q7 = _mm_sub_epi16(in0, in7); - // Work on first four results - { - // Add/subtract - const __m128i r0 = _mm_add_epi16(q0, q3); - const __m128i r1 = _mm_add_epi16(q1, q2); - const __m128i r2 = _mm_sub_epi16(q1, q2); - const __m128i r3 = _mm_sub_epi16(q0, q3); - // Interleave to do the multiply by constants which gets us into 32bits - const __m128i t0 = _mm_unpacklo_epi16(r0, r1); - const __m128i t1 = _mm_unpackhi_epi16(r0, r1); - const __m128i t2 = _mm_unpacklo_epi16(r2, r3); - const __m128i t3 = _mm_unpackhi_epi16(r2, r3); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); - const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); - const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); - const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); - const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); - const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); - const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); - const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); - const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); - const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); - const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); - // Combine - res0 = _mm_packs_epi32(w0, w1); - res4 = _mm_packs_epi32(w2, w3); - res2 = _mm_packs_epi32(w4, w5); - res6 = _mm_packs_epi32(w6, w7); - } - // Work on next four results - { - // Interleave to do the multiply by constants which gets us into 32bits - const __m128i d0 = _mm_unpacklo_epi16(q6, q5); - const __m128i d1 = _mm_unpackhi_epi16(q6, q5); - const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); - const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); - const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); - const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); - // dct_const_round_shift - const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); - const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); - const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); - const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); - const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); - const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); - const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); - const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); - // Combine - const __m128i r0 = _mm_packs_epi32(s0, s1); - const __m128i r1 = _mm_packs_epi32(s2, s3); - // Add/subtract - const __m128i x0 = _mm_add_epi16(q4, r0); - const __m128i x1 = _mm_sub_epi16(q4, r0); - const __m128i x2 = _mm_sub_epi16(q7, r1); - const __m128i x3 = _mm_add_epi16(q7, r1); - // Interleave to do the multiply by constants which gets us into 32bits - const __m128i t0 = _mm_unpacklo_epi16(x0, x3); - const __m128i t1 = _mm_unpackhi_epi16(x0, x3); - const __m128i t2 = _mm_unpacklo_epi16(x1, x2); - const __m128i t3 = _mm_unpackhi_epi16(x1, x2); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); - const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); - const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); - const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); - const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); - const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); - const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); - const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); - const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); - const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); - const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); - // Combine - res1 = _mm_packs_epi32(w0, w1); - res7 = _mm_packs_epi32(w2, w3); - res5 = _mm_packs_epi32(w4, w5); - res3 = _mm_packs_epi32(w6, w7); - } - // Transpose the 8x8. - { - // 00 01 02 03 04 05 06 07 - // 10 11 12 13 14 15 16 17 - // 20 21 22 23 24 25 26 27 - // 30 31 32 33 34 35 36 37 - // 40 41 42 43 44 45 46 47 - // 50 51 52 53 54 55 56 57 - // 60 61 62 63 64 65 66 67 - // 70 71 72 73 74 75 76 77 - const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); - const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3); - const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1); - const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3); - const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5); - const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7); - const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5); - const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7); - // 00 10 01 11 02 12 03 13 - // 20 30 21 31 22 32 23 33 - // 04 14 05 15 06 16 07 17 - // 24 34 25 35 26 36 27 37 - // 40 50 41 51 42 52 43 53 - // 60 70 61 71 62 72 63 73 - // 54 54 55 55 56 56 57 57 - // 64 74 65 75 66 76 67 77 - const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); - const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); - const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); - const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); - const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); - const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); - const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); - const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); - // 00 10 20 30 01 11 21 31 - // 40 50 60 70 41 51 61 71 - // 02 12 22 32 03 13 23 33 - // 42 52 62 72 43 53 63 73 - // 04 14 24 34 05 15 21 36 - // 44 54 64 74 45 55 61 76 - // 06 16 26 36 07 17 27 37 - // 46 56 66 76 47 57 67 77 - in0 = _mm_unpacklo_epi64(tr1_0, tr1_4); - in1 = _mm_unpackhi_epi64(tr1_0, tr1_4); - in2 = _mm_unpacklo_epi64(tr1_2, tr1_6); - in3 = _mm_unpackhi_epi64(tr1_2, tr1_6); - in4 = _mm_unpacklo_epi64(tr1_1, tr1_5); - in5 = _mm_unpackhi_epi64(tr1_1, tr1_5); - in6 = _mm_unpacklo_epi64(tr1_3, tr1_7); - in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); - // 00 10 20 30 40 50 60 70 - // 01 11 21 31 41 51 61 71 - // 02 12 22 32 42 52 62 72 - // 03 13 23 33 43 53 63 73 - // 04 14 24 34 44 54 64 74 - // 05 15 25 35 45 55 65 75 - // 06 16 26 36 46 56 66 76 - // 07 17 27 37 47 57 67 77 - } - } - // Post-condition output and store it - { - // Post-condition (division by two) - // division of two 16 bits signed numbers using shifts - // n / 2 = (n - (n >> 15)) >> 1 - const __m128i sign_in0 = _mm_srai_epi16(in0, 15); - const __m128i sign_in1 = _mm_srai_epi16(in1, 15); - const __m128i sign_in2 = _mm_srai_epi16(in2, 15); - const __m128i sign_in3 = _mm_srai_epi16(in3, 15); - const __m128i sign_in4 = _mm_srai_epi16(in4, 15); - const __m128i sign_in5 = _mm_srai_epi16(in5, 15); - const __m128i sign_in6 = _mm_srai_epi16(in6, 15); - const __m128i sign_in7 = _mm_srai_epi16(in7, 15); - in0 = _mm_sub_epi16(in0, sign_in0); - in1 = _mm_sub_epi16(in1, sign_in1); - in2 = _mm_sub_epi16(in2, sign_in2); - in3 = _mm_sub_epi16(in3, sign_in3); - in4 = _mm_sub_epi16(in4, sign_in4); - in5 = _mm_sub_epi16(in5, sign_in5); - in6 = _mm_sub_epi16(in6, sign_in6); - in7 = _mm_sub_epi16(in7, sign_in7); - in0 = _mm_srai_epi16(in0, 1); - in1 = _mm_srai_epi16(in1, 1); - in2 = _mm_srai_epi16(in2, 1); - in3 = _mm_srai_epi16(in3, 1); - in4 = _mm_srai_epi16(in4, 1); - in5 = _mm_srai_epi16(in5, 1); - in6 = _mm_srai_epi16(in6, 1); - in7 = _mm_srai_epi16(in7, 1); - } - - iscan_ptr += n_coeffs; - qcoeff_ptr += n_coeffs; - dqcoeff_ptr += n_coeffs; - n_coeffs = -n_coeffs; - zero = _mm_setzero_si128(); - - if (!skip_block) { - __m128i eob; - __m128i round, quant, dequant; - { - __m128i coeff0, coeff1; - - // Setup global values - { - round = _mm_load_si128((const __m128i *)round_ptr); - quant = _mm_load_si128((const __m128i *)quant_ptr); - dequant = _mm_load_si128((const __m128i *)dequant_ptr); - } - - { - __m128i coeff0_sign, coeff1_sign; - __m128i qcoeff0, qcoeff1; - __m128i qtmp0, qtmp1; - // Do DC and first 15 AC - coeff0 = *in[0]; - coeff1 = *in[1]; - - // Poor man's sign extract - coeff0_sign = _mm_srai_epi16(coeff0, 15); - coeff1_sign = _mm_srai_epi16(coeff1, 15); - qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); - qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); - - qcoeff0 = _mm_adds_epi16(qcoeff0, round); - round = _mm_unpackhi_epi64(round, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); - quant = _mm_unpackhi_epi64(quant, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - // Reinsert signs - qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign); - qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); - - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0); - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1); - - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); - dequant = _mm_unpackhi_epi64(dequant, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); - - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0); - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1); - } - - { - // Scan for eob - __m128i zero_coeff0, zero_coeff1; - __m128i nzero_coeff0, nzero_coeff1; - __m128i iscan0, iscan1; - __m128i eob1; - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); - nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); - // Add one to convert from indices to counts - iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0); - iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1); - eob = _mm_and_si128(iscan0, nzero_coeff0); - eob1 = _mm_and_si128(iscan1, nzero_coeff1); - eob = _mm_max_epi16(eob, eob1); - } - n_coeffs += 8 * 2; - } - - // AC only loop - index = 2; - while (n_coeffs < 0) { - __m128i coeff0, coeff1; - { - __m128i coeff0_sign, coeff1_sign; - __m128i qcoeff0, qcoeff1; - __m128i qtmp0, qtmp1; - - assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1); - coeff0 = *in[index]; - coeff1 = *in[index + 1]; - - // Poor man's sign extract - coeff0_sign = _mm_srai_epi16(coeff0, 15); - coeff1_sign = _mm_srai_epi16(coeff1, 15); - qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); - qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); - - qcoeff0 = _mm_adds_epi16(qcoeff0, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - // Reinsert signs - qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign); - qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); - - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0); - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1); - - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); - - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0); - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1); - } - - { - // Scan for eob - __m128i zero_coeff0, zero_coeff1; - __m128i nzero_coeff0, nzero_coeff1; - __m128i iscan0, iscan1; - __m128i eob0, eob1; - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); - nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); - // Add one to convert from indices to counts - iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0); - iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1); - eob0 = _mm_and_si128(iscan0, nzero_coeff0); - eob1 = _mm_and_si128(iscan1, nzero_coeff1); - eob0 = _mm_max_epi16(eob0, eob1); - eob = _mm_max_epi16(eob, eob0); - } - n_coeffs += 8 * 2; - index += 2; - } - - // Accumulate EOB - { - __m128i eob_shuffled; - eob_shuffled = _mm_shuffle_epi32(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); - eob = _mm_max_epi16(eob, eob_shuffled); - *eob_ptr = _mm_extract_epi16(eob, 1); - } - } else { - do { - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero); - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero); - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero); - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero); - n_coeffs += 8 * 2; - } while (n_coeffs < 0); - *eob_ptr = 0; - } -} - // load 8x8 array static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in, int stride, int flipud, int fliplr) { @@ -1307,7 +866,7 @@ static void fidtx8_sse2(__m128i *in) { void av1_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i in[8]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -2344,7 +1903,7 @@ static void fidtx16_sse2(__m128i *in0, __m128i *in1) { void av1_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i in0[16], in1[16]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -2564,7 +2123,7 @@ static INLINE void write_buffer_4x8(tran_low_t *output, __m128i *res) { void av1_fht4x8_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i in[8]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -2742,7 +2301,7 @@ static INLINE void write_buffer_8x4(tran_low_t *output, __m128i *res) { void av1_fht8x4_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i in[8]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -2886,7 +2445,7 @@ static void row_8x16_rounding(__m128i *in, int bits) { void av1_fht8x16_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i in[16]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -3071,7 +2630,7 @@ static INLINE void load_buffer_16x8(const int16_t *input, __m128i *in, void av1_fht16x8_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i in[16]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -3385,7 +2944,7 @@ static INLINE void fhalfright32_16col(__m128i *tl, __m128i *tr, __m128i *bl, void av1_fht16x32_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i intl[16], intr[16], inbl[16], inbr[16]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -3578,7 +3137,7 @@ static INLINE void write_buffer_32x16(tran_low_t *output, __m128i *res0, void av1_fht32x16_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i in0[16], in1[16], in2[16], in3[16]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -3822,7 +3381,7 @@ static INLINE void write_buffer_32x32(__m128i *in0, __m128i *in1, __m128i *in2, void av1_fht32x32_sse2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m128i in0[32], in1[32], in2[32], in3[32]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "No 32x32 sse2 MRC_DCT implementation"); #endif diff --git a/third_party/aom/av1/encoder/x86/dct_ssse3.c b/third_party/aom/av1/encoder/x86/dct_ssse3.c deleted file mode 100644 index 717a99af8..000000000 --- a/third_party/aom/av1/encoder/x86/dct_ssse3.c +++ /dev/null @@ -1,469 +0,0 @@ -/* - * Copyright (c) 2016, Alliance for Open Media. All rights reserved - * - * This source code is subject to the terms of the BSD 2 Clause License and - * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License - * was not distributed with this source code in the LICENSE file, you can - * obtain it at www.aomedia.org/license/software. If the Alliance for Open - * Media Patent License 1.0 was not distributed with this source code in the - * PATENTS file, you can obtain it at www.aomedia.org/license/patent. - */ - -#include -#if defined(_MSC_VER) && _MSC_VER <= 1500 -// Need to include math.h before calling tmmintrin.h/intrin.h -// in certain versions of MSVS. -#include -#endif -#include // SSSE3 - -#include "./av1_rtcd.h" -#include "aom_dsp/x86/inv_txfm_sse2.h" -#include "aom_dsp/x86/txfm_common_sse2.h" - -void av1_fdct8x8_quant_ssse3( - const int16_t *input, int stride, int16_t *coeff_ptr, intptr_t n_coeffs, - int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, - const int16_t *quant_ptr, const int16_t *quant_shift_ptr, - int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, - uint16_t *eob_ptr, const int16_t *scan_ptr, const int16_t *iscan_ptr) { - __m128i zero; - int pass; - // Constants - // When we use them, in one case, they are all the same. In all others - // it's a pair of them that we need to repeat four times. This is done - // by constructing the 32 bit constant corresponding to that pair. - const __m128i k__dual_p16_p16 = dual_set_epi16(23170, 23170); - const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); - const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); - const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); - const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); - const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); - const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); - const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); - const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); - const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); - // Load input - __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); - __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); - __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); - __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride)); - __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride)); - __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride)); - __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride)); - __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); - __m128i *in[8]; - int index = 0; - - (void)scan_ptr; - (void)zbin_ptr; - (void)quant_shift_ptr; - (void)coeff_ptr; - - // Pre-condition input (shift by two) - in0 = _mm_slli_epi16(in0, 2); - in1 = _mm_slli_epi16(in1, 2); - in2 = _mm_slli_epi16(in2, 2); - in3 = _mm_slli_epi16(in3, 2); - in4 = _mm_slli_epi16(in4, 2); - in5 = _mm_slli_epi16(in5, 2); - in6 = _mm_slli_epi16(in6, 2); - in7 = _mm_slli_epi16(in7, 2); - - in[0] = &in0; - in[1] = &in1; - in[2] = &in2; - in[3] = &in3; - in[4] = &in4; - in[5] = &in5; - in[6] = &in6; - in[7] = &in7; - - // We do two passes, first the columns, then the rows. The results of the - // first pass are transposed so that the same column code can be reused. The - // results of the second pass are also transposed so that the rows (processed - // as columns) are put back in row positions. - for (pass = 0; pass < 2; pass++) { - // To store results of each pass before the transpose. - __m128i res0, res1, res2, res3, res4, res5, res6, res7; - // Add/subtract - const __m128i q0 = _mm_add_epi16(in0, in7); - const __m128i q1 = _mm_add_epi16(in1, in6); - const __m128i q2 = _mm_add_epi16(in2, in5); - const __m128i q3 = _mm_add_epi16(in3, in4); - const __m128i q4 = _mm_sub_epi16(in3, in4); - const __m128i q5 = _mm_sub_epi16(in2, in5); - const __m128i q6 = _mm_sub_epi16(in1, in6); - const __m128i q7 = _mm_sub_epi16(in0, in7); - // Work on first four results - { - // Add/subtract - const __m128i r0 = _mm_add_epi16(q0, q3); - const __m128i r1 = _mm_add_epi16(q1, q2); - const __m128i r2 = _mm_sub_epi16(q1, q2); - const __m128i r3 = _mm_sub_epi16(q0, q3); - // Interleave to do the multiply by constants which gets us into 32bits - const __m128i t0 = _mm_unpacklo_epi16(r0, r1); - const __m128i t1 = _mm_unpackhi_epi16(r0, r1); - const __m128i t2 = _mm_unpacklo_epi16(r2, r3); - const __m128i t3 = _mm_unpackhi_epi16(r2, r3); - - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); - - const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); - const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); - const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); - const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); - // dct_const_round_shift - - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - - const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); - const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); - const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); - const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); - - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - - const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); - const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); - const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); - const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); - // Combine - - res0 = _mm_packs_epi32(w0, w1); - res4 = _mm_packs_epi32(w2, w3); - res2 = _mm_packs_epi32(w4, w5); - res6 = _mm_packs_epi32(w6, w7); - } - // Work on next four results - { - // Interleave to do the multiply by constants which gets us into 32bits - const __m128i d0 = _mm_sub_epi16(q6, q5); - const __m128i d1 = _mm_add_epi16(q6, q5); - const __m128i r0 = _mm_mulhrs_epi16(d0, k__dual_p16_p16); - const __m128i r1 = _mm_mulhrs_epi16(d1, k__dual_p16_p16); - - // Add/subtract - const __m128i x0 = _mm_add_epi16(q4, r0); - const __m128i x1 = _mm_sub_epi16(q4, r0); - const __m128i x2 = _mm_sub_epi16(q7, r1); - const __m128i x3 = _mm_add_epi16(q7, r1); - // Interleave to do the multiply by constants which gets us into 32bits - const __m128i t0 = _mm_unpacklo_epi16(x0, x3); - const __m128i t1 = _mm_unpackhi_epi16(x0, x3); - const __m128i t2 = _mm_unpacklo_epi16(x1, x2); - const __m128i t3 = _mm_unpackhi_epi16(x1, x2); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); - const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); - const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); - const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); - const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); - const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); - const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); - // dct_const_round_shift - const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); - const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); - const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); - const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); - const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); - const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); - const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); - const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); - const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); - const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); - const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); - const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); - const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); - const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); - // Combine - res1 = _mm_packs_epi32(w0, w1); - res7 = _mm_packs_epi32(w2, w3); - res5 = _mm_packs_epi32(w4, w5); - res3 = _mm_packs_epi32(w6, w7); - } - // Transpose the 8x8. - { - // 00 01 02 03 04 05 06 07 - // 10 11 12 13 14 15 16 17 - // 20 21 22 23 24 25 26 27 - // 30 31 32 33 34 35 36 37 - // 40 41 42 43 44 45 46 47 - // 50 51 52 53 54 55 56 57 - // 60 61 62 63 64 65 66 67 - // 70 71 72 73 74 75 76 77 - const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); - const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3); - const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1); - const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3); - const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5); - const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7); - const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5); - const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7); - // 00 10 01 11 02 12 03 13 - // 20 30 21 31 22 32 23 33 - // 04 14 05 15 06 16 07 17 - // 24 34 25 35 26 36 27 37 - // 40 50 41 51 42 52 43 53 - // 60 70 61 71 62 72 63 73 - // 54 54 55 55 56 56 57 57 - // 64 74 65 75 66 76 67 77 - const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); - const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); - const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); - const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); - const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); - const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); - const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); - const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); - // 00 10 20 30 01 11 21 31 - // 40 50 60 70 41 51 61 71 - // 02 12 22 32 03 13 23 33 - // 42 52 62 72 43 53 63 73 - // 04 14 24 34 05 15 21 36 - // 44 54 64 74 45 55 61 76 - // 06 16 26 36 07 17 27 37 - // 46 56 66 76 47 57 67 77 - in0 = _mm_unpacklo_epi64(tr1_0, tr1_4); - in1 = _mm_unpackhi_epi64(tr1_0, tr1_4); - in2 = _mm_unpacklo_epi64(tr1_2, tr1_6); - in3 = _mm_unpackhi_epi64(tr1_2, tr1_6); - in4 = _mm_unpacklo_epi64(tr1_1, tr1_5); - in5 = _mm_unpackhi_epi64(tr1_1, tr1_5); - in6 = _mm_unpacklo_epi64(tr1_3, tr1_7); - in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); - // 00 10 20 30 40 50 60 70 - // 01 11 21 31 41 51 61 71 - // 02 12 22 32 42 52 62 72 - // 03 13 23 33 43 53 63 73 - // 04 14 24 34 44 54 64 74 - // 05 15 25 35 45 55 65 75 - // 06 16 26 36 46 56 66 76 - // 07 17 27 37 47 57 67 77 - } - } - // Post-condition output and store it - { - // Post-condition (division by two) - // division of two 16 bits signed numbers using shifts - // n / 2 = (n - (n >> 15)) >> 1 - const __m128i sign_in0 = _mm_srai_epi16(in0, 15); - const __m128i sign_in1 = _mm_srai_epi16(in1, 15); - const __m128i sign_in2 = _mm_srai_epi16(in2, 15); - const __m128i sign_in3 = _mm_srai_epi16(in3, 15); - const __m128i sign_in4 = _mm_srai_epi16(in4, 15); - const __m128i sign_in5 = _mm_srai_epi16(in5, 15); - const __m128i sign_in6 = _mm_srai_epi16(in6, 15); - const __m128i sign_in7 = _mm_srai_epi16(in7, 15); - in0 = _mm_sub_epi16(in0, sign_in0); - in1 = _mm_sub_epi16(in1, sign_in1); - in2 = _mm_sub_epi16(in2, sign_in2); - in3 = _mm_sub_epi16(in3, sign_in3); - in4 = _mm_sub_epi16(in4, sign_in4); - in5 = _mm_sub_epi16(in5, sign_in5); - in6 = _mm_sub_epi16(in6, sign_in6); - in7 = _mm_sub_epi16(in7, sign_in7); - in0 = _mm_srai_epi16(in0, 1); - in1 = _mm_srai_epi16(in1, 1); - in2 = _mm_srai_epi16(in2, 1); - in3 = _mm_srai_epi16(in3, 1); - in4 = _mm_srai_epi16(in4, 1); - in5 = _mm_srai_epi16(in5, 1); - in6 = _mm_srai_epi16(in6, 1); - in7 = _mm_srai_epi16(in7, 1); - } - - iscan_ptr += n_coeffs; - qcoeff_ptr += n_coeffs; - dqcoeff_ptr += n_coeffs; - n_coeffs = -n_coeffs; - zero = _mm_setzero_si128(); - - if (!skip_block) { - __m128i eob; - __m128i round, quant, dequant, thr; - int16_t nzflag; - { - __m128i coeff0, coeff1; - - // Setup global values - { - round = _mm_load_si128((const __m128i *)round_ptr); - quant = _mm_load_si128((const __m128i *)quant_ptr); - dequant = _mm_load_si128((const __m128i *)dequant_ptr); - } - - { - __m128i coeff0_sign, coeff1_sign; - __m128i qcoeff0, qcoeff1; - __m128i qtmp0, qtmp1; - // Do DC and first 15 AC - coeff0 = *in[0]; - coeff1 = *in[1]; - - // Poor man's sign extract - coeff0_sign = _mm_srai_epi16(coeff0, 15); - coeff1_sign = _mm_srai_epi16(coeff1, 15); - qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); - qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); - - qcoeff0 = _mm_adds_epi16(qcoeff0, round); - round = _mm_unpackhi_epi64(round, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); - quant = _mm_unpackhi_epi64(quant, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - // Reinsert signs - qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign); - qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); - - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0); - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1); - - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); - dequant = _mm_unpackhi_epi64(dequant, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); - - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0); - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1); - } - - { - // Scan for eob - __m128i zero_coeff0, zero_coeff1; - __m128i nzero_coeff0, nzero_coeff1; - __m128i iscan0, iscan1; - __m128i eob1; - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); - nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); - // Add one to convert from indices to counts - iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0); - iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1); - eob = _mm_and_si128(iscan0, nzero_coeff0); - eob1 = _mm_and_si128(iscan1, nzero_coeff1); - eob = _mm_max_epi16(eob, eob1); - } - n_coeffs += 8 * 2; - } - - // AC only loop - index = 2; - thr = _mm_srai_epi16(dequant, 1); - while (n_coeffs < 0) { - __m128i coeff0, coeff1; - { - __m128i coeff0_sign, coeff1_sign; - __m128i qcoeff0, qcoeff1; - __m128i qtmp0, qtmp1; - - assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1); - coeff0 = *in[index]; - coeff1 = *in[index + 1]; - - // Poor man's sign extract - coeff0_sign = _mm_srai_epi16(coeff0, 15); - coeff1_sign = _mm_srai_epi16(coeff1, 15); - qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); - qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); - - nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) | - _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr)); - - if (nzflag) { - qcoeff0 = _mm_adds_epi16(qcoeff0, round); - qcoeff1 = _mm_adds_epi16(qcoeff1, round); - qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); - qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); - - // Reinsert signs - qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign); - qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign); - qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); - qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); - - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0); - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1); - - coeff0 = _mm_mullo_epi16(qcoeff0, dequant); - coeff1 = _mm_mullo_epi16(qcoeff1, dequant); - - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0); - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1); - } else { - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero); - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero); - - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero); - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero); - } - } - - if (nzflag) { - // Scan for eob - __m128i zero_coeff0, zero_coeff1; - __m128i nzero_coeff0, nzero_coeff1; - __m128i iscan0, iscan1; - __m128i eob0, eob1; - zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); - zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); - nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); - nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero); - iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); - iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); - // Add one to convert from indices to counts - iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0); - iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1); - eob0 = _mm_and_si128(iscan0, nzero_coeff0); - eob1 = _mm_and_si128(iscan1, nzero_coeff1); - eob0 = _mm_max_epi16(eob0, eob1); - eob = _mm_max_epi16(eob, eob0); - } - n_coeffs += 8 * 2; - index += 2; - } - - // Accumulate EOB - { - __m128i eob_shuffled; - eob_shuffled = _mm_shuffle_epi32(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); - eob = _mm_max_epi16(eob, eob_shuffled); - eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); - eob = _mm_max_epi16(eob, eob_shuffled); - *eob_ptr = _mm_extract_epi16(eob, 1); - } - } else { - do { - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero); - _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero); - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero); - _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero); - n_coeffs += 8 * 2; - } while (n_coeffs < 0); - *eob_ptr = 0; - } -} diff --git a/third_party/aom/av1/encoder/x86/error_intrin_avx2.c b/third_party/aom/av1/encoder/x86/error_intrin_avx2.c index 20ba4149c..6599630d0 100644 --- a/third_party/aom/av1/encoder/x86/error_intrin_avx2.c +++ b/third_party/aom/av1/encoder/x86/error_intrin_avx2.c @@ -17,14 +17,15 @@ static INLINE void read_coeff(const tran_low_t *coeff, intptr_t offset, __m256i *c) { const tran_low_t *addr = coeff + offset; -#if CONFIG_HIGHBITDEPTH - const __m256i x0 = _mm256_loadu_si256((const __m256i *)addr); - const __m256i x1 = _mm256_loadu_si256((const __m256i *)addr + 1); - const __m256i y = _mm256_packs_epi32(x0, x1); - *c = _mm256_permute4x64_epi64(y, 0xD8); -#else - *c = _mm256_loadu_si256((const __m256i *)addr); -#endif + + if (sizeof(tran_low_t) == 4) { + const __m256i x0 = _mm256_loadu_si256((const __m256i *)addr); + const __m256i x1 = _mm256_loadu_si256((const __m256i *)addr + 1); + const __m256i y = _mm256_packs_epi32(x0, x1); + *c = _mm256_permute4x64_epi64(y, 0xD8); + } else { + *c = _mm256_loadu_si256((const __m256i *)addr); + } } int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, diff --git a/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c b/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c index cab36f2bd..b684f7a3a 100644 --- a/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c +++ b/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c @@ -195,7 +195,7 @@ static void fadst4x4_sse4_1(__m128i *in, int bit) { } void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *coeff, - int input_stride, int tx_type, int bd) { + int input_stride, TX_TYPE tx_type, int bd) { __m128i in[4]; const TXFM_1D_CFG *row_cfg = NULL; const TXFM_1D_CFG *col_cfg = NULL; @@ -926,7 +926,7 @@ static void fadst8x8_sse4_1(__m128i *in, __m128i *out, int bit) { } void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *coeff, int stride, - int tx_type, int bd) { + TX_TYPE tx_type, int bd) { __m128i in[16], out[16]; const TXFM_1D_CFG *row_cfg = NULL; const TXFM_1D_CFG *col_cfg = NULL; @@ -1800,7 +1800,7 @@ static void write_buffer_16x16(const __m128i *in, int32_t *output) { } void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *coeff, - int stride, int tx_type, int bd) { + int stride, TX_TYPE tx_type, int bd) { __m128i in[64], out[64]; const TXFM_1D_CFG *row_cfg = NULL; const TXFM_1D_CFG *col_cfg = NULL; diff --git a/third_party/aom/av1/encoder/x86/hybrid_fwd_txfm_avx2.c b/third_party/aom/av1/encoder/x86/hybrid_fwd_txfm_avx2.c index af8e9a5f4..88621c82b 100644 --- a/third_party/aom/av1/encoder/x86/hybrid_fwd_txfm_avx2.c +++ b/third_party/aom/av1/encoder/x86/hybrid_fwd_txfm_avx2.c @@ -916,7 +916,7 @@ static void fidtx16_avx2(__m256i *in) { void av1_fht16x16_avx2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m256i in[16]; - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "Invalid tx type for tx size"); #endif @@ -1516,7 +1516,7 @@ void av1_fht32x32_avx2(const int16_t *input, tran_low_t *output, int stride, TxfmParam *txfm_param) { __m256i in0[32]; // left 32 columns __m256i in1[32]; // right 32 columns - int tx_type = txfm_param->tx_type; + const TX_TYPE tx_type = txfm_param->tx_type; #if CONFIG_MRC_TX assert(tx_type != MRC_DCT && "No avx2 32x32 implementation of MRC_DCT"); #endif -- cgit v1.2.3