michael@0: /* michael@0: * Copyright (c) 2012 The WebM project authors. All Rights Reserved. michael@0: * michael@0: * Use of this source code is governed by a BSD-style license michael@0: * that can be found in the LICENSE file in the root of the source michael@0: * tree. An additional intellectual property rights grant can be found michael@0: * in the file PATENTS. All contributing project authors may michael@0: * be found in the AUTHORS file in the root of the source tree. michael@0: */ michael@0: michael@0: #include // SSE2 michael@0: #include "vp9/common/vp9_idct.h" // for cospi constants michael@0: #include "vpx_ports/mem.h" michael@0: michael@0: void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride) { michael@0: // The 2D transform is done with two passes which are actually pretty michael@0: // similar. In the first one, we transform the columns and transpose michael@0: // the results. In the second one, we transform the rows. To achieve that, michael@0: // as the first pass results are transposed, we tranpose the columns (that michael@0: // is the transposed rows) and transpose the results (so that it goes back michael@0: // in normal/row positions). michael@0: int pass; michael@0: // Constants michael@0: // When we use them, in one case, they are all the same. In all others michael@0: // it's a pair of them that we need to repeat four times. This is done michael@0: // by constructing the 32 bit constant corresponding to that pair. michael@0: const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); michael@0: const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); michael@0: const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1); michael@0: const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0); michael@0: const __m128i kOne = _mm_set1_epi16(1); michael@0: __m128i in0, in1, in2, in3; michael@0: // Load inputs. michael@0: { michael@0: in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); michael@0: in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); michael@0: in2 = _mm_loadl_epi64((const __m128i *)(input + 2 * stride)); michael@0: in3 = _mm_loadl_epi64((const __m128i *)(input + 3 * stride)); michael@0: // x = x << 4 michael@0: in0 = _mm_slli_epi16(in0, 4); michael@0: in1 = _mm_slli_epi16(in1, 4); michael@0: in2 = _mm_slli_epi16(in2, 4); michael@0: in3 = _mm_slli_epi16(in3, 4); michael@0: // if (i == 0 && input[0]) input[0] += 1; michael@0: { michael@0: // The mask will only contain wether the first value is zero, all michael@0: // other comparison will fail as something shifted by 4 (above << 4) michael@0: // can never be equal to one. To increment in the non-zero case, we michael@0: // add the mask and one for the first element: michael@0: // - if zero, mask = -1, v = v - 1 + 1 = v michael@0: // - if non-zero, mask = 0, v = v + 0 + 1 = v + 1 michael@0: __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a); michael@0: in0 = _mm_add_epi16(in0, mask); michael@0: in0 = _mm_add_epi16(in0, k__nonzero_bias_b); michael@0: } michael@0: } michael@0: // Do the two transform/transpose passes michael@0: for (pass = 0; pass < 2; ++pass) { michael@0: // Transform 1/2: Add/substract michael@0: const __m128i r0 = _mm_add_epi16(in0, in3); michael@0: const __m128i r1 = _mm_add_epi16(in1, in2); michael@0: const __m128i r2 = _mm_sub_epi16(in1, in2); michael@0: const __m128i r3 = _mm_sub_epi16(in0, in3); michael@0: // Transform 1/2: Interleave to do the multiply by constants which gets us michael@0: // into 32 bits. michael@0: const __m128i t0 = _mm_unpacklo_epi16(r0, r1); michael@0: const __m128i t2 = _mm_unpacklo_epi16(r2, r3); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); michael@0: const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); michael@0: const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); michael@0: const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); michael@0: const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); michael@0: const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); michael@0: // Combine and transpose michael@0: const __m128i res0 = _mm_packs_epi32(w0, w2); michael@0: const __m128i res1 = _mm_packs_epi32(w4, w6); michael@0: // 00 01 02 03 20 21 22 23 michael@0: // 10 11 12 13 30 31 32 33 michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); michael@0: const __m128i tr0_1 = _mm_unpackhi_epi16(res0, res1); michael@0: // 00 10 01 11 02 12 03 13 michael@0: // 20 30 21 31 22 32 23 33 michael@0: in0 = _mm_unpacklo_epi32(tr0_0, tr0_1); michael@0: in2 = _mm_unpackhi_epi32(tr0_0, tr0_1); michael@0: // 00 10 20 30 01 11 21 31 in0 contains 0 followed by 1 michael@0: // 02 12 22 32 03 13 23 33 in2 contains 2 followed by 3 michael@0: if (0 == pass) { michael@0: // Extract values in the high part for second pass as transform code michael@0: // only uses the first four values. michael@0: in1 = _mm_unpackhi_epi64(in0, in0); michael@0: in3 = _mm_unpackhi_epi64(in2, in2); michael@0: } else { michael@0: // Post-condition output and store it (v + 1) >> 2, taking advantage michael@0: // of the fact 1/3 are stored just after 0/2. michael@0: __m128i out01 = _mm_add_epi16(in0, kOne); michael@0: __m128i out23 = _mm_add_epi16(in2, kOne); michael@0: out01 = _mm_srai_epi16(out01, 2); michael@0: out23 = _mm_srai_epi16(out23, 2); michael@0: _mm_storeu_si128((__m128i *)(output + 0 * 4), out01); michael@0: _mm_storeu_si128((__m128i *)(output + 2 * 4), out23); michael@0: } michael@0: } michael@0: } michael@0: michael@0: static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in, michael@0: int stride) { michael@0: const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1); michael@0: const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0); michael@0: __m128i mask; michael@0: michael@0: in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); michael@0: in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); michael@0: in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride)); michael@0: in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride)); michael@0: michael@0: in[0] = _mm_slli_epi16(in[0], 4); michael@0: in[1] = _mm_slli_epi16(in[1], 4); michael@0: in[2] = _mm_slli_epi16(in[2], 4); michael@0: in[3] = _mm_slli_epi16(in[3], 4); michael@0: michael@0: mask = _mm_cmpeq_epi16(in[0], k__nonzero_bias_a); michael@0: in[0] = _mm_add_epi16(in[0], mask); michael@0: in[0] = _mm_add_epi16(in[0], k__nonzero_bias_b); michael@0: } michael@0: michael@0: static INLINE void write_buffer_4x4(int16_t *output, __m128i *res) { michael@0: const __m128i kOne = _mm_set1_epi16(1); michael@0: __m128i in01 = _mm_unpacklo_epi64(res[0], res[1]); michael@0: __m128i in23 = _mm_unpacklo_epi64(res[2], res[3]); michael@0: __m128i out01 = _mm_add_epi16(in01, kOne); michael@0: __m128i out23 = _mm_add_epi16(in23, kOne); michael@0: out01 = _mm_srai_epi16(out01, 2); michael@0: out23 = _mm_srai_epi16(out23, 2); michael@0: _mm_store_si128((__m128i *)(output + 0 * 8), out01); michael@0: _mm_store_si128((__m128i *)(output + 1 * 8), out23); michael@0: } michael@0: michael@0: static INLINE void transpose_4x4(__m128i *res) { michael@0: // Combine and transpose michael@0: // 00 01 02 03 20 21 22 23 michael@0: // 10 11 12 13 30 31 32 33 michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]); michael@0: const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]); michael@0: michael@0: // 00 10 01 11 02 12 03 13 michael@0: // 20 30 21 31 22 32 23 33 michael@0: res[0] = _mm_unpacklo_epi32(tr0_0, tr0_1); michael@0: res[2] = _mm_unpackhi_epi32(tr0_0, tr0_1); michael@0: michael@0: // 00 10 20 30 01 11 21 31 michael@0: // 02 12 22 32 03 13 23 33 michael@0: // only use the first 4 16-bit integers michael@0: res[1] = _mm_unpackhi_epi64(res[0], res[0]); michael@0: res[3] = _mm_unpackhi_epi64(res[2], res[2]); michael@0: } michael@0: michael@0: void fdct4_1d_sse2(__m128i *in) { michael@0: const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); michael@0: const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: michael@0: __m128i u[4], v[4]; michael@0: u[0]=_mm_unpacklo_epi16(in[0], in[1]); michael@0: u[1]=_mm_unpacklo_epi16(in[3], in[2]); michael@0: michael@0: v[0] = _mm_add_epi16(u[0], u[1]); michael@0: v[1] = _mm_sub_epi16(u[0], u[1]); michael@0: michael@0: u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0 michael@0: u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); // 2 michael@0: u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24); // 1 michael@0: u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08); // 3 michael@0: michael@0: v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); michael@0: v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); michael@0: v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); michael@0: v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); michael@0: u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); michael@0: michael@0: in[0] = _mm_packs_epi32(u[0], u[1]); michael@0: in[1] = _mm_packs_epi32(u[2], u[3]); michael@0: transpose_4x4(in); michael@0: } michael@0: michael@0: void fadst4_1d_sse2(__m128i *in) { michael@0: const __m128i k__sinpi_p01_p02 = pair_set_epi16(sinpi_1_9, sinpi_2_9); michael@0: const __m128i k__sinpi_p04_m01 = pair_set_epi16(sinpi_4_9, -sinpi_1_9); michael@0: const __m128i k__sinpi_p03_p04 = pair_set_epi16(sinpi_3_9, sinpi_4_9); michael@0: const __m128i k__sinpi_m03_p02 = pair_set_epi16(-sinpi_3_9, sinpi_2_9); michael@0: const __m128i k__sinpi_p03_p03 = _mm_set1_epi16(sinpi_3_9); michael@0: const __m128i kZero = _mm_set1_epi16(0); michael@0: const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: __m128i u[8], v[8]; michael@0: __m128i in7 = _mm_add_epi16(in[0], in[1]); michael@0: michael@0: u[0] = _mm_unpacklo_epi16(in[0], in[1]); michael@0: u[1] = _mm_unpacklo_epi16(in[2], in[3]); michael@0: u[2] = _mm_unpacklo_epi16(in7, kZero); michael@0: u[3] = _mm_unpacklo_epi16(in[2], kZero); michael@0: u[4] = _mm_unpacklo_epi16(in[3], kZero); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02); // s0 + s2 michael@0: v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04); // s4 + s5 michael@0: v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03); // x1 michael@0: v[3] = _mm_madd_epi16(u[0], k__sinpi_p04_m01); // s1 - s3 michael@0: v[4] = _mm_madd_epi16(u[1], k__sinpi_m03_p02); // -s4 + s6 michael@0: v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03); // s4 michael@0: v[6] = _mm_madd_epi16(u[4], k__sinpi_p03_p03); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], v[1]); michael@0: u[1] = _mm_sub_epi32(v[2], v[6]); michael@0: u[2] = _mm_add_epi32(v[3], v[4]); michael@0: u[3] = _mm_sub_epi32(u[2], u[0]); michael@0: u[4] = _mm_slli_epi32(v[5], 2); michael@0: u[5] = _mm_sub_epi32(u[4], v[5]); michael@0: u[6] = _mm_add_epi32(u[3], u[5]); michael@0: michael@0: v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); michael@0: v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); michael@0: v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); michael@0: v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); michael@0: michael@0: u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); michael@0: michael@0: in[0] = _mm_packs_epi32(u[0], u[2]); michael@0: in[1] = _mm_packs_epi32(u[1], u[3]); michael@0: transpose_4x4(in); michael@0: } michael@0: michael@0: void vp9_short_fht4x4_sse2(const int16_t *input, int16_t *output, michael@0: int stride, int tx_type) { michael@0: __m128i in[4]; michael@0: load_buffer_4x4(input, in, stride); michael@0: switch (tx_type) { michael@0: case 0: // DCT_DCT michael@0: fdct4_1d_sse2(in); michael@0: fdct4_1d_sse2(in); michael@0: break; michael@0: case 1: // ADST_DCT michael@0: fadst4_1d_sse2(in); michael@0: fdct4_1d_sse2(in); michael@0: break; michael@0: case 2: // DCT_ADST michael@0: fdct4_1d_sse2(in); michael@0: fadst4_1d_sse2(in); michael@0: break; michael@0: case 3: // ADST_ADST michael@0: fadst4_1d_sse2(in); michael@0: fadst4_1d_sse2(in); michael@0: break; michael@0: default: michael@0: assert(0); michael@0: break; michael@0: } michael@0: write_buffer_4x4(output, in); michael@0: } michael@0: michael@0: void vp9_fdct8x8_sse2(const int16_t *input, int16_t *output, int stride) { michael@0: int pass; michael@0: // Constants michael@0: // When we use them, in one case, they are all the same. In all others michael@0: // it's a pair of them that we need to repeat four times. This is done michael@0: // by constructing the 32 bit constant corresponding to that pair. michael@0: const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); michael@0: const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); michael@0: const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); michael@0: const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); michael@0: const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); michael@0: const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); michael@0: const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: // Load input michael@0: __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); michael@0: __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); michael@0: __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); michael@0: __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride)); michael@0: __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride)); michael@0: __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride)); michael@0: __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride)); michael@0: __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); michael@0: // Pre-condition input (shift by two) michael@0: in0 = _mm_slli_epi16(in0, 2); michael@0: in1 = _mm_slli_epi16(in1, 2); michael@0: in2 = _mm_slli_epi16(in2, 2); michael@0: in3 = _mm_slli_epi16(in3, 2); michael@0: in4 = _mm_slli_epi16(in4, 2); michael@0: in5 = _mm_slli_epi16(in5, 2); michael@0: in6 = _mm_slli_epi16(in6, 2); michael@0: in7 = _mm_slli_epi16(in7, 2); michael@0: michael@0: // We do two passes, first the columns, then the rows. The results of the michael@0: // first pass are transposed so that the same column code can be reused. The michael@0: // results of the second pass are also transposed so that the rows (processed michael@0: // as columns) are put back in row positions. michael@0: for (pass = 0; pass < 2; pass++) { michael@0: // To store results of each pass before the transpose. michael@0: __m128i res0, res1, res2, res3, res4, res5, res6, res7; michael@0: // Add/substract michael@0: const __m128i q0 = _mm_add_epi16(in0, in7); michael@0: const __m128i q1 = _mm_add_epi16(in1, in6); michael@0: const __m128i q2 = _mm_add_epi16(in2, in5); michael@0: const __m128i q3 = _mm_add_epi16(in3, in4); michael@0: const __m128i q4 = _mm_sub_epi16(in3, in4); michael@0: const __m128i q5 = _mm_sub_epi16(in2, in5); michael@0: const __m128i q6 = _mm_sub_epi16(in1, in6); michael@0: const __m128i q7 = _mm_sub_epi16(in0, in7); michael@0: // Work on first four results michael@0: { michael@0: // Add/substract michael@0: const __m128i r0 = _mm_add_epi16(q0, q3); michael@0: const __m128i r1 = _mm_add_epi16(q1, q2); michael@0: const __m128i r2 = _mm_sub_epi16(q1, q2); michael@0: const __m128i r3 = _mm_sub_epi16(q0, q3); michael@0: // Interleave to do the multiply by constants which gets us into 32bits michael@0: const __m128i t0 = _mm_unpacklo_epi16(r0, r1); michael@0: const __m128i t1 = _mm_unpackhi_epi16(r0, r1); michael@0: const __m128i t2 = _mm_unpacklo_epi16(r2, r3); michael@0: const __m128i t3 = _mm_unpackhi_epi16(r2, r3); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); michael@0: const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); michael@0: const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); michael@0: const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); michael@0: const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); michael@0: const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); michael@0: const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); michael@0: const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); michael@0: const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); michael@0: const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); michael@0: const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); michael@0: const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); michael@0: const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); michael@0: // Combine michael@0: res0 = _mm_packs_epi32(w0, w1); michael@0: res4 = _mm_packs_epi32(w2, w3); michael@0: res2 = _mm_packs_epi32(w4, w5); michael@0: res6 = _mm_packs_epi32(w6, w7); michael@0: } michael@0: // Work on next four results michael@0: { michael@0: // Interleave to do the multiply by constants which gets us into 32bits michael@0: const __m128i d0 = _mm_unpacklo_epi16(q6, q5); michael@0: const __m128i d1 = _mm_unpackhi_epi16(q6, q5); michael@0: const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); michael@0: const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); michael@0: const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); michael@0: const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); michael@0: // dct_const_round_shift michael@0: const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); michael@0: const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); michael@0: const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); michael@0: const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); michael@0: const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); michael@0: const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); michael@0: const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); michael@0: const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); michael@0: // Combine michael@0: const __m128i r0 = _mm_packs_epi32(s0, s1); michael@0: const __m128i r1 = _mm_packs_epi32(s2, s3); michael@0: // Add/substract michael@0: const __m128i x0 = _mm_add_epi16(q4, r0); michael@0: const __m128i x1 = _mm_sub_epi16(q4, r0); michael@0: const __m128i x2 = _mm_sub_epi16(q7, r1); michael@0: const __m128i x3 = _mm_add_epi16(q7, r1); michael@0: // Interleave to do the multiply by constants which gets us into 32bits michael@0: const __m128i t0 = _mm_unpacklo_epi16(x0, x3); michael@0: const __m128i t1 = _mm_unpackhi_epi16(x0, x3); michael@0: const __m128i t2 = _mm_unpacklo_epi16(x1, x2); michael@0: const __m128i t3 = _mm_unpackhi_epi16(x1, x2); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); michael@0: const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); michael@0: const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); michael@0: const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); michael@0: const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); michael@0: const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); michael@0: const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); michael@0: const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); michael@0: const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); michael@0: const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); michael@0: const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); michael@0: const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); michael@0: const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); michael@0: // Combine michael@0: res1 = _mm_packs_epi32(w0, w1); michael@0: res7 = _mm_packs_epi32(w2, w3); michael@0: res5 = _mm_packs_epi32(w4, w5); michael@0: res3 = _mm_packs_epi32(w6, w7); michael@0: } michael@0: // Transpose the 8x8. michael@0: { michael@0: // 00 01 02 03 04 05 06 07 michael@0: // 10 11 12 13 14 15 16 17 michael@0: // 20 21 22 23 24 25 26 27 michael@0: // 30 31 32 33 34 35 36 37 michael@0: // 40 41 42 43 44 45 46 47 michael@0: // 50 51 52 53 54 55 56 57 michael@0: // 60 61 62 63 64 65 66 67 michael@0: // 70 71 72 73 74 75 76 77 michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); michael@0: const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3); michael@0: const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1); michael@0: const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3); michael@0: const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5); michael@0: const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7); michael@0: const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5); michael@0: const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7); michael@0: // 00 10 01 11 02 12 03 13 michael@0: // 20 30 21 31 22 32 23 33 michael@0: // 04 14 05 15 06 16 07 17 michael@0: // 24 34 25 35 26 36 27 37 michael@0: // 40 50 41 51 42 52 43 53 michael@0: // 60 70 61 71 62 72 63 73 michael@0: // 54 54 55 55 56 56 57 57 michael@0: // 64 74 65 75 66 76 67 77 michael@0: const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); michael@0: const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); michael@0: const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); michael@0: const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); michael@0: const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); michael@0: const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); michael@0: const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); michael@0: const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); michael@0: // 00 10 20 30 01 11 21 31 michael@0: // 40 50 60 70 41 51 61 71 michael@0: // 02 12 22 32 03 13 23 33 michael@0: // 42 52 62 72 43 53 63 73 michael@0: // 04 14 24 34 05 15 21 36 michael@0: // 44 54 64 74 45 55 61 76 michael@0: // 06 16 26 36 07 17 27 37 michael@0: // 46 56 66 76 47 57 67 77 michael@0: in0 = _mm_unpacklo_epi64(tr1_0, tr1_4); michael@0: in1 = _mm_unpackhi_epi64(tr1_0, tr1_4); michael@0: in2 = _mm_unpacklo_epi64(tr1_2, tr1_6); michael@0: in3 = _mm_unpackhi_epi64(tr1_2, tr1_6); michael@0: in4 = _mm_unpacklo_epi64(tr1_1, tr1_5); michael@0: in5 = _mm_unpackhi_epi64(tr1_1, tr1_5); michael@0: in6 = _mm_unpacklo_epi64(tr1_3, tr1_7); michael@0: in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); michael@0: // 00 10 20 30 40 50 60 70 michael@0: // 01 11 21 31 41 51 61 71 michael@0: // 02 12 22 32 42 52 62 72 michael@0: // 03 13 23 33 43 53 63 73 michael@0: // 04 14 24 34 44 54 64 74 michael@0: // 05 15 25 35 45 55 65 75 michael@0: // 06 16 26 36 46 56 66 76 michael@0: // 07 17 27 37 47 57 67 77 michael@0: } michael@0: } michael@0: // Post-condition output and store it michael@0: { michael@0: // Post-condition (division by two) michael@0: // division of two 16 bits signed numbers using shifts michael@0: // n / 2 = (n - (n >> 15)) >> 1 michael@0: const __m128i sign_in0 = _mm_srai_epi16(in0, 15); michael@0: const __m128i sign_in1 = _mm_srai_epi16(in1, 15); michael@0: const __m128i sign_in2 = _mm_srai_epi16(in2, 15); michael@0: const __m128i sign_in3 = _mm_srai_epi16(in3, 15); michael@0: const __m128i sign_in4 = _mm_srai_epi16(in4, 15); michael@0: const __m128i sign_in5 = _mm_srai_epi16(in5, 15); michael@0: const __m128i sign_in6 = _mm_srai_epi16(in6, 15); michael@0: const __m128i sign_in7 = _mm_srai_epi16(in7, 15); michael@0: in0 = _mm_sub_epi16(in0, sign_in0); michael@0: in1 = _mm_sub_epi16(in1, sign_in1); michael@0: in2 = _mm_sub_epi16(in2, sign_in2); michael@0: in3 = _mm_sub_epi16(in3, sign_in3); michael@0: in4 = _mm_sub_epi16(in4, sign_in4); michael@0: in5 = _mm_sub_epi16(in5, sign_in5); michael@0: in6 = _mm_sub_epi16(in6, sign_in6); michael@0: in7 = _mm_sub_epi16(in7, sign_in7); michael@0: in0 = _mm_srai_epi16(in0, 1); michael@0: in1 = _mm_srai_epi16(in1, 1); michael@0: in2 = _mm_srai_epi16(in2, 1); michael@0: in3 = _mm_srai_epi16(in3, 1); michael@0: in4 = _mm_srai_epi16(in4, 1); michael@0: in5 = _mm_srai_epi16(in5, 1); michael@0: in6 = _mm_srai_epi16(in6, 1); michael@0: in7 = _mm_srai_epi16(in7, 1); michael@0: // store results michael@0: _mm_store_si128((__m128i *)(output + 0 * 8), in0); michael@0: _mm_store_si128((__m128i *)(output + 1 * 8), in1); michael@0: _mm_store_si128((__m128i *)(output + 2 * 8), in2); michael@0: _mm_store_si128((__m128i *)(output + 3 * 8), in3); michael@0: _mm_store_si128((__m128i *)(output + 4 * 8), in4); michael@0: _mm_store_si128((__m128i *)(output + 5 * 8), in5); michael@0: _mm_store_si128((__m128i *)(output + 6 * 8), in6); michael@0: _mm_store_si128((__m128i *)(output + 7 * 8), in7); michael@0: } michael@0: } michael@0: michael@0: // load 8x8 array michael@0: static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in, michael@0: int stride) { michael@0: in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride)); michael@0: in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride)); michael@0: in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride)); michael@0: in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride)); michael@0: in[4] = _mm_load_si128((const __m128i *)(input + 4 * stride)); michael@0: in[5] = _mm_load_si128((const __m128i *)(input + 5 * stride)); michael@0: in[6] = _mm_load_si128((const __m128i *)(input + 6 * stride)); michael@0: in[7] = _mm_load_si128((const __m128i *)(input + 7 * stride)); michael@0: michael@0: in[0] = _mm_slli_epi16(in[0], 2); michael@0: in[1] = _mm_slli_epi16(in[1], 2); michael@0: in[2] = _mm_slli_epi16(in[2], 2); michael@0: in[3] = _mm_slli_epi16(in[3], 2); michael@0: in[4] = _mm_slli_epi16(in[4], 2); michael@0: in[5] = _mm_slli_epi16(in[5], 2); michael@0: in[6] = _mm_slli_epi16(in[6], 2); michael@0: in[7] = _mm_slli_epi16(in[7], 2); michael@0: } michael@0: michael@0: // right shift and rounding michael@0: static INLINE void right_shift_8x8(__m128i *res, int const bit) { michael@0: const __m128i kOne = _mm_set1_epi16(1); michael@0: const int bit_m02 = bit - 2; michael@0: __m128i sign0 = _mm_srai_epi16(res[0], 15); michael@0: __m128i sign1 = _mm_srai_epi16(res[1], 15); michael@0: __m128i sign2 = _mm_srai_epi16(res[2], 15); michael@0: __m128i sign3 = _mm_srai_epi16(res[3], 15); michael@0: __m128i sign4 = _mm_srai_epi16(res[4], 15); michael@0: __m128i sign5 = _mm_srai_epi16(res[5], 15); michael@0: __m128i sign6 = _mm_srai_epi16(res[6], 15); michael@0: __m128i sign7 = _mm_srai_epi16(res[7], 15); michael@0: michael@0: if (bit_m02 >= 0) { michael@0: __m128i k_const_rounding = _mm_slli_epi16(kOne, bit_m02); michael@0: res[0] = _mm_add_epi16(res[0], k_const_rounding); michael@0: res[1] = _mm_add_epi16(res[1], k_const_rounding); michael@0: res[2] = _mm_add_epi16(res[2], k_const_rounding); michael@0: res[3] = _mm_add_epi16(res[3], k_const_rounding); michael@0: res[4] = _mm_add_epi16(res[4], k_const_rounding); michael@0: res[5] = _mm_add_epi16(res[5], k_const_rounding); michael@0: res[6] = _mm_add_epi16(res[6], k_const_rounding); michael@0: res[7] = _mm_add_epi16(res[7], k_const_rounding); michael@0: } michael@0: michael@0: res[0] = _mm_sub_epi16(res[0], sign0); michael@0: res[1] = _mm_sub_epi16(res[1], sign1); michael@0: res[2] = _mm_sub_epi16(res[2], sign2); michael@0: res[3] = _mm_sub_epi16(res[3], sign3); michael@0: res[4] = _mm_sub_epi16(res[4], sign4); michael@0: res[5] = _mm_sub_epi16(res[5], sign5); michael@0: res[6] = _mm_sub_epi16(res[6], sign6); michael@0: res[7] = _mm_sub_epi16(res[7], sign7); michael@0: michael@0: res[0] = _mm_srai_epi16(res[0], bit); michael@0: res[1] = _mm_srai_epi16(res[1], bit); michael@0: res[2] = _mm_srai_epi16(res[2], bit); michael@0: res[3] = _mm_srai_epi16(res[3], bit); michael@0: res[4] = _mm_srai_epi16(res[4], bit); michael@0: res[5] = _mm_srai_epi16(res[5], bit); michael@0: res[6] = _mm_srai_epi16(res[6], bit); michael@0: res[7] = _mm_srai_epi16(res[7], bit); michael@0: } michael@0: michael@0: // write 8x8 array michael@0: static INLINE void write_buffer_8x8(int16_t *output, __m128i *res, int stride) { michael@0: _mm_store_si128((__m128i *)(output + 0 * stride), res[0]); michael@0: _mm_store_si128((__m128i *)(output + 1 * stride), res[1]); michael@0: _mm_store_si128((__m128i *)(output + 2 * stride), res[2]); michael@0: _mm_store_si128((__m128i *)(output + 3 * stride), res[3]); michael@0: _mm_store_si128((__m128i *)(output + 4 * stride), res[4]); michael@0: _mm_store_si128((__m128i *)(output + 5 * stride), res[5]); michael@0: _mm_store_si128((__m128i *)(output + 6 * stride), res[6]); michael@0: _mm_store_si128((__m128i *)(output + 7 * stride), res[7]); michael@0: } michael@0: michael@0: // perform in-place transpose michael@0: static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) { michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]); michael@0: const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]); michael@0: const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]); michael@0: const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]); michael@0: const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]); michael@0: const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]); michael@0: const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]); michael@0: const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]); michael@0: // 00 10 01 11 02 12 03 13 michael@0: // 20 30 21 31 22 32 23 33 michael@0: // 04 14 05 15 06 16 07 17 michael@0: // 24 34 25 35 26 36 27 37 michael@0: // 40 50 41 51 42 52 43 53 michael@0: // 60 70 61 71 62 72 63 73 michael@0: // 44 54 45 55 46 56 47 57 michael@0: // 64 74 65 75 66 76 67 77 michael@0: const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); michael@0: const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5); michael@0: const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); michael@0: const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5); michael@0: const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3); michael@0: const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); michael@0: const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3); michael@0: const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); michael@0: // 00 10 20 30 01 11 21 31 michael@0: // 40 50 60 70 41 51 61 71 michael@0: // 02 12 22 32 03 13 23 33 michael@0: // 42 52 62 72 43 53 63 73 michael@0: // 04 14 24 34 05 15 25 35 michael@0: // 44 54 64 74 45 55 65 75 michael@0: // 06 16 26 36 07 17 27 37 michael@0: // 46 56 66 76 47 57 67 77 michael@0: res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1); michael@0: res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1); michael@0: res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3); michael@0: res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3); michael@0: res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5); michael@0: res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5); michael@0: res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7); michael@0: res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7); michael@0: // 00 10 20 30 40 50 60 70 michael@0: // 01 11 21 31 41 51 61 71 michael@0: // 02 12 22 32 42 52 62 72 michael@0: // 03 13 23 33 43 53 63 73 michael@0: // 04 14 24 34 44 54 64 74 michael@0: // 05 15 25 35 45 55 65 75 michael@0: // 06 16 26 36 46 56 66 76 michael@0: // 07 17 27 37 47 57 67 77 michael@0: } michael@0: michael@0: void fdct8_1d_sse2(__m128i *in) { michael@0: // constants michael@0: const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); michael@0: const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); michael@0: const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); michael@0: const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); michael@0: const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); michael@0: const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); michael@0: const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: __m128i u0, u1, u2, u3, u4, u5, u6, u7; michael@0: __m128i v0, v1, v2, v3, v4, v5, v6, v7; michael@0: __m128i s0, s1, s2, s3, s4, s5, s6, s7; michael@0: michael@0: // stage 1 michael@0: s0 = _mm_add_epi16(in[0], in[7]); michael@0: s1 = _mm_add_epi16(in[1], in[6]); michael@0: s2 = _mm_add_epi16(in[2], in[5]); michael@0: s3 = _mm_add_epi16(in[3], in[4]); michael@0: s4 = _mm_sub_epi16(in[3], in[4]); michael@0: s5 = _mm_sub_epi16(in[2], in[5]); michael@0: s6 = _mm_sub_epi16(in[1], in[6]); michael@0: s7 = _mm_sub_epi16(in[0], in[7]); michael@0: michael@0: u0 = _mm_add_epi16(s0, s3); michael@0: u1 = _mm_add_epi16(s1, s2); michael@0: u2 = _mm_sub_epi16(s1, s2); michael@0: u3 = _mm_sub_epi16(s0, s3); michael@0: // interleave and perform butterfly multiplication/addition michael@0: v0 = _mm_unpacklo_epi16(u0, u1); michael@0: v1 = _mm_unpackhi_epi16(u0, u1); michael@0: v2 = _mm_unpacklo_epi16(u2, u3); michael@0: v3 = _mm_unpackhi_epi16(u2, u3); michael@0: michael@0: u0 = _mm_madd_epi16(v0, k__cospi_p16_p16); michael@0: u1 = _mm_madd_epi16(v1, k__cospi_p16_p16); michael@0: u2 = _mm_madd_epi16(v0, k__cospi_p16_m16); michael@0: u3 = _mm_madd_epi16(v1, k__cospi_p16_m16); michael@0: u4 = _mm_madd_epi16(v2, k__cospi_p24_p08); michael@0: u5 = _mm_madd_epi16(v3, k__cospi_p24_p08); michael@0: u6 = _mm_madd_epi16(v2, k__cospi_m08_p24); michael@0: u7 = _mm_madd_epi16(v3, k__cospi_m08_p24); michael@0: michael@0: // shift and rounding michael@0: v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); michael@0: v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); michael@0: v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); michael@0: v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); michael@0: michael@0: u0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: u1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: u2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: u3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: u4 = _mm_srai_epi32(v4, DCT_CONST_BITS); michael@0: u5 = _mm_srai_epi32(v5, DCT_CONST_BITS); michael@0: u6 = _mm_srai_epi32(v6, DCT_CONST_BITS); michael@0: u7 = _mm_srai_epi32(v7, DCT_CONST_BITS); michael@0: michael@0: in[0] = _mm_packs_epi32(u0, u1); michael@0: in[2] = _mm_packs_epi32(u4, u5); michael@0: in[4] = _mm_packs_epi32(u2, u3); michael@0: in[6] = _mm_packs_epi32(u6, u7); michael@0: michael@0: // stage 2 michael@0: // interleave and perform butterfly multiplication/addition michael@0: u0 = _mm_unpacklo_epi16(s6, s5); michael@0: u1 = _mm_unpackhi_epi16(s6, s5); michael@0: v0 = _mm_madd_epi16(u0, k__cospi_p16_m16); michael@0: v1 = _mm_madd_epi16(u1, k__cospi_p16_m16); michael@0: v2 = _mm_madd_epi16(u0, k__cospi_p16_p16); michael@0: v3 = _mm_madd_epi16(u1, k__cospi_p16_p16); michael@0: michael@0: // shift and rounding michael@0: u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING); michael@0: u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING); michael@0: u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING); michael@0: u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING); michael@0: michael@0: v0 = _mm_srai_epi32(u0, DCT_CONST_BITS); michael@0: v1 = _mm_srai_epi32(u1, DCT_CONST_BITS); michael@0: v2 = _mm_srai_epi32(u2, DCT_CONST_BITS); michael@0: v3 = _mm_srai_epi32(u3, DCT_CONST_BITS); michael@0: michael@0: u0 = _mm_packs_epi32(v0, v1); michael@0: u1 = _mm_packs_epi32(v2, v3); michael@0: michael@0: // stage 3 michael@0: s0 = _mm_add_epi16(s4, u0); michael@0: s1 = _mm_sub_epi16(s4, u0); michael@0: s2 = _mm_sub_epi16(s7, u1); michael@0: s3 = _mm_add_epi16(s7, u1); michael@0: michael@0: // stage 4 michael@0: u0 = _mm_unpacklo_epi16(s0, s3); michael@0: u1 = _mm_unpackhi_epi16(s0, s3); michael@0: u2 = _mm_unpacklo_epi16(s1, s2); michael@0: u3 = _mm_unpackhi_epi16(s1, s2); michael@0: michael@0: v0 = _mm_madd_epi16(u0, k__cospi_p28_p04); michael@0: v1 = _mm_madd_epi16(u1, k__cospi_p28_p04); michael@0: v2 = _mm_madd_epi16(u2, k__cospi_p12_p20); michael@0: v3 = _mm_madd_epi16(u3, k__cospi_p12_p20); michael@0: v4 = _mm_madd_epi16(u2, k__cospi_m20_p12); michael@0: v5 = _mm_madd_epi16(u3, k__cospi_m20_p12); michael@0: v6 = _mm_madd_epi16(u0, k__cospi_m04_p28); michael@0: v7 = _mm_madd_epi16(u1, k__cospi_m04_p28); michael@0: michael@0: // shift and rounding michael@0: u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING); michael@0: u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING); michael@0: u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING); michael@0: u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING); michael@0: u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING); michael@0: u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING); michael@0: u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING); michael@0: u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING); michael@0: michael@0: v0 = _mm_srai_epi32(u0, DCT_CONST_BITS); michael@0: v1 = _mm_srai_epi32(u1, DCT_CONST_BITS); michael@0: v2 = _mm_srai_epi32(u2, DCT_CONST_BITS); michael@0: v3 = _mm_srai_epi32(u3, DCT_CONST_BITS); michael@0: v4 = _mm_srai_epi32(u4, DCT_CONST_BITS); michael@0: v5 = _mm_srai_epi32(u5, DCT_CONST_BITS); michael@0: v6 = _mm_srai_epi32(u6, DCT_CONST_BITS); michael@0: v7 = _mm_srai_epi32(u7, DCT_CONST_BITS); michael@0: michael@0: in[1] = _mm_packs_epi32(v0, v1); michael@0: in[3] = _mm_packs_epi32(v4, v5); michael@0: in[5] = _mm_packs_epi32(v2, v3); michael@0: in[7] = _mm_packs_epi32(v6, v7); michael@0: michael@0: // transpose michael@0: array_transpose_8x8(in, in); michael@0: } michael@0: michael@0: void fadst8_1d_sse2(__m128i *in) { michael@0: // Constants michael@0: const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64); michael@0: const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64); michael@0: const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64); michael@0: const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64); michael@0: const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64); michael@0: const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64); michael@0: const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64); michael@0: const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64); michael@0: const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64); michael@0: const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); michael@0: const __m128i k__const_0 = _mm_set1_epi16(0); michael@0: const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: michael@0: __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15; michael@0: __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15; michael@0: __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15; michael@0: __m128i s0, s1, s2, s3, s4, s5, s6, s7; michael@0: __m128i in0, in1, in2, in3, in4, in5, in6, in7; michael@0: michael@0: // properly aligned for butterfly input michael@0: in0 = in[7]; michael@0: in1 = in[0]; michael@0: in2 = in[5]; michael@0: in3 = in[2]; michael@0: in4 = in[3]; michael@0: in5 = in[4]; michael@0: in6 = in[1]; michael@0: in7 = in[6]; michael@0: michael@0: // column transformation michael@0: // stage 1 michael@0: // interleave and multiply/add into 32-bit integer michael@0: s0 = _mm_unpacklo_epi16(in0, in1); michael@0: s1 = _mm_unpackhi_epi16(in0, in1); michael@0: s2 = _mm_unpacklo_epi16(in2, in3); michael@0: s3 = _mm_unpackhi_epi16(in2, in3); michael@0: s4 = _mm_unpacklo_epi16(in4, in5); michael@0: s5 = _mm_unpackhi_epi16(in4, in5); michael@0: s6 = _mm_unpacklo_epi16(in6, in7); michael@0: s7 = _mm_unpackhi_epi16(in6, in7); michael@0: michael@0: u0 = _mm_madd_epi16(s0, k__cospi_p02_p30); michael@0: u1 = _mm_madd_epi16(s1, k__cospi_p02_p30); michael@0: u2 = _mm_madd_epi16(s0, k__cospi_p30_m02); michael@0: u3 = _mm_madd_epi16(s1, k__cospi_p30_m02); michael@0: u4 = _mm_madd_epi16(s2, k__cospi_p10_p22); michael@0: u5 = _mm_madd_epi16(s3, k__cospi_p10_p22); michael@0: u6 = _mm_madd_epi16(s2, k__cospi_p22_m10); michael@0: u7 = _mm_madd_epi16(s3, k__cospi_p22_m10); michael@0: u8 = _mm_madd_epi16(s4, k__cospi_p18_p14); michael@0: u9 = _mm_madd_epi16(s5, k__cospi_p18_p14); michael@0: u10 = _mm_madd_epi16(s4, k__cospi_p14_m18); michael@0: u11 = _mm_madd_epi16(s5, k__cospi_p14_m18); michael@0: u12 = _mm_madd_epi16(s6, k__cospi_p26_p06); michael@0: u13 = _mm_madd_epi16(s7, k__cospi_p26_p06); michael@0: u14 = _mm_madd_epi16(s6, k__cospi_p06_m26); michael@0: u15 = _mm_madd_epi16(s7, k__cospi_p06_m26); michael@0: michael@0: // addition michael@0: w0 = _mm_add_epi32(u0, u8); michael@0: w1 = _mm_add_epi32(u1, u9); michael@0: w2 = _mm_add_epi32(u2, u10); michael@0: w3 = _mm_add_epi32(u3, u11); michael@0: w4 = _mm_add_epi32(u4, u12); michael@0: w5 = _mm_add_epi32(u5, u13); michael@0: w6 = _mm_add_epi32(u6, u14); michael@0: w7 = _mm_add_epi32(u7, u15); michael@0: w8 = _mm_sub_epi32(u0, u8); michael@0: w9 = _mm_sub_epi32(u1, u9); michael@0: w10 = _mm_sub_epi32(u2, u10); michael@0: w11 = _mm_sub_epi32(u3, u11); michael@0: w12 = _mm_sub_epi32(u4, u12); michael@0: w13 = _mm_sub_epi32(u5, u13); michael@0: w14 = _mm_sub_epi32(u6, u14); michael@0: w15 = _mm_sub_epi32(u7, u15); michael@0: michael@0: // shift and rounding michael@0: v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING); michael@0: v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING); michael@0: v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING); michael@0: v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING); michael@0: v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING); michael@0: v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING); michael@0: v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING); michael@0: v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING); michael@0: v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING); michael@0: v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING); michael@0: v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING); michael@0: v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING); michael@0: v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING); michael@0: v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING); michael@0: v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING); michael@0: v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING); michael@0: michael@0: u0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: u1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: u2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: u3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: u4 = _mm_srai_epi32(v4, DCT_CONST_BITS); michael@0: u5 = _mm_srai_epi32(v5, DCT_CONST_BITS); michael@0: u6 = _mm_srai_epi32(v6, DCT_CONST_BITS); michael@0: u7 = _mm_srai_epi32(v7, DCT_CONST_BITS); michael@0: u8 = _mm_srai_epi32(v8, DCT_CONST_BITS); michael@0: u9 = _mm_srai_epi32(v9, DCT_CONST_BITS); michael@0: u10 = _mm_srai_epi32(v10, DCT_CONST_BITS); michael@0: u11 = _mm_srai_epi32(v11, DCT_CONST_BITS); michael@0: u12 = _mm_srai_epi32(v12, DCT_CONST_BITS); michael@0: u13 = _mm_srai_epi32(v13, DCT_CONST_BITS); michael@0: u14 = _mm_srai_epi32(v14, DCT_CONST_BITS); michael@0: u15 = _mm_srai_epi32(v15, DCT_CONST_BITS); michael@0: michael@0: // back to 16-bit and pack 8 integers into __m128i michael@0: in[0] = _mm_packs_epi32(u0, u1); michael@0: in[1] = _mm_packs_epi32(u2, u3); michael@0: in[2] = _mm_packs_epi32(u4, u5); michael@0: in[3] = _mm_packs_epi32(u6, u7); michael@0: in[4] = _mm_packs_epi32(u8, u9); michael@0: in[5] = _mm_packs_epi32(u10, u11); michael@0: in[6] = _mm_packs_epi32(u12, u13); michael@0: in[7] = _mm_packs_epi32(u14, u15); michael@0: michael@0: // stage 2 michael@0: s0 = _mm_add_epi16(in[0], in[2]); michael@0: s1 = _mm_add_epi16(in[1], in[3]); michael@0: s2 = _mm_sub_epi16(in[0], in[2]); michael@0: s3 = _mm_sub_epi16(in[1], in[3]); michael@0: u0 = _mm_unpacklo_epi16(in[4], in[5]); michael@0: u1 = _mm_unpackhi_epi16(in[4], in[5]); michael@0: u2 = _mm_unpacklo_epi16(in[6], in[7]); michael@0: u3 = _mm_unpackhi_epi16(in[6], in[7]); michael@0: michael@0: v0 = _mm_madd_epi16(u0, k__cospi_p08_p24); michael@0: v1 = _mm_madd_epi16(u1, k__cospi_p08_p24); michael@0: v2 = _mm_madd_epi16(u0, k__cospi_p24_m08); michael@0: v3 = _mm_madd_epi16(u1, k__cospi_p24_m08); michael@0: v4 = _mm_madd_epi16(u2, k__cospi_m24_p08); michael@0: v5 = _mm_madd_epi16(u3, k__cospi_m24_p08); michael@0: v6 = _mm_madd_epi16(u2, k__cospi_p08_p24); michael@0: v7 = _mm_madd_epi16(u3, k__cospi_p08_p24); michael@0: michael@0: w0 = _mm_add_epi32(v0, v4); michael@0: w1 = _mm_add_epi32(v1, v5); michael@0: w2 = _mm_add_epi32(v2, v6); michael@0: w3 = _mm_add_epi32(v3, v7); michael@0: w4 = _mm_sub_epi32(v0, v4); michael@0: w5 = _mm_sub_epi32(v1, v5); michael@0: w6 = _mm_sub_epi32(v2, v6); michael@0: w7 = _mm_sub_epi32(v3, v7); michael@0: michael@0: v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING); michael@0: v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING); michael@0: v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING); michael@0: v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING); michael@0: v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING); michael@0: v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING); michael@0: v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING); michael@0: v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING); michael@0: michael@0: u0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: u1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: u2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: u3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: u4 = _mm_srai_epi32(v4, DCT_CONST_BITS); michael@0: u5 = _mm_srai_epi32(v5, DCT_CONST_BITS); michael@0: u6 = _mm_srai_epi32(v6, DCT_CONST_BITS); michael@0: u7 = _mm_srai_epi32(v7, DCT_CONST_BITS); michael@0: michael@0: // back to 16-bit intergers michael@0: s4 = _mm_packs_epi32(u0, u1); michael@0: s5 = _mm_packs_epi32(u2, u3); michael@0: s6 = _mm_packs_epi32(u4, u5); michael@0: s7 = _mm_packs_epi32(u6, u7); michael@0: michael@0: // stage 3 michael@0: u0 = _mm_unpacklo_epi16(s2, s3); michael@0: u1 = _mm_unpackhi_epi16(s2, s3); michael@0: u2 = _mm_unpacklo_epi16(s6, s7); michael@0: u3 = _mm_unpackhi_epi16(s6, s7); michael@0: michael@0: v0 = _mm_madd_epi16(u0, k__cospi_p16_p16); michael@0: v1 = _mm_madd_epi16(u1, k__cospi_p16_p16); michael@0: v2 = _mm_madd_epi16(u0, k__cospi_p16_m16); michael@0: v3 = _mm_madd_epi16(u1, k__cospi_p16_m16); michael@0: v4 = _mm_madd_epi16(u2, k__cospi_p16_p16); michael@0: v5 = _mm_madd_epi16(u3, k__cospi_p16_p16); michael@0: v6 = _mm_madd_epi16(u2, k__cospi_p16_m16); michael@0: v7 = _mm_madd_epi16(u3, k__cospi_p16_m16); michael@0: michael@0: u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING); michael@0: u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING); michael@0: u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING); michael@0: u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING); michael@0: u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING); michael@0: u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING); michael@0: u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING); michael@0: u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING); michael@0: michael@0: v0 = _mm_srai_epi32(u0, DCT_CONST_BITS); michael@0: v1 = _mm_srai_epi32(u1, DCT_CONST_BITS); michael@0: v2 = _mm_srai_epi32(u2, DCT_CONST_BITS); michael@0: v3 = _mm_srai_epi32(u3, DCT_CONST_BITS); michael@0: v4 = _mm_srai_epi32(u4, DCT_CONST_BITS); michael@0: v5 = _mm_srai_epi32(u5, DCT_CONST_BITS); michael@0: v6 = _mm_srai_epi32(u6, DCT_CONST_BITS); michael@0: v7 = _mm_srai_epi32(u7, DCT_CONST_BITS); michael@0: michael@0: s2 = _mm_packs_epi32(v0, v1); michael@0: s3 = _mm_packs_epi32(v2, v3); michael@0: s6 = _mm_packs_epi32(v4, v5); michael@0: s7 = _mm_packs_epi32(v6, v7); michael@0: michael@0: // FIXME(jingning): do subtract using bit inversion? michael@0: in[0] = s0; michael@0: in[1] = _mm_sub_epi16(k__const_0, s4); michael@0: in[2] = s6; michael@0: in[3] = _mm_sub_epi16(k__const_0, s2); michael@0: in[4] = s3; michael@0: in[5] = _mm_sub_epi16(k__const_0, s7); michael@0: in[6] = s5; michael@0: in[7] = _mm_sub_epi16(k__const_0, s1); michael@0: michael@0: // transpose michael@0: array_transpose_8x8(in, in); michael@0: } michael@0: michael@0: void vp9_short_fht8x8_sse2(const int16_t *input, int16_t *output, michael@0: int stride, int tx_type) { michael@0: __m128i in[8]; michael@0: load_buffer_8x8(input, in, stride); michael@0: switch (tx_type) { michael@0: case 0: // DCT_DCT michael@0: fdct8_1d_sse2(in); michael@0: fdct8_1d_sse2(in); michael@0: break; michael@0: case 1: // ADST_DCT michael@0: fadst8_1d_sse2(in); michael@0: fdct8_1d_sse2(in); michael@0: break; michael@0: case 2: // DCT_ADST michael@0: fdct8_1d_sse2(in); michael@0: fadst8_1d_sse2(in); michael@0: break; michael@0: case 3: // ADST_ADST michael@0: fadst8_1d_sse2(in); michael@0: fadst8_1d_sse2(in); michael@0: break; michael@0: default: michael@0: assert(0); michael@0: break; michael@0: } michael@0: right_shift_8x8(in, 1); michael@0: write_buffer_8x8(output, in, 8); michael@0: } michael@0: michael@0: void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride) { michael@0: // The 2D transform is done with two passes which are actually pretty michael@0: // similar. In the first one, we transform the columns and transpose michael@0: // the results. In the second one, we transform the rows. To achieve that, michael@0: // as the first pass results are transposed, we tranpose the columns (that michael@0: // is the transposed rows) and transpose the results (so that it goes back michael@0: // in normal/row positions). michael@0: int pass; michael@0: // We need an intermediate buffer between passes. michael@0: DECLARE_ALIGNED_ARRAY(16, int16_t, intermediate, 256); michael@0: const int16_t *in = input; michael@0: int16_t *out = intermediate; michael@0: // Constants michael@0: // When we use them, in one case, they are all the same. In all others michael@0: // it's a pair of them that we need to repeat four times. This is done michael@0: // by constructing the 32 bit constant corresponding to that pair. michael@0: const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); michael@0: const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); michael@0: const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64); michael@0: const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); michael@0: const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); michael@0: const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); michael@0: const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); michael@0: const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64); michael@0: const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64); michael@0: const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64); michael@0: const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64); michael@0: const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64); michael@0: const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64); michael@0: const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64); michael@0: const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64); michael@0: const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i kOne = _mm_set1_epi16(1); michael@0: // Do the two transform/transpose passes michael@0: for (pass = 0; pass < 2; ++pass) { michael@0: // We process eight columns (transposed rows in second pass) at a time. michael@0: int column_start; michael@0: for (column_start = 0; column_start < 16; column_start += 8) { michael@0: __m128i in00, in01, in02, in03, in04, in05, in06, in07; michael@0: __m128i in08, in09, in10, in11, in12, in13, in14, in15; michael@0: __m128i input0, input1, input2, input3, input4, input5, input6, input7; michael@0: __m128i step1_0, step1_1, step1_2, step1_3; michael@0: __m128i step1_4, step1_5, step1_6, step1_7; michael@0: __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6; michael@0: __m128i step3_0, step3_1, step3_2, step3_3; michael@0: __m128i step3_4, step3_5, step3_6, step3_7; michael@0: __m128i res00, res01, res02, res03, res04, res05, res06, res07; michael@0: __m128i res08, res09, res10, res11, res12, res13, res14, res15; michael@0: // Load and pre-condition input. michael@0: if (0 == pass) { michael@0: in00 = _mm_load_si128((const __m128i *)(in + 0 * stride)); michael@0: in01 = _mm_load_si128((const __m128i *)(in + 1 * stride)); michael@0: in02 = _mm_load_si128((const __m128i *)(in + 2 * stride)); michael@0: in03 = _mm_load_si128((const __m128i *)(in + 3 * stride)); michael@0: in04 = _mm_load_si128((const __m128i *)(in + 4 * stride)); michael@0: in05 = _mm_load_si128((const __m128i *)(in + 5 * stride)); michael@0: in06 = _mm_load_si128((const __m128i *)(in + 6 * stride)); michael@0: in07 = _mm_load_si128((const __m128i *)(in + 7 * stride)); michael@0: in08 = _mm_load_si128((const __m128i *)(in + 8 * stride)); michael@0: in09 = _mm_load_si128((const __m128i *)(in + 9 * stride)); michael@0: in10 = _mm_load_si128((const __m128i *)(in + 10 * stride)); michael@0: in11 = _mm_load_si128((const __m128i *)(in + 11 * stride)); michael@0: in12 = _mm_load_si128((const __m128i *)(in + 12 * stride)); michael@0: in13 = _mm_load_si128((const __m128i *)(in + 13 * stride)); michael@0: in14 = _mm_load_si128((const __m128i *)(in + 14 * stride)); michael@0: in15 = _mm_load_si128((const __m128i *)(in + 15 * stride)); michael@0: // x = x << 2 michael@0: in00 = _mm_slli_epi16(in00, 2); michael@0: in01 = _mm_slli_epi16(in01, 2); michael@0: in02 = _mm_slli_epi16(in02, 2); michael@0: in03 = _mm_slli_epi16(in03, 2); michael@0: in04 = _mm_slli_epi16(in04, 2); michael@0: in05 = _mm_slli_epi16(in05, 2); michael@0: in06 = _mm_slli_epi16(in06, 2); michael@0: in07 = _mm_slli_epi16(in07, 2); michael@0: in08 = _mm_slli_epi16(in08, 2); michael@0: in09 = _mm_slli_epi16(in09, 2); michael@0: in10 = _mm_slli_epi16(in10, 2); michael@0: in11 = _mm_slli_epi16(in11, 2); michael@0: in12 = _mm_slli_epi16(in12, 2); michael@0: in13 = _mm_slli_epi16(in13, 2); michael@0: in14 = _mm_slli_epi16(in14, 2); michael@0: in15 = _mm_slli_epi16(in15, 2); michael@0: } else { michael@0: in00 = _mm_load_si128((const __m128i *)(in + 0 * 16)); michael@0: in01 = _mm_load_si128((const __m128i *)(in + 1 * 16)); michael@0: in02 = _mm_load_si128((const __m128i *)(in + 2 * 16)); michael@0: in03 = _mm_load_si128((const __m128i *)(in + 3 * 16)); michael@0: in04 = _mm_load_si128((const __m128i *)(in + 4 * 16)); michael@0: in05 = _mm_load_si128((const __m128i *)(in + 5 * 16)); michael@0: in06 = _mm_load_si128((const __m128i *)(in + 6 * 16)); michael@0: in07 = _mm_load_si128((const __m128i *)(in + 7 * 16)); michael@0: in08 = _mm_load_si128((const __m128i *)(in + 8 * 16)); michael@0: in09 = _mm_load_si128((const __m128i *)(in + 9 * 16)); michael@0: in10 = _mm_load_si128((const __m128i *)(in + 10 * 16)); michael@0: in11 = _mm_load_si128((const __m128i *)(in + 11 * 16)); michael@0: in12 = _mm_load_si128((const __m128i *)(in + 12 * 16)); michael@0: in13 = _mm_load_si128((const __m128i *)(in + 13 * 16)); michael@0: in14 = _mm_load_si128((const __m128i *)(in + 14 * 16)); michael@0: in15 = _mm_load_si128((const __m128i *)(in + 15 * 16)); michael@0: // x = (x + 1) >> 2 michael@0: in00 = _mm_add_epi16(in00, kOne); michael@0: in01 = _mm_add_epi16(in01, kOne); michael@0: in02 = _mm_add_epi16(in02, kOne); michael@0: in03 = _mm_add_epi16(in03, kOne); michael@0: in04 = _mm_add_epi16(in04, kOne); michael@0: in05 = _mm_add_epi16(in05, kOne); michael@0: in06 = _mm_add_epi16(in06, kOne); michael@0: in07 = _mm_add_epi16(in07, kOne); michael@0: in08 = _mm_add_epi16(in08, kOne); michael@0: in09 = _mm_add_epi16(in09, kOne); michael@0: in10 = _mm_add_epi16(in10, kOne); michael@0: in11 = _mm_add_epi16(in11, kOne); michael@0: in12 = _mm_add_epi16(in12, kOne); michael@0: in13 = _mm_add_epi16(in13, kOne); michael@0: in14 = _mm_add_epi16(in14, kOne); michael@0: in15 = _mm_add_epi16(in15, kOne); michael@0: in00 = _mm_srai_epi16(in00, 2); michael@0: in01 = _mm_srai_epi16(in01, 2); michael@0: in02 = _mm_srai_epi16(in02, 2); michael@0: in03 = _mm_srai_epi16(in03, 2); michael@0: in04 = _mm_srai_epi16(in04, 2); michael@0: in05 = _mm_srai_epi16(in05, 2); michael@0: in06 = _mm_srai_epi16(in06, 2); michael@0: in07 = _mm_srai_epi16(in07, 2); michael@0: in08 = _mm_srai_epi16(in08, 2); michael@0: in09 = _mm_srai_epi16(in09, 2); michael@0: in10 = _mm_srai_epi16(in10, 2); michael@0: in11 = _mm_srai_epi16(in11, 2); michael@0: in12 = _mm_srai_epi16(in12, 2); michael@0: in13 = _mm_srai_epi16(in13, 2); michael@0: in14 = _mm_srai_epi16(in14, 2); michael@0: in15 = _mm_srai_epi16(in15, 2); michael@0: } michael@0: in += 8; michael@0: // Calculate input for the first 8 results. michael@0: { michael@0: input0 = _mm_add_epi16(in00, in15); michael@0: input1 = _mm_add_epi16(in01, in14); michael@0: input2 = _mm_add_epi16(in02, in13); michael@0: input3 = _mm_add_epi16(in03, in12); michael@0: input4 = _mm_add_epi16(in04, in11); michael@0: input5 = _mm_add_epi16(in05, in10); michael@0: input6 = _mm_add_epi16(in06, in09); michael@0: input7 = _mm_add_epi16(in07, in08); michael@0: } michael@0: // Calculate input for the next 8 results. michael@0: { michael@0: step1_0 = _mm_sub_epi16(in07, in08); michael@0: step1_1 = _mm_sub_epi16(in06, in09); michael@0: step1_2 = _mm_sub_epi16(in05, in10); michael@0: step1_3 = _mm_sub_epi16(in04, in11); michael@0: step1_4 = _mm_sub_epi16(in03, in12); michael@0: step1_5 = _mm_sub_epi16(in02, in13); michael@0: step1_6 = _mm_sub_epi16(in01, in14); michael@0: step1_7 = _mm_sub_epi16(in00, in15); michael@0: } michael@0: // Work on the first eight values; fdct8_1d(input, even_results); michael@0: { michael@0: // Add/substract michael@0: const __m128i q0 = _mm_add_epi16(input0, input7); michael@0: const __m128i q1 = _mm_add_epi16(input1, input6); michael@0: const __m128i q2 = _mm_add_epi16(input2, input5); michael@0: const __m128i q3 = _mm_add_epi16(input3, input4); michael@0: const __m128i q4 = _mm_sub_epi16(input3, input4); michael@0: const __m128i q5 = _mm_sub_epi16(input2, input5); michael@0: const __m128i q6 = _mm_sub_epi16(input1, input6); michael@0: const __m128i q7 = _mm_sub_epi16(input0, input7); michael@0: // Work on first four results michael@0: { michael@0: // Add/substract michael@0: const __m128i r0 = _mm_add_epi16(q0, q3); michael@0: const __m128i r1 = _mm_add_epi16(q1, q2); michael@0: const __m128i r2 = _mm_sub_epi16(q1, q2); michael@0: const __m128i r3 = _mm_sub_epi16(q0, q3); michael@0: // Interleave to do the multiply by constants which gets us michael@0: // into 32 bits. michael@0: const __m128i t0 = _mm_unpacklo_epi16(r0, r1); michael@0: const __m128i t1 = _mm_unpackhi_epi16(r0, r1); michael@0: const __m128i t2 = _mm_unpacklo_epi16(r2, r3); michael@0: const __m128i t3 = _mm_unpackhi_epi16(r2, r3); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); michael@0: const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); michael@0: const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); michael@0: const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); michael@0: const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); michael@0: const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); michael@0: const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); michael@0: const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); michael@0: const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); michael@0: const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); michael@0: const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); michael@0: const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); michael@0: const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); michael@0: // Combine michael@0: res00 = _mm_packs_epi32(w0, w1); michael@0: res08 = _mm_packs_epi32(w2, w3); michael@0: res04 = _mm_packs_epi32(w4, w5); michael@0: res12 = _mm_packs_epi32(w6, w7); michael@0: } michael@0: // Work on next four results michael@0: { michael@0: // Interleave to do the multiply by constants which gets us michael@0: // into 32 bits. michael@0: const __m128i d0 = _mm_unpacklo_epi16(q6, q5); michael@0: const __m128i d1 = _mm_unpackhi_epi16(q6, q5); michael@0: const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); michael@0: const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); michael@0: const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); michael@0: const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); michael@0: // dct_const_round_shift michael@0: const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); michael@0: const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); michael@0: const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); michael@0: const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); michael@0: const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); michael@0: const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); michael@0: const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); michael@0: const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); michael@0: // Combine michael@0: const __m128i r0 = _mm_packs_epi32(s0, s1); michael@0: const __m128i r1 = _mm_packs_epi32(s2, s3); michael@0: // Add/substract michael@0: const __m128i x0 = _mm_add_epi16(q4, r0); michael@0: const __m128i x1 = _mm_sub_epi16(q4, r0); michael@0: const __m128i x2 = _mm_sub_epi16(q7, r1); michael@0: const __m128i x3 = _mm_add_epi16(q7, r1); michael@0: // Interleave to do the multiply by constants which gets us michael@0: // into 32 bits. michael@0: const __m128i t0 = _mm_unpacklo_epi16(x0, x3); michael@0: const __m128i t1 = _mm_unpackhi_epi16(x0, x3); michael@0: const __m128i t2 = _mm_unpacklo_epi16(x1, x2); michael@0: const __m128i t3 = _mm_unpackhi_epi16(x1, x2); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); michael@0: const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); michael@0: const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); michael@0: const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); michael@0: const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); michael@0: const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); michael@0: const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); michael@0: const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); michael@0: const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); michael@0: const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); michael@0: const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); michael@0: const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); michael@0: const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); michael@0: // Combine michael@0: res02 = _mm_packs_epi32(w0, w1); michael@0: res14 = _mm_packs_epi32(w2, w3); michael@0: res10 = _mm_packs_epi32(w4, w5); michael@0: res06 = _mm_packs_epi32(w6, w7); michael@0: } michael@0: } michael@0: // Work on the next eight values; step1 -> odd_results michael@0: { michael@0: // step 2 michael@0: { michael@0: const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2); michael@0: const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2); michael@0: const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3); michael@0: const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_m16); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_m16); michael@0: const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_m16); michael@0: const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_m16); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: // Combine michael@0: step2_2 = _mm_packs_epi32(w0, w1); michael@0: step2_3 = _mm_packs_epi32(w2, w3); michael@0: } michael@0: { michael@0: const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2); michael@0: const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2); michael@0: const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3); michael@0: const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); michael@0: const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_p16); michael@0: const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_p16); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: // Combine michael@0: step2_5 = _mm_packs_epi32(w0, w1); michael@0: step2_4 = _mm_packs_epi32(w2, w3); michael@0: } michael@0: // step 3 michael@0: { michael@0: step3_0 = _mm_add_epi16(step1_0, step2_3); michael@0: step3_1 = _mm_add_epi16(step1_1, step2_2); michael@0: step3_2 = _mm_sub_epi16(step1_1, step2_2); michael@0: step3_3 = _mm_sub_epi16(step1_0, step2_3); michael@0: step3_4 = _mm_sub_epi16(step1_7, step2_4); michael@0: step3_5 = _mm_sub_epi16(step1_6, step2_5); michael@0: step3_6 = _mm_add_epi16(step1_6, step2_5); michael@0: step3_7 = _mm_add_epi16(step1_7, step2_4); michael@0: } michael@0: // step 4 michael@0: { michael@0: const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6); michael@0: const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6); michael@0: const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5); michael@0: const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m08_p24); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m08_p24); michael@0: const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m24_m08); michael@0: const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m24_m08); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: // Combine michael@0: step2_1 = _mm_packs_epi32(w0, w1); michael@0: step2_2 = _mm_packs_epi32(w2, w3); michael@0: } michael@0: { michael@0: const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6); michael@0: const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6); michael@0: const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5); michael@0: const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p24_p08); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p24_p08); michael@0: const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m08_p24); michael@0: const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m08_p24); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: // Combine michael@0: step2_6 = _mm_packs_epi32(w0, w1); michael@0: step2_5 = _mm_packs_epi32(w2, w3); michael@0: } michael@0: // step 5 michael@0: { michael@0: step1_0 = _mm_add_epi16(step3_0, step2_1); michael@0: step1_1 = _mm_sub_epi16(step3_0, step2_1); michael@0: step1_2 = _mm_sub_epi16(step3_3, step2_2); michael@0: step1_3 = _mm_add_epi16(step3_3, step2_2); michael@0: step1_4 = _mm_add_epi16(step3_4, step2_5); michael@0: step1_5 = _mm_sub_epi16(step3_4, step2_5); michael@0: step1_6 = _mm_sub_epi16(step3_7, step2_6); michael@0: step1_7 = _mm_add_epi16(step3_7, step2_6); michael@0: } michael@0: // step 6 michael@0: { michael@0: const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7); michael@0: const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7); michael@0: const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6); michael@0: const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p30_p02); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p30_p02); michael@0: const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p14_p18); michael@0: const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p14_p18); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: // Combine michael@0: res01 = _mm_packs_epi32(w0, w1); michael@0: res09 = _mm_packs_epi32(w2, w3); michael@0: } michael@0: { michael@0: const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5); michael@0: const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5); michael@0: const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4); michael@0: const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p22_p10); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p22_p10); michael@0: const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p06_p26); michael@0: const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p06_p26); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: // Combine michael@0: res05 = _mm_packs_epi32(w0, w1); michael@0: res13 = _mm_packs_epi32(w2, w3); michael@0: } michael@0: { michael@0: const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5); michael@0: const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5); michael@0: const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4); michael@0: const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m10_p22); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m10_p22); michael@0: const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m26_p06); michael@0: const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m26_p06); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: // Combine michael@0: res11 = _mm_packs_epi32(w0, w1); michael@0: res03 = _mm_packs_epi32(w2, w3); michael@0: } michael@0: { michael@0: const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7); michael@0: const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7); michael@0: const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6); michael@0: const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6); michael@0: const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m02_p30); michael@0: const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m02_p30); michael@0: const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m18_p14); michael@0: const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m18_p14); michael@0: // dct_const_round_shift michael@0: const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); michael@0: const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); michael@0: const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); michael@0: const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); michael@0: const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); michael@0: const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); michael@0: const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); michael@0: const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); michael@0: // Combine michael@0: res15 = _mm_packs_epi32(w0, w1); michael@0: res07 = _mm_packs_epi32(w2, w3); michael@0: } michael@0: } michael@0: // Transpose the results, do it as two 8x8 transposes. michael@0: { michael@0: // 00 01 02 03 04 05 06 07 michael@0: // 10 11 12 13 14 15 16 17 michael@0: // 20 21 22 23 24 25 26 27 michael@0: // 30 31 32 33 34 35 36 37 michael@0: // 40 41 42 43 44 45 46 47 michael@0: // 50 51 52 53 54 55 56 57 michael@0: // 60 61 62 63 64 65 66 67 michael@0: // 70 71 72 73 74 75 76 77 michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(res00, res01); michael@0: const __m128i tr0_1 = _mm_unpacklo_epi16(res02, res03); michael@0: const __m128i tr0_2 = _mm_unpackhi_epi16(res00, res01); michael@0: const __m128i tr0_3 = _mm_unpackhi_epi16(res02, res03); michael@0: const __m128i tr0_4 = _mm_unpacklo_epi16(res04, res05); michael@0: const __m128i tr0_5 = _mm_unpacklo_epi16(res06, res07); michael@0: const __m128i tr0_6 = _mm_unpackhi_epi16(res04, res05); michael@0: const __m128i tr0_7 = _mm_unpackhi_epi16(res06, res07); michael@0: // 00 10 01 11 02 12 03 13 michael@0: // 20 30 21 31 22 32 23 33 michael@0: // 04 14 05 15 06 16 07 17 michael@0: // 24 34 25 35 26 36 27 37 michael@0: // 40 50 41 51 42 52 43 53 michael@0: // 60 70 61 71 62 72 63 73 michael@0: // 54 54 55 55 56 56 57 57 michael@0: // 64 74 65 75 66 76 67 77 michael@0: const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); michael@0: const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); michael@0: const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); michael@0: const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); michael@0: const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); michael@0: const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); michael@0: const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); michael@0: const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); michael@0: // 00 10 20 30 01 11 21 31 michael@0: // 40 50 60 70 41 51 61 71 michael@0: // 02 12 22 32 03 13 23 33 michael@0: // 42 52 62 72 43 53 63 73 michael@0: // 04 14 24 34 05 15 21 36 michael@0: // 44 54 64 74 45 55 61 76 michael@0: // 06 16 26 36 07 17 27 37 michael@0: // 46 56 66 76 47 57 67 77 michael@0: const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4); michael@0: const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4); michael@0: const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6); michael@0: const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6); michael@0: const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5); michael@0: const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5); michael@0: const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7); michael@0: const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7); michael@0: // 00 10 20 30 40 50 60 70 michael@0: // 01 11 21 31 41 51 61 71 michael@0: // 02 12 22 32 42 52 62 72 michael@0: // 03 13 23 33 43 53 63 73 michael@0: // 04 14 24 34 44 54 64 74 michael@0: // 05 15 25 35 45 55 65 75 michael@0: // 06 16 26 36 46 56 66 76 michael@0: // 07 17 27 37 47 57 67 77 michael@0: _mm_storeu_si128((__m128i *)(out + 0 * 16), tr2_0); michael@0: _mm_storeu_si128((__m128i *)(out + 1 * 16), tr2_1); michael@0: _mm_storeu_si128((__m128i *)(out + 2 * 16), tr2_2); michael@0: _mm_storeu_si128((__m128i *)(out + 3 * 16), tr2_3); michael@0: _mm_storeu_si128((__m128i *)(out + 4 * 16), tr2_4); michael@0: _mm_storeu_si128((__m128i *)(out + 5 * 16), tr2_5); michael@0: _mm_storeu_si128((__m128i *)(out + 6 * 16), tr2_6); michael@0: _mm_storeu_si128((__m128i *)(out + 7 * 16), tr2_7); michael@0: } michael@0: { michael@0: // 00 01 02 03 04 05 06 07 michael@0: // 10 11 12 13 14 15 16 17 michael@0: // 20 21 22 23 24 25 26 27 michael@0: // 30 31 32 33 34 35 36 37 michael@0: // 40 41 42 43 44 45 46 47 michael@0: // 50 51 52 53 54 55 56 57 michael@0: // 60 61 62 63 64 65 66 67 michael@0: // 70 71 72 73 74 75 76 77 michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(res08, res09); michael@0: const __m128i tr0_1 = _mm_unpacklo_epi16(res10, res11); michael@0: const __m128i tr0_2 = _mm_unpackhi_epi16(res08, res09); michael@0: const __m128i tr0_3 = _mm_unpackhi_epi16(res10, res11); michael@0: const __m128i tr0_4 = _mm_unpacklo_epi16(res12, res13); michael@0: const __m128i tr0_5 = _mm_unpacklo_epi16(res14, res15); michael@0: const __m128i tr0_6 = _mm_unpackhi_epi16(res12, res13); michael@0: const __m128i tr0_7 = _mm_unpackhi_epi16(res14, res15); michael@0: // 00 10 01 11 02 12 03 13 michael@0: // 20 30 21 31 22 32 23 33 michael@0: // 04 14 05 15 06 16 07 17 michael@0: // 24 34 25 35 26 36 27 37 michael@0: // 40 50 41 51 42 52 43 53 michael@0: // 60 70 61 71 62 72 63 73 michael@0: // 54 54 55 55 56 56 57 57 michael@0: // 64 74 65 75 66 76 67 77 michael@0: const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); michael@0: const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); michael@0: const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); michael@0: const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); michael@0: const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); michael@0: const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); michael@0: const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); michael@0: const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); michael@0: // 00 10 20 30 01 11 21 31 michael@0: // 40 50 60 70 41 51 61 71 michael@0: // 02 12 22 32 03 13 23 33 michael@0: // 42 52 62 72 43 53 63 73 michael@0: // 04 14 24 34 05 15 21 36 michael@0: // 44 54 64 74 45 55 61 76 michael@0: // 06 16 26 36 07 17 27 37 michael@0: // 46 56 66 76 47 57 67 77 michael@0: const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4); michael@0: const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4); michael@0: const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6); michael@0: const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6); michael@0: const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5); michael@0: const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5); michael@0: const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7); michael@0: const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7); michael@0: // 00 10 20 30 40 50 60 70 michael@0: // 01 11 21 31 41 51 61 71 michael@0: // 02 12 22 32 42 52 62 72 michael@0: // 03 13 23 33 43 53 63 73 michael@0: // 04 14 24 34 44 54 64 74 michael@0: // 05 15 25 35 45 55 65 75 michael@0: // 06 16 26 36 46 56 66 76 michael@0: // 07 17 27 37 47 57 67 77 michael@0: // Store results michael@0: _mm_store_si128((__m128i *)(out + 8 + 0 * 16), tr2_0); michael@0: _mm_store_si128((__m128i *)(out + 8 + 1 * 16), tr2_1); michael@0: _mm_store_si128((__m128i *)(out + 8 + 2 * 16), tr2_2); michael@0: _mm_store_si128((__m128i *)(out + 8 + 3 * 16), tr2_3); michael@0: _mm_store_si128((__m128i *)(out + 8 + 4 * 16), tr2_4); michael@0: _mm_store_si128((__m128i *)(out + 8 + 5 * 16), tr2_5); michael@0: _mm_store_si128((__m128i *)(out + 8 + 6 * 16), tr2_6); michael@0: _mm_store_si128((__m128i *)(out + 8 + 7 * 16), tr2_7); michael@0: } michael@0: out += 8*16; michael@0: } michael@0: // Setup in/out for next pass. michael@0: in = intermediate; michael@0: out = output; michael@0: } michael@0: } michael@0: michael@0: static INLINE void load_buffer_16x16(const int16_t* input, __m128i *in0, michael@0: __m128i *in1, int stride) { michael@0: // load first 8 columns michael@0: load_buffer_8x8(input, in0, stride); michael@0: load_buffer_8x8(input + 8 * stride, in0 + 8, stride); michael@0: michael@0: input += 8; michael@0: // load second 8 columns michael@0: load_buffer_8x8(input, in1, stride); michael@0: load_buffer_8x8(input + 8 * stride, in1 + 8, stride); michael@0: } michael@0: michael@0: static INLINE void write_buffer_16x16(int16_t *output, __m128i *in0, michael@0: __m128i *in1, int stride) { michael@0: // write first 8 columns michael@0: write_buffer_8x8(output, in0, stride); michael@0: write_buffer_8x8(output + 8 * stride, in0 + 8, stride); michael@0: // write second 8 columns michael@0: output += 8; michael@0: write_buffer_8x8(output, in1, stride); michael@0: write_buffer_8x8(output + 8 * stride, in1 + 8, stride); michael@0: } michael@0: michael@0: static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) { michael@0: __m128i tbuf[8]; michael@0: array_transpose_8x8(res0, res0); michael@0: array_transpose_8x8(res1, tbuf); michael@0: array_transpose_8x8(res0 + 8, res1); michael@0: array_transpose_8x8(res1 + 8, res1 + 8); michael@0: michael@0: res0[8] = tbuf[0]; michael@0: res0[9] = tbuf[1]; michael@0: res0[10] = tbuf[2]; michael@0: res0[11] = tbuf[3]; michael@0: res0[12] = tbuf[4]; michael@0: res0[13] = tbuf[5]; michael@0: res0[14] = tbuf[6]; michael@0: res0[15] = tbuf[7]; michael@0: } michael@0: michael@0: static INLINE void right_shift_16x16(__m128i *res0, __m128i *res1) { michael@0: // perform rounding operations michael@0: right_shift_8x8(res0, 2); michael@0: right_shift_8x8(res0 + 8, 2); michael@0: right_shift_8x8(res1, 2); michael@0: right_shift_8x8(res1 + 8, 2); michael@0: } michael@0: michael@0: void fdct16_1d_8col(__m128i *in) { michael@0: // perform 16x16 1-D DCT for 8 columns michael@0: __m128i i[8], s[8], p[8], t[8], u[16], v[16]; michael@0: const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); michael@0: const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64); michael@0: const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); michael@0: const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64); michael@0: const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); michael@0: const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); michael@0: const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); michael@0: const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); michael@0: const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64); michael@0: const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64); michael@0: const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64); michael@0: const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64); michael@0: const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64); michael@0: const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64); michael@0: const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64); michael@0: const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64); michael@0: const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: michael@0: // stage 1 michael@0: i[0] = _mm_add_epi16(in[0], in[15]); michael@0: i[1] = _mm_add_epi16(in[1], in[14]); michael@0: i[2] = _mm_add_epi16(in[2], in[13]); michael@0: i[3] = _mm_add_epi16(in[3], in[12]); michael@0: i[4] = _mm_add_epi16(in[4], in[11]); michael@0: i[5] = _mm_add_epi16(in[5], in[10]); michael@0: i[6] = _mm_add_epi16(in[6], in[9]); michael@0: i[7] = _mm_add_epi16(in[7], in[8]); michael@0: michael@0: s[0] = _mm_sub_epi16(in[7], in[8]); michael@0: s[1] = _mm_sub_epi16(in[6], in[9]); michael@0: s[2] = _mm_sub_epi16(in[5], in[10]); michael@0: s[3] = _mm_sub_epi16(in[4], in[11]); michael@0: s[4] = _mm_sub_epi16(in[3], in[12]); michael@0: s[5] = _mm_sub_epi16(in[2], in[13]); michael@0: s[6] = _mm_sub_epi16(in[1], in[14]); michael@0: s[7] = _mm_sub_epi16(in[0], in[15]); michael@0: michael@0: p[0] = _mm_add_epi16(i[0], i[7]); michael@0: p[1] = _mm_add_epi16(i[1], i[6]); michael@0: p[2] = _mm_add_epi16(i[2], i[5]); michael@0: p[3] = _mm_add_epi16(i[3], i[4]); michael@0: p[4] = _mm_sub_epi16(i[3], i[4]); michael@0: p[5] = _mm_sub_epi16(i[2], i[5]); michael@0: p[6] = _mm_sub_epi16(i[1], i[6]); michael@0: p[7] = _mm_sub_epi16(i[0], i[7]); michael@0: michael@0: u[0] = _mm_add_epi16(p[0], p[3]); michael@0: u[1] = _mm_add_epi16(p[1], p[2]); michael@0: u[2] = _mm_sub_epi16(p[1], p[2]); michael@0: u[3] = _mm_sub_epi16(p[0], p[3]); michael@0: michael@0: v[0] = _mm_unpacklo_epi16(u[0], u[1]); michael@0: v[1] = _mm_unpackhi_epi16(u[0], u[1]); michael@0: v[2] = _mm_unpacklo_epi16(u[2], u[3]); michael@0: v[3] = _mm_unpackhi_epi16(u[2], u[3]); michael@0: michael@0: u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); michael@0: u[1] = _mm_madd_epi16(v[1], k__cospi_p16_p16); michael@0: u[2] = _mm_madd_epi16(v[0], k__cospi_p16_m16); michael@0: u[3] = _mm_madd_epi16(v[1], k__cospi_p16_m16); michael@0: u[4] = _mm_madd_epi16(v[2], k__cospi_p24_p08); michael@0: u[5] = _mm_madd_epi16(v[3], k__cospi_p24_p08); michael@0: u[6] = _mm_madd_epi16(v[2], k__cospi_m08_p24); michael@0: u[7] = _mm_madd_epi16(v[3], k__cospi_m08_p24); michael@0: michael@0: v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); michael@0: v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); michael@0: v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); michael@0: v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); michael@0: v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); michael@0: v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); michael@0: v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); michael@0: v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); michael@0: michael@0: u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); michael@0: u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); michael@0: u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); michael@0: u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); michael@0: u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); michael@0: michael@0: in[0] = _mm_packs_epi32(u[0], u[1]); michael@0: in[4] = _mm_packs_epi32(u[4], u[5]); michael@0: in[8] = _mm_packs_epi32(u[2], u[3]); michael@0: in[12] = _mm_packs_epi32(u[6], u[7]); michael@0: michael@0: u[0] = _mm_unpacklo_epi16(p[5], p[6]); michael@0: u[1] = _mm_unpackhi_epi16(p[5], p[6]); michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16); michael@0: v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16); michael@0: v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); michael@0: u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); michael@0: u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); michael@0: u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); michael@0: michael@0: v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: michael@0: u[0] = _mm_packs_epi32(v[0], v[1]); michael@0: u[1] = _mm_packs_epi32(v[2], v[3]); michael@0: michael@0: t[0] = _mm_add_epi16(p[4], u[0]); michael@0: t[1] = _mm_sub_epi16(p[4], u[0]); michael@0: t[2] = _mm_sub_epi16(p[7], u[1]); michael@0: t[3] = _mm_add_epi16(p[7], u[1]); michael@0: michael@0: u[0] = _mm_unpacklo_epi16(t[0], t[3]); michael@0: u[1] = _mm_unpackhi_epi16(t[0], t[3]); michael@0: u[2] = _mm_unpacklo_epi16(t[1], t[2]); michael@0: u[3] = _mm_unpackhi_epi16(t[1], t[2]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_p28_p04); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_p28_p04); michael@0: v[2] = _mm_madd_epi16(u[2], k__cospi_p12_p20); michael@0: v[3] = _mm_madd_epi16(u[3], k__cospi_p12_p20); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_m20_p12); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_m20_p12); michael@0: v[6] = _mm_madd_epi16(u[0], k__cospi_m04_p28); michael@0: v[7] = _mm_madd_epi16(u[1], k__cospi_m04_p28); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); michael@0: u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); michael@0: u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); michael@0: u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); michael@0: u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); michael@0: u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); michael@0: u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); michael@0: u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); michael@0: michael@0: v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: michael@0: in[2] = _mm_packs_epi32(v[0], v[1]); michael@0: in[6] = _mm_packs_epi32(v[4], v[5]); michael@0: in[10] = _mm_packs_epi32(v[2], v[3]); michael@0: in[14] = _mm_packs_epi32(v[6], v[7]); michael@0: michael@0: // stage 2 michael@0: u[0] = _mm_unpacklo_epi16(s[2], s[5]); michael@0: u[1] = _mm_unpackhi_epi16(s[2], s[5]); michael@0: u[2] = _mm_unpacklo_epi16(s[3], s[4]); michael@0: u[3] = _mm_unpackhi_epi16(s[3], s[4]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16); michael@0: v[2] = _mm_madd_epi16(u[2], k__cospi_m16_p16); michael@0: v[3] = _mm_madd_epi16(u[3], k__cospi_m16_p16); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16); michael@0: v[6] = _mm_madd_epi16(u[0], k__cospi_p16_p16); michael@0: v[7] = _mm_madd_epi16(u[1], k__cospi_p16_p16); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); michael@0: u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); michael@0: u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); michael@0: u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); michael@0: u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); michael@0: u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); michael@0: u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); michael@0: u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); michael@0: michael@0: v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: michael@0: t[2] = _mm_packs_epi32(v[0], v[1]); michael@0: t[3] = _mm_packs_epi32(v[2], v[3]); michael@0: t[4] = _mm_packs_epi32(v[4], v[5]); michael@0: t[5] = _mm_packs_epi32(v[6], v[7]); michael@0: michael@0: // stage 3 michael@0: p[0] = _mm_add_epi16(s[0], t[3]); michael@0: p[1] = _mm_add_epi16(s[1], t[2]); michael@0: p[2] = _mm_sub_epi16(s[1], t[2]); michael@0: p[3] = _mm_sub_epi16(s[0], t[3]); michael@0: p[4] = _mm_sub_epi16(s[7], t[4]); michael@0: p[5] = _mm_sub_epi16(s[6], t[5]); michael@0: p[6] = _mm_add_epi16(s[6], t[5]); michael@0: p[7] = _mm_add_epi16(s[7], t[4]); michael@0: michael@0: // stage 4 michael@0: u[0] = _mm_unpacklo_epi16(p[1], p[6]); michael@0: u[1] = _mm_unpackhi_epi16(p[1], p[6]); michael@0: u[2] = _mm_unpacklo_epi16(p[2], p[5]); michael@0: u[3] = _mm_unpackhi_epi16(p[2], p[5]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24); michael@0: v[2] = _mm_madd_epi16(u[2], k__cospi_m24_m08); michael@0: v[3] = _mm_madd_epi16(u[3], k__cospi_m24_m08); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_m08_p24); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_m08_p24); michael@0: v[6] = _mm_madd_epi16(u[0], k__cospi_p24_p08); michael@0: v[7] = _mm_madd_epi16(u[1], k__cospi_p24_p08); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); michael@0: u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); michael@0: u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); michael@0: u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); michael@0: u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); michael@0: u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); michael@0: u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); michael@0: u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); michael@0: michael@0: v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: michael@0: t[1] = _mm_packs_epi32(v[0], v[1]); michael@0: t[2] = _mm_packs_epi32(v[2], v[3]); michael@0: t[5] = _mm_packs_epi32(v[4], v[5]); michael@0: t[6] = _mm_packs_epi32(v[6], v[7]); michael@0: michael@0: // stage 5 michael@0: s[0] = _mm_add_epi16(p[0], t[1]); michael@0: s[1] = _mm_sub_epi16(p[0], t[1]); michael@0: s[2] = _mm_sub_epi16(p[3], t[2]); michael@0: s[3] = _mm_add_epi16(p[3], t[2]); michael@0: s[4] = _mm_add_epi16(p[4], t[5]); michael@0: s[5] = _mm_sub_epi16(p[4], t[5]); michael@0: s[6] = _mm_sub_epi16(p[7], t[6]); michael@0: s[7] = _mm_add_epi16(p[7], t[6]); michael@0: michael@0: // stage 6 michael@0: u[0] = _mm_unpacklo_epi16(s[0], s[7]); michael@0: u[1] = _mm_unpackhi_epi16(s[0], s[7]); michael@0: u[2] = _mm_unpacklo_epi16(s[1], s[6]); michael@0: u[3] = _mm_unpackhi_epi16(s[1], s[6]); michael@0: u[4] = _mm_unpacklo_epi16(s[2], s[5]); michael@0: u[5] = _mm_unpackhi_epi16(s[2], s[5]); michael@0: u[6] = _mm_unpacklo_epi16(s[3], s[4]); michael@0: u[7] = _mm_unpackhi_epi16(s[3], s[4]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_p30_p02); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_p30_p02); michael@0: v[2] = _mm_madd_epi16(u[2], k__cospi_p14_p18); michael@0: v[3] = _mm_madd_epi16(u[3], k__cospi_p14_p18); michael@0: v[4] = _mm_madd_epi16(u[4], k__cospi_p22_p10); michael@0: v[5] = _mm_madd_epi16(u[5], k__cospi_p22_p10); michael@0: v[6] = _mm_madd_epi16(u[6], k__cospi_p06_p26); michael@0: v[7] = _mm_madd_epi16(u[7], k__cospi_p06_p26); michael@0: v[8] = _mm_madd_epi16(u[6], k__cospi_m26_p06); michael@0: v[9] = _mm_madd_epi16(u[7], k__cospi_m26_p06); michael@0: v[10] = _mm_madd_epi16(u[4], k__cospi_m10_p22); michael@0: v[11] = _mm_madd_epi16(u[5], k__cospi_m10_p22); michael@0: v[12] = _mm_madd_epi16(u[2], k__cospi_m18_p14); michael@0: v[13] = _mm_madd_epi16(u[3], k__cospi_m18_p14); michael@0: v[14] = _mm_madd_epi16(u[0], k__cospi_m02_p30); michael@0: v[15] = _mm_madd_epi16(u[1], k__cospi_m02_p30); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); michael@0: u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); michael@0: u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); michael@0: u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); michael@0: u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); michael@0: u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); michael@0: u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); michael@0: u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); michael@0: u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING); michael@0: u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING); michael@0: u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING); michael@0: u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING); michael@0: u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING); michael@0: u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING); michael@0: u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING); michael@0: u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING); michael@0: michael@0: v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS); michael@0: v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS); michael@0: v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS); michael@0: v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS); michael@0: v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS); michael@0: v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS); michael@0: v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS); michael@0: v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS); michael@0: michael@0: in[1] = _mm_packs_epi32(v[0], v[1]); michael@0: in[9] = _mm_packs_epi32(v[2], v[3]); michael@0: in[5] = _mm_packs_epi32(v[4], v[5]); michael@0: in[13] = _mm_packs_epi32(v[6], v[7]); michael@0: in[3] = _mm_packs_epi32(v[8], v[9]); michael@0: in[11] = _mm_packs_epi32(v[10], v[11]); michael@0: in[7] = _mm_packs_epi32(v[12], v[13]); michael@0: in[15] = _mm_packs_epi32(v[14], v[15]); michael@0: } michael@0: michael@0: void fadst16_1d_8col(__m128i *in) { michael@0: // perform 16x16 1-D ADST for 8 columns michael@0: __m128i s[16], x[16], u[32], v[32]; michael@0: const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64); michael@0: const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64); michael@0: const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64); michael@0: const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64); michael@0: const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64); michael@0: const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64); michael@0: const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64); michael@0: const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64); michael@0: const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64); michael@0: const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64); michael@0: const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64); michael@0: const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64); michael@0: const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64); michael@0: const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64); michael@0: const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64); michael@0: const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64); michael@0: const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64); michael@0: const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64); michael@0: const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64); michael@0: const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64); michael@0: const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64); michael@0: const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64); michael@0: const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64); michael@0: const __m128i k__cospi_m16_m16 = _mm_set1_epi16(-cospi_16_64); michael@0: const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); michael@0: const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64); michael@0: const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i kZero = _mm_set1_epi16(0); michael@0: michael@0: u[0] = _mm_unpacklo_epi16(in[15], in[0]); michael@0: u[1] = _mm_unpackhi_epi16(in[15], in[0]); michael@0: u[2] = _mm_unpacklo_epi16(in[13], in[2]); michael@0: u[3] = _mm_unpackhi_epi16(in[13], in[2]); michael@0: u[4] = _mm_unpacklo_epi16(in[11], in[4]); michael@0: u[5] = _mm_unpackhi_epi16(in[11], in[4]); michael@0: u[6] = _mm_unpacklo_epi16(in[9], in[6]); michael@0: u[7] = _mm_unpackhi_epi16(in[9], in[6]); michael@0: u[8] = _mm_unpacklo_epi16(in[7], in[8]); michael@0: u[9] = _mm_unpackhi_epi16(in[7], in[8]); michael@0: u[10] = _mm_unpacklo_epi16(in[5], in[10]); michael@0: u[11] = _mm_unpackhi_epi16(in[5], in[10]); michael@0: u[12] = _mm_unpacklo_epi16(in[3], in[12]); michael@0: u[13] = _mm_unpackhi_epi16(in[3], in[12]); michael@0: u[14] = _mm_unpacklo_epi16(in[1], in[14]); michael@0: u[15] = _mm_unpackhi_epi16(in[1], in[14]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31); michael@0: v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01); michael@0: v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27); michael@0: v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05); michael@0: v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05); michael@0: v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23); michael@0: v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23); michael@0: v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09); michael@0: v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09); michael@0: v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19); michael@0: v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19); michael@0: v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13); michael@0: v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13); michael@0: v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15); michael@0: v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15); michael@0: v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17); michael@0: v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17); michael@0: v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11); michael@0: v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11); michael@0: v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21); michael@0: v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21); michael@0: v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07); michael@0: v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07); michael@0: v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25); michael@0: v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25); michael@0: v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03); michael@0: v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03); michael@0: v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29); michael@0: v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], v[16]); michael@0: u[1] = _mm_add_epi32(v[1], v[17]); michael@0: u[2] = _mm_add_epi32(v[2], v[18]); michael@0: u[3] = _mm_add_epi32(v[3], v[19]); michael@0: u[4] = _mm_add_epi32(v[4], v[20]); michael@0: u[5] = _mm_add_epi32(v[5], v[21]); michael@0: u[6] = _mm_add_epi32(v[6], v[22]); michael@0: u[7] = _mm_add_epi32(v[7], v[23]); michael@0: u[8] = _mm_add_epi32(v[8], v[24]); michael@0: u[9] = _mm_add_epi32(v[9], v[25]); michael@0: u[10] = _mm_add_epi32(v[10], v[26]); michael@0: u[11] = _mm_add_epi32(v[11], v[27]); michael@0: u[12] = _mm_add_epi32(v[12], v[28]); michael@0: u[13] = _mm_add_epi32(v[13], v[29]); michael@0: u[14] = _mm_add_epi32(v[14], v[30]); michael@0: u[15] = _mm_add_epi32(v[15], v[31]); michael@0: u[16] = _mm_sub_epi32(v[0], v[16]); michael@0: u[17] = _mm_sub_epi32(v[1], v[17]); michael@0: u[18] = _mm_sub_epi32(v[2], v[18]); michael@0: u[19] = _mm_sub_epi32(v[3], v[19]); michael@0: u[20] = _mm_sub_epi32(v[4], v[20]); michael@0: u[21] = _mm_sub_epi32(v[5], v[21]); michael@0: u[22] = _mm_sub_epi32(v[6], v[22]); michael@0: u[23] = _mm_sub_epi32(v[7], v[23]); michael@0: u[24] = _mm_sub_epi32(v[8], v[24]); michael@0: u[25] = _mm_sub_epi32(v[9], v[25]); michael@0: u[26] = _mm_sub_epi32(v[10], v[26]); michael@0: u[27] = _mm_sub_epi32(v[11], v[27]); michael@0: u[28] = _mm_sub_epi32(v[12], v[28]); michael@0: u[29] = _mm_sub_epi32(v[13], v[29]); michael@0: u[30] = _mm_sub_epi32(v[14], v[30]); michael@0: u[31] = _mm_sub_epi32(v[15], v[31]); michael@0: michael@0: v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); michael@0: v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); michael@0: v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); michael@0: v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); michael@0: v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); michael@0: v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); michael@0: v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); michael@0: v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); michael@0: v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); michael@0: v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); michael@0: v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); michael@0: v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); michael@0: v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); michael@0: v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); michael@0: v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); michael@0: v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); michael@0: v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING); michael@0: v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING); michael@0: v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING); michael@0: v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING); michael@0: v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING); michael@0: v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING); michael@0: v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING); michael@0: v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING); michael@0: v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING); michael@0: v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING); michael@0: v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING); michael@0: v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING); michael@0: v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING); michael@0: v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING); michael@0: v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING); michael@0: v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING); michael@0: michael@0: u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); michael@0: u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); michael@0: u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); michael@0: u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); michael@0: u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); michael@0: u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS); michael@0: u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS); michael@0: u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS); michael@0: u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS); michael@0: u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS); michael@0: u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS); michael@0: u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS); michael@0: u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS); michael@0: u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS); michael@0: u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS); michael@0: u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS); michael@0: u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS); michael@0: u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS); michael@0: u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS); michael@0: u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS); michael@0: u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS); michael@0: u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS); michael@0: u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS); michael@0: u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS); michael@0: u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS); michael@0: u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS); michael@0: u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS); michael@0: u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS); michael@0: u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS); michael@0: michael@0: s[0] = _mm_packs_epi32(u[0], u[1]); michael@0: s[1] = _mm_packs_epi32(u[2], u[3]); michael@0: s[2] = _mm_packs_epi32(u[4], u[5]); michael@0: s[3] = _mm_packs_epi32(u[6], u[7]); michael@0: s[4] = _mm_packs_epi32(u[8], u[9]); michael@0: s[5] = _mm_packs_epi32(u[10], u[11]); michael@0: s[6] = _mm_packs_epi32(u[12], u[13]); michael@0: s[7] = _mm_packs_epi32(u[14], u[15]); michael@0: s[8] = _mm_packs_epi32(u[16], u[17]); michael@0: s[9] = _mm_packs_epi32(u[18], u[19]); michael@0: s[10] = _mm_packs_epi32(u[20], u[21]); michael@0: s[11] = _mm_packs_epi32(u[22], u[23]); michael@0: s[12] = _mm_packs_epi32(u[24], u[25]); michael@0: s[13] = _mm_packs_epi32(u[26], u[27]); michael@0: s[14] = _mm_packs_epi32(u[28], u[29]); michael@0: s[15] = _mm_packs_epi32(u[30], u[31]); michael@0: michael@0: // stage 2 michael@0: u[0] = _mm_unpacklo_epi16(s[8], s[9]); michael@0: u[1] = _mm_unpackhi_epi16(s[8], s[9]); michael@0: u[2] = _mm_unpacklo_epi16(s[10], s[11]); michael@0: u[3] = _mm_unpackhi_epi16(s[10], s[11]); michael@0: u[4] = _mm_unpacklo_epi16(s[12], s[13]); michael@0: u[5] = _mm_unpackhi_epi16(s[12], s[13]); michael@0: u[6] = _mm_unpacklo_epi16(s[14], s[15]); michael@0: u[7] = _mm_unpackhi_epi16(s[14], s[15]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28); michael@0: v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04); michael@0: v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12); michael@0: v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20); michael@0: v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20); michael@0: v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04); michael@0: v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04); michael@0: v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28); michael@0: v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28); michael@0: v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20); michael@0: v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20); michael@0: v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12); michael@0: v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], v[8]); michael@0: u[1] = _mm_add_epi32(v[1], v[9]); michael@0: u[2] = _mm_add_epi32(v[2], v[10]); michael@0: u[3] = _mm_add_epi32(v[3], v[11]); michael@0: u[4] = _mm_add_epi32(v[4], v[12]); michael@0: u[5] = _mm_add_epi32(v[5], v[13]); michael@0: u[6] = _mm_add_epi32(v[6], v[14]); michael@0: u[7] = _mm_add_epi32(v[7], v[15]); michael@0: u[8] = _mm_sub_epi32(v[0], v[8]); michael@0: u[9] = _mm_sub_epi32(v[1], v[9]); michael@0: u[10] = _mm_sub_epi32(v[2], v[10]); michael@0: u[11] = _mm_sub_epi32(v[3], v[11]); michael@0: u[12] = _mm_sub_epi32(v[4], v[12]); michael@0: u[13] = _mm_sub_epi32(v[5], v[13]); michael@0: u[14] = _mm_sub_epi32(v[6], v[14]); michael@0: u[15] = _mm_sub_epi32(v[7], v[15]); michael@0: michael@0: v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); michael@0: v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); michael@0: v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); michael@0: v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); michael@0: v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); michael@0: v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); michael@0: v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); michael@0: v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); michael@0: v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); michael@0: v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); michael@0: v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); michael@0: v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); michael@0: v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); michael@0: v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); michael@0: v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); michael@0: v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); michael@0: michael@0: u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); michael@0: u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); michael@0: u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); michael@0: u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); michael@0: u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); michael@0: u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS); michael@0: u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS); michael@0: u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS); michael@0: u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS); michael@0: u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS); michael@0: u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS); michael@0: u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS); michael@0: u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS); michael@0: michael@0: x[0] = _mm_add_epi16(s[0], s[4]); michael@0: x[1] = _mm_add_epi16(s[1], s[5]); michael@0: x[2] = _mm_add_epi16(s[2], s[6]); michael@0: x[3] = _mm_add_epi16(s[3], s[7]); michael@0: x[4] = _mm_sub_epi16(s[0], s[4]); michael@0: x[5] = _mm_sub_epi16(s[1], s[5]); michael@0: x[6] = _mm_sub_epi16(s[2], s[6]); michael@0: x[7] = _mm_sub_epi16(s[3], s[7]); michael@0: x[8] = _mm_packs_epi32(u[0], u[1]); michael@0: x[9] = _mm_packs_epi32(u[2], u[3]); michael@0: x[10] = _mm_packs_epi32(u[4], u[5]); michael@0: x[11] = _mm_packs_epi32(u[6], u[7]); michael@0: x[12] = _mm_packs_epi32(u[8], u[9]); michael@0: x[13] = _mm_packs_epi32(u[10], u[11]); michael@0: x[14] = _mm_packs_epi32(u[12], u[13]); michael@0: x[15] = _mm_packs_epi32(u[14], u[15]); michael@0: michael@0: // stage 3 michael@0: u[0] = _mm_unpacklo_epi16(x[4], x[5]); michael@0: u[1] = _mm_unpackhi_epi16(x[4], x[5]); michael@0: u[2] = _mm_unpacklo_epi16(x[6], x[7]); michael@0: u[3] = _mm_unpackhi_epi16(x[6], x[7]); michael@0: u[4] = _mm_unpacklo_epi16(x[12], x[13]); michael@0: u[5] = _mm_unpackhi_epi16(x[12], x[13]); michael@0: u[6] = _mm_unpacklo_epi16(x[14], x[15]); michael@0: u[7] = _mm_unpackhi_epi16(x[14], x[15]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24); michael@0: v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08); michael@0: v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08); michael@0: v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24); michael@0: v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24); michael@0: v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24); michael@0: v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24); michael@0: v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08); michael@0: v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08); michael@0: v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08); michael@0: v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08); michael@0: v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24); michael@0: v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], v[4]); michael@0: u[1] = _mm_add_epi32(v[1], v[5]); michael@0: u[2] = _mm_add_epi32(v[2], v[6]); michael@0: u[3] = _mm_add_epi32(v[3], v[7]); michael@0: u[4] = _mm_sub_epi32(v[0], v[4]); michael@0: u[5] = _mm_sub_epi32(v[1], v[5]); michael@0: u[6] = _mm_sub_epi32(v[2], v[6]); michael@0: u[7] = _mm_sub_epi32(v[3], v[7]); michael@0: u[8] = _mm_add_epi32(v[8], v[12]); michael@0: u[9] = _mm_add_epi32(v[9], v[13]); michael@0: u[10] = _mm_add_epi32(v[10], v[14]); michael@0: u[11] = _mm_add_epi32(v[11], v[15]); michael@0: u[12] = _mm_sub_epi32(v[8], v[12]); michael@0: u[13] = _mm_sub_epi32(v[9], v[13]); michael@0: u[14] = _mm_sub_epi32(v[10], v[14]); michael@0: u[15] = _mm_sub_epi32(v[11], v[15]); michael@0: michael@0: u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); michael@0: u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); michael@0: u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); michael@0: u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); michael@0: u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); michael@0: u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); michael@0: u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); michael@0: u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); michael@0: u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); michael@0: u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); michael@0: u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); michael@0: u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); michael@0: u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); michael@0: u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); michael@0: u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); michael@0: u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); michael@0: michael@0: v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS); michael@0: v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS); michael@0: v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS); michael@0: v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS); michael@0: v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS); michael@0: v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS); michael@0: v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS); michael@0: v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS); michael@0: michael@0: s[0] = _mm_add_epi16(x[0], x[2]); michael@0: s[1] = _mm_add_epi16(x[1], x[3]); michael@0: s[2] = _mm_sub_epi16(x[0], x[2]); michael@0: s[3] = _mm_sub_epi16(x[1], x[3]); michael@0: s[4] = _mm_packs_epi32(v[0], v[1]); michael@0: s[5] = _mm_packs_epi32(v[2], v[3]); michael@0: s[6] = _mm_packs_epi32(v[4], v[5]); michael@0: s[7] = _mm_packs_epi32(v[6], v[7]); michael@0: s[8] = _mm_add_epi16(x[8], x[10]); michael@0: s[9] = _mm_add_epi16(x[9], x[11]); michael@0: s[10] = _mm_sub_epi16(x[8], x[10]); michael@0: s[11] = _mm_sub_epi16(x[9], x[11]); michael@0: s[12] = _mm_packs_epi32(v[8], v[9]); michael@0: s[13] = _mm_packs_epi32(v[10], v[11]); michael@0: s[14] = _mm_packs_epi32(v[12], v[13]); michael@0: s[15] = _mm_packs_epi32(v[14], v[15]); michael@0: michael@0: // stage 4 michael@0: u[0] = _mm_unpacklo_epi16(s[2], s[3]); michael@0: u[1] = _mm_unpackhi_epi16(s[2], s[3]); michael@0: u[2] = _mm_unpacklo_epi16(s[6], s[7]); michael@0: u[3] = _mm_unpackhi_epi16(s[6], s[7]); michael@0: u[4] = _mm_unpacklo_epi16(s[10], s[11]); michael@0: u[5] = _mm_unpackhi_epi16(s[10], s[11]); michael@0: u[6] = _mm_unpacklo_epi16(s[14], s[15]); michael@0: u[7] = _mm_unpackhi_epi16(s[14], s[15]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16); michael@0: v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16); michael@0: v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16); michael@0: v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16); michael@0: v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16); michael@0: v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16); michael@0: v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16); michael@0: v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16); michael@0: v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16); michael@0: v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16); michael@0: v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16); michael@0: v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16); michael@0: v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16); michael@0: michael@0: u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); michael@0: u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); michael@0: u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); michael@0: u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); michael@0: u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); michael@0: u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); michael@0: u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); michael@0: u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); michael@0: u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING); michael@0: u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING); michael@0: u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING); michael@0: u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING); michael@0: u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING); michael@0: u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING); michael@0: u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING); michael@0: u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING); michael@0: michael@0: v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS); michael@0: v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS); michael@0: v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS); michael@0: v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS); michael@0: v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS); michael@0: v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS); michael@0: v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS); michael@0: v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS); michael@0: michael@0: in[0] = s[0]; michael@0: in[1] = _mm_sub_epi16(kZero, s[8]); michael@0: in[2] = s[12]; michael@0: in[3] = _mm_sub_epi16(kZero, s[4]); michael@0: in[4] = _mm_packs_epi32(v[4], v[5]); michael@0: in[5] = _mm_packs_epi32(v[12], v[13]); michael@0: in[6] = _mm_packs_epi32(v[8], v[9]); michael@0: in[7] = _mm_packs_epi32(v[0], v[1]); michael@0: in[8] = _mm_packs_epi32(v[2], v[3]); michael@0: in[9] = _mm_packs_epi32(v[10], v[11]); michael@0: in[10] = _mm_packs_epi32(v[14], v[15]); michael@0: in[11] = _mm_packs_epi32(v[6], v[7]); michael@0: in[12] = s[5]; michael@0: in[13] = _mm_sub_epi16(kZero, s[13]); michael@0: in[14] = s[9]; michael@0: in[15] = _mm_sub_epi16(kZero, s[1]); michael@0: } michael@0: michael@0: void fdct16_1d_sse2(__m128i *in0, __m128i *in1) { michael@0: fdct16_1d_8col(in0); michael@0: fdct16_1d_8col(in1); michael@0: array_transpose_16x16(in0, in1); michael@0: } michael@0: michael@0: void fadst16_1d_sse2(__m128i *in0, __m128i *in1) { michael@0: fadst16_1d_8col(in0); michael@0: fadst16_1d_8col(in1); michael@0: array_transpose_16x16(in0, in1); michael@0: } michael@0: michael@0: void vp9_short_fht16x16_sse2(const int16_t *input, int16_t *output, michael@0: int stride, int tx_type) { michael@0: __m128i in0[16], in1[16]; michael@0: load_buffer_16x16(input, in0, in1, stride); michael@0: switch (tx_type) { michael@0: case 0: // DCT_DCT michael@0: fdct16_1d_sse2(in0, in1); michael@0: right_shift_16x16(in0, in1); michael@0: fdct16_1d_sse2(in0, in1); michael@0: break; michael@0: case 1: // ADST_DCT michael@0: fadst16_1d_sse2(in0, in1); michael@0: right_shift_16x16(in0, in1); michael@0: fdct16_1d_sse2(in0, in1); michael@0: break; michael@0: case 2: // DCT_ADST michael@0: fdct16_1d_sse2(in0, in1); michael@0: right_shift_16x16(in0, in1); michael@0: fadst16_1d_sse2(in0, in1); michael@0: break; michael@0: case 3: // ADST_ADST michael@0: fadst16_1d_sse2(in0, in1); michael@0: right_shift_16x16(in0, in1); michael@0: fadst16_1d_sse2(in0, in1); michael@0: break; michael@0: default: michael@0: assert(0); michael@0: break; michael@0: } michael@0: write_buffer_16x16(output, in0, in1, 16); michael@0: } michael@0: michael@0: #define FDCT32x32_2D vp9_fdct32x32_rd_sse2 michael@0: #define FDCT32x32_HIGH_PRECISION 0 michael@0: #include "vp9/encoder/x86/vp9_dct32x32_sse2.c" michael@0: #undef FDCT32x32_2D michael@0: #undef FDCT32x32_HIGH_PRECISION michael@0: michael@0: #define FDCT32x32_2D vp9_fdct32x32_sse2 michael@0: #define FDCT32x32_HIGH_PRECISION 1 michael@0: #include "vp9/encoder/x86/vp9_dct32x32_sse2.c" // NOLINT michael@0: #undef FDCT32x32_2D michael@0: #undef FDCT32x32_HIGH_PRECISION