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 michael@0: #include // SSE2 michael@0: #include "./vpx_config.h" michael@0: #include "vpx/vpx_integer.h" michael@0: #include "vp9/common/vp9_common.h" michael@0: #include "vp9/common/vp9_idct.h" michael@0: michael@0: #define RECON_AND_STORE4X4(dest, in_x) \ michael@0: { \ michael@0: __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); \ michael@0: d0 = _mm_unpacklo_epi8(d0, zero); \ michael@0: d0 = _mm_add_epi16(in_x, d0); \ michael@0: d0 = _mm_packus_epi16(d0, d0); \ michael@0: *(int *)dest = _mm_cvtsi128_si32(d0); \ michael@0: dest += stride; \ michael@0: } michael@0: michael@0: void vp9_idct4x4_16_add_sse2(const int16_t *input, uint8_t *dest, int stride) { michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: const __m128i eight = _mm_set1_epi16(8); michael@0: const __m128i cst = _mm_setr_epi16((int16_t)cospi_16_64, (int16_t)cospi_16_64, michael@0: (int16_t)cospi_16_64, (int16_t)-cospi_16_64, michael@0: (int16_t)cospi_24_64, (int16_t)-cospi_8_64, michael@0: (int16_t)cospi_8_64, (int16_t)cospi_24_64); michael@0: const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: __m128i input0, input1, input2, input3; michael@0: michael@0: // Rows michael@0: input0 = _mm_load_si128((const __m128i *)input); michael@0: input2 = _mm_load_si128((const __m128i *)(input + 8)); michael@0: michael@0: // Construct i3, i1, i3, i1, i2, i0, i2, i0 michael@0: input0 = _mm_shufflelo_epi16(input0, 0xd8); michael@0: input0 = _mm_shufflehi_epi16(input0, 0xd8); michael@0: input2 = _mm_shufflelo_epi16(input2, 0xd8); michael@0: input2 = _mm_shufflehi_epi16(input2, 0xd8); michael@0: michael@0: input1 = _mm_unpackhi_epi32(input0, input0); michael@0: input0 = _mm_unpacklo_epi32(input0, input0); michael@0: input3 = _mm_unpackhi_epi32(input2, input2); michael@0: input2 = _mm_unpacklo_epi32(input2, input2); michael@0: michael@0: // Stage 1 michael@0: input0 = _mm_madd_epi16(input0, cst); michael@0: input1 = _mm_madd_epi16(input1, cst); michael@0: input2 = _mm_madd_epi16(input2, cst); michael@0: input3 = _mm_madd_epi16(input3, cst); michael@0: michael@0: input0 = _mm_add_epi32(input0, rounding); michael@0: input1 = _mm_add_epi32(input1, rounding); michael@0: input2 = _mm_add_epi32(input2, rounding); michael@0: input3 = _mm_add_epi32(input3, rounding); michael@0: michael@0: input0 = _mm_srai_epi32(input0, DCT_CONST_BITS); michael@0: input1 = _mm_srai_epi32(input1, DCT_CONST_BITS); michael@0: input2 = _mm_srai_epi32(input2, DCT_CONST_BITS); michael@0: input3 = _mm_srai_epi32(input3, DCT_CONST_BITS); michael@0: michael@0: // Stage 2 michael@0: input0 = _mm_packs_epi32(input0, input1); michael@0: input1 = _mm_packs_epi32(input2, input3); michael@0: michael@0: // Transpose michael@0: input2 = _mm_unpacklo_epi16(input0, input1); michael@0: input3 = _mm_unpackhi_epi16(input0, input1); michael@0: input0 = _mm_unpacklo_epi32(input2, input3); michael@0: input1 = _mm_unpackhi_epi32(input2, input3); michael@0: michael@0: // Switch column2, column 3, and then, we got: michael@0: // input2: column1, column 0; input3: column2, column 3. michael@0: input1 = _mm_shuffle_epi32(input1, 0x4e); michael@0: input2 = _mm_add_epi16(input0, input1); michael@0: input3 = _mm_sub_epi16(input0, input1); michael@0: michael@0: // Columns michael@0: // Construct i3, i1, i3, i1, i2, i0, i2, i0 michael@0: input0 = _mm_unpacklo_epi32(input2, input2); michael@0: input1 = _mm_unpackhi_epi32(input2, input2); michael@0: input2 = _mm_unpackhi_epi32(input3, input3); michael@0: input3 = _mm_unpacklo_epi32(input3, input3); michael@0: michael@0: // Stage 1 michael@0: input0 = _mm_madd_epi16(input0, cst); michael@0: input1 = _mm_madd_epi16(input1, cst); michael@0: input2 = _mm_madd_epi16(input2, cst); michael@0: input3 = _mm_madd_epi16(input3, cst); michael@0: michael@0: input0 = _mm_add_epi32(input0, rounding); michael@0: input1 = _mm_add_epi32(input1, rounding); michael@0: input2 = _mm_add_epi32(input2, rounding); michael@0: input3 = _mm_add_epi32(input3, rounding); michael@0: michael@0: input0 = _mm_srai_epi32(input0, DCT_CONST_BITS); michael@0: input1 = _mm_srai_epi32(input1, DCT_CONST_BITS); michael@0: input2 = _mm_srai_epi32(input2, DCT_CONST_BITS); michael@0: input3 = _mm_srai_epi32(input3, DCT_CONST_BITS); michael@0: michael@0: // Stage 2 michael@0: input0 = _mm_packs_epi32(input0, input2); michael@0: input1 = _mm_packs_epi32(input1, input3); michael@0: michael@0: // Transpose michael@0: input2 = _mm_unpacklo_epi16(input0, input1); michael@0: input3 = _mm_unpackhi_epi16(input0, input1); michael@0: input0 = _mm_unpacklo_epi32(input2, input3); michael@0: input1 = _mm_unpackhi_epi32(input2, input3); michael@0: michael@0: // Switch column2, column 3, and then, we got: michael@0: // input2: column1, column 0; input3: column2, column 3. michael@0: input1 = _mm_shuffle_epi32(input1, 0x4e); michael@0: input2 = _mm_add_epi16(input0, input1); michael@0: input3 = _mm_sub_epi16(input0, input1); michael@0: michael@0: // Final round and shift michael@0: input2 = _mm_add_epi16(input2, eight); michael@0: input3 = _mm_add_epi16(input3, eight); michael@0: michael@0: input2 = _mm_srai_epi16(input2, 4); michael@0: input3 = _mm_srai_epi16(input3, 4); michael@0: michael@0: // Reconstruction and Store michael@0: { michael@0: __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); michael@0: __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2)); michael@0: d0 = _mm_unpacklo_epi32(d0, michael@0: _mm_cvtsi32_si128(*(const int *) (dest + stride))); michael@0: d2 = _mm_unpacklo_epi32(_mm_cvtsi32_si128( michael@0: *(const int *) (dest + stride * 3)), d2); michael@0: d0 = _mm_unpacklo_epi8(d0, zero); michael@0: d2 = _mm_unpacklo_epi8(d2, zero); michael@0: d0 = _mm_add_epi16(d0, input2); michael@0: d2 = _mm_add_epi16(d2, input3); michael@0: d0 = _mm_packus_epi16(d0, d2); michael@0: // store input0 michael@0: *(int *)dest = _mm_cvtsi128_si32(d0); michael@0: // store input1 michael@0: d0 = _mm_srli_si128(d0, 4); michael@0: *(int *)(dest + stride) = _mm_cvtsi128_si32(d0); michael@0: // store input2 michael@0: d0 = _mm_srli_si128(d0, 4); michael@0: *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0); michael@0: // store input3 michael@0: d0 = _mm_srli_si128(d0, 4); michael@0: *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0); michael@0: } michael@0: } michael@0: michael@0: void vp9_idct4x4_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) { michael@0: __m128i dc_value; michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: int a; michael@0: michael@0: a = dct_const_round_shift(input[0] * cospi_16_64); michael@0: a = dct_const_round_shift(a * cospi_16_64); michael@0: a = ROUND_POWER_OF_TWO(a, 4); michael@0: michael@0: dc_value = _mm_set1_epi16(a); michael@0: michael@0: RECON_AND_STORE4X4(dest, dc_value); michael@0: RECON_AND_STORE4X4(dest, dc_value); michael@0: RECON_AND_STORE4X4(dest, dc_value); michael@0: RECON_AND_STORE4X4(dest, dc_value); michael@0: } michael@0: michael@0: static INLINE void transpose_4x4(__m128i *res) { michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]); michael@0: const __m128i tr0_1 = _mm_unpacklo_epi16(res[2], res[3]); 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: 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: static void idct4_1d_sse2(__m128i *in) { michael@0: const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, 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_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i k__cospi_p08_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: __m128i u[8], v[8]; michael@0: michael@0: transpose_4x4(in); michael@0: // stage 1 michael@0: u[0] = _mm_unpacklo_epi16(in[0], in[2]); michael@0: u[1] = _mm_unpacklo_epi16(in[1], in[3]); michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16); michael@0: v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16); michael@0: v[2] = _mm_madd_epi16(u[1], k__cospi_p24_m08); michael@0: v[3] = _mm_madd_epi16(u[1], k__cospi_p08_p24); 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[2]); michael@0: u[1] = _mm_packs_epi32(v[1], v[3]); michael@0: u[2] = _mm_unpackhi_epi64(u[0], u[0]); michael@0: u[3] = _mm_unpackhi_epi64(u[1], u[1]); michael@0: michael@0: // stage 2 michael@0: in[0] = _mm_add_epi16(u[0], u[3]); michael@0: in[1] = _mm_add_epi16(u[1], u[2]); michael@0: in[2] = _mm_sub_epi16(u[1], u[2]); michael@0: in[3] = _mm_sub_epi16(u[0], u[3]); michael@0: } michael@0: michael@0: static void iadst4_1d_sse2(__m128i *in) { michael@0: const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9); michael@0: const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9); michael@0: const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9); michael@0: const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_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], in7; michael@0: michael@0: transpose_4x4(in); michael@0: in7 = _mm_add_epi16(in[0], in[3]); michael@0: in7 = _mm_sub_epi16(in7, in[2]); michael@0: michael@0: u[0] = _mm_unpacklo_epi16(in[0], in[2]); michael@0: u[1] = _mm_unpacklo_epi16(in[1], in[3]); michael@0: u[2] = _mm_unpacklo_epi16(in7, kZero); michael@0: u[3] = _mm_unpacklo_epi16(in[1], kZero); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04); // s0 + s3 michael@0: v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02); // s2 + s5 michael@0: v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03); // x2 michael@0: v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01); // s1 - s4 michael@0: v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04); // s2 - s6 michael@0: v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03); // s2 michael@0: michael@0: u[0] = _mm_add_epi32(v[0], v[1]); michael@0: u[1] = _mm_add_epi32(v[3], v[4]); michael@0: u[2] = v[2]; michael@0: u[3] = _mm_add_epi32(u[0], u[1]); michael@0: u[4] = _mm_slli_epi32(v[5], 2); michael@0: u[5] = _mm_add_epi32(u[3], v[5]); michael@0: u[6] = _mm_sub_epi32(u[5], u[4]); 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: in[2] = _mm_unpackhi_epi64(in[0], in[0]); michael@0: in[3] = _mm_unpackhi_epi64(in[1], in[1]); michael@0: } michael@0: michael@0: void vp9_iht4x4_16_add_sse2(const int16_t *input, uint8_t *dest, int stride, michael@0: int tx_type) { michael@0: __m128i in[4]; michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: const __m128i eight = _mm_set1_epi16(8); michael@0: michael@0: in[0] = _mm_loadl_epi64((const __m128i *)input); michael@0: in[1] = _mm_loadl_epi64((const __m128i *)(input + 4)); michael@0: in[2] = _mm_loadl_epi64((const __m128i *)(input + 8)); michael@0: in[3] = _mm_loadl_epi64((const __m128i *)(input + 12)); michael@0: michael@0: switch (tx_type) { michael@0: case 0: // DCT_DCT michael@0: idct4_1d_sse2(in); michael@0: idct4_1d_sse2(in); michael@0: break; michael@0: case 1: // ADST_DCT michael@0: idct4_1d_sse2(in); michael@0: iadst4_1d_sse2(in); michael@0: break; michael@0: case 2: // DCT_ADST michael@0: iadst4_1d_sse2(in); michael@0: idct4_1d_sse2(in); michael@0: break; michael@0: case 3: // ADST_ADST michael@0: iadst4_1d_sse2(in); michael@0: iadst4_1d_sse2(in); michael@0: break; michael@0: default: michael@0: assert(0); michael@0: break; michael@0: } michael@0: michael@0: // Final round and shift michael@0: in[0] = _mm_add_epi16(in[0], eight); michael@0: in[1] = _mm_add_epi16(in[1], eight); michael@0: in[2] = _mm_add_epi16(in[2], eight); michael@0: in[3] = _mm_add_epi16(in[3], eight); michael@0: michael@0: in[0] = _mm_srai_epi16(in[0], 4); michael@0: in[1] = _mm_srai_epi16(in[1], 4); michael@0: in[2] = _mm_srai_epi16(in[2], 4); michael@0: in[3] = _mm_srai_epi16(in[3], 4); michael@0: michael@0: RECON_AND_STORE4X4(dest, in[0]); michael@0: RECON_AND_STORE4X4(dest, in[1]); michael@0: RECON_AND_STORE4X4(dest, in[2]); michael@0: RECON_AND_STORE4X4(dest, in[3]); michael@0: } michael@0: michael@0: #define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, \ michael@0: out0, out1, out2, out3, out4, out5, out6, out7) \ michael@0: { \ michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \ michael@0: const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \ michael@0: const __m128i tr0_2 = _mm_unpackhi_epi16(in0, in1); \ michael@0: const __m128i tr0_3 = _mm_unpackhi_epi16(in2, in3); \ michael@0: const __m128i tr0_4 = _mm_unpacklo_epi16(in4, in5); \ michael@0: const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7); \ michael@0: const __m128i tr0_6 = _mm_unpackhi_epi16(in4, in5); \ michael@0: const __m128i tr0_7 = _mm_unpackhi_epi16(in6, in7); \ michael@0: \ 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: \ michael@0: out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \ michael@0: out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \ michael@0: out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \ michael@0: out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \ michael@0: out4 = _mm_unpacklo_epi64(tr1_1, tr1_5); \ michael@0: out5 = _mm_unpackhi_epi64(tr1_1, tr1_5); \ michael@0: out6 = _mm_unpacklo_epi64(tr1_3, tr1_7); \ michael@0: out7 = _mm_unpackhi_epi64(tr1_3, tr1_7); \ michael@0: } michael@0: michael@0: #define TRANSPOSE_4X8(in0, in1, in2, in3, in4, in5, in6, in7, \ michael@0: out0, out1, out2, out3, out4, out5, out6, out7) \ michael@0: { \ michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \ michael@0: const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \ michael@0: const __m128i tr0_4 = _mm_unpacklo_epi16(in4, in5); \ michael@0: const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7); \ michael@0: \ michael@0: const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \ michael@0: const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \ michael@0: const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \ michael@0: const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \ michael@0: \ michael@0: out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \ michael@0: out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \ michael@0: out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \ michael@0: out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \ michael@0: out4 = out5 = out6 = out7 = zero; \ michael@0: } michael@0: michael@0: #define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1, out2, out3) \ michael@0: { \ michael@0: const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \ michael@0: const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \ michael@0: const __m128i tr0_2 = _mm_unpackhi_epi16(in0, in1); \ michael@0: const __m128i tr0_3 = _mm_unpackhi_epi16(in2, in3); \ michael@0: \ michael@0: in0 = _mm_unpacklo_epi32(tr0_0, tr0_1); /* i1 i0 */ \ michael@0: in1 = _mm_unpackhi_epi32(tr0_0, tr0_1); /* i3 i2 */ \ michael@0: in2 = _mm_unpacklo_epi32(tr0_2, tr0_3); /* i5 i4 */ \ michael@0: in3 = _mm_unpackhi_epi32(tr0_2, tr0_3); /* i7 i6 */ \ michael@0: } michael@0: michael@0: // Define Macro for multiplying elements by constants and adding them together. michael@0: #define MULTIPLICATION_AND_ADD(lo_0, hi_0, lo_1, hi_1, \ michael@0: cst0, cst1, cst2, cst3, res0, res1, res2, res3) \ michael@0: { \ michael@0: tmp0 = _mm_madd_epi16(lo_0, cst0); \ michael@0: tmp1 = _mm_madd_epi16(hi_0, cst0); \ michael@0: tmp2 = _mm_madd_epi16(lo_0, cst1); \ michael@0: tmp3 = _mm_madd_epi16(hi_0, cst1); \ michael@0: tmp4 = _mm_madd_epi16(lo_1, cst2); \ michael@0: tmp5 = _mm_madd_epi16(hi_1, cst2); \ michael@0: tmp6 = _mm_madd_epi16(lo_1, cst3); \ michael@0: tmp7 = _mm_madd_epi16(hi_1, cst3); \ michael@0: \ michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); \ michael@0: tmp1 = _mm_add_epi32(tmp1, rounding); \ michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); \ michael@0: tmp3 = _mm_add_epi32(tmp3, rounding); \ michael@0: tmp4 = _mm_add_epi32(tmp4, rounding); \ michael@0: tmp5 = _mm_add_epi32(tmp5, rounding); \ michael@0: tmp6 = _mm_add_epi32(tmp6, rounding); \ michael@0: tmp7 = _mm_add_epi32(tmp7, rounding); \ michael@0: \ michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ michael@0: tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ michael@0: tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ michael@0: tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); \ michael@0: tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); \ michael@0: tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); \ michael@0: tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); \ michael@0: \ michael@0: res0 = _mm_packs_epi32(tmp0, tmp1); \ michael@0: res1 = _mm_packs_epi32(tmp2, tmp3); \ michael@0: res2 = _mm_packs_epi32(tmp4, tmp5); \ michael@0: res3 = _mm_packs_epi32(tmp6, tmp7); \ michael@0: } michael@0: michael@0: #define IDCT8_1D \ michael@0: /* Stage1 */ \ michael@0: { \ michael@0: const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); \ michael@0: const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); \ michael@0: const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5); \ michael@0: const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_17, hi_17, lo_35, hi_35, stg1_0, \ michael@0: stg1_1, stg1_2, stg1_3, stp1_4, \ michael@0: stp1_7, stp1_5, stp1_6) \ michael@0: } \ michael@0: \ michael@0: /* Stage2 */ \ michael@0: { \ michael@0: const __m128i lo_04 = _mm_unpacklo_epi16(in0, in4); \ michael@0: const __m128i hi_04 = _mm_unpackhi_epi16(in0, in4); \ michael@0: const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6); \ michael@0: const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_04, hi_04, lo_26, hi_26, stg2_0, \ michael@0: stg2_1, stg2_2, stg2_3, stp2_0, \ michael@0: stp2_1, stp2_2, stp2_3) \ michael@0: \ michael@0: stp2_4 = _mm_adds_epi16(stp1_4, stp1_5); \ michael@0: stp2_5 = _mm_subs_epi16(stp1_4, stp1_5); \ michael@0: stp2_6 = _mm_subs_epi16(stp1_7, stp1_6); \ michael@0: stp2_7 = _mm_adds_epi16(stp1_7, stp1_6); \ michael@0: } \ michael@0: \ michael@0: /* Stage3 */ \ michael@0: { \ michael@0: const __m128i lo_56 = _mm_unpacklo_epi16(stp2_6, stp2_5); \ michael@0: const __m128i hi_56 = _mm_unpackhi_epi16(stp2_6, stp2_5); \ michael@0: \ michael@0: stp1_0 = _mm_adds_epi16(stp2_0, stp2_3); \ michael@0: stp1_1 = _mm_adds_epi16(stp2_1, stp2_2); \ michael@0: stp1_2 = _mm_subs_epi16(stp2_1, stp2_2); \ michael@0: stp1_3 = _mm_subs_epi16(stp2_0, stp2_3); \ michael@0: \ michael@0: tmp0 = _mm_madd_epi16(lo_56, stg2_1); \ michael@0: tmp1 = _mm_madd_epi16(hi_56, stg2_1); \ michael@0: tmp2 = _mm_madd_epi16(lo_56, stg2_0); \ michael@0: tmp3 = _mm_madd_epi16(hi_56, stg2_0); \ michael@0: \ michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); \ michael@0: tmp1 = _mm_add_epi32(tmp1, rounding); \ michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); \ michael@0: tmp3 = _mm_add_epi32(tmp3, rounding); \ michael@0: \ michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ michael@0: tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ michael@0: tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ michael@0: \ michael@0: stp1_5 = _mm_packs_epi32(tmp0, tmp1); \ michael@0: stp1_6 = _mm_packs_epi32(tmp2, tmp3); \ michael@0: } \ michael@0: \ michael@0: /* Stage4 */ \ michael@0: in0 = _mm_adds_epi16(stp1_0, stp2_7); \ michael@0: in1 = _mm_adds_epi16(stp1_1, stp1_6); \ michael@0: in2 = _mm_adds_epi16(stp1_2, stp1_5); \ michael@0: in3 = _mm_adds_epi16(stp1_3, stp2_4); \ michael@0: in4 = _mm_subs_epi16(stp1_3, stp2_4); \ michael@0: in5 = _mm_subs_epi16(stp1_2, stp1_5); \ michael@0: in6 = _mm_subs_epi16(stp1_1, stp1_6); \ michael@0: in7 = _mm_subs_epi16(stp1_0, stp2_7); michael@0: michael@0: #define RECON_AND_STORE(dest, in_x) \ michael@0: { \ michael@0: __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \ michael@0: d0 = _mm_unpacklo_epi8(d0, zero); \ michael@0: d0 = _mm_add_epi16(in_x, d0); \ michael@0: d0 = _mm_packus_epi16(d0, d0); \ michael@0: _mm_storel_epi64((__m128i *)(dest), d0); \ michael@0: dest += stride; \ michael@0: } michael@0: michael@0: void vp9_idct8x8_64_add_sse2(const int16_t *input, uint8_t *dest, int stride) { michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i final_rounding = _mm_set1_epi16(1<<4); michael@0: const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); michael@0: const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64); michael@0: const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64); michael@0: const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64); michael@0: const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64); michael@0: const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: michael@0: __m128i in0, in1, in2, in3, in4, in5, in6, in7; michael@0: __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7; michael@0: __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7; michael@0: __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; michael@0: int i; michael@0: michael@0: // Load input data. michael@0: in0 = _mm_load_si128((const __m128i *)input); michael@0: in1 = _mm_load_si128((const __m128i *)(input + 8 * 1)); michael@0: in2 = _mm_load_si128((const __m128i *)(input + 8 * 2)); michael@0: in3 = _mm_load_si128((const __m128i *)(input + 8 * 3)); michael@0: in4 = _mm_load_si128((const __m128i *)(input + 8 * 4)); michael@0: in5 = _mm_load_si128((const __m128i *)(input + 8 * 5)); michael@0: in6 = _mm_load_si128((const __m128i *)(input + 8 * 6)); michael@0: in7 = _mm_load_si128((const __m128i *)(input + 8 * 7)); michael@0: michael@0: // 2-D michael@0: for (i = 0; i < 2; i++) { michael@0: // 8x8 Transpose is copied from vp9_fdct8x8_sse2() michael@0: TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, michael@0: in4, in5, in6, in7); michael@0: michael@0: // 4-stage 1D idct8x8 michael@0: IDCT8_1D michael@0: } michael@0: michael@0: // Final rounding and shift michael@0: in0 = _mm_adds_epi16(in0, final_rounding); michael@0: in1 = _mm_adds_epi16(in1, final_rounding); michael@0: in2 = _mm_adds_epi16(in2, final_rounding); michael@0: in3 = _mm_adds_epi16(in3, final_rounding); michael@0: in4 = _mm_adds_epi16(in4, final_rounding); michael@0: in5 = _mm_adds_epi16(in5, final_rounding); michael@0: in6 = _mm_adds_epi16(in6, final_rounding); michael@0: in7 = _mm_adds_epi16(in7, final_rounding); michael@0: michael@0: in0 = _mm_srai_epi16(in0, 5); michael@0: in1 = _mm_srai_epi16(in1, 5); michael@0: in2 = _mm_srai_epi16(in2, 5); michael@0: in3 = _mm_srai_epi16(in3, 5); michael@0: in4 = _mm_srai_epi16(in4, 5); michael@0: in5 = _mm_srai_epi16(in5, 5); michael@0: in6 = _mm_srai_epi16(in6, 5); michael@0: in7 = _mm_srai_epi16(in7, 5); michael@0: michael@0: RECON_AND_STORE(dest, in0); michael@0: RECON_AND_STORE(dest, in1); michael@0: RECON_AND_STORE(dest, in2); michael@0: RECON_AND_STORE(dest, in3); michael@0: RECON_AND_STORE(dest, in4); michael@0: RECON_AND_STORE(dest, in5); michael@0: RECON_AND_STORE(dest, in6); michael@0: RECON_AND_STORE(dest, in7); michael@0: } michael@0: michael@0: void vp9_idct8x8_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) { michael@0: __m128i dc_value; michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: int a; michael@0: michael@0: a = dct_const_round_shift(input[0] * cospi_16_64); michael@0: a = dct_const_round_shift(a * cospi_16_64); michael@0: a = ROUND_POWER_OF_TWO(a, 5); michael@0: michael@0: dc_value = _mm_set1_epi16(a); michael@0: michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: } michael@0: michael@0: // perform 8x8 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: 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: 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: } michael@0: michael@0: static void idct8_1d_sse2(__m128i *in) { michael@0: const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); michael@0: const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64); michael@0: const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64); michael@0: const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64); michael@0: const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64); michael@0: const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: michael@0: __m128i in0, in1, in2, in3, in4, in5, in6, in7; michael@0: __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7; michael@0: __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7; michael@0: __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; michael@0: michael@0: in0 = in[0]; michael@0: in1 = in[1]; michael@0: in2 = in[2]; michael@0: in3 = in[3]; michael@0: in4 = in[4]; michael@0: in5 = in[5]; michael@0: in6 = in[6]; michael@0: in7 = in[7]; michael@0: michael@0: // 8x8 Transpose is copied from vp9_fdct8x8_sse2() michael@0: TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, michael@0: in4, in5, in6, in7); michael@0: michael@0: // 4-stage 1D idct8x8 michael@0: IDCT8_1D michael@0: in[0] = in0; michael@0: in[1] = in1; michael@0: in[2] = in2; michael@0: in[3] = in3; michael@0: in[4] = in4; michael@0: in[5] = in5; michael@0: in[6] = in6; michael@0: in[7] = in7; michael@0: } michael@0: michael@0: static void iadst8_1d_sse2(__m128i *in) { 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: // transpose michael@0: array_transpose_8x8(in, in); 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: 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: michael@0: michael@0: void vp9_iht8x8_64_add_sse2(const int16_t *input, uint8_t *dest, int stride, michael@0: int tx_type) { michael@0: __m128i in[8]; michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: const __m128i final_rounding = _mm_set1_epi16(1<<4); michael@0: michael@0: // load input data michael@0: in[0] = _mm_load_si128((const __m128i *)input); michael@0: in[1] = _mm_load_si128((const __m128i *)(input + 8 * 1)); michael@0: in[2] = _mm_load_si128((const __m128i *)(input + 8 * 2)); michael@0: in[3] = _mm_load_si128((const __m128i *)(input + 8 * 3)); michael@0: in[4] = _mm_load_si128((const __m128i *)(input + 8 * 4)); michael@0: in[5] = _mm_load_si128((const __m128i *)(input + 8 * 5)); michael@0: in[6] = _mm_load_si128((const __m128i *)(input + 8 * 6)); michael@0: in[7] = _mm_load_si128((const __m128i *)(input + 8 * 7)); michael@0: michael@0: switch (tx_type) { michael@0: case 0: // DCT_DCT michael@0: idct8_1d_sse2(in); michael@0: idct8_1d_sse2(in); michael@0: break; michael@0: case 1: // ADST_DCT michael@0: idct8_1d_sse2(in); michael@0: iadst8_1d_sse2(in); michael@0: break; michael@0: case 2: // DCT_ADST michael@0: iadst8_1d_sse2(in); michael@0: idct8_1d_sse2(in); michael@0: break; michael@0: case 3: // ADST_ADST michael@0: iadst8_1d_sse2(in); michael@0: iadst8_1d_sse2(in); michael@0: break; michael@0: default: michael@0: assert(0); michael@0: break; michael@0: } michael@0: michael@0: // Final rounding and shift michael@0: in[0] = _mm_adds_epi16(in[0], final_rounding); michael@0: in[1] = _mm_adds_epi16(in[1], final_rounding); michael@0: in[2] = _mm_adds_epi16(in[2], final_rounding); michael@0: in[3] = _mm_adds_epi16(in[3], final_rounding); michael@0: in[4] = _mm_adds_epi16(in[4], final_rounding); michael@0: in[5] = _mm_adds_epi16(in[5], final_rounding); michael@0: in[6] = _mm_adds_epi16(in[6], final_rounding); michael@0: in[7] = _mm_adds_epi16(in[7], final_rounding); michael@0: michael@0: in[0] = _mm_srai_epi16(in[0], 5); michael@0: in[1] = _mm_srai_epi16(in[1], 5); michael@0: in[2] = _mm_srai_epi16(in[2], 5); michael@0: in[3] = _mm_srai_epi16(in[3], 5); michael@0: in[4] = _mm_srai_epi16(in[4], 5); michael@0: in[5] = _mm_srai_epi16(in[5], 5); michael@0: in[6] = _mm_srai_epi16(in[6], 5); michael@0: in[7] = _mm_srai_epi16(in[7], 5); michael@0: michael@0: RECON_AND_STORE(dest, in[0]); michael@0: RECON_AND_STORE(dest, in[1]); michael@0: RECON_AND_STORE(dest, in[2]); michael@0: RECON_AND_STORE(dest, in[3]); michael@0: RECON_AND_STORE(dest, in[4]); michael@0: RECON_AND_STORE(dest, in[5]); michael@0: RECON_AND_STORE(dest, in[6]); michael@0: RECON_AND_STORE(dest, in[7]); michael@0: } michael@0: michael@0: void vp9_idct8x8_10_add_sse2(const int16_t *input, uint8_t *dest, int stride) { michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i final_rounding = _mm_set1_epi16(1<<4); michael@0: const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); michael@0: const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64); michael@0: const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64); michael@0: const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64); michael@0: const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64); michael@0: const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); michael@0: michael@0: __m128i in0, in1, in2, in3, in4, in5, in6, in7; michael@0: __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7; michael@0: __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7; michael@0: __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; michael@0: michael@0: // Rows. Load 4-row input data. michael@0: in0 = _mm_load_si128((const __m128i *)input); michael@0: in1 = _mm_load_si128((const __m128i *)(input + 8 * 1)); michael@0: in2 = _mm_load_si128((const __m128i *)(input + 8 * 2)); michael@0: in3 = _mm_load_si128((const __m128i *)(input + 8 * 3)); michael@0: michael@0: // 8x4 Transpose michael@0: TRANSPOSE_8X4(in0, in1, in2, in3, in0, in1, in2, in3) michael@0: michael@0: // Stage1 michael@0: { //NOLINT michael@0: const __m128i lo_17 = _mm_unpackhi_epi16(in0, in3); michael@0: const __m128i lo_35 = _mm_unpackhi_epi16(in1, in2); michael@0: michael@0: tmp0 = _mm_madd_epi16(lo_17, stg1_0); michael@0: tmp2 = _mm_madd_epi16(lo_17, stg1_1); michael@0: tmp4 = _mm_madd_epi16(lo_35, stg1_2); michael@0: tmp6 = _mm_madd_epi16(lo_35, stg1_3); michael@0: michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); michael@0: tmp4 = _mm_add_epi32(tmp4, rounding); michael@0: tmp6 = _mm_add_epi32(tmp6, rounding); michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); michael@0: tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); michael@0: tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); michael@0: michael@0: stp1_4 = _mm_packs_epi32(tmp0, zero); michael@0: stp1_7 = _mm_packs_epi32(tmp2, zero); michael@0: stp1_5 = _mm_packs_epi32(tmp4, zero); michael@0: stp1_6 = _mm_packs_epi32(tmp6, zero); michael@0: } michael@0: michael@0: // Stage2 michael@0: { //NOLINT michael@0: const __m128i lo_04 = _mm_unpacklo_epi16(in0, in2); michael@0: const __m128i lo_26 = _mm_unpacklo_epi16(in1, in3); michael@0: michael@0: tmp0 = _mm_madd_epi16(lo_04, stg2_0); michael@0: tmp2 = _mm_madd_epi16(lo_04, stg2_1); michael@0: tmp4 = _mm_madd_epi16(lo_26, stg2_2); michael@0: tmp6 = _mm_madd_epi16(lo_26, stg2_3); michael@0: michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); michael@0: tmp4 = _mm_add_epi32(tmp4, rounding); michael@0: tmp6 = _mm_add_epi32(tmp6, rounding); michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); michael@0: tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); michael@0: tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); michael@0: michael@0: stp2_0 = _mm_packs_epi32(tmp0, zero); michael@0: stp2_1 = _mm_packs_epi32(tmp2, zero); michael@0: stp2_2 = _mm_packs_epi32(tmp4, zero); michael@0: stp2_3 = _mm_packs_epi32(tmp6, zero); michael@0: michael@0: stp2_4 = _mm_adds_epi16(stp1_4, stp1_5); michael@0: stp2_5 = _mm_subs_epi16(stp1_4, stp1_5); michael@0: stp2_6 = _mm_subs_epi16(stp1_7, stp1_6); michael@0: stp2_7 = _mm_adds_epi16(stp1_7, stp1_6); michael@0: } michael@0: michael@0: // Stage3 michael@0: { //NOLINT michael@0: const __m128i lo_56 = _mm_unpacklo_epi16(stp2_5, stp2_6); michael@0: stp1_0 = _mm_adds_epi16(stp2_0, stp2_3); michael@0: stp1_1 = _mm_adds_epi16(stp2_1, stp2_2); michael@0: stp1_2 = _mm_subs_epi16(stp2_1, stp2_2); michael@0: stp1_3 = _mm_subs_epi16(stp2_0, stp2_3); michael@0: michael@0: tmp0 = _mm_madd_epi16(lo_56, stg3_0); michael@0: tmp2 = _mm_madd_epi16(lo_56, stg2_0); // stg3_1 = stg2_0 michael@0: michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); michael@0: michael@0: stp1_5 = _mm_packs_epi32(tmp0, zero); michael@0: stp1_6 = _mm_packs_epi32(tmp2, zero); michael@0: } michael@0: michael@0: // Stage4 michael@0: in0 = _mm_adds_epi16(stp1_0, stp2_7); michael@0: in1 = _mm_adds_epi16(stp1_1, stp1_6); michael@0: in2 = _mm_adds_epi16(stp1_2, stp1_5); michael@0: in3 = _mm_adds_epi16(stp1_3, stp2_4); michael@0: in4 = _mm_subs_epi16(stp1_3, stp2_4); michael@0: in5 = _mm_subs_epi16(stp1_2, stp1_5); michael@0: in6 = _mm_subs_epi16(stp1_1, stp1_6); michael@0: in7 = _mm_subs_epi16(stp1_0, stp2_7); michael@0: michael@0: // Columns. 4x8 Transpose michael@0: TRANSPOSE_4X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, michael@0: in4, in5, in6, in7) michael@0: michael@0: // 1D idct8x8 michael@0: IDCT8_1D michael@0: michael@0: // Final rounding and shift michael@0: in0 = _mm_adds_epi16(in0, final_rounding); michael@0: in1 = _mm_adds_epi16(in1, final_rounding); michael@0: in2 = _mm_adds_epi16(in2, final_rounding); michael@0: in3 = _mm_adds_epi16(in3, final_rounding); michael@0: in4 = _mm_adds_epi16(in4, final_rounding); michael@0: in5 = _mm_adds_epi16(in5, final_rounding); michael@0: in6 = _mm_adds_epi16(in6, final_rounding); michael@0: in7 = _mm_adds_epi16(in7, final_rounding); michael@0: michael@0: in0 = _mm_srai_epi16(in0, 5); michael@0: in1 = _mm_srai_epi16(in1, 5); michael@0: in2 = _mm_srai_epi16(in2, 5); michael@0: in3 = _mm_srai_epi16(in3, 5); michael@0: in4 = _mm_srai_epi16(in4, 5); michael@0: in5 = _mm_srai_epi16(in5, 5); michael@0: in6 = _mm_srai_epi16(in6, 5); michael@0: in7 = _mm_srai_epi16(in7, 5); michael@0: michael@0: RECON_AND_STORE(dest, in0); michael@0: RECON_AND_STORE(dest, in1); michael@0: RECON_AND_STORE(dest, in2); michael@0: RECON_AND_STORE(dest, in3); michael@0: RECON_AND_STORE(dest, in4); michael@0: RECON_AND_STORE(dest, in5); michael@0: RECON_AND_STORE(dest, in6); michael@0: RECON_AND_STORE(dest, in7); michael@0: } michael@0: michael@0: #define IDCT16_1D \ michael@0: /* Stage2 */ \ michael@0: { \ michael@0: const __m128i lo_1_15 = _mm_unpacklo_epi16(in1, in15); \ michael@0: const __m128i hi_1_15 = _mm_unpackhi_epi16(in1, in15); \ michael@0: const __m128i lo_9_7 = _mm_unpacklo_epi16(in9, in7); \ michael@0: const __m128i hi_9_7 = _mm_unpackhi_epi16(in9, in7); \ michael@0: const __m128i lo_5_11 = _mm_unpacklo_epi16(in5, in11); \ michael@0: const __m128i hi_5_11 = _mm_unpackhi_epi16(in5, in11); \ michael@0: const __m128i lo_13_3 = _mm_unpacklo_epi16(in13, in3); \ michael@0: const __m128i hi_13_3 = _mm_unpackhi_epi16(in13, in3); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_9_7, hi_9_7, \ michael@0: stg2_0, stg2_1, stg2_2, stg2_3, \ michael@0: stp2_8, stp2_15, stp2_9, stp2_14) \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_5_11, hi_5_11, lo_13_3, hi_13_3, \ michael@0: stg2_4, stg2_5, stg2_6, stg2_7, \ michael@0: stp2_10, stp2_13, stp2_11, stp2_12) \ michael@0: } \ michael@0: \ michael@0: /* Stage3 */ \ michael@0: { \ michael@0: const __m128i lo_2_14 = _mm_unpacklo_epi16(in2, in14); \ michael@0: const __m128i hi_2_14 = _mm_unpackhi_epi16(in2, in14); \ michael@0: const __m128i lo_10_6 = _mm_unpacklo_epi16(in10, in6); \ michael@0: const __m128i hi_10_6 = _mm_unpackhi_epi16(in10, in6); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_2_14, hi_2_14, lo_10_6, hi_10_6, \ michael@0: stg3_0, stg3_1, stg3_2, stg3_3, \ michael@0: stp1_4, stp1_7, stp1_5, stp1_6) \ michael@0: \ michael@0: stp1_8_0 = _mm_add_epi16(stp2_8, stp2_9); \ michael@0: stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \ michael@0: stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \ michael@0: stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \ michael@0: \ michael@0: stp1_12_0 = _mm_add_epi16(stp2_12, stp2_13); \ michael@0: stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \ michael@0: stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \ michael@0: stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \ michael@0: } \ michael@0: \ michael@0: /* Stage4 */ \ michael@0: { \ michael@0: const __m128i lo_0_8 = _mm_unpacklo_epi16(in0, in8); \ michael@0: const __m128i hi_0_8 = _mm_unpackhi_epi16(in0, in8); \ michael@0: const __m128i lo_4_12 = _mm_unpacklo_epi16(in4, in12); \ michael@0: const __m128i hi_4_12 = _mm_unpackhi_epi16(in4, in12); \ michael@0: \ michael@0: const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \ michael@0: const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \ michael@0: const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ michael@0: const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_0_8, hi_0_8, lo_4_12, hi_4_12, \ michael@0: stg4_0, stg4_1, stg4_2, stg4_3, \ michael@0: stp2_0, stp2_1, stp2_2, stp2_3) \ michael@0: \ michael@0: stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \ michael@0: stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \ michael@0: stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \ michael@0: stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, \ michael@0: stg4_4, stg4_5, stg4_6, stg4_7, \ michael@0: stp2_9, stp2_14, stp2_10, stp2_13) \ michael@0: } \ michael@0: \ michael@0: /* Stage5 */ \ michael@0: { \ michael@0: const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \ michael@0: const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \ michael@0: \ michael@0: stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \ michael@0: stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \ michael@0: stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \ michael@0: stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \ michael@0: \ michael@0: tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \ michael@0: tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \ michael@0: tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \ michael@0: tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \ michael@0: \ michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); \ michael@0: tmp1 = _mm_add_epi32(tmp1, rounding); \ michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); \ michael@0: tmp3 = _mm_add_epi32(tmp3, rounding); \ michael@0: \ michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ michael@0: tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ michael@0: tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ michael@0: \ michael@0: stp1_5 = _mm_packs_epi32(tmp0, tmp1); \ michael@0: stp1_6 = _mm_packs_epi32(tmp2, tmp3); \ michael@0: \ michael@0: stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11); \ michael@0: stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \ michael@0: stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \ michael@0: stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \ michael@0: \ michael@0: stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \ michael@0: stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \ michael@0: stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \ michael@0: stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \ michael@0: } \ michael@0: \ michael@0: /* Stage6 */ \ michael@0: { \ michael@0: const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ michael@0: const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ michael@0: const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \ michael@0: const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \ michael@0: \ michael@0: stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \ michael@0: stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \ michael@0: stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \ michael@0: stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \ michael@0: stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \ michael@0: stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \ michael@0: stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \ michael@0: stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \ michael@0: stg6_0, stg4_0, stg6_0, stg4_0, \ michael@0: stp2_10, stp2_13, stp2_11, stp2_12) \ michael@0: } michael@0: michael@0: void vp9_idct16x16_256_add_sse2(const int16_t *input, uint8_t *dest, michael@0: int stride) { michael@0: const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i final_rounding = _mm_set1_epi16(1<<5); michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: michael@0: const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); michael@0: const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); michael@0: const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64); michael@0: const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64); michael@0: const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64); michael@0: const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64); michael@0: const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); michael@0: const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); michael@0: michael@0: const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); michael@0: const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); michael@0: const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64); michael@0: const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64); michael@0: michael@0: const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); michael@0: const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); michael@0: const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); michael@0: const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: michael@0: const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); michael@0: michael@0: __m128i in0 = zero, in1 = zero, in2 = zero, in3 = zero, in4 = zero, michael@0: in5 = zero, in6 = zero, in7 = zero, in8 = zero, in9 = zero, michael@0: in10 = zero, in11 = zero, in12 = zero, in13 = zero, michael@0: in14 = zero, in15 = zero; michael@0: __m128i l0 = zero, l1 = zero, l2 = zero, l3 = zero, l4 = zero, l5 = zero, michael@0: l6 = zero, l7 = zero, l8 = zero, l9 = zero, l10 = zero, l11 = zero, michael@0: l12 = zero, l13 = zero, l14 = zero, l15 = zero; michael@0: __m128i r0 = zero, r1 = zero, r2 = zero, r3 = zero, r4 = zero, r5 = zero, michael@0: r6 = zero, r7 = zero, r8 = zero, r9 = zero, r10 = zero, r11 = zero, michael@0: r12 = zero, r13 = zero, r14 = zero, r15 = zero; michael@0: __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7, michael@0: stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, michael@0: stp1_8_0, stp1_12_0; michael@0: __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7, michael@0: stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15; michael@0: __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; michael@0: int i; michael@0: michael@0: // We work on a 8x16 block each time, and loop 4 times for 2-D 16x16 idct. michael@0: for (i = 0; i < 4; i++) { michael@0: // 1-D idct michael@0: if (i < 2) { michael@0: if (i == 1) input += 128; michael@0: michael@0: // Load input data. michael@0: in0 = _mm_load_si128((const __m128i *)input); michael@0: in8 = _mm_load_si128((const __m128i *)(input + 8 * 1)); michael@0: in1 = _mm_load_si128((const __m128i *)(input + 8 * 2)); michael@0: in9 = _mm_load_si128((const __m128i *)(input + 8 * 3)); michael@0: in2 = _mm_load_si128((const __m128i *)(input + 8 * 4)); michael@0: in10 = _mm_load_si128((const __m128i *)(input + 8 * 5)); michael@0: in3 = _mm_load_si128((const __m128i *)(input + 8 * 6)); michael@0: in11 = _mm_load_si128((const __m128i *)(input + 8 * 7)); michael@0: in4 = _mm_load_si128((const __m128i *)(input + 8 * 8)); michael@0: in12 = _mm_load_si128((const __m128i *)(input + 8 * 9)); michael@0: in5 = _mm_load_si128((const __m128i *)(input + 8 * 10)); michael@0: in13 = _mm_load_si128((const __m128i *)(input + 8 * 11)); michael@0: in6 = _mm_load_si128((const __m128i *)(input + 8 * 12)); michael@0: in14 = _mm_load_si128((const __m128i *)(input + 8 * 13)); michael@0: in7 = _mm_load_si128((const __m128i *)(input + 8 * 14)); michael@0: in15 = _mm_load_si128((const __m128i *)(input + 8 * 15)); michael@0: michael@0: TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, michael@0: in4, in5, in6, in7); michael@0: TRANSPOSE_8X8(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9, michael@0: in10, in11, in12, in13, in14, in15); michael@0: } michael@0: michael@0: if (i == 2) { michael@0: TRANSPOSE_8X8(l0, l1, l2, l3, l4, l5, l6, l7, in0, in1, in2, in3, in4, michael@0: in5, in6, in7); michael@0: TRANSPOSE_8X8(r0, r1, r2, r3, r4, r5, r6, r7, in8, in9, in10, in11, in12, michael@0: in13, in14, in15); michael@0: } michael@0: michael@0: if (i == 3) { michael@0: TRANSPOSE_8X8(l8, l9, l10, l11, l12, l13, l14, l15, in0, in1, in2, in3, michael@0: in4, in5, in6, in7); michael@0: TRANSPOSE_8X8(r8, r9, r10, r11, r12, r13, r14, r15, in8, in9, in10, in11, michael@0: in12, in13, in14, in15); michael@0: } michael@0: michael@0: IDCT16_1D michael@0: michael@0: // Stage7 michael@0: if (i == 0) { michael@0: // Left 8x16 michael@0: l0 = _mm_add_epi16(stp2_0, stp1_15); michael@0: l1 = _mm_add_epi16(stp2_1, stp1_14); michael@0: l2 = _mm_add_epi16(stp2_2, stp2_13); michael@0: l3 = _mm_add_epi16(stp2_3, stp2_12); michael@0: l4 = _mm_add_epi16(stp2_4, stp2_11); michael@0: l5 = _mm_add_epi16(stp2_5, stp2_10); michael@0: l6 = _mm_add_epi16(stp2_6, stp1_9); michael@0: l7 = _mm_add_epi16(stp2_7, stp1_8); michael@0: l8 = _mm_sub_epi16(stp2_7, stp1_8); michael@0: l9 = _mm_sub_epi16(stp2_6, stp1_9); michael@0: l10 = _mm_sub_epi16(stp2_5, stp2_10); michael@0: l11 = _mm_sub_epi16(stp2_4, stp2_11); michael@0: l12 = _mm_sub_epi16(stp2_3, stp2_12); michael@0: l13 = _mm_sub_epi16(stp2_2, stp2_13); michael@0: l14 = _mm_sub_epi16(stp2_1, stp1_14); michael@0: l15 = _mm_sub_epi16(stp2_0, stp1_15); michael@0: } else if (i == 1) { michael@0: // Right 8x16 michael@0: r0 = _mm_add_epi16(stp2_0, stp1_15); michael@0: r1 = _mm_add_epi16(stp2_1, stp1_14); michael@0: r2 = _mm_add_epi16(stp2_2, stp2_13); michael@0: r3 = _mm_add_epi16(stp2_3, stp2_12); michael@0: r4 = _mm_add_epi16(stp2_4, stp2_11); michael@0: r5 = _mm_add_epi16(stp2_5, stp2_10); michael@0: r6 = _mm_add_epi16(stp2_6, stp1_9); michael@0: r7 = _mm_add_epi16(stp2_7, stp1_8); michael@0: r8 = _mm_sub_epi16(stp2_7, stp1_8); michael@0: r9 = _mm_sub_epi16(stp2_6, stp1_9); michael@0: r10 = _mm_sub_epi16(stp2_5, stp2_10); michael@0: r11 = _mm_sub_epi16(stp2_4, stp2_11); michael@0: r12 = _mm_sub_epi16(stp2_3, stp2_12); michael@0: r13 = _mm_sub_epi16(stp2_2, stp2_13); michael@0: r14 = _mm_sub_epi16(stp2_1, stp1_14); michael@0: r15 = _mm_sub_epi16(stp2_0, stp1_15); michael@0: } else { michael@0: // 2-D michael@0: in0 = _mm_add_epi16(stp2_0, stp1_15); michael@0: in1 = _mm_add_epi16(stp2_1, stp1_14); michael@0: in2 = _mm_add_epi16(stp2_2, stp2_13); michael@0: in3 = _mm_add_epi16(stp2_3, stp2_12); michael@0: in4 = _mm_add_epi16(stp2_4, stp2_11); michael@0: in5 = _mm_add_epi16(stp2_5, stp2_10); michael@0: in6 = _mm_add_epi16(stp2_6, stp1_9); michael@0: in7 = _mm_add_epi16(stp2_7, stp1_8); michael@0: in8 = _mm_sub_epi16(stp2_7, stp1_8); michael@0: in9 = _mm_sub_epi16(stp2_6, stp1_9); michael@0: in10 = _mm_sub_epi16(stp2_5, stp2_10); michael@0: in11 = _mm_sub_epi16(stp2_4, stp2_11); michael@0: in12 = _mm_sub_epi16(stp2_3, stp2_12); michael@0: in13 = _mm_sub_epi16(stp2_2, stp2_13); michael@0: in14 = _mm_sub_epi16(stp2_1, stp1_14); michael@0: in15 = _mm_sub_epi16(stp2_0, stp1_15); michael@0: michael@0: // Final rounding and shift michael@0: in0 = _mm_adds_epi16(in0, final_rounding); michael@0: in1 = _mm_adds_epi16(in1, final_rounding); michael@0: in2 = _mm_adds_epi16(in2, final_rounding); michael@0: in3 = _mm_adds_epi16(in3, final_rounding); michael@0: in4 = _mm_adds_epi16(in4, final_rounding); michael@0: in5 = _mm_adds_epi16(in5, final_rounding); michael@0: in6 = _mm_adds_epi16(in6, final_rounding); michael@0: in7 = _mm_adds_epi16(in7, final_rounding); michael@0: in8 = _mm_adds_epi16(in8, final_rounding); michael@0: in9 = _mm_adds_epi16(in9, final_rounding); michael@0: in10 = _mm_adds_epi16(in10, final_rounding); michael@0: in11 = _mm_adds_epi16(in11, final_rounding); michael@0: in12 = _mm_adds_epi16(in12, final_rounding); michael@0: in13 = _mm_adds_epi16(in13, final_rounding); michael@0: in14 = _mm_adds_epi16(in14, final_rounding); michael@0: in15 = _mm_adds_epi16(in15, final_rounding); michael@0: michael@0: in0 = _mm_srai_epi16(in0, 6); michael@0: in1 = _mm_srai_epi16(in1, 6); michael@0: in2 = _mm_srai_epi16(in2, 6); michael@0: in3 = _mm_srai_epi16(in3, 6); michael@0: in4 = _mm_srai_epi16(in4, 6); michael@0: in5 = _mm_srai_epi16(in5, 6); michael@0: in6 = _mm_srai_epi16(in6, 6); michael@0: in7 = _mm_srai_epi16(in7, 6); michael@0: in8 = _mm_srai_epi16(in8, 6); michael@0: in9 = _mm_srai_epi16(in9, 6); michael@0: in10 = _mm_srai_epi16(in10, 6); michael@0: in11 = _mm_srai_epi16(in11, 6); michael@0: in12 = _mm_srai_epi16(in12, 6); michael@0: in13 = _mm_srai_epi16(in13, 6); michael@0: in14 = _mm_srai_epi16(in14, 6); michael@0: in15 = _mm_srai_epi16(in15, 6); michael@0: michael@0: RECON_AND_STORE(dest, in0); michael@0: RECON_AND_STORE(dest, in1); michael@0: RECON_AND_STORE(dest, in2); michael@0: RECON_AND_STORE(dest, in3); michael@0: RECON_AND_STORE(dest, in4); michael@0: RECON_AND_STORE(dest, in5); michael@0: RECON_AND_STORE(dest, in6); michael@0: RECON_AND_STORE(dest, in7); michael@0: RECON_AND_STORE(dest, in8); michael@0: RECON_AND_STORE(dest, in9); michael@0: RECON_AND_STORE(dest, in10); michael@0: RECON_AND_STORE(dest, in11); michael@0: RECON_AND_STORE(dest, in12); michael@0: RECON_AND_STORE(dest, in13); michael@0: RECON_AND_STORE(dest, in14); michael@0: RECON_AND_STORE(dest, in15); michael@0: michael@0: dest += 8 - (stride * 16); michael@0: } michael@0: } michael@0: } michael@0: michael@0: void vp9_idct16x16_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) { michael@0: __m128i dc_value; michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: int a, i; michael@0: michael@0: a = dct_const_round_shift(input[0] * cospi_16_64); michael@0: a = dct_const_round_shift(a * cospi_16_64); michael@0: a = ROUND_POWER_OF_TWO(a, 6); michael@0: michael@0: dc_value = _mm_set1_epi16(a); michael@0: michael@0: for (i = 0; i < 2; ++i) { michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: dest += 8 - (stride * 16); michael@0: } 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 void iadst16_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: static void idct16_1d_8col(__m128i *in) { michael@0: const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64); michael@0: const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64); michael@0: const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64); michael@0: const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64); michael@0: const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64); michael@0: const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64); michael@0: const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64); michael@0: const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64); michael@0: const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64); michael@0: const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64); michael@0: const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64); michael@0: const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_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_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_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_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: __m128i v[16], u[16], s[16], t[16]; michael@0: michael@0: // stage 1 michael@0: s[0] = in[0]; michael@0: s[1] = in[8]; michael@0: s[2] = in[4]; michael@0: s[3] = in[12]; michael@0: s[4] = in[2]; michael@0: s[5] = in[10]; michael@0: s[6] = in[6]; michael@0: s[7] = in[14]; michael@0: s[8] = in[1]; michael@0: s[9] = in[9]; michael@0: s[10] = in[5]; michael@0: s[11] = in[13]; michael@0: s[12] = in[3]; michael@0: s[13] = in[11]; michael@0: s[14] = in[7]; michael@0: s[15] = in[15]; michael@0: michael@0: // stage 2 michael@0: u[0] = _mm_unpacklo_epi16(s[8], s[15]); michael@0: u[1] = _mm_unpackhi_epi16(s[8], s[15]); michael@0: u[2] = _mm_unpacklo_epi16(s[9], s[14]); michael@0: u[3] = _mm_unpackhi_epi16(s[9], s[14]); michael@0: u[4] = _mm_unpacklo_epi16(s[10], s[13]); michael@0: u[5] = _mm_unpackhi_epi16(s[10], s[13]); michael@0: u[6] = _mm_unpacklo_epi16(s[11], s[12]); michael@0: u[7] = _mm_unpackhi_epi16(s[11], s[12]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02); michael@0: v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30); michael@0: v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18); michael@0: v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14); michael@0: v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14); michael@0: v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10); michael@0: v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10); michael@0: v[10] = _mm_madd_epi16(u[4], k__cospi_p10_p22); michael@0: v[11] = _mm_madd_epi16(u[5], k__cospi_p10_p22); michael@0: v[12] = _mm_madd_epi16(u[6], k__cospi_p06_m26); michael@0: v[13] = _mm_madd_epi16(u[7], k__cospi_p06_m26); michael@0: v[14] = _mm_madd_epi16(u[6], k__cospi_p26_p06); michael@0: v[15] = _mm_madd_epi16(u[7], k__cospi_p26_p06); 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: u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS); michael@0: u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS); michael@0: u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS); michael@0: u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS); michael@0: u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS); michael@0: u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS); michael@0: u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS); michael@0: u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS); michael@0: michael@0: s[8] = _mm_packs_epi32(u[0], u[1]); michael@0: s[15] = _mm_packs_epi32(u[2], u[3]); michael@0: s[9] = _mm_packs_epi32(u[4], u[5]); michael@0: s[14] = _mm_packs_epi32(u[6], u[7]); michael@0: s[10] = _mm_packs_epi32(u[8], u[9]); michael@0: s[13] = _mm_packs_epi32(u[10], u[11]); michael@0: s[11] = _mm_packs_epi32(u[12], u[13]); michael@0: s[12] = _mm_packs_epi32(u[14], u[15]); michael@0: michael@0: // stage 3 michael@0: t[0] = s[0]; michael@0: t[1] = s[1]; michael@0: t[2] = s[2]; michael@0: t[3] = s[3]; michael@0: u[0] = _mm_unpacklo_epi16(s[4], s[7]); michael@0: u[1] = _mm_unpackhi_epi16(s[4], s[7]); michael@0: u[2] = _mm_unpacklo_epi16(s[5], s[6]); michael@0: u[3] = _mm_unpackhi_epi16(s[5], s[6]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_p28_m04); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_p28_m04); michael@0: v[2] = _mm_madd_epi16(u[0], k__cospi_p04_p28); michael@0: v[3] = _mm_madd_epi16(u[1], k__cospi_p04_p28); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_p12_m20); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_p12_m20); michael@0: v[6] = _mm_madd_epi16(u[2], k__cospi_p20_p12); michael@0: v[7] = _mm_madd_epi16(u[3], k__cospi_p20_p12); 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: u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: michael@0: t[4] = _mm_packs_epi32(u[0], u[1]); michael@0: t[7] = _mm_packs_epi32(u[2], u[3]); michael@0: t[5] = _mm_packs_epi32(u[4], u[5]); michael@0: t[6] = _mm_packs_epi32(u[6], u[7]); michael@0: t[8] = _mm_add_epi16(s[8], s[9]); michael@0: t[9] = _mm_sub_epi16(s[8], s[9]); michael@0: t[10] = _mm_sub_epi16(s[11], s[10]); michael@0: t[11] = _mm_add_epi16(s[10], s[11]); michael@0: t[12] = _mm_add_epi16(s[12], s[13]); michael@0: t[13] = _mm_sub_epi16(s[12], s[13]); michael@0: t[14] = _mm_sub_epi16(s[15], s[14]); michael@0: t[15] = _mm_add_epi16(s[14], s[15]); michael@0: michael@0: // stage 4 michael@0: u[0] = _mm_unpacklo_epi16(t[0], t[1]); michael@0: u[1] = _mm_unpackhi_epi16(t[0], t[1]); michael@0: u[2] = _mm_unpacklo_epi16(t[2], t[3]); michael@0: u[3] = _mm_unpackhi_epi16(t[2], t[3]); michael@0: u[4] = _mm_unpacklo_epi16(t[9], t[14]); michael@0: u[5] = _mm_unpackhi_epi16(t[9], t[14]); michael@0: u[6] = _mm_unpacklo_epi16(t[10], t[13]); michael@0: u[7] = _mm_unpackhi_epi16(t[10], t[13]); michael@0: michael@0: v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16); michael@0: v[1] = _mm_madd_epi16(u[1], k__cospi_p16_p16); 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_p24_m08); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_p24_m08); 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_m08_p24); michael@0: v[9] = _mm_madd_epi16(u[5], k__cospi_m08_p24); michael@0: v[10] = _mm_madd_epi16(u[4], k__cospi_p24_p08); michael@0: v[11] = _mm_madd_epi16(u[5], k__cospi_p24_p08); michael@0: v[12] = _mm_madd_epi16(u[6], k__cospi_m24_m08); michael@0: v[13] = _mm_madd_epi16(u[7], k__cospi_m24_m08); michael@0: v[14] = _mm_madd_epi16(u[6], k__cospi_m08_p24); michael@0: v[15] = _mm_madd_epi16(u[7], k__cospi_m08_p24); 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: u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS); michael@0: u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS); michael@0: u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS); michael@0: u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS); michael@0: u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS); michael@0: u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS); michael@0: u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS); michael@0: u[15] = _mm_srai_epi32(u[15], 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_add_epi16(t[4], t[5]); michael@0: s[5] = _mm_sub_epi16(t[4], t[5]); michael@0: s[6] = _mm_sub_epi16(t[7], t[6]); michael@0: s[7] = _mm_add_epi16(t[6], t[7]); michael@0: s[8] = t[8]; michael@0: s[15] = t[15]; michael@0: s[9] = _mm_packs_epi32(u[8], u[9]); michael@0: s[14] = _mm_packs_epi32(u[10], u[11]); michael@0: s[10] = _mm_packs_epi32(u[12], u[13]); michael@0: s[13] = _mm_packs_epi32(u[14], u[15]); michael@0: s[11] = t[11]; michael@0: s[12] = t[12]; michael@0: michael@0: // stage 5 michael@0: t[0] = _mm_add_epi16(s[0], s[3]); michael@0: t[1] = _mm_add_epi16(s[1], s[2]); michael@0: t[2] = _mm_sub_epi16(s[1], s[2]); michael@0: t[3] = _mm_sub_epi16(s[0], s[3]); michael@0: t[4] = s[4]; michael@0: t[7] = s[7]; michael@0: michael@0: u[0] = _mm_unpacklo_epi16(s[5], s[6]); michael@0: u[1] = _mm_unpackhi_epi16(s[5], s[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: 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[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: t[5] = _mm_packs_epi32(u[0], u[1]); michael@0: t[6] = _mm_packs_epi32(u[2], u[3]); michael@0: michael@0: t[8] = _mm_add_epi16(s[8], s[11]); michael@0: t[9] = _mm_add_epi16(s[9], s[10]); michael@0: t[10] = _mm_sub_epi16(s[9], s[10]); michael@0: t[11] = _mm_sub_epi16(s[8], s[11]); michael@0: t[12] = _mm_sub_epi16(s[15], s[12]); michael@0: t[13] = _mm_sub_epi16(s[14], s[13]); michael@0: t[14] = _mm_add_epi16(s[13], s[14]); michael@0: t[15] = _mm_add_epi16(s[12], s[15]); michael@0: michael@0: // stage 6 michael@0: s[0] = _mm_add_epi16(t[0], t[7]); michael@0: s[1] = _mm_add_epi16(t[1], t[6]); michael@0: s[2] = _mm_add_epi16(t[2], t[5]); michael@0: s[3] = _mm_add_epi16(t[3], t[4]); michael@0: s[4] = _mm_sub_epi16(t[3], t[4]); michael@0: s[5] = _mm_sub_epi16(t[2], t[5]); michael@0: s[6] = _mm_sub_epi16(t[1], t[6]); michael@0: s[7] = _mm_sub_epi16(t[0], t[7]); michael@0: s[8] = t[8]; michael@0: s[9] = t[9]; michael@0: michael@0: u[0] = _mm_unpacklo_epi16(t[10], t[13]); michael@0: u[1] = _mm_unpackhi_epi16(t[10], t[13]); michael@0: u[2] = _mm_unpacklo_epi16(t[11], t[12]); michael@0: u[3] = _mm_unpackhi_epi16(t[11], t[12]); 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[0], k__cospi_p16_p16); michael@0: v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16); michael@0: v[4] = _mm_madd_epi16(u[2], k__cospi_m16_p16); michael@0: v[5] = _mm_madd_epi16(u[3], k__cospi_m16_p16); michael@0: v[6] = _mm_madd_epi16(u[2], k__cospi_p16_p16); michael@0: v[7] = _mm_madd_epi16(u[3], 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: u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); michael@0: u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); michael@0: u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); michael@0: u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); michael@0: u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); michael@0: u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); michael@0: u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); michael@0: u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); michael@0: michael@0: s[10] = _mm_packs_epi32(u[0], u[1]); michael@0: s[13] = _mm_packs_epi32(u[2], u[3]); michael@0: s[11] = _mm_packs_epi32(u[4], u[5]); michael@0: s[12] = _mm_packs_epi32(u[6], u[7]); michael@0: s[14] = t[14]; michael@0: s[15] = t[15]; michael@0: michael@0: // stage 7 michael@0: in[0] = _mm_add_epi16(s[0], s[15]); michael@0: in[1] = _mm_add_epi16(s[1], s[14]); michael@0: in[2] = _mm_add_epi16(s[2], s[13]); michael@0: in[3] = _mm_add_epi16(s[3], s[12]); michael@0: in[4] = _mm_add_epi16(s[4], s[11]); michael@0: in[5] = _mm_add_epi16(s[5], s[10]); michael@0: in[6] = _mm_add_epi16(s[6], s[9]); michael@0: in[7] = _mm_add_epi16(s[7], s[8]); michael@0: in[8] = _mm_sub_epi16(s[7], s[8]); michael@0: in[9] = _mm_sub_epi16(s[6], s[9]); michael@0: in[10] = _mm_sub_epi16(s[5], s[10]); michael@0: in[11] = _mm_sub_epi16(s[4], s[11]); michael@0: in[12] = _mm_sub_epi16(s[3], s[12]); michael@0: in[13] = _mm_sub_epi16(s[2], s[13]); michael@0: in[14] = _mm_sub_epi16(s[1], s[14]); michael@0: in[15] = _mm_sub_epi16(s[0], s[15]); michael@0: } michael@0: michael@0: static void idct16_1d_sse2(__m128i *in0, __m128i *in1) { michael@0: array_transpose_16x16(in0, in1); michael@0: idct16_1d_8col(in0); michael@0: idct16_1d_8col(in1); michael@0: } michael@0: michael@0: static void iadst16_1d_sse2(__m128i *in0, __m128i *in1) { michael@0: array_transpose_16x16(in0, in1); michael@0: iadst16_1d_8col(in0); michael@0: iadst16_1d_8col(in1); michael@0: } michael@0: michael@0: static INLINE void load_buffer_8x16(const int16_t *input, __m128i *in) { michael@0: in[0] = _mm_load_si128((const __m128i *)(input + 0 * 16)); michael@0: in[1] = _mm_load_si128((const __m128i *)(input + 1 * 16)); michael@0: in[2] = _mm_load_si128((const __m128i *)(input + 2 * 16)); michael@0: in[3] = _mm_load_si128((const __m128i *)(input + 3 * 16)); michael@0: in[4] = _mm_load_si128((const __m128i *)(input + 4 * 16)); michael@0: in[5] = _mm_load_si128((const __m128i *)(input + 5 * 16)); michael@0: in[6] = _mm_load_si128((const __m128i *)(input + 6 * 16)); michael@0: in[7] = _mm_load_si128((const __m128i *)(input + 7 * 16)); michael@0: michael@0: in[8] = _mm_load_si128((const __m128i *)(input + 8 * 16)); michael@0: in[9] = _mm_load_si128((const __m128i *)(input + 9 * 16)); michael@0: in[10] = _mm_load_si128((const __m128i *)(input + 10 * 16)); michael@0: in[11] = _mm_load_si128((const __m128i *)(input + 11 * 16)); michael@0: in[12] = _mm_load_si128((const __m128i *)(input + 12 * 16)); michael@0: in[13] = _mm_load_si128((const __m128i *)(input + 13 * 16)); michael@0: in[14] = _mm_load_si128((const __m128i *)(input + 14 * 16)); michael@0: in[15] = _mm_load_si128((const __m128i *)(input + 15 * 16)); michael@0: } michael@0: michael@0: static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) { michael@0: const __m128i final_rounding = _mm_set1_epi16(1<<5); michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: // Final rounding and shift michael@0: in[0] = _mm_adds_epi16(in[0], final_rounding); michael@0: in[1] = _mm_adds_epi16(in[1], final_rounding); michael@0: in[2] = _mm_adds_epi16(in[2], final_rounding); michael@0: in[3] = _mm_adds_epi16(in[3], final_rounding); michael@0: in[4] = _mm_adds_epi16(in[4], final_rounding); michael@0: in[5] = _mm_adds_epi16(in[5], final_rounding); michael@0: in[6] = _mm_adds_epi16(in[6], final_rounding); michael@0: in[7] = _mm_adds_epi16(in[7], final_rounding); michael@0: in[8] = _mm_adds_epi16(in[8], final_rounding); michael@0: in[9] = _mm_adds_epi16(in[9], final_rounding); michael@0: in[10] = _mm_adds_epi16(in[10], final_rounding); michael@0: in[11] = _mm_adds_epi16(in[11], final_rounding); michael@0: in[12] = _mm_adds_epi16(in[12], final_rounding); michael@0: in[13] = _mm_adds_epi16(in[13], final_rounding); michael@0: in[14] = _mm_adds_epi16(in[14], final_rounding); michael@0: in[15] = _mm_adds_epi16(in[15], final_rounding); michael@0: michael@0: in[0] = _mm_srai_epi16(in[0], 6); michael@0: in[1] = _mm_srai_epi16(in[1], 6); michael@0: in[2] = _mm_srai_epi16(in[2], 6); michael@0: in[3] = _mm_srai_epi16(in[3], 6); michael@0: in[4] = _mm_srai_epi16(in[4], 6); michael@0: in[5] = _mm_srai_epi16(in[5], 6); michael@0: in[6] = _mm_srai_epi16(in[6], 6); michael@0: in[7] = _mm_srai_epi16(in[7], 6); michael@0: in[8] = _mm_srai_epi16(in[8], 6); michael@0: in[9] = _mm_srai_epi16(in[9], 6); michael@0: in[10] = _mm_srai_epi16(in[10], 6); michael@0: in[11] = _mm_srai_epi16(in[11], 6); michael@0: in[12] = _mm_srai_epi16(in[12], 6); michael@0: in[13] = _mm_srai_epi16(in[13], 6); michael@0: in[14] = _mm_srai_epi16(in[14], 6); michael@0: in[15] = _mm_srai_epi16(in[15], 6); michael@0: michael@0: RECON_AND_STORE(dest, in[0]); michael@0: RECON_AND_STORE(dest, in[1]); michael@0: RECON_AND_STORE(dest, in[2]); michael@0: RECON_AND_STORE(dest, in[3]); michael@0: RECON_AND_STORE(dest, in[4]); michael@0: RECON_AND_STORE(dest, in[5]); michael@0: RECON_AND_STORE(dest, in[6]); michael@0: RECON_AND_STORE(dest, in[7]); michael@0: RECON_AND_STORE(dest, in[8]); michael@0: RECON_AND_STORE(dest, in[9]); michael@0: RECON_AND_STORE(dest, in[10]); michael@0: RECON_AND_STORE(dest, in[11]); michael@0: RECON_AND_STORE(dest, in[12]); michael@0: RECON_AND_STORE(dest, in[13]); michael@0: RECON_AND_STORE(dest, in[14]); michael@0: RECON_AND_STORE(dest, in[15]); michael@0: } michael@0: michael@0: void vp9_iht16x16_256_add_sse2(const int16_t *input, uint8_t *dest, int stride, michael@0: int tx_type) { michael@0: __m128i in0[16], in1[16]; michael@0: michael@0: load_buffer_8x16(input, in0); michael@0: input += 8; michael@0: load_buffer_8x16(input, in1); michael@0: michael@0: switch (tx_type) { michael@0: case 0: // DCT_DCT michael@0: idct16_1d_sse2(in0, in1); michael@0: idct16_1d_sse2(in0, in1); michael@0: break; michael@0: case 1: // ADST_DCT michael@0: idct16_1d_sse2(in0, in1); michael@0: iadst16_1d_sse2(in0, in1); michael@0: break; michael@0: case 2: // DCT_ADST michael@0: iadst16_1d_sse2(in0, in1); michael@0: idct16_1d_sse2(in0, in1); michael@0: break; michael@0: case 3: // ADST_ADST michael@0: iadst16_1d_sse2(in0, in1); michael@0: iadst16_1d_sse2(in0, in1); michael@0: break; michael@0: default: michael@0: assert(0); michael@0: break; michael@0: } michael@0: michael@0: write_buffer_8x16(dest, in0, stride); michael@0: dest += 8; michael@0: write_buffer_8x16(dest, in1, stride); michael@0: } michael@0: michael@0: void vp9_idct16x16_10_add_sse2(const int16_t *input, uint8_t *dest, michael@0: int stride) { michael@0: const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i final_rounding = _mm_set1_epi16(1<<5); michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: michael@0: const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); michael@0: const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); michael@0: const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64); michael@0: const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64); michael@0: const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64); michael@0: const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64); michael@0: const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); michael@0: const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); michael@0: michael@0: const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); michael@0: const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); michael@0: const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64); michael@0: const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64); michael@0: michael@0: const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); michael@0: const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); michael@0: const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); michael@0: const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: michael@0: const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); michael@0: michael@0: __m128i in0 = zero, in1 = zero, in2 = zero, in3 = zero, in4 = zero, michael@0: in5 = zero, in6 = zero, in7 = zero, in8 = zero, in9 = zero, michael@0: in10 = zero, in11 = zero, in12 = zero, in13 = zero, michael@0: in14 = zero, in15 = zero; michael@0: __m128i l0 = zero, l1 = zero, l2 = zero, l3 = zero, l4 = zero, l5 = zero, michael@0: l6 = zero, l7 = zero, l8 = zero, l9 = zero, l10 = zero, l11 = zero, michael@0: l12 = zero, l13 = zero, l14 = zero, l15 = zero; michael@0: michael@0: __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7, michael@0: stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, michael@0: stp1_8_0, stp1_12_0; michael@0: __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7, michael@0: stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15; michael@0: __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; michael@0: int i; michael@0: // 1-D idct. Load input data. michael@0: in0 = _mm_load_si128((const __m128i *)input); michael@0: in8 = _mm_load_si128((const __m128i *)(input + 8 * 1)); michael@0: in1 = _mm_load_si128((const __m128i *)(input + 8 * 2)); michael@0: in9 = _mm_load_si128((const __m128i *)(input + 8 * 3)); michael@0: in2 = _mm_load_si128((const __m128i *)(input + 8 * 4)); michael@0: in10 = _mm_load_si128((const __m128i *)(input + 8 * 5)); michael@0: in3 = _mm_load_si128((const __m128i *)(input + 8 * 6)); michael@0: in11 = _mm_load_si128((const __m128i *)(input + 8 * 7)); michael@0: michael@0: TRANSPOSE_8X4(in0, in1, in2, in3, in0, in1, in2, in3); michael@0: TRANSPOSE_8X4(in8, in9, in10, in11, in8, in9, in10, in11); michael@0: michael@0: // Stage2 michael@0: { michael@0: const __m128i lo_1_15 = _mm_unpackhi_epi16(in0, in11); michael@0: const __m128i lo_9_7 = _mm_unpackhi_epi16(in8, in3); michael@0: const __m128i lo_5_11 = _mm_unpackhi_epi16(in2, in9); michael@0: const __m128i lo_13_3 = _mm_unpackhi_epi16(in10, in1); michael@0: michael@0: tmp0 = _mm_madd_epi16(lo_1_15, stg2_0); michael@0: tmp2 = _mm_madd_epi16(lo_1_15, stg2_1); michael@0: tmp4 = _mm_madd_epi16(lo_9_7, stg2_2); michael@0: tmp6 = _mm_madd_epi16(lo_9_7, stg2_3); michael@0: tmp1 = _mm_madd_epi16(lo_5_11, stg2_4); michael@0: tmp3 = _mm_madd_epi16(lo_5_11, stg2_5); michael@0: tmp5 = _mm_madd_epi16(lo_13_3, stg2_6); michael@0: tmp7 = _mm_madd_epi16(lo_13_3, stg2_7); michael@0: michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); michael@0: tmp4 = _mm_add_epi32(tmp4, rounding); michael@0: tmp6 = _mm_add_epi32(tmp6, rounding); michael@0: tmp1 = _mm_add_epi32(tmp1, rounding); michael@0: tmp3 = _mm_add_epi32(tmp3, rounding); michael@0: tmp5 = _mm_add_epi32(tmp5, rounding); michael@0: tmp7 = _mm_add_epi32(tmp7, rounding); michael@0: michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); michael@0: tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); michael@0: tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); michael@0: tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); michael@0: tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); michael@0: tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); michael@0: tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); michael@0: michael@0: stp2_8 = _mm_packs_epi32(tmp0, zero); michael@0: stp2_15 = _mm_packs_epi32(tmp2, zero); michael@0: stp2_9 = _mm_packs_epi32(tmp4, zero); michael@0: stp2_14 = _mm_packs_epi32(tmp6, zero); michael@0: michael@0: stp2_10 = _mm_packs_epi32(tmp1, zero); michael@0: stp2_13 = _mm_packs_epi32(tmp3, zero); michael@0: stp2_11 = _mm_packs_epi32(tmp5, zero); michael@0: stp2_12 = _mm_packs_epi32(tmp7, zero); michael@0: } michael@0: michael@0: // Stage3 michael@0: { michael@0: const __m128i lo_2_14 = _mm_unpacklo_epi16(in1, in11); michael@0: const __m128i lo_10_6 = _mm_unpacklo_epi16(in9, in3); michael@0: michael@0: tmp0 = _mm_madd_epi16(lo_2_14, stg3_0); michael@0: tmp2 = _mm_madd_epi16(lo_2_14, stg3_1); michael@0: tmp4 = _mm_madd_epi16(lo_10_6, stg3_2); michael@0: tmp6 = _mm_madd_epi16(lo_10_6, stg3_3); michael@0: michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); michael@0: tmp4 = _mm_add_epi32(tmp4, rounding); michael@0: tmp6 = _mm_add_epi32(tmp6, rounding); michael@0: michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); michael@0: tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); michael@0: tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); michael@0: michael@0: stp1_4 = _mm_packs_epi32(tmp0, zero); michael@0: stp1_7 = _mm_packs_epi32(tmp2, zero); michael@0: stp1_5 = _mm_packs_epi32(tmp4, zero); michael@0: stp1_6 = _mm_packs_epi32(tmp6, zero); michael@0: michael@0: stp1_8_0 = _mm_add_epi16(stp2_8, stp2_9); michael@0: stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); michael@0: stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); michael@0: stp1_11 = _mm_add_epi16(stp2_11, stp2_10); michael@0: michael@0: stp1_12_0 = _mm_add_epi16(stp2_12, stp2_13); michael@0: stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); michael@0: stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); michael@0: stp1_15 = _mm_add_epi16(stp2_15, stp2_14); michael@0: } michael@0: michael@0: // Stage4 michael@0: { michael@0: const __m128i lo_0_8 = _mm_unpacklo_epi16(in0, in8); michael@0: const __m128i lo_4_12 = _mm_unpacklo_epi16(in2, in10); michael@0: const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); michael@0: const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); michael@0: michael@0: tmp0 = _mm_madd_epi16(lo_0_8, stg4_0); michael@0: tmp2 = _mm_madd_epi16(lo_0_8, stg4_1); michael@0: tmp4 = _mm_madd_epi16(lo_4_12, stg4_2); michael@0: tmp6 = _mm_madd_epi16(lo_4_12, stg4_3); michael@0: tmp1 = _mm_madd_epi16(lo_9_14, stg4_4); michael@0: tmp3 = _mm_madd_epi16(lo_9_14, stg4_5); michael@0: tmp5 = _mm_madd_epi16(lo_10_13, stg4_6); michael@0: tmp7 = _mm_madd_epi16(lo_10_13, stg4_7); michael@0: michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); michael@0: tmp4 = _mm_add_epi32(tmp4, rounding); michael@0: tmp6 = _mm_add_epi32(tmp6, rounding); michael@0: tmp1 = _mm_add_epi32(tmp1, rounding); michael@0: tmp3 = _mm_add_epi32(tmp3, rounding); michael@0: tmp5 = _mm_add_epi32(tmp5, rounding); michael@0: tmp7 = _mm_add_epi32(tmp7, rounding); michael@0: michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); michael@0: tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); michael@0: tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); michael@0: tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); michael@0: tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); michael@0: tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); michael@0: tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); michael@0: michael@0: stp2_0 = _mm_packs_epi32(tmp0, zero); michael@0: stp2_1 = _mm_packs_epi32(tmp2, zero); michael@0: stp2_2 = _mm_packs_epi32(tmp4, zero); michael@0: stp2_3 = _mm_packs_epi32(tmp6, zero); michael@0: stp2_9 = _mm_packs_epi32(tmp1, zero); michael@0: stp2_14 = _mm_packs_epi32(tmp3, zero); michael@0: stp2_10 = _mm_packs_epi32(tmp5, zero); michael@0: stp2_13 = _mm_packs_epi32(tmp7, zero); michael@0: michael@0: stp2_4 = _mm_add_epi16(stp1_4, stp1_5); michael@0: stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); michael@0: stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); michael@0: stp2_7 = _mm_add_epi16(stp1_7, stp1_6); michael@0: } michael@0: michael@0: // Stage5 and Stage6 michael@0: { michael@0: stp1_0 = _mm_add_epi16(stp2_0, stp2_3); michael@0: stp1_1 = _mm_add_epi16(stp2_1, stp2_2); michael@0: stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); michael@0: stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); michael@0: michael@0: stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11); michael@0: stp1_9 = _mm_add_epi16(stp2_9, stp2_10); michael@0: stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); michael@0: stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); michael@0: michael@0: stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); michael@0: stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); michael@0: stp1_14 = _mm_add_epi16(stp2_14, stp2_13); michael@0: stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); michael@0: } michael@0: michael@0: // Stage6 michael@0: { michael@0: const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); michael@0: const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); michael@0: const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); michael@0: michael@0: tmp1 = _mm_madd_epi16(lo_6_5, stg4_1); michael@0: tmp3 = _mm_madd_epi16(lo_6_5, stg4_0); michael@0: tmp0 = _mm_madd_epi16(lo_10_13, stg6_0); michael@0: tmp2 = _mm_madd_epi16(lo_10_13, stg4_0); michael@0: tmp4 = _mm_madd_epi16(lo_11_12, stg6_0); michael@0: tmp6 = _mm_madd_epi16(lo_11_12, stg4_0); michael@0: michael@0: tmp1 = _mm_add_epi32(tmp1, rounding); michael@0: tmp3 = _mm_add_epi32(tmp3, rounding); michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); michael@0: tmp4 = _mm_add_epi32(tmp4, rounding); michael@0: tmp6 = _mm_add_epi32(tmp6, rounding); michael@0: michael@0: tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); michael@0: tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); michael@0: tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); michael@0: tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); michael@0: michael@0: stp1_5 = _mm_packs_epi32(tmp1, zero); michael@0: stp1_6 = _mm_packs_epi32(tmp3, zero); michael@0: stp2_10 = _mm_packs_epi32(tmp0, zero); michael@0: stp2_13 = _mm_packs_epi32(tmp2, zero); michael@0: stp2_11 = _mm_packs_epi32(tmp4, zero); michael@0: stp2_12 = _mm_packs_epi32(tmp6, zero); michael@0: michael@0: stp2_0 = _mm_add_epi16(stp1_0, stp2_7); michael@0: stp2_1 = _mm_add_epi16(stp1_1, stp1_6); michael@0: stp2_2 = _mm_add_epi16(stp1_2, stp1_5); michael@0: stp2_3 = _mm_add_epi16(stp1_3, stp2_4); michael@0: stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); michael@0: stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); michael@0: stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); michael@0: stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); michael@0: } michael@0: michael@0: // Stage7. Left 8x16 only. michael@0: l0 = _mm_add_epi16(stp2_0, stp1_15); michael@0: l1 = _mm_add_epi16(stp2_1, stp1_14); michael@0: l2 = _mm_add_epi16(stp2_2, stp2_13); michael@0: l3 = _mm_add_epi16(stp2_3, stp2_12); michael@0: l4 = _mm_add_epi16(stp2_4, stp2_11); michael@0: l5 = _mm_add_epi16(stp2_5, stp2_10); michael@0: l6 = _mm_add_epi16(stp2_6, stp1_9); michael@0: l7 = _mm_add_epi16(stp2_7, stp1_8); michael@0: l8 = _mm_sub_epi16(stp2_7, stp1_8); michael@0: l9 = _mm_sub_epi16(stp2_6, stp1_9); michael@0: l10 = _mm_sub_epi16(stp2_5, stp2_10); michael@0: l11 = _mm_sub_epi16(stp2_4, stp2_11); michael@0: l12 = _mm_sub_epi16(stp2_3, stp2_12); michael@0: l13 = _mm_sub_epi16(stp2_2, stp2_13); michael@0: l14 = _mm_sub_epi16(stp2_1, stp1_14); michael@0: l15 = _mm_sub_epi16(stp2_0, stp1_15); michael@0: michael@0: // 2-D idct. We do 2 8x16 blocks. michael@0: for (i = 0; i < 2; i++) { michael@0: if (i == 0) michael@0: TRANSPOSE_4X8(l0, l1, l2, l3, l4, l5, l6, l7, in0, in1, in2, in3, in4, michael@0: in5, in6, in7); michael@0: michael@0: if (i == 1) michael@0: TRANSPOSE_4X8(l8, l9, l10, l11, l12, l13, l14, l15, in0, in1, in2, in3, michael@0: in4, in5, in6, in7); michael@0: michael@0: in8 = in9 = in10 = in11 = in12 = in13 = in14 = in15 = zero; michael@0: michael@0: IDCT16_1D michael@0: michael@0: // Stage7 michael@0: in0 = _mm_add_epi16(stp2_0, stp1_15); michael@0: in1 = _mm_add_epi16(stp2_1, stp1_14); michael@0: in2 = _mm_add_epi16(stp2_2, stp2_13); michael@0: in3 = _mm_add_epi16(stp2_3, stp2_12); michael@0: in4 = _mm_add_epi16(stp2_4, stp2_11); michael@0: in5 = _mm_add_epi16(stp2_5, stp2_10); michael@0: in6 = _mm_add_epi16(stp2_6, stp1_9); michael@0: in7 = _mm_add_epi16(stp2_7, stp1_8); michael@0: in8 = _mm_sub_epi16(stp2_7, stp1_8); michael@0: in9 = _mm_sub_epi16(stp2_6, stp1_9); michael@0: in10 = _mm_sub_epi16(stp2_5, stp2_10); michael@0: in11 = _mm_sub_epi16(stp2_4, stp2_11); michael@0: in12 = _mm_sub_epi16(stp2_3, stp2_12); michael@0: in13 = _mm_sub_epi16(stp2_2, stp2_13); michael@0: in14 = _mm_sub_epi16(stp2_1, stp1_14); michael@0: in15 = _mm_sub_epi16(stp2_0, stp1_15); michael@0: michael@0: // Final rounding and shift michael@0: in0 = _mm_adds_epi16(in0, final_rounding); michael@0: in1 = _mm_adds_epi16(in1, final_rounding); michael@0: in2 = _mm_adds_epi16(in2, final_rounding); michael@0: in3 = _mm_adds_epi16(in3, final_rounding); michael@0: in4 = _mm_adds_epi16(in4, final_rounding); michael@0: in5 = _mm_adds_epi16(in5, final_rounding); michael@0: in6 = _mm_adds_epi16(in6, final_rounding); michael@0: in7 = _mm_adds_epi16(in7, final_rounding); michael@0: in8 = _mm_adds_epi16(in8, final_rounding); michael@0: in9 = _mm_adds_epi16(in9, final_rounding); michael@0: in10 = _mm_adds_epi16(in10, final_rounding); michael@0: in11 = _mm_adds_epi16(in11, final_rounding); michael@0: in12 = _mm_adds_epi16(in12, final_rounding); michael@0: in13 = _mm_adds_epi16(in13, final_rounding); michael@0: in14 = _mm_adds_epi16(in14, final_rounding); michael@0: in15 = _mm_adds_epi16(in15, final_rounding); michael@0: michael@0: in0 = _mm_srai_epi16(in0, 6); michael@0: in1 = _mm_srai_epi16(in1, 6); michael@0: in2 = _mm_srai_epi16(in2, 6); michael@0: in3 = _mm_srai_epi16(in3, 6); michael@0: in4 = _mm_srai_epi16(in4, 6); michael@0: in5 = _mm_srai_epi16(in5, 6); michael@0: in6 = _mm_srai_epi16(in6, 6); michael@0: in7 = _mm_srai_epi16(in7, 6); michael@0: in8 = _mm_srai_epi16(in8, 6); michael@0: in9 = _mm_srai_epi16(in9, 6); michael@0: in10 = _mm_srai_epi16(in10, 6); michael@0: in11 = _mm_srai_epi16(in11, 6); michael@0: in12 = _mm_srai_epi16(in12, 6); michael@0: in13 = _mm_srai_epi16(in13, 6); michael@0: in14 = _mm_srai_epi16(in14, 6); michael@0: in15 = _mm_srai_epi16(in15, 6); michael@0: michael@0: RECON_AND_STORE(dest, in0); michael@0: RECON_AND_STORE(dest, in1); michael@0: RECON_AND_STORE(dest, in2); michael@0: RECON_AND_STORE(dest, in3); michael@0: RECON_AND_STORE(dest, in4); michael@0: RECON_AND_STORE(dest, in5); michael@0: RECON_AND_STORE(dest, in6); michael@0: RECON_AND_STORE(dest, in7); michael@0: RECON_AND_STORE(dest, in8); michael@0: RECON_AND_STORE(dest, in9); michael@0: RECON_AND_STORE(dest, in10); michael@0: RECON_AND_STORE(dest, in11); michael@0: RECON_AND_STORE(dest, in12); michael@0: RECON_AND_STORE(dest, in13); michael@0: RECON_AND_STORE(dest, in14); michael@0: RECON_AND_STORE(dest, in15); michael@0: michael@0: dest += 8 - (stride * 16); michael@0: } michael@0: } michael@0: michael@0: #define LOAD_DQCOEFF(reg, input) \ michael@0: { \ michael@0: reg = _mm_load_si128((const __m128i *) input); \ michael@0: input += 8; \ michael@0: } \ michael@0: michael@0: #define IDCT32_1D \ michael@0: /* Stage1 */ \ michael@0: { \ michael@0: const __m128i lo_1_31 = _mm_unpacklo_epi16(in1, in31); \ michael@0: const __m128i hi_1_31 = _mm_unpackhi_epi16(in1, in31); \ michael@0: const __m128i lo_17_15 = _mm_unpacklo_epi16(in17, in15); \ michael@0: const __m128i hi_17_15 = _mm_unpackhi_epi16(in17, in15); \ michael@0: \ michael@0: const __m128i lo_9_23 = _mm_unpacklo_epi16(in9, in23); \ michael@0: const __m128i hi_9_23 = _mm_unpackhi_epi16(in9, in23); \ michael@0: const __m128i lo_25_7= _mm_unpacklo_epi16(in25, in7); \ michael@0: const __m128i hi_25_7 = _mm_unpackhi_epi16(in25, in7); \ michael@0: \ michael@0: const __m128i lo_5_27 = _mm_unpacklo_epi16(in5, in27); \ michael@0: const __m128i hi_5_27 = _mm_unpackhi_epi16(in5, in27); \ michael@0: const __m128i lo_21_11 = _mm_unpacklo_epi16(in21, in11); \ michael@0: const __m128i hi_21_11 = _mm_unpackhi_epi16(in21, in11); \ michael@0: \ michael@0: const __m128i lo_13_19 = _mm_unpacklo_epi16(in13, in19); \ michael@0: const __m128i hi_13_19 = _mm_unpackhi_epi16(in13, in19); \ michael@0: const __m128i lo_29_3 = _mm_unpacklo_epi16(in29, in3); \ michael@0: const __m128i hi_29_3 = _mm_unpackhi_epi16(in29, in3); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_1_31, hi_1_31, lo_17_15, hi_17_15, stg1_0, \ michael@0: stg1_1, stg1_2, stg1_3, stp1_16, stp1_31, \ michael@0: stp1_17, stp1_30) \ michael@0: MULTIPLICATION_AND_ADD(lo_9_23, hi_9_23, lo_25_7, hi_25_7, stg1_4, \ michael@0: stg1_5, stg1_6, stg1_7, stp1_18, stp1_29, \ michael@0: stp1_19, stp1_28) \ michael@0: MULTIPLICATION_AND_ADD(lo_5_27, hi_5_27, lo_21_11, hi_21_11, stg1_8, \ michael@0: stg1_9, stg1_10, stg1_11, stp1_20, stp1_27, \ michael@0: stp1_21, stp1_26) \ michael@0: MULTIPLICATION_AND_ADD(lo_13_19, hi_13_19, lo_29_3, hi_29_3, stg1_12, \ michael@0: stg1_13, stg1_14, stg1_15, stp1_22, stp1_25, \ michael@0: stp1_23, stp1_24) \ michael@0: } \ michael@0: \ michael@0: /* Stage2 */ \ michael@0: { \ michael@0: const __m128i lo_2_30 = _mm_unpacklo_epi16(in2, in30); \ michael@0: const __m128i hi_2_30 = _mm_unpackhi_epi16(in2, in30); \ michael@0: const __m128i lo_18_14 = _mm_unpacklo_epi16(in18, in14); \ michael@0: const __m128i hi_18_14 = _mm_unpackhi_epi16(in18, in14); \ michael@0: \ michael@0: const __m128i lo_10_22 = _mm_unpacklo_epi16(in10, in22); \ michael@0: const __m128i hi_10_22 = _mm_unpackhi_epi16(in10, in22); \ michael@0: const __m128i lo_26_6 = _mm_unpacklo_epi16(in26, in6); \ michael@0: const __m128i hi_26_6 = _mm_unpackhi_epi16(in26, in6); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_2_30, hi_2_30, lo_18_14, hi_18_14, stg2_0, \ michael@0: stg2_1, stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, \ michael@0: stp2_14) \ michael@0: MULTIPLICATION_AND_ADD(lo_10_22, hi_10_22, lo_26_6, hi_26_6, stg2_4, \ michael@0: stg2_5, stg2_6, stg2_7, stp2_10, stp2_13, \ michael@0: stp2_11, stp2_12) \ michael@0: \ michael@0: stp2_16 = _mm_add_epi16(stp1_16, stp1_17); \ michael@0: stp2_17 = _mm_sub_epi16(stp1_16, stp1_17); \ michael@0: stp2_18 = _mm_sub_epi16(stp1_19, stp1_18); \ michael@0: stp2_19 = _mm_add_epi16(stp1_19, stp1_18); \ michael@0: \ michael@0: stp2_20 = _mm_add_epi16(stp1_20, stp1_21); \ michael@0: stp2_21 = _mm_sub_epi16(stp1_20, stp1_21); \ michael@0: stp2_22 = _mm_sub_epi16(stp1_23, stp1_22); \ michael@0: stp2_23 = _mm_add_epi16(stp1_23, stp1_22); \ michael@0: \ michael@0: stp2_24 = _mm_add_epi16(stp1_24, stp1_25); \ michael@0: stp2_25 = _mm_sub_epi16(stp1_24, stp1_25); \ michael@0: stp2_26 = _mm_sub_epi16(stp1_27, stp1_26); \ michael@0: stp2_27 = _mm_add_epi16(stp1_27, stp1_26); \ michael@0: \ michael@0: stp2_28 = _mm_add_epi16(stp1_28, stp1_29); \ michael@0: stp2_29 = _mm_sub_epi16(stp1_28, stp1_29); \ michael@0: stp2_30 = _mm_sub_epi16(stp1_31, stp1_30); \ michael@0: stp2_31 = _mm_add_epi16(stp1_31, stp1_30); \ michael@0: } \ michael@0: \ michael@0: /* Stage3 */ \ michael@0: { \ michael@0: const __m128i lo_4_28 = _mm_unpacklo_epi16(in4, in28); \ michael@0: const __m128i hi_4_28 = _mm_unpackhi_epi16(in4, in28); \ michael@0: const __m128i lo_20_12 = _mm_unpacklo_epi16(in20, in12); \ michael@0: const __m128i hi_20_12 = _mm_unpackhi_epi16(in20, in12); \ michael@0: \ michael@0: const __m128i lo_17_30 = _mm_unpacklo_epi16(stp2_17, stp2_30); \ michael@0: const __m128i hi_17_30 = _mm_unpackhi_epi16(stp2_17, stp2_30); \ michael@0: const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \ michael@0: const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \ michael@0: \ michael@0: const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \ michael@0: const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \ michael@0: const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \ michael@0: const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_4_28, hi_4_28, lo_20_12, hi_20_12, stg3_0, \ michael@0: stg3_1, stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, \ michael@0: stp1_6) \ michael@0: \ michael@0: stp1_8 = _mm_add_epi16(stp2_8, stp2_9); \ michael@0: stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \ michael@0: stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \ michael@0: stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \ michael@0: stp1_12 = _mm_add_epi16(stp2_12, stp2_13); \ michael@0: stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \ michael@0: stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \ michael@0: stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \ michael@0: stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, \ michael@0: stp1_18, stp1_29) \ michael@0: MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \ michael@0: stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, \ michael@0: stp1_22, stp1_25) \ michael@0: \ michael@0: stp1_16 = stp2_16; \ michael@0: stp1_31 = stp2_31; \ michael@0: stp1_19 = stp2_19; \ michael@0: stp1_20 = stp2_20; \ michael@0: stp1_23 = stp2_23; \ michael@0: stp1_24 = stp2_24; \ michael@0: stp1_27 = stp2_27; \ michael@0: stp1_28 = stp2_28; \ michael@0: } \ michael@0: \ michael@0: /* Stage4 */ \ michael@0: { \ michael@0: const __m128i lo_0_16 = _mm_unpacklo_epi16(in0, in16); \ michael@0: const __m128i hi_0_16 = _mm_unpackhi_epi16(in0, in16); \ michael@0: const __m128i lo_8_24 = _mm_unpacklo_epi16(in8, in24); \ michael@0: const __m128i hi_8_24 = _mm_unpackhi_epi16(in8, in24); \ michael@0: \ michael@0: const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \ michael@0: const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \ michael@0: const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ michael@0: const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_0_16, hi_0_16, lo_8_24, hi_8_24, stg4_0, \ michael@0: stg4_1, stg4_2, stg4_3, stp2_0, stp2_1, \ michael@0: stp2_2, stp2_3) \ michael@0: \ michael@0: stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \ michael@0: stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \ michael@0: stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \ michael@0: stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \ michael@0: stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, \ michael@0: stp2_10, stp2_13) \ michael@0: \ michael@0: stp2_8 = stp1_8; \ michael@0: stp2_15 = stp1_15; \ michael@0: stp2_11 = stp1_11; \ michael@0: stp2_12 = stp1_12; \ michael@0: \ michael@0: stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \ michael@0: stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \ michael@0: stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \ michael@0: stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \ michael@0: stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \ michael@0: stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \ michael@0: stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \ michael@0: stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \ michael@0: \ michael@0: stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \ michael@0: stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \ michael@0: stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \ michael@0: stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \ michael@0: stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \ michael@0: stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \ michael@0: stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \ michael@0: stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \ michael@0: } \ michael@0: \ michael@0: /* Stage5 */ \ michael@0: { \ michael@0: const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \ michael@0: const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \ michael@0: const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \ michael@0: const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \ michael@0: \ michael@0: const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \ michael@0: const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \ michael@0: const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \ michael@0: const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \ michael@0: \ michael@0: const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \ michael@0: const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \ michael@0: \ michael@0: stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \ michael@0: stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \ michael@0: stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \ michael@0: stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \ michael@0: \ michael@0: tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \ michael@0: tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \ michael@0: tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \ michael@0: tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \ michael@0: \ michael@0: tmp0 = _mm_add_epi32(tmp0, rounding); \ michael@0: tmp1 = _mm_add_epi32(tmp1, rounding); \ michael@0: tmp2 = _mm_add_epi32(tmp2, rounding); \ michael@0: tmp3 = _mm_add_epi32(tmp3, rounding); \ michael@0: \ michael@0: tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ michael@0: tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ michael@0: tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ michael@0: tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ michael@0: \ michael@0: stp1_5 = _mm_packs_epi32(tmp0, tmp1); \ michael@0: stp1_6 = _mm_packs_epi32(tmp2, tmp3); \ michael@0: \ michael@0: stp1_4 = stp2_4; \ michael@0: stp1_7 = stp2_7; \ michael@0: \ michael@0: stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \ michael@0: stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \ michael@0: stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \ michael@0: stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \ michael@0: stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \ michael@0: stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \ michael@0: stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \ michael@0: stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \ michael@0: \ michael@0: stp1_16 = stp2_16; \ michael@0: stp1_17 = stp2_17; \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \ michael@0: stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, \ michael@0: stp1_19, stp1_28) \ michael@0: MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \ michael@0: stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, \ michael@0: stp1_21, stp1_26) \ michael@0: \ michael@0: stp1_22 = stp2_22; \ michael@0: stp1_23 = stp2_23; \ michael@0: stp1_24 = stp2_24; \ michael@0: stp1_25 = stp2_25; \ michael@0: stp1_30 = stp2_30; \ michael@0: stp1_31 = stp2_31; \ michael@0: } \ michael@0: \ michael@0: /* Stage6 */ \ michael@0: { \ michael@0: const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ michael@0: const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ michael@0: const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \ michael@0: const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \ michael@0: \ michael@0: stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \ michael@0: stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \ michael@0: stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \ michael@0: stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \ michael@0: stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \ michael@0: stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \ michael@0: stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \ michael@0: stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \ michael@0: \ michael@0: stp2_8 = stp1_8; \ michael@0: stp2_9 = stp1_9; \ michael@0: stp2_14 = stp1_14; \ michael@0: stp2_15 = stp1_15; \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \ michael@0: stg6_0, stg4_0, stg6_0, stg4_0, stp2_10, \ michael@0: stp2_13, stp2_11, stp2_12) \ michael@0: \ michael@0: stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \ michael@0: stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \ michael@0: stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \ michael@0: stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \ michael@0: stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \ michael@0: stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \ michael@0: stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \ michael@0: stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \ michael@0: \ michael@0: stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \ michael@0: stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \ michael@0: stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \ michael@0: stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \ michael@0: stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \ michael@0: stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \ michael@0: stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \ michael@0: stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \ michael@0: } \ michael@0: \ michael@0: /* Stage7 */ \ michael@0: { \ michael@0: const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \ michael@0: const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \ michael@0: const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \ michael@0: const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \ michael@0: \ michael@0: const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \ michael@0: const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \ michael@0: const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \ michael@0: const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \ michael@0: \ michael@0: stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \ michael@0: stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \ michael@0: stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \ michael@0: stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \ michael@0: stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \ michael@0: stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \ michael@0: stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \ michael@0: stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \ michael@0: stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \ michael@0: stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \ michael@0: stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \ michael@0: stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \ michael@0: stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \ michael@0: stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \ michael@0: stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \ michael@0: stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \ michael@0: \ michael@0: stp1_16 = stp2_16; \ michael@0: stp1_17 = stp2_17; \ michael@0: stp1_18 = stp2_18; \ michael@0: stp1_19 = stp2_19; \ michael@0: \ michael@0: MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \ michael@0: stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, \ michael@0: stp1_21, stp1_26) \ michael@0: MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \ michael@0: stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, \ michael@0: stp1_23, stp1_24) \ michael@0: \ michael@0: stp1_28 = stp2_28; \ michael@0: stp1_29 = stp2_29; \ michael@0: stp1_30 = stp2_30; \ michael@0: stp1_31 = stp2_31; \ michael@0: } michael@0: michael@0: // Only upper-left 8x8 has non-zero coeff michael@0: void vp9_idct32x32_34_add_sse2(const int16_t *input, uint8_t *dest, michael@0: int stride) { michael@0: const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i final_rounding = _mm_set1_epi16(1<<5); michael@0: michael@0: // idct constants for each stage michael@0: const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64); michael@0: const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64); michael@0: const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64); michael@0: const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64); michael@0: const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64); michael@0: const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64); michael@0: const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64); michael@0: const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64); michael@0: const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64); michael@0: const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64); michael@0: const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64); michael@0: const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64); michael@0: const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64); michael@0: const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64); michael@0: const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64); michael@0: const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64); michael@0: michael@0: const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); michael@0: const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); michael@0: const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64); michael@0: const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64); michael@0: const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64); michael@0: const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64); michael@0: const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); michael@0: const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); michael@0: michael@0: const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); michael@0: const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); michael@0: const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64); michael@0: const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64); michael@0: const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64); michael@0: const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64); michael@0: const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64); michael@0: const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64); michael@0: const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64); michael@0: const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64); michael@0: michael@0: const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); michael@0: const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); michael@0: const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); michael@0: michael@0: const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); michael@0: michael@0: __m128i in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11, in12, michael@0: in13, in14, in15, in16, in17, in18, in19, in20, in21, in22, in23, michael@0: in24, in25, in26, in27, in28, in29, in30, in31; michael@0: __m128i col[128]; michael@0: __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7, michael@0: stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, michael@0: stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, michael@0: stp1_23, stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, michael@0: stp1_30, stp1_31; michael@0: __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7, michael@0: stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15, michael@0: stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, michael@0: stp2_23, stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, michael@0: stp2_30, stp2_31; michael@0: __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; michael@0: int i, j, i32; michael@0: michael@0: // We work on a 8x32 block each time, and loop 8 times for 2-D 32x32 idct. michael@0: for (i = 0; i < 8; i++) { michael@0: i32 = (i << 5); michael@0: if (i == 0) { michael@0: // First 1-D idct: first 8 rows michael@0: // Load input data. michael@0: LOAD_DQCOEFF(in0, input); michael@0: LOAD_DQCOEFF(in8, input); michael@0: LOAD_DQCOEFF(in16, input); michael@0: LOAD_DQCOEFF(in24, input); michael@0: LOAD_DQCOEFF(in1, input); michael@0: LOAD_DQCOEFF(in9, input); michael@0: LOAD_DQCOEFF(in17, input); michael@0: LOAD_DQCOEFF(in25, input); michael@0: LOAD_DQCOEFF(in2, input); michael@0: LOAD_DQCOEFF(in10, input); michael@0: LOAD_DQCOEFF(in18, input); michael@0: LOAD_DQCOEFF(in26, input); michael@0: LOAD_DQCOEFF(in3, input); michael@0: LOAD_DQCOEFF(in11, input); michael@0: LOAD_DQCOEFF(in19, input); michael@0: LOAD_DQCOEFF(in27, input); michael@0: michael@0: LOAD_DQCOEFF(in4, input); michael@0: LOAD_DQCOEFF(in12, input); michael@0: LOAD_DQCOEFF(in20, input); michael@0: LOAD_DQCOEFF(in28, input); michael@0: LOAD_DQCOEFF(in5, input); michael@0: LOAD_DQCOEFF(in13, input); michael@0: LOAD_DQCOEFF(in21, input); michael@0: LOAD_DQCOEFF(in29, input); michael@0: LOAD_DQCOEFF(in6, input); michael@0: LOAD_DQCOEFF(in14, input); michael@0: LOAD_DQCOEFF(in22, input); michael@0: LOAD_DQCOEFF(in30, input); michael@0: LOAD_DQCOEFF(in7, input); michael@0: LOAD_DQCOEFF(in15, input); michael@0: LOAD_DQCOEFF(in23, input); michael@0: LOAD_DQCOEFF(in31, input); michael@0: michael@0: // Transpose 32x8 block to 8x32 block michael@0: TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, michael@0: in4, in5, in6, in7); michael@0: TRANSPOSE_8X8(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9, michael@0: in10, in11, in12, in13, in14, in15); michael@0: TRANSPOSE_8X8(in16, in17, in18, in19, in20, in21, in22, in23, in16, in17, michael@0: in18, in19, in20, in21, in22, in23); michael@0: TRANSPOSE_8X8(in24, in25, in26, in27, in28, in29, in30, in31, in24, in25, michael@0: in26, in27, in28, in29, in30, in31); michael@0: } else if (i < 4) { michael@0: // First 1-D idct: next 24 zero-coeff rows michael@0: col[i32 + 0] = _mm_setzero_si128(); michael@0: col[i32 + 1] = _mm_setzero_si128(); michael@0: col[i32 + 2] = _mm_setzero_si128(); michael@0: col[i32 + 3] = _mm_setzero_si128(); michael@0: col[i32 + 4] = _mm_setzero_si128(); michael@0: col[i32 + 5] = _mm_setzero_si128(); michael@0: col[i32 + 6] = _mm_setzero_si128(); michael@0: col[i32 + 7] = _mm_setzero_si128(); michael@0: col[i32 + 8] = _mm_setzero_si128(); michael@0: col[i32 + 9] = _mm_setzero_si128(); michael@0: col[i32 + 10] = _mm_setzero_si128(); michael@0: col[i32 + 11] = _mm_setzero_si128(); michael@0: col[i32 + 12] = _mm_setzero_si128(); michael@0: col[i32 + 13] = _mm_setzero_si128(); michael@0: col[i32 + 14] = _mm_setzero_si128(); michael@0: col[i32 + 15] = _mm_setzero_si128(); michael@0: col[i32 + 16] = _mm_setzero_si128(); michael@0: col[i32 + 17] = _mm_setzero_si128(); michael@0: col[i32 + 18] = _mm_setzero_si128(); michael@0: col[i32 + 19] = _mm_setzero_si128(); michael@0: col[i32 + 20] = _mm_setzero_si128(); michael@0: col[i32 + 21] = _mm_setzero_si128(); michael@0: col[i32 + 22] = _mm_setzero_si128(); michael@0: col[i32 + 23] = _mm_setzero_si128(); michael@0: col[i32 + 24] = _mm_setzero_si128(); michael@0: col[i32 + 25] = _mm_setzero_si128(); michael@0: col[i32 + 26] = _mm_setzero_si128(); michael@0: col[i32 + 27] = _mm_setzero_si128(); michael@0: col[i32 + 28] = _mm_setzero_si128(); michael@0: col[i32 + 29] = _mm_setzero_si128(); michael@0: col[i32 + 30] = _mm_setzero_si128(); michael@0: col[i32 + 31] = _mm_setzero_si128(); michael@0: continue; michael@0: } else { michael@0: // Second 1-D idct michael@0: j = i - 4; michael@0: michael@0: // Transpose 32x8 block to 8x32 block michael@0: TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2], michael@0: col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5], michael@0: col[j * 8 + 6], col[j * 8 + 7], in0, in1, in2, in3, in4, michael@0: in5, in6, in7); michael@0: j += 4; michael@0: TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2], michael@0: col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5], michael@0: col[j * 8 + 6], col[j * 8 + 7], in8, in9, in10, michael@0: in11, in12, in13, in14, in15); michael@0: j += 4; michael@0: TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2], michael@0: col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5], michael@0: col[j * 8 + 6], col[j * 8 + 7], in16, in17, in18, michael@0: in19, in20, in21, in22, in23); michael@0: j += 4; michael@0: TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2], michael@0: col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5], michael@0: col[j * 8 + 6], col[j * 8 + 7], in24, in25, in26, in27, michael@0: in28, in29, in30, in31); michael@0: } michael@0: michael@0: IDCT32_1D michael@0: michael@0: // final stage michael@0: if (i < 4) { michael@0: // 1_D: Store 32 intermediate results for each 8x32 block. michael@0: col[i32 + 0] = _mm_add_epi16(stp1_0, stp1_31); michael@0: col[i32 + 1] = _mm_add_epi16(stp1_1, stp1_30); michael@0: col[i32 + 2] = _mm_add_epi16(stp1_2, stp1_29); michael@0: col[i32 + 3] = _mm_add_epi16(stp1_3, stp1_28); michael@0: col[i32 + 4] = _mm_add_epi16(stp1_4, stp1_27); michael@0: col[i32 + 5] = _mm_add_epi16(stp1_5, stp1_26); michael@0: col[i32 + 6] = _mm_add_epi16(stp1_6, stp1_25); michael@0: col[i32 + 7] = _mm_add_epi16(stp1_7, stp1_24); michael@0: col[i32 + 8] = _mm_add_epi16(stp1_8, stp1_23); michael@0: col[i32 + 9] = _mm_add_epi16(stp1_9, stp1_22); michael@0: col[i32 + 10] = _mm_add_epi16(stp1_10, stp1_21); michael@0: col[i32 + 11] = _mm_add_epi16(stp1_11, stp1_20); michael@0: col[i32 + 12] = _mm_add_epi16(stp1_12, stp1_19); michael@0: col[i32 + 13] = _mm_add_epi16(stp1_13, stp1_18); michael@0: col[i32 + 14] = _mm_add_epi16(stp1_14, stp1_17); michael@0: col[i32 + 15] = _mm_add_epi16(stp1_15, stp1_16); michael@0: col[i32 + 16] = _mm_sub_epi16(stp1_15, stp1_16); michael@0: col[i32 + 17] = _mm_sub_epi16(stp1_14, stp1_17); michael@0: col[i32 + 18] = _mm_sub_epi16(stp1_13, stp1_18); michael@0: col[i32 + 19] = _mm_sub_epi16(stp1_12, stp1_19); michael@0: col[i32 + 20] = _mm_sub_epi16(stp1_11, stp1_20); michael@0: col[i32 + 21] = _mm_sub_epi16(stp1_10, stp1_21); michael@0: col[i32 + 22] = _mm_sub_epi16(stp1_9, stp1_22); michael@0: col[i32 + 23] = _mm_sub_epi16(stp1_8, stp1_23); michael@0: col[i32 + 24] = _mm_sub_epi16(stp1_7, stp1_24); michael@0: col[i32 + 25] = _mm_sub_epi16(stp1_6, stp1_25); michael@0: col[i32 + 26] = _mm_sub_epi16(stp1_5, stp1_26); michael@0: col[i32 + 27] = _mm_sub_epi16(stp1_4, stp1_27); michael@0: col[i32 + 28] = _mm_sub_epi16(stp1_3, stp1_28); michael@0: col[i32 + 29] = _mm_sub_epi16(stp1_2, stp1_29); michael@0: col[i32 + 30] = _mm_sub_epi16(stp1_1, stp1_30); michael@0: col[i32 + 31] = _mm_sub_epi16(stp1_0, stp1_31); michael@0: } else { michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: michael@0: // 2_D: Calculate the results and store them to destination. michael@0: in0 = _mm_add_epi16(stp1_0, stp1_31); michael@0: in1 = _mm_add_epi16(stp1_1, stp1_30); michael@0: in2 = _mm_add_epi16(stp1_2, stp1_29); michael@0: in3 = _mm_add_epi16(stp1_3, stp1_28); michael@0: in4 = _mm_add_epi16(stp1_4, stp1_27); michael@0: in5 = _mm_add_epi16(stp1_5, stp1_26); michael@0: in6 = _mm_add_epi16(stp1_6, stp1_25); michael@0: in7 = _mm_add_epi16(stp1_7, stp1_24); michael@0: in8 = _mm_add_epi16(stp1_8, stp1_23); michael@0: in9 = _mm_add_epi16(stp1_9, stp1_22); michael@0: in10 = _mm_add_epi16(stp1_10, stp1_21); michael@0: in11 = _mm_add_epi16(stp1_11, stp1_20); michael@0: in12 = _mm_add_epi16(stp1_12, stp1_19); michael@0: in13 = _mm_add_epi16(stp1_13, stp1_18); michael@0: in14 = _mm_add_epi16(stp1_14, stp1_17); michael@0: in15 = _mm_add_epi16(stp1_15, stp1_16); michael@0: in16 = _mm_sub_epi16(stp1_15, stp1_16); michael@0: in17 = _mm_sub_epi16(stp1_14, stp1_17); michael@0: in18 = _mm_sub_epi16(stp1_13, stp1_18); michael@0: in19 = _mm_sub_epi16(stp1_12, stp1_19); michael@0: in20 = _mm_sub_epi16(stp1_11, stp1_20); michael@0: in21 = _mm_sub_epi16(stp1_10, stp1_21); michael@0: in22 = _mm_sub_epi16(stp1_9, stp1_22); michael@0: in23 = _mm_sub_epi16(stp1_8, stp1_23); michael@0: in24 = _mm_sub_epi16(stp1_7, stp1_24); michael@0: in25 = _mm_sub_epi16(stp1_6, stp1_25); michael@0: in26 = _mm_sub_epi16(stp1_5, stp1_26); michael@0: in27 = _mm_sub_epi16(stp1_4, stp1_27); michael@0: in28 = _mm_sub_epi16(stp1_3, stp1_28); michael@0: in29 = _mm_sub_epi16(stp1_2, stp1_29); michael@0: in30 = _mm_sub_epi16(stp1_1, stp1_30); michael@0: in31 = _mm_sub_epi16(stp1_0, stp1_31); michael@0: michael@0: // Final rounding and shift michael@0: in0 = _mm_adds_epi16(in0, final_rounding); michael@0: in1 = _mm_adds_epi16(in1, final_rounding); michael@0: in2 = _mm_adds_epi16(in2, final_rounding); michael@0: in3 = _mm_adds_epi16(in3, final_rounding); michael@0: in4 = _mm_adds_epi16(in4, final_rounding); michael@0: in5 = _mm_adds_epi16(in5, final_rounding); michael@0: in6 = _mm_adds_epi16(in6, final_rounding); michael@0: in7 = _mm_adds_epi16(in7, final_rounding); michael@0: in8 = _mm_adds_epi16(in8, final_rounding); michael@0: in9 = _mm_adds_epi16(in9, final_rounding); michael@0: in10 = _mm_adds_epi16(in10, final_rounding); michael@0: in11 = _mm_adds_epi16(in11, final_rounding); michael@0: in12 = _mm_adds_epi16(in12, final_rounding); michael@0: in13 = _mm_adds_epi16(in13, final_rounding); michael@0: in14 = _mm_adds_epi16(in14, final_rounding); michael@0: in15 = _mm_adds_epi16(in15, final_rounding); michael@0: in16 = _mm_adds_epi16(in16, final_rounding); michael@0: in17 = _mm_adds_epi16(in17, final_rounding); michael@0: in18 = _mm_adds_epi16(in18, final_rounding); michael@0: in19 = _mm_adds_epi16(in19, final_rounding); michael@0: in20 = _mm_adds_epi16(in20, final_rounding); michael@0: in21 = _mm_adds_epi16(in21, final_rounding); michael@0: in22 = _mm_adds_epi16(in22, final_rounding); michael@0: in23 = _mm_adds_epi16(in23, final_rounding); michael@0: in24 = _mm_adds_epi16(in24, final_rounding); michael@0: in25 = _mm_adds_epi16(in25, final_rounding); michael@0: in26 = _mm_adds_epi16(in26, final_rounding); michael@0: in27 = _mm_adds_epi16(in27, final_rounding); michael@0: in28 = _mm_adds_epi16(in28, final_rounding); michael@0: in29 = _mm_adds_epi16(in29, final_rounding); michael@0: in30 = _mm_adds_epi16(in30, final_rounding); michael@0: in31 = _mm_adds_epi16(in31, final_rounding); michael@0: michael@0: in0 = _mm_srai_epi16(in0, 6); michael@0: in1 = _mm_srai_epi16(in1, 6); michael@0: in2 = _mm_srai_epi16(in2, 6); michael@0: in3 = _mm_srai_epi16(in3, 6); michael@0: in4 = _mm_srai_epi16(in4, 6); michael@0: in5 = _mm_srai_epi16(in5, 6); michael@0: in6 = _mm_srai_epi16(in6, 6); michael@0: in7 = _mm_srai_epi16(in7, 6); michael@0: in8 = _mm_srai_epi16(in8, 6); michael@0: in9 = _mm_srai_epi16(in9, 6); michael@0: in10 = _mm_srai_epi16(in10, 6); michael@0: in11 = _mm_srai_epi16(in11, 6); michael@0: in12 = _mm_srai_epi16(in12, 6); michael@0: in13 = _mm_srai_epi16(in13, 6); michael@0: in14 = _mm_srai_epi16(in14, 6); michael@0: in15 = _mm_srai_epi16(in15, 6); michael@0: in16 = _mm_srai_epi16(in16, 6); michael@0: in17 = _mm_srai_epi16(in17, 6); michael@0: in18 = _mm_srai_epi16(in18, 6); michael@0: in19 = _mm_srai_epi16(in19, 6); michael@0: in20 = _mm_srai_epi16(in20, 6); michael@0: in21 = _mm_srai_epi16(in21, 6); michael@0: in22 = _mm_srai_epi16(in22, 6); michael@0: in23 = _mm_srai_epi16(in23, 6); michael@0: in24 = _mm_srai_epi16(in24, 6); michael@0: in25 = _mm_srai_epi16(in25, 6); michael@0: in26 = _mm_srai_epi16(in26, 6); michael@0: in27 = _mm_srai_epi16(in27, 6); michael@0: in28 = _mm_srai_epi16(in28, 6); michael@0: in29 = _mm_srai_epi16(in29, 6); michael@0: in30 = _mm_srai_epi16(in30, 6); michael@0: in31 = _mm_srai_epi16(in31, 6); michael@0: michael@0: RECON_AND_STORE(dest, in0); michael@0: RECON_AND_STORE(dest, in1); michael@0: RECON_AND_STORE(dest, in2); michael@0: RECON_AND_STORE(dest, in3); michael@0: RECON_AND_STORE(dest, in4); michael@0: RECON_AND_STORE(dest, in5); michael@0: RECON_AND_STORE(dest, in6); michael@0: RECON_AND_STORE(dest, in7); michael@0: RECON_AND_STORE(dest, in8); michael@0: RECON_AND_STORE(dest, in9); michael@0: RECON_AND_STORE(dest, in10); michael@0: RECON_AND_STORE(dest, in11); michael@0: RECON_AND_STORE(dest, in12); michael@0: RECON_AND_STORE(dest, in13); michael@0: RECON_AND_STORE(dest, in14); michael@0: RECON_AND_STORE(dest, in15); michael@0: RECON_AND_STORE(dest, in16); michael@0: RECON_AND_STORE(dest, in17); michael@0: RECON_AND_STORE(dest, in18); michael@0: RECON_AND_STORE(dest, in19); michael@0: RECON_AND_STORE(dest, in20); michael@0: RECON_AND_STORE(dest, in21); michael@0: RECON_AND_STORE(dest, in22); michael@0: RECON_AND_STORE(dest, in23); michael@0: RECON_AND_STORE(dest, in24); michael@0: RECON_AND_STORE(dest, in25); michael@0: RECON_AND_STORE(dest, in26); michael@0: RECON_AND_STORE(dest, in27); michael@0: RECON_AND_STORE(dest, in28); michael@0: RECON_AND_STORE(dest, in29); michael@0: RECON_AND_STORE(dest, in30); michael@0: RECON_AND_STORE(dest, in31); michael@0: michael@0: dest += 8 - (stride * 32); michael@0: } michael@0: } michael@0: } michael@0: michael@0: void vp9_idct32x32_1024_add_sse2(const int16_t *input, uint8_t *dest, michael@0: int stride) { michael@0: const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); michael@0: const __m128i final_rounding = _mm_set1_epi16(1<<5); michael@0: michael@0: // idct constants for each stage michael@0: const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64); michael@0: const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64); michael@0: const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64); michael@0: const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64); michael@0: const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64); michael@0: const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64); michael@0: const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64); michael@0: const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64); michael@0: const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64); michael@0: const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64); michael@0: const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64); michael@0: const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64); michael@0: const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64); michael@0: const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64); michael@0: const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64); michael@0: const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64); michael@0: michael@0: const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); michael@0: const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); michael@0: const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64); michael@0: const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64); michael@0: const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64); michael@0: const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64); michael@0: const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); michael@0: const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); michael@0: michael@0: const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); michael@0: const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); michael@0: const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64); michael@0: const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64); michael@0: const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64); michael@0: const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64); michael@0: const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64); michael@0: const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64); michael@0: const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64); michael@0: const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64); michael@0: michael@0: const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); michael@0: const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); michael@0: const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); michael@0: const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64); michael@0: const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); michael@0: const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); michael@0: const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); michael@0: michael@0: const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); michael@0: michael@0: __m128i in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11, in12, michael@0: in13, in14, in15, in16, in17, in18, in19, in20, in21, in22, in23, michael@0: in24, in25, in26, in27, in28, in29, in30, in31; michael@0: __m128i col[128]; michael@0: __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7, michael@0: stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, michael@0: stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, michael@0: stp1_23, stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, michael@0: stp1_30, stp1_31; michael@0: __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7, michael@0: stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15, michael@0: stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, michael@0: stp2_23, stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, michael@0: stp2_30, stp2_31; michael@0: __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; michael@0: int i, j, i32; michael@0: __m128i zero_idx[16]; michael@0: int zero_flag[2]; michael@0: michael@0: // We work on a 8x32 block each time, and loop 8 times for 2-D 32x32 idct. michael@0: for (i = 0; i < 8; i++) { michael@0: i32 = (i << 5); michael@0: if (i < 4) { michael@0: // First 1-D idct michael@0: // Load input data. michael@0: LOAD_DQCOEFF(in0, input); michael@0: LOAD_DQCOEFF(in8, input); michael@0: LOAD_DQCOEFF(in16, input); michael@0: LOAD_DQCOEFF(in24, input); michael@0: LOAD_DQCOEFF(in1, input); michael@0: LOAD_DQCOEFF(in9, input); michael@0: LOAD_DQCOEFF(in17, input); michael@0: LOAD_DQCOEFF(in25, input); michael@0: LOAD_DQCOEFF(in2, input); michael@0: LOAD_DQCOEFF(in10, input); michael@0: LOAD_DQCOEFF(in18, input); michael@0: LOAD_DQCOEFF(in26, input); michael@0: LOAD_DQCOEFF(in3, input); michael@0: LOAD_DQCOEFF(in11, input); michael@0: LOAD_DQCOEFF(in19, input); michael@0: LOAD_DQCOEFF(in27, input); michael@0: michael@0: LOAD_DQCOEFF(in4, input); michael@0: LOAD_DQCOEFF(in12, input); michael@0: LOAD_DQCOEFF(in20, input); michael@0: LOAD_DQCOEFF(in28, input); michael@0: LOAD_DQCOEFF(in5, input); michael@0: LOAD_DQCOEFF(in13, input); michael@0: LOAD_DQCOEFF(in21, input); michael@0: LOAD_DQCOEFF(in29, input); michael@0: LOAD_DQCOEFF(in6, input); michael@0: LOAD_DQCOEFF(in14, input); michael@0: LOAD_DQCOEFF(in22, input); michael@0: LOAD_DQCOEFF(in30, input); michael@0: LOAD_DQCOEFF(in7, input); michael@0: LOAD_DQCOEFF(in15, input); michael@0: LOAD_DQCOEFF(in23, input); michael@0: LOAD_DQCOEFF(in31, input); michael@0: michael@0: // checking if all entries are zero michael@0: zero_idx[0] = _mm_or_si128(in0, in1); michael@0: zero_idx[1] = _mm_or_si128(in2, in3); michael@0: zero_idx[2] = _mm_or_si128(in4, in5); michael@0: zero_idx[3] = _mm_or_si128(in6, in7); michael@0: zero_idx[4] = _mm_or_si128(in8, in9); michael@0: zero_idx[5] = _mm_or_si128(in10, in11); michael@0: zero_idx[6] = _mm_or_si128(in12, in13); michael@0: zero_idx[7] = _mm_or_si128(in14, in15); michael@0: zero_idx[8] = _mm_or_si128(in16, in17); michael@0: zero_idx[9] = _mm_or_si128(in18, in19); michael@0: zero_idx[10] = _mm_or_si128(in20, in21); michael@0: zero_idx[11] = _mm_or_si128(in22, in23); michael@0: zero_idx[12] = _mm_or_si128(in24, in25); michael@0: zero_idx[13] = _mm_or_si128(in26, in27); michael@0: zero_idx[14] = _mm_or_si128(in28, in29); michael@0: zero_idx[15] = _mm_or_si128(in30, in31); michael@0: michael@0: zero_idx[0] = _mm_or_si128(zero_idx[0], zero_idx[1]); michael@0: zero_idx[1] = _mm_or_si128(zero_idx[2], zero_idx[3]); michael@0: zero_idx[2] = _mm_or_si128(zero_idx[4], zero_idx[5]); michael@0: zero_idx[3] = _mm_or_si128(zero_idx[6], zero_idx[7]); michael@0: zero_idx[4] = _mm_or_si128(zero_idx[8], zero_idx[9]); michael@0: zero_idx[5] = _mm_or_si128(zero_idx[10], zero_idx[11]); michael@0: zero_idx[6] = _mm_or_si128(zero_idx[12], zero_idx[13]); michael@0: zero_idx[7] = _mm_or_si128(zero_idx[14], zero_idx[15]); michael@0: michael@0: zero_idx[8] = _mm_or_si128(zero_idx[0], zero_idx[1]); michael@0: zero_idx[9] = _mm_or_si128(zero_idx[2], zero_idx[3]); michael@0: zero_idx[10] = _mm_or_si128(zero_idx[4], zero_idx[5]); michael@0: zero_idx[11] = _mm_or_si128(zero_idx[6], zero_idx[7]); michael@0: zero_idx[12] = _mm_or_si128(zero_idx[8], zero_idx[9]); michael@0: zero_idx[13] = _mm_or_si128(zero_idx[10], zero_idx[11]); michael@0: zero_idx[14] = _mm_or_si128(zero_idx[12], zero_idx[13]); michael@0: michael@0: zero_idx[0] = _mm_unpackhi_epi64(zero_idx[14], zero_idx[14]); michael@0: zero_idx[1] = _mm_or_si128(zero_idx[0], zero_idx[14]); michael@0: zero_idx[2] = _mm_srli_epi64(zero_idx[1], 32); michael@0: zero_flag[0] = _mm_cvtsi128_si32(zero_idx[1]); michael@0: zero_flag[1] = _mm_cvtsi128_si32(zero_idx[2]); michael@0: michael@0: if (!zero_flag[0] && !zero_flag[1]) { michael@0: col[i32 + 0] = _mm_setzero_si128(); michael@0: col[i32 + 1] = _mm_setzero_si128(); michael@0: col[i32 + 2] = _mm_setzero_si128(); michael@0: col[i32 + 3] = _mm_setzero_si128(); michael@0: col[i32 + 4] = _mm_setzero_si128(); michael@0: col[i32 + 5] = _mm_setzero_si128(); michael@0: col[i32 + 6] = _mm_setzero_si128(); michael@0: col[i32 + 7] = _mm_setzero_si128(); michael@0: col[i32 + 8] = _mm_setzero_si128(); michael@0: col[i32 + 9] = _mm_setzero_si128(); michael@0: col[i32 + 10] = _mm_setzero_si128(); michael@0: col[i32 + 11] = _mm_setzero_si128(); michael@0: col[i32 + 12] = _mm_setzero_si128(); michael@0: col[i32 + 13] = _mm_setzero_si128(); michael@0: col[i32 + 14] = _mm_setzero_si128(); michael@0: col[i32 + 15] = _mm_setzero_si128(); michael@0: col[i32 + 16] = _mm_setzero_si128(); michael@0: col[i32 + 17] = _mm_setzero_si128(); michael@0: col[i32 + 18] = _mm_setzero_si128(); michael@0: col[i32 + 19] = _mm_setzero_si128(); michael@0: col[i32 + 20] = _mm_setzero_si128(); michael@0: col[i32 + 21] = _mm_setzero_si128(); michael@0: col[i32 + 22] = _mm_setzero_si128(); michael@0: col[i32 + 23] = _mm_setzero_si128(); michael@0: col[i32 + 24] = _mm_setzero_si128(); michael@0: col[i32 + 25] = _mm_setzero_si128(); michael@0: col[i32 + 26] = _mm_setzero_si128(); michael@0: col[i32 + 27] = _mm_setzero_si128(); michael@0: col[i32 + 28] = _mm_setzero_si128(); michael@0: col[i32 + 29] = _mm_setzero_si128(); michael@0: col[i32 + 30] = _mm_setzero_si128(); michael@0: col[i32 + 31] = _mm_setzero_si128(); michael@0: continue; michael@0: } michael@0: michael@0: // Transpose 32x8 block to 8x32 block michael@0: TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, michael@0: in4, in5, in6, in7); michael@0: TRANSPOSE_8X8(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9, michael@0: in10, in11, in12, in13, in14, in15); michael@0: TRANSPOSE_8X8(in16, in17, in18, in19, in20, in21, in22, in23, in16, in17, michael@0: in18, in19, in20, in21, in22, in23); michael@0: TRANSPOSE_8X8(in24, in25, in26, in27, in28, in29, in30, in31, in24, in25, michael@0: in26, in27, in28, in29, in30, in31); michael@0: } else { michael@0: // Second 1-D idct michael@0: j = i - 4; michael@0: michael@0: // Transpose 32x8 block to 8x32 block michael@0: TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2], michael@0: col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5], michael@0: col[j * 8 + 6], col[j * 8 + 7], in0, in1, in2, in3, in4, michael@0: in5, in6, in7); michael@0: j += 4; michael@0: TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2], michael@0: col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5], michael@0: col[j * 8 + 6], col[j * 8 + 7], in8, in9, in10, michael@0: in11, in12, in13, in14, in15); michael@0: j += 4; michael@0: TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2], michael@0: col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5], michael@0: col[j * 8 + 6], col[j * 8 + 7], in16, in17, in18, michael@0: in19, in20, in21, in22, in23); michael@0: j += 4; michael@0: TRANSPOSE_8X8(col[j * 8 + 0], col[j * 8 + 1], col[j * 8 + 2], michael@0: col[j * 8 + 3], col[j * 8 + 4], col[j * 8 + 5], michael@0: col[j * 8 + 6], col[j * 8 + 7], in24, in25, in26, in27, michael@0: in28, in29, in30, in31); michael@0: } michael@0: michael@0: IDCT32_1D michael@0: michael@0: // final stage michael@0: if (i < 4) { michael@0: // 1_D: Store 32 intermediate results for each 8x32 block. michael@0: col[i32 + 0] = _mm_add_epi16(stp1_0, stp1_31); michael@0: col[i32 + 1] = _mm_add_epi16(stp1_1, stp1_30); michael@0: col[i32 + 2] = _mm_add_epi16(stp1_2, stp1_29); michael@0: col[i32 + 3] = _mm_add_epi16(stp1_3, stp1_28); michael@0: col[i32 + 4] = _mm_add_epi16(stp1_4, stp1_27); michael@0: col[i32 + 5] = _mm_add_epi16(stp1_5, stp1_26); michael@0: col[i32 + 6] = _mm_add_epi16(stp1_6, stp1_25); michael@0: col[i32 + 7] = _mm_add_epi16(stp1_7, stp1_24); michael@0: col[i32 + 8] = _mm_add_epi16(stp1_8, stp1_23); michael@0: col[i32 + 9] = _mm_add_epi16(stp1_9, stp1_22); michael@0: col[i32 + 10] = _mm_add_epi16(stp1_10, stp1_21); michael@0: col[i32 + 11] = _mm_add_epi16(stp1_11, stp1_20); michael@0: col[i32 + 12] = _mm_add_epi16(stp1_12, stp1_19); michael@0: col[i32 + 13] = _mm_add_epi16(stp1_13, stp1_18); michael@0: col[i32 + 14] = _mm_add_epi16(stp1_14, stp1_17); michael@0: col[i32 + 15] = _mm_add_epi16(stp1_15, stp1_16); michael@0: col[i32 + 16] = _mm_sub_epi16(stp1_15, stp1_16); michael@0: col[i32 + 17] = _mm_sub_epi16(stp1_14, stp1_17); michael@0: col[i32 + 18] = _mm_sub_epi16(stp1_13, stp1_18); michael@0: col[i32 + 19] = _mm_sub_epi16(stp1_12, stp1_19); michael@0: col[i32 + 20] = _mm_sub_epi16(stp1_11, stp1_20); michael@0: col[i32 + 21] = _mm_sub_epi16(stp1_10, stp1_21); michael@0: col[i32 + 22] = _mm_sub_epi16(stp1_9, stp1_22); michael@0: col[i32 + 23] = _mm_sub_epi16(stp1_8, stp1_23); michael@0: col[i32 + 24] = _mm_sub_epi16(stp1_7, stp1_24); michael@0: col[i32 + 25] = _mm_sub_epi16(stp1_6, stp1_25); michael@0: col[i32 + 26] = _mm_sub_epi16(stp1_5, stp1_26); michael@0: col[i32 + 27] = _mm_sub_epi16(stp1_4, stp1_27); michael@0: col[i32 + 28] = _mm_sub_epi16(stp1_3, stp1_28); michael@0: col[i32 + 29] = _mm_sub_epi16(stp1_2, stp1_29); michael@0: col[i32 + 30] = _mm_sub_epi16(stp1_1, stp1_30); michael@0: col[i32 + 31] = _mm_sub_epi16(stp1_0, stp1_31); michael@0: } else { michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: michael@0: // 2_D: Calculate the results and store them to destination. michael@0: in0 = _mm_add_epi16(stp1_0, stp1_31); michael@0: in1 = _mm_add_epi16(stp1_1, stp1_30); michael@0: in2 = _mm_add_epi16(stp1_2, stp1_29); michael@0: in3 = _mm_add_epi16(stp1_3, stp1_28); michael@0: in4 = _mm_add_epi16(stp1_4, stp1_27); michael@0: in5 = _mm_add_epi16(stp1_5, stp1_26); michael@0: in6 = _mm_add_epi16(stp1_6, stp1_25); michael@0: in7 = _mm_add_epi16(stp1_7, stp1_24); michael@0: in8 = _mm_add_epi16(stp1_8, stp1_23); michael@0: in9 = _mm_add_epi16(stp1_9, stp1_22); michael@0: in10 = _mm_add_epi16(stp1_10, stp1_21); michael@0: in11 = _mm_add_epi16(stp1_11, stp1_20); michael@0: in12 = _mm_add_epi16(stp1_12, stp1_19); michael@0: in13 = _mm_add_epi16(stp1_13, stp1_18); michael@0: in14 = _mm_add_epi16(stp1_14, stp1_17); michael@0: in15 = _mm_add_epi16(stp1_15, stp1_16); michael@0: in16 = _mm_sub_epi16(stp1_15, stp1_16); michael@0: in17 = _mm_sub_epi16(stp1_14, stp1_17); michael@0: in18 = _mm_sub_epi16(stp1_13, stp1_18); michael@0: in19 = _mm_sub_epi16(stp1_12, stp1_19); michael@0: in20 = _mm_sub_epi16(stp1_11, stp1_20); michael@0: in21 = _mm_sub_epi16(stp1_10, stp1_21); michael@0: in22 = _mm_sub_epi16(stp1_9, stp1_22); michael@0: in23 = _mm_sub_epi16(stp1_8, stp1_23); michael@0: in24 = _mm_sub_epi16(stp1_7, stp1_24); michael@0: in25 = _mm_sub_epi16(stp1_6, stp1_25); michael@0: in26 = _mm_sub_epi16(stp1_5, stp1_26); michael@0: in27 = _mm_sub_epi16(stp1_4, stp1_27); michael@0: in28 = _mm_sub_epi16(stp1_3, stp1_28); michael@0: in29 = _mm_sub_epi16(stp1_2, stp1_29); michael@0: in30 = _mm_sub_epi16(stp1_1, stp1_30); michael@0: in31 = _mm_sub_epi16(stp1_0, stp1_31); michael@0: michael@0: // Final rounding and shift michael@0: in0 = _mm_adds_epi16(in0, final_rounding); michael@0: in1 = _mm_adds_epi16(in1, final_rounding); michael@0: in2 = _mm_adds_epi16(in2, final_rounding); michael@0: in3 = _mm_adds_epi16(in3, final_rounding); michael@0: in4 = _mm_adds_epi16(in4, final_rounding); michael@0: in5 = _mm_adds_epi16(in5, final_rounding); michael@0: in6 = _mm_adds_epi16(in6, final_rounding); michael@0: in7 = _mm_adds_epi16(in7, final_rounding); michael@0: in8 = _mm_adds_epi16(in8, final_rounding); michael@0: in9 = _mm_adds_epi16(in9, final_rounding); michael@0: in10 = _mm_adds_epi16(in10, final_rounding); michael@0: in11 = _mm_adds_epi16(in11, final_rounding); michael@0: in12 = _mm_adds_epi16(in12, final_rounding); michael@0: in13 = _mm_adds_epi16(in13, final_rounding); michael@0: in14 = _mm_adds_epi16(in14, final_rounding); michael@0: in15 = _mm_adds_epi16(in15, final_rounding); michael@0: in16 = _mm_adds_epi16(in16, final_rounding); michael@0: in17 = _mm_adds_epi16(in17, final_rounding); michael@0: in18 = _mm_adds_epi16(in18, final_rounding); michael@0: in19 = _mm_adds_epi16(in19, final_rounding); michael@0: in20 = _mm_adds_epi16(in20, final_rounding); michael@0: in21 = _mm_adds_epi16(in21, final_rounding); michael@0: in22 = _mm_adds_epi16(in22, final_rounding); michael@0: in23 = _mm_adds_epi16(in23, final_rounding); michael@0: in24 = _mm_adds_epi16(in24, final_rounding); michael@0: in25 = _mm_adds_epi16(in25, final_rounding); michael@0: in26 = _mm_adds_epi16(in26, final_rounding); michael@0: in27 = _mm_adds_epi16(in27, final_rounding); michael@0: in28 = _mm_adds_epi16(in28, final_rounding); michael@0: in29 = _mm_adds_epi16(in29, final_rounding); michael@0: in30 = _mm_adds_epi16(in30, final_rounding); michael@0: in31 = _mm_adds_epi16(in31, final_rounding); michael@0: michael@0: in0 = _mm_srai_epi16(in0, 6); michael@0: in1 = _mm_srai_epi16(in1, 6); michael@0: in2 = _mm_srai_epi16(in2, 6); michael@0: in3 = _mm_srai_epi16(in3, 6); michael@0: in4 = _mm_srai_epi16(in4, 6); michael@0: in5 = _mm_srai_epi16(in5, 6); michael@0: in6 = _mm_srai_epi16(in6, 6); michael@0: in7 = _mm_srai_epi16(in7, 6); michael@0: in8 = _mm_srai_epi16(in8, 6); michael@0: in9 = _mm_srai_epi16(in9, 6); michael@0: in10 = _mm_srai_epi16(in10, 6); michael@0: in11 = _mm_srai_epi16(in11, 6); michael@0: in12 = _mm_srai_epi16(in12, 6); michael@0: in13 = _mm_srai_epi16(in13, 6); michael@0: in14 = _mm_srai_epi16(in14, 6); michael@0: in15 = _mm_srai_epi16(in15, 6); michael@0: in16 = _mm_srai_epi16(in16, 6); michael@0: in17 = _mm_srai_epi16(in17, 6); michael@0: in18 = _mm_srai_epi16(in18, 6); michael@0: in19 = _mm_srai_epi16(in19, 6); michael@0: in20 = _mm_srai_epi16(in20, 6); michael@0: in21 = _mm_srai_epi16(in21, 6); michael@0: in22 = _mm_srai_epi16(in22, 6); michael@0: in23 = _mm_srai_epi16(in23, 6); michael@0: in24 = _mm_srai_epi16(in24, 6); michael@0: in25 = _mm_srai_epi16(in25, 6); michael@0: in26 = _mm_srai_epi16(in26, 6); michael@0: in27 = _mm_srai_epi16(in27, 6); michael@0: in28 = _mm_srai_epi16(in28, 6); michael@0: in29 = _mm_srai_epi16(in29, 6); michael@0: in30 = _mm_srai_epi16(in30, 6); michael@0: in31 = _mm_srai_epi16(in31, 6); michael@0: michael@0: RECON_AND_STORE(dest, in0); michael@0: RECON_AND_STORE(dest, in1); michael@0: RECON_AND_STORE(dest, in2); michael@0: RECON_AND_STORE(dest, in3); michael@0: RECON_AND_STORE(dest, in4); michael@0: RECON_AND_STORE(dest, in5); michael@0: RECON_AND_STORE(dest, in6); michael@0: RECON_AND_STORE(dest, in7); michael@0: RECON_AND_STORE(dest, in8); michael@0: RECON_AND_STORE(dest, in9); michael@0: RECON_AND_STORE(dest, in10); michael@0: RECON_AND_STORE(dest, in11); michael@0: RECON_AND_STORE(dest, in12); michael@0: RECON_AND_STORE(dest, in13); michael@0: RECON_AND_STORE(dest, in14); michael@0: RECON_AND_STORE(dest, in15); michael@0: RECON_AND_STORE(dest, in16); michael@0: RECON_AND_STORE(dest, in17); michael@0: RECON_AND_STORE(dest, in18); michael@0: RECON_AND_STORE(dest, in19); michael@0: RECON_AND_STORE(dest, in20); michael@0: RECON_AND_STORE(dest, in21); michael@0: RECON_AND_STORE(dest, in22); michael@0: RECON_AND_STORE(dest, in23); michael@0: RECON_AND_STORE(dest, in24); michael@0: RECON_AND_STORE(dest, in25); michael@0: RECON_AND_STORE(dest, in26); michael@0: RECON_AND_STORE(dest, in27); michael@0: RECON_AND_STORE(dest, in28); michael@0: RECON_AND_STORE(dest, in29); michael@0: RECON_AND_STORE(dest, in30); michael@0: RECON_AND_STORE(dest, in31); michael@0: michael@0: dest += 8 - (stride * 32); michael@0: } michael@0: } michael@0: } //NOLINT michael@0: michael@0: void vp9_idct32x32_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) { michael@0: __m128i dc_value; michael@0: const __m128i zero = _mm_setzero_si128(); michael@0: int a, i; michael@0: michael@0: a = dct_const_round_shift(input[0] * cospi_16_64); michael@0: a = dct_const_round_shift(a * cospi_16_64); michael@0: a = ROUND_POWER_OF_TWO(a, 6); michael@0: michael@0: dc_value = _mm_set1_epi16(a); michael@0: michael@0: for (i = 0; i < 4; ++i) { michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: RECON_AND_STORE(dest, dc_value); michael@0: dest += 8 - (stride * 32); michael@0: } michael@0: }