1f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; 2f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; Copyright (c) 2013 The WebM project authors. All Rights Reserved. 3f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; 4f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; Use of this source code is governed by a BSD-style license 5f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; that can be found in the LICENSE file in the root of the source 6f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; tree. An additional intellectual property rights grant can be found 7f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; in the file PATENTS. All contributing project authors may 8f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; be found in the AUTHORS file in the root of the source tree. 9f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; 10f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 11ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org EXPORT |vp9_iht8x8_64_add_neon| 12f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ARM 13f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org REQUIRE8 14f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org PRESERVE8 15f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 16f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org AREA ||.text||, CODE, READONLY, ALIGN=2 17f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 18f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; Generate IADST constants in r0 - r12 for the IADST. 19f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MACRO 20f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org GENERATE_IADST_CONSTANTS 21f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_2_64 = 16305 22f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r0, #0x3f00 23f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r0, #0xb1 24f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 25f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_30_64 = 1606 26f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r1, #0x600 27f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r1, #0x46 28f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 29f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_10_64 = 14449 30f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r2, #0x3800 31f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r2, #0x71 32f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 33f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_22_64 = 7723 34f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r3, #0x1e00 35f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r3, #0x2b 36f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 37f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_18_64 = 10394 38f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r4, #0x2800 39f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r4, #0x9a 40f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 41f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_14_64 = 12665 42f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r5, #0x3100 43f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r5, #0x79 44f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 45f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_26_64 = 4756 46f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r6, #0x1200 47f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r6, #0x94 48f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 49f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_6_64 = 15679 50f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r7, #0x3d00 51f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r7, #0x3f 52f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 53f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_8_64 = 15137 54f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r8, #0x3b00 55f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r8, #0x21 56f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 57f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_24_64 = 6270 58f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r9, #0x1800 59f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r9, #0x7e 60f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 61f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate 0 62f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r10, #0 63f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 64f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_16_64 = 11585 65f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r12, #0x2d00 66f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r12, #0x41 67f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MEND 68f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 69f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; Generate IDCT constants in r3 - r9 for the IDCT. 70f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MACRO 71f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org GENERATE_IDCT_CONSTANTS 72f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_28_64 = 3196 73f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r3, #0x0c00 74f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r3, #0x7c 75f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 76f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_4_64 = 16069 77f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r4, #0x3e00 78f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r4, #0xc5 79f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 80f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_12_64 = 13623 81f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r5, #0x3500 82f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r5, #0x37 83f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 84f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_20_64 = 9102 85f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r6, #0x2300 86f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r6, #0x8e 87f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 88f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_16_64 = 11585 89f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r7, #0x2d00 90f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r7, #0x41 91f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 92f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_24_64 = 6270 93f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r8, #0x1800 94f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r8, #0x7e 95f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 96f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate cospi_8_64 = 15137 97f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r9, #0x3b00 98f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org add r9, #0x21 99f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MEND 100f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 101f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; Transpose a 8x8 16bits data matrix. Datas are loaded in q8-q15. 102f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MACRO 103f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org TRANSPOSE8X8 104f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vswp d17, d24 105f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vswp d23, d30 106f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vswp d21, d28 107f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vswp d19, d26 108f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vtrn.32 q8, q10 109f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vtrn.32 q9, q11 110f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vtrn.32 q12, q14 111f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vtrn.32 q13, q15 112f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vtrn.16 q8, q9 113f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vtrn.16 q10, q11 114f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vtrn.16 q12, q13 115f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vtrn.16 q14, q15 116f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MEND 117f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 118f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; Parallel 1D IDCT on all the columns of a 8x8 16bits data matrix which are 119f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; loaded in q8-q15. The IDCT constants are loaded in r3 - r9. The output 120f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; will be stored back into q8-q15 registers. This macro will touch q0-q7 121f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; registers and use them as buffer during calculation. 122f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MACRO 123f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org IDCT8x8_1D 124f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; stage 1 125f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d0, r3 ; duplicate cospi_28_64 126f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d1, r4 ; duplicate cospi_4_64 127f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d2, r5 ; duplicate cospi_12_64 128f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d3, r6 ; duplicate cospi_20_64 129f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 130f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[1] * cospi_28_64 131f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q2, d18, d0 132f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q3, d19, d0 133f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 134f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[5] * cospi_12_64 135f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q5, d26, d2 136f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q6, d27, d2 137f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 138f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[1]*cospi_28_64-input[7]*cospi_4_64 139f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q2, d30, d1 140f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q3, d31, d1 141f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 142f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[5] * cospi_12_64 - input[3] * cospi_20_64 143f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q5, d22, d3 144f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q6, d23, d3 145f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 146f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 147f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d8, q2, #14 ; >> 14 148f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d9, q3, #14 ; >> 14 149f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 150f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 151f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d10, q5, #14 ; >> 14 152f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d11, q6, #14 ; >> 14 153f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 154f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[1] * cospi_4_64 155f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q2, d18, d1 156f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q3, d19, d1 157f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 158f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[5] * cospi_20_64 159f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q9, d26, d3 160f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q13, d27, d3 161f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 162f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[1]*cospi_4_64+input[7]*cospi_28_64 163f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q2, d30, d0 164f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q3, d31, d0 165f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 166f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[5] * cospi_20_64 + input[3] * cospi_12_64 167f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q9, d22, d2 168f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q13, d23, d2 169f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 170f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 171f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d14, q2, #14 ; >> 14 172f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d15, q3, #14 ; >> 14 173f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 174f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; stage 2 & stage 3 - even half 175f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d0, r7 ; duplicate cospi_16_64 176f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 177f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 178f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d12, q9, #14 ; >> 14 179f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d13, q13, #14 ; >> 14 180f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 181f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[0] * cospi_16_64 182f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q2, d16, d0 183f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q3, d17, d0 184f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 185f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[0] * cospi_16_64 186f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q13, d16, d0 187f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q15, d17, d0 188f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 189f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (input[0] + input[2]) * cospi_16_64 190f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q2, d24, d0 191f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q3, d25, d0 192f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 193f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (input[0] - input[2]) * cospi_16_64 194f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q13, d24, d0 195f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q15, d25, d0 196f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 197f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d0, r8 ; duplicate cospi_24_64 198f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d1, r9 ; duplicate cospi_8_64 199f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 200f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 201f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d18, q2, #14 ; >> 14 202f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d19, q3, #14 ; >> 14 203f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 204f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 205f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d22, q13, #14 ; >> 14 206f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d23, q15, #14 ; >> 14 207f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 208f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[1] * cospi_24_64 209f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q2, d20, d0 210f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q3, d21, d0 211f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 212f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[1] * cospi_8_64 213f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q8, d20, d1 214f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q12, d21, d1 215f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 216f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[1] * cospi_24_64 - input[3] * cospi_8_64 217f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q2, d28, d1 218f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q3, d29, d1 219f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 220f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; input[1] * cospi_8_64 + input[3] * cospi_24_64 221f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q8, d28, d0 222f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q12, d29, d0 223f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 224f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 225f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d26, q2, #14 ; >> 14 226f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d27, q3, #14 ; >> 14 227f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 228f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 229f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d30, q8, #14 ; >> 14 230f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d31, q12, #14 ; >> 14 231f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 232f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q0, q9, q15 ; output[0] = step[0] + step[3] 233f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q1, q11, q13 ; output[1] = step[1] + step[2] 234f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q2, q11, q13 ; output[2] = step[1] - step[2] 235f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q3, q9, q15 ; output[3] = step[0] - step[3] 236f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 237f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; stage 3 -odd half 238f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d16, r7 ; duplicate cospi_16_64 239f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 240f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; stage 2 - odd half 241f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q13, q4, q5 ; step2[5] = step1[4] - step1[5] 242f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q4, q4, q5 ; step2[4] = step1[4] + step1[5] 243f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q14, q7, q6 ; step2[6] = -step1[6] + step1[7] 244f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q7, q7, q6 ; step2[7] = step1[6] + step1[7] 245f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 246f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; step2[6] * cospi_16_64 247f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q9, d28, d16 248f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q10, d29, d16 249f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 250f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; step2[6] * cospi_16_64 251f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q11, d28, d16 252f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q12, d29, d16 253f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 254f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (step2[6] - step2[5]) * cospi_16_64 255f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q9, d26, d16 256f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q10, d27, d16 257f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 258f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (step2[5] + step2[6]) * cospi_16_64 259f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q11, d26, d16 260f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q12, d27, d16 261f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 262f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 263f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d10, q9, #14 ; >> 14 264f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d11, q10, #14 ; >> 14 265f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 266f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; dct_const_round_shift(input_dc * cospi_16_64) 267f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d12, q11, #14 ; >> 14 268f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d13, q12, #14 ; >> 14 269f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 270f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; stage 4 271f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q8, q0, q7 ; output[0] = step1[0] + step1[7]; 272f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q9, q1, q6 ; output[1] = step1[1] + step1[6]; 273f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q10, q2, q5 ; output[2] = step1[2] + step1[5]; 274f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q11, q3, q4 ; output[3] = step1[3] + step1[4]; 275f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q12, q3, q4 ; output[4] = step1[3] - step1[4]; 276f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q13, q2, q5 ; output[5] = step1[2] - step1[5]; 277f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q14, q1, q6 ; output[6] = step1[1] - step1[6]; 278f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q15, q0, q7 ; output[7] = step1[0] - step1[7]; 279f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MEND 280f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 281f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; Parallel 1D IADST on all the columns of a 8x8 16bits data matrix which 282f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; loaded in q8-q15. IADST constants are loaded in r0 - r12 registers. The 283f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; output will be stored back into q8-q15 registers. This macro will touch 284f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; q0 - q7 registers and use them as buffer during calculation. 285f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MACRO 286f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org IADST8X8_1D 287f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d14, r0 ; duplicate cospi_2_64 288f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d15, r1 ; duplicate cospi_30_64 289f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 290f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_2_64 * x0 291f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q1, d30, d14 292f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q2, d31, d14 293f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 294f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_30_64 * x0 295f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q3, d30, d15 296f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q4, d31, d15 297f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 2989c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vdup.16 d30, r4 ; duplicate cospi_18_64 2999c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vdup.16 d31, r5 ; duplicate cospi_14_64 3009c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 3019c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; s0 = cospi_2_64 * x0 + cospi_30_64 * x1; 3029c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q1, d16, d15 3039c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q2, d17, d15 3049c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 305f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; s1 = cospi_30_64 * x0 - cospi_2_64 * x1 306f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q3, d16, d14 307f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q4, d17, d14 308f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 309f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_18_64 * x4 310f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q5, d22, d30 311f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q6, d23, d30 312f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 313f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_14_64 * x4 314f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q7, d22, d31 315f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q8, d23, d31 316f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 3179c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; s4 = cospi_18_64 * x4 + cospi_14_64 * x5; 3189c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q5, d24, d31 3199c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q6, d25, d31 3209c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 321f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; s5 = cospi_14_64 * x4 - cospi_18_64 * x5 322f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q7, d24, d30 323f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q8, d25, d30 324f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 325f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s0 + s4) 326f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q11, q1, q5 327f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q12, q2, q6 328f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 3299c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vdup.16 d0, r2 ; duplicate cospi_10_64 3309c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vdup.16 d1, r3 ; duplicate cospi_22_64 331f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 332f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s0 - s4) 333f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q1, q1, q5 334f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q2, q2, q6 335f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 3369c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; x0 = dct_const_round_shift(s0 + s4); 3379c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d22, q11, #14 ; >> 14 3389c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d23, q12, #14 ; >> 14 339f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 340f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s1 + s5) 341f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q12, q3, q7 342f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q15, q4, q8 343f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 344f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s1 - s5) 345f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q3, q3, q7 346f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q4, q4, q8 347f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 3489c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; x4 = dct_const_round_shift(s0 - s4); 3499c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d2, q1, #14 ; >> 14 3509c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d3, q2, #14 ; >> 14 3519c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 3529c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; x1 = dct_const_round_shift(s1 + s5); 3539c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d24, q12, #14 ; >> 14 3549c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d25, q15, #14 ; >> 14 3559c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 356f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; x5 = dct_const_round_shift(s1 - s5); 357f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d6, q3, #14 ; >> 14 358f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d7, q4, #14 ; >> 14 359f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 360f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_10_64 * x2 3619c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmull.s16 q4, d26, d0 3629c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmull.s16 q5, d27, d0 363f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 364f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_22_64 * x2 3659c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmull.s16 q2, d26, d1 3669c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmull.s16 q6, d27, d1 367f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 368f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d30, r6 ; duplicate cospi_26_64 369f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d31, r7 ; duplicate cospi_6_64 370f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 3719c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; s2 = cospi_10_64 * x2 + cospi_22_64 * x3; 3729c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q4, d20, d1 3739c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q5, d21, d1 3749c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 3759c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; s3 = cospi_22_64 * x2 - cospi_10_64 * x3; 3769c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlsl.s16 q2, d20, d0 3779c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlsl.s16 q6, d21, d0 3789c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 379f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_26_64 * x6 380f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q0, d18, d30 381f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q13, d19, d30 382f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 383f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; s6 = cospi_26_64 * x6 + cospi_6_64 * x7; 384f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q0, d28, d31 385f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q13, d29, d31 386f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 387f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_6_64 * x6 388f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q10, d18, d31 389f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q9, d19, d31 390f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 391f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; s7 = cospi_6_64 * x6 - cospi_26_64 * x7; 392f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q10, d28, d30 393f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q9, d29, d30 394f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 395f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s3 + s7) 396f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q14, q2, q10 397f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q15, q6, q9 398f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 399f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s3 - s7) 400f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q2, q2, q10 401f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q6, q6, q9 402f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 4039c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; x3 = dct_const_round_shift(s3 + s7); 4049c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d28, q14, #14 ; >> 14 4059c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d29, q15, #14 ; >> 14 4069c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 407f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; x7 = dct_const_round_shift(s3 - s7); 408f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d4, q2, #14 ; >> 14 409f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d5, q6, #14 ; >> 14 410f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 411f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s2 + s6) 412f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q9, q4, q0 413f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q10, q5, q13 414f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 415f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s2 - s6) 416f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q4, q4, q0 417f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q5, q5, q13 418f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 4199c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vdup.16 d30, r8 ; duplicate cospi_8_64 4209c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vdup.16 d31, r9 ; duplicate cospi_24_64 4219c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 4229c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; x2 = dct_const_round_shift(s2 + s6); 4239c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d18, q9, #14 ; >> 14 4249c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d19, q10, #14 ; >> 14 4259c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 426f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; x6 = dct_const_round_shift(s2 - s6); 427f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d8, q4, #14 ; >> 14 428f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d9, q5, #14 ; >> 14 429f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 430f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_8_64 * x4 431f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q5, d2, d30 432f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q6, d3, d30 433f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 434f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_24_64 * x4 435f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q7, d2, d31 436f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q0, d3, d31 437f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 4389c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; s4 = cospi_8_64 * x4 + cospi_24_64 * x5; 4399c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q5, d6, d31 4409c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q6, d7, d31 4419c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 442f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; s5 = cospi_24_64 * x4 - cospi_8_64 * x5; 443f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q7, d6, d30 444f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q0, d7, d30 445f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 446f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_8_64 * x7 447f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q1, d4, d30 448f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q3, d5, d30 449f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 450f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_24_64 * x7 451f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q10, d4, d31 452f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q2, d5, d31 453f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 4549c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; s6 = -cospi_24_64 * x6 + cospi_8_64 * x7; 4559c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlsl.s16 q1, d8, d31 4569c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlsl.s16 q3, d9, d31 4579c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 458f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; s7 = cospi_8_64 * x6 + cospi_24_64 * x7; 459f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q10, d8, d30 460f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q2, d9, d30 461f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 462f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q8, q11, q9 ; x0 = s0 + s2; 463f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 464f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q11, q11, q9 ; x2 = s0 - s2; 465f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 466f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s16 q4, q12, q14 ; x1 = s1 + s3; 467f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 468f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q12, q12, q14 ; x3 = s1 - s3; 469f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 470f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s4 + s6) 471f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q14, q5, q1 472f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q15, q6, q3 473f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 474f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s4 - s6) 475f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q5, q5, q1 476f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q6, q6, q3 477f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 4789c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; x4 = dct_const_round_shift(s4 + s6); 4799c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d18, q14, #14 ; >> 14 4809c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d19, q15, #14 ; >> 14 4819c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 482f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; x6 = dct_const_round_shift(s4 - s6); 483f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d10, q5, #14 ; >> 14 484f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d11, q6, #14 ; >> 14 485f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 486f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s5 + s7) 487f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q1, q7, q10 488f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vadd.s32 q3, q0, q2 489f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 490f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; (s5 - s7)) 491f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q7, q7, q10 492f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s32 q0, q0, q2 493f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 4949c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; x5 = dct_const_round_shift(s5 + s7); 4959c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d28, q1, #14 ; >> 14 4969c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d29, q3, #14 ; >> 14 4979c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 498f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; x7 = dct_const_round_shift(s5 - s7); 499f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d14, q7, #14 ; >> 14 500f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d15, q0, #14 ; >> 14 501f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 502f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 d30, r12 ; duplicate cospi_16_64 503f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 504f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_16_64 * x2 505f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q2, d22, d30 506f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q3, d23, d30 507f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 508f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_6_64 * x6 509f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q13, d22, d30 510f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q1, d23, d30 511f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 5129c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; cospi_16_64 * x2 + cospi_16_64 * x3; 5139c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q2, d24, d30 5149c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlal.s16 q3, d25, d30 5159c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 516f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_16_64 * x2 - cospi_16_64 * x3; 517f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q13, d24, d30 518f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlsl.s16 q1, d25, d30 519f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 5209c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; x2 = dct_const_round_shift(s2); 5219c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d4, q2, #14 ; >> 14 5229c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d5, q3, #14 ; >> 14 5239c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 524f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ;x3 = dct_const_round_shift(s3); 525f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d24, q13, #14 ; >> 14 526f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d25, q1, #14 ; >> 14 527f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 528f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_16_64 * x6 529f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q13, d10, d30 530f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmull.s16 q1, d11, d30 531f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 5329c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; cospi_6_64 * x6 5339c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmull.s16 q11, d10, d30 5349c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmull.s16 q0, d11, d30 5359c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 536f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; cospi_16_64 * x6 + cospi_16_64 * x7; 537f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q13, d14, d30 538f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vmlal.s16 q1, d15, d30 539f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 5409c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org ; cospi_16_64 * x6 - cospi_16_64 * x7; 5419c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlsl.s16 q11, d14, d30 5429c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vmlsl.s16 q0, d15, d30 5439c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org 544f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; x6 = dct_const_round_shift(s6); 545f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d20, q13, #14 ; >> 14 546f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqrshrn.s32 d21, q1, #14 ; >> 14 547f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 548f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ;x7 = dct_const_round_shift(s7); 5499c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d12, q11, #14 ; >> 14 5509c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vqrshrn.s32 d13, q0, #14 ; >> 14 551f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 552f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vdup.16 q5, r10 ; duplicate 0 553f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 554f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q9, q5, q9 ; output[1] = -x4; 555f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q11, q5, q2 ; output[3] = -x2; 5569c920af5cd2f78ab30bb06f01f4a4d9d30d5c92bfgalligan@chromium.org vsub.s16 q13, q5, q6 ; output[5] = -x7; 557f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vsub.s16 q15, q5, q4 ; output[7] = -x1; 558f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org MEND 559f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 560f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 561f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org AREA Block, CODE, READONLY ; name this block of code 562ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org;void vp9_iht8x8_64_add_neon(int16_t *input, uint8_t *dest, 563f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; int dest_stride, int tx_type) 564f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; 565f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; r0 int16_t input 566f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; r1 uint8_t *dest 567f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; r2 int dest_stride 568f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; r3 int tx_type) 569f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org; This function will only handle tx_type of 1,2,3. 570ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org|vp9_iht8x8_64_add_neon| PROC 571f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 572f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; load the inputs into d16-d19 573f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.s16 {q8,q9}, [r0]! 574f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.s16 {q10,q11}, [r0]! 575f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.s16 {q12,q13}, [r0]! 576f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.s16 {q14,q15}, [r0]! 577f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 578f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org push {r0-r10} 579411971f94253c85e1866c281860d6344f6aa0c78fgalligan@chromium.org vpush {d8-d15} 580f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 581f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; transpose the input data 582f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org TRANSPOSE8X8 583f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 584f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; decide the type of transform 585f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org cmp r3, #2 586f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org beq idct_iadst 587f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org cmp r3, #3 588f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org beq iadst_iadst 589f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 590f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.orgiadst_idct 591f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate IDCT constants 592f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org GENERATE_IDCT_CONSTANTS 593f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 594f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; first transform rows 595f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org IDCT8x8_1D 596f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 597f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; transpose the matrix 598f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org TRANSPOSE8X8 599f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 600f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate IADST constants 601f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org GENERATE_IADST_CONSTANTS 602f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 603f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; then transform columns 604f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org IADST8X8_1D 605f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 606ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org b end_vp9_iht8x8_64_add_neon 607f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 608f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.orgidct_iadst 609f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate IADST constants 610f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org GENERATE_IADST_CONSTANTS 611f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 612f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; first transform rows 613f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org IADST8X8_1D 614f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 615f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; transpose the matrix 616f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org TRANSPOSE8X8 617f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 618f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate IDCT constants 619f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org GENERATE_IDCT_CONSTANTS 620f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 621f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; then transform columns 622f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org IDCT8x8_1D 623f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 624ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org b end_vp9_iht8x8_64_add_neon 625f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 626f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.orgiadst_iadst 627f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; generate IADST constants 628f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org GENERATE_IADST_CONSTANTS 629f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 630f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; first transform rows 631f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org IADST8X8_1D 632f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 633f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; transpose the matrix 634f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org TRANSPOSE8X8 635f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 636f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; then transform columns 637f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org IADST8X8_1D 638f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 639ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.orgend_vp9_iht8x8_64_add_neon 640411971f94253c85e1866c281860d6344f6aa0c78fgalligan@chromium.org vpop {d8-d15} 641f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org pop {r0-r10} 642f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 643f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; ROUND_POWER_OF_TWO(temp_out[j], 5) 644f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vrshr.s16 q8, q8, #5 645f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vrshr.s16 q9, q9, #5 646f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vrshr.s16 q10, q10, #5 647f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vrshr.s16 q11, q11, #5 648f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vrshr.s16 q12, q12, #5 649f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vrshr.s16 q13, q13, #5 650f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vrshr.s16 q14, q14, #5 651f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vrshr.s16 q15, q15, #5 652f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 653f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; save dest pointer 654f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org mov r0, r1 655f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 656f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; load destination data 657f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.64 {d0}, [r1], r2 658f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.64 {d1}, [r1], r2 659f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.64 {d2}, [r1], r2 660f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.64 {d3}, [r1], r2 661f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.64 {d4}, [r1], r2 662f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.64 {d5}, [r1], r2 663f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.64 {d6}, [r1], r2 664f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vld1.64 {d7}, [r1] 665f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 666f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i] 667f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vaddw.u8 q8, q8, d0 668f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vaddw.u8 q9, q9, d1 669f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vaddw.u8 q10, q10, d2 670f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vaddw.u8 q11, q11, d3 671f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vaddw.u8 q12, q12, d4 672f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vaddw.u8 q13, q13, d5 673f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vaddw.u8 q14, q14, d6 674f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vaddw.u8 q15, q15, d7 675f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 676f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; clip_pixel 677f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqmovun.s16 d0, q8 678f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqmovun.s16 d1, q9 679f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqmovun.s16 d2, q10 680f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqmovun.s16 d3, q11 681f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqmovun.s16 d4, q12 682f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqmovun.s16 d5, q13 683f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqmovun.s16 d6, q14 684f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vqmovun.s16 d7, q15 685f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 686f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org ; store the data 687f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vst1.64 {d0}, [r0], r2 688f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vst1.64 {d1}, [r0], r2 689f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vst1.64 {d2}, [r0], r2 690f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vst1.64 {d3}, [r0], r2 691f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vst1.64 {d4}, [r0], r2 692f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vst1.64 {d5}, [r0], r2 693f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vst1.64 {d6}, [r0], r2 694f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org vst1.64 {d7}, [r0], r2 695f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org bx lr 696ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ENDP ; |vp9_iht8x8_64_add_neon| 697f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org 698f9586bb54d74c97d07b09eb2512f8569c9c1c025fgalligan@chromium.org END 699