From 78182538d64d2e919c2729fd7f10dd77ded27bec Mon Sep 17 00:00:00 2001 From: Christian Duvivier Date: Sun, 4 Aug 2013 17:37:05 -0700 Subject: [PATCH] Neon version of vp9_short_idct4x4_add. Change-Id: Idec4cae0cb9b3a29835fd2750d354c1393d47aa4 --- .../arm/neon/vp9_short_idct4x4_add_neon.asm | 188 ++++++++++++++++++ vp9/common/vp9_rtcd_defs.sh | 2 +- vp9/vp9_common.mk | 1 + 3 files changed, 190 insertions(+), 1 deletion(-) create mode 100644 vp9/common/arm/neon/vp9_short_idct4x4_add_neon.asm diff --git a/vp9/common/arm/neon/vp9_short_idct4x4_add_neon.asm b/vp9/common/arm/neon/vp9_short_idct4x4_add_neon.asm new file mode 100644 index 000000000..433032237 --- /dev/null +++ b/vp9/common/arm/neon/vp9_short_idct4x4_add_neon.asm @@ -0,0 +1,188 @@ +; +; Copyright (c) 2013 The WebM project authors. All Rights Reserved. +; +; Use of this source code is governed by a BSD-style license +; that can be found in the LICENSE file in the root of the source +; tree. An additional intellectual property rights grant can be found +; in the file PATENTS. All contributing project authors may +; be found in the AUTHORS file in the root of the source tree. +; + + EXPORT |vp9_short_idct4x4_add_neon| + ARM + REQUIRE8 + PRESERVE8 + + AREA ||.text||, CODE, READONLY, ALIGN=2 + + AREA Block, CODE, READONLY ; name this block of code +;void vp9_short_idct4x4_add_neon(int16_t *input, uint8_t *dest, int dest_stride) +; +; r0 int16_t input +; r1 uint8_t *dest +; r2 int dest_stride) + +|vp9_short_idct4x4_add_neon| PROC + + ; The 2D transform is done with two passes which are actually pretty + ; similar. We first transform the rows. This is done by transposing + ; the inputs, doing an SIMD column transform (the columns are the + ; transposed rows) and then transpose the results (so that it goes back + ; in normal/row positions). Then, we transform the columns by doing + ; another SIMD column transform. + ; So, two passes of a transpose followed by a column transform. + + ; load the inputs into q8-q9, d16-d19 + vld1.s16 {q8}, [r0]! + vld1.s16 {q9}, [r0]! + + ; generate scalar constants + ; cospi_8_64 = 15137 = 0x3b21 + mov r0, #0x3b00 + add r0, #0x21 + ; cospi_16_64 = 11585 = 0x2d41 + mov r3, #0x2d00 + add r3, #0x41 + ; cospi_24_64 = 6270 = 0x 187e + mov r12, #0x1800 + add r12, #0x7e + ; generate constant vectors + vdup.16 d20, r0; ; replicate cospi_8_64 + vdup.16 d21, r3; ; replicate cospi_16_64 + vdup.16 d22, r12; ; replicate cospi_24_64 + + ; transpose the input data + ; 00 01 02 03 d16 + ; 10 11 12 13 d17 + ; 20 21 22 23 d18 + ; 30 31 32 33 d19 + vtrn.16 d16, d17 + vtrn.16 d18, d19 + ; 00 10 02 12 d16 + ; 01 11 03 13 d17 + ; 20 30 22 32 d18 + ; 21 31 23 33 d19 + vtrn.32 q8, q9 + ; 00 10 20 30 d16 + ; 01 11 21 31 d17 + ; 02 12 22 32 d18 + ; 03 13 23 33 d19 + + ; do the transform on transposed rows + + ; stage 1 + ; (input[0] + input[2]) * cospi_16_64; + ; (input[0] - input[2]) * cospi_16_64; + vadd.s16 d23, d16, d18 + vsub.s16 d24, d16, d18 + vmull.s16 q13, d23, d21 + vmull.s16 q14, d24, d21 + + ; input[1] * cospi_24_64 - input[3] * cospi_8_64; + ; input[1] * cospi_8_64 + input[3] * cospi_24_64; + vmull.s16 q15, d17, d22 + vmull.s16 q0, d19, d20 + vmull.s16 q1, d17, d20 + vmull.s16 q2, d19, d22 + vsub.s32 q3, q15, q0 + vadd.s32 q8, q1, q2 + + ; dct_const_round_shift + vqrshrn.s32 d26, q13, #14 + vqrshrn.s32 d27, q14, #14 + vqrshrn.s32 d29, q3, #14 + vqrshrn.s32 d28, q8, #14 + + ; stage 2 + ; output[0] = step[0] + step[3]; + ; output[1] = step[1] + step[2]; + ; output[3] = step[0] - step[3]; + ; output[2] = step[1] - step[2]; + vadd.s16 q8, q13, q14 + vsub.s16 q9, q13, q14 + vswp d18, d19 + + ; transpose the results + ; 00 01 02 03 d16 + ; 10 11 12 13 d17 + ; 20 21 22 23 d18 + ; 30 31 32 33 d19 + vtrn.16 d16, d17 + vtrn.16 d18, d19 + ; 00 10 02 12 d16 + ; 01 11 03 13 d17 + ; 20 30 22 32 d18 + ; 21 31 23 33 d19 + vtrn.32 q8, q9 + ; 00 10 20 30 d16 + ; 01 11 21 31 d17 + ; 02 12 22 32 d18 + ; 03 13 23 33 d19 + + ; do the transform on columns + + ; stage 1 + ; (input[0] + input[2]) * cospi_16_64; + ; (input[0] - input[2]) * cospi_16_64; + vadd.s16 d23, d16, d18 + vsub.s16 d24, d16, d18 + vmull.s16 q13, d23, d21 + vmull.s16 q14, d24, d21 + + ; input[1] * cospi_24_64 - input[3] * cospi_8_64; + ; input[1] * cospi_8_64 + input[3] * cospi_24_64; + vmull.s16 q15, d17, d22 + vmull.s16 q0, d19, d20 + vmull.s16 q1, d17, d20 + vmull.s16 q2, d19, d22 + vsub.s32 q3, q15, q0 + vadd.s32 q8, q1, q2 + + ; dct_const_round_shift + vqrshrn.s32 d26, q13, #14 + vqrshrn.s32 d27, q14, #14 + vqrshrn.s32 d29, q3, #14 + vqrshrn.s32 d28, q8, #14 + + ; stage 2 + ; output[0] = step[0] + step[3]; + ; output[1] = step[1] + step[2]; + ; output[3] = step[0] - step[3]; + ; output[2] = step[1] - step[2]; + vadd.s16 q8, q13, q14 + vsub.s16 q9, q13, q14 + + ; The results are in two registers, one of them being swapped. This will + ; be taken care of by loading the 'dest' value in a swapped fashion and + ; also storing them in the same swapped fashion. + ; temp_out[0, 1] = d16, d17 = q8 + ; temp_out[2, 3] = d19, d18 = q9 swapped + + ; ROUND_POWER_OF_TWO(temp_out[j], 4) + vrshr.s16 q8, q8, #4 + vrshr.s16 q9, q9, #4 + + vld1.32 {d26[0]}, [r1], r2 + vld1.32 {d26[1]}, [r1], r2 + vld1.32 {d27[1]}, [r1], r2 + vld1.32 {d27[0]}, [r1] ; no post-increment + + ; ROUND_POWER_OF_TWO(temp_out[j], 4) + dest[j * dest_stride + i] + vaddw.u8 q8, q8, d26 + vaddw.u8 q9, q9, d27 + + ; clip_pixel + vqmovun.s16 d26, q8 + vqmovun.s16 d27, q9 + + ; do the stores in reverse order with negative post-increment, by changing + ; the sign of the stride + rsb r2, r2, #0 + vst1.32 {d27[0]}, [r1], r2 + vst1.32 {d27[1]}, [r1], r2 + vst1.32 {d26[1]}, [r1], r2 + vst1.32 {d26[0]}, [r1] ; no post-increment + bx lr + ENDP ; |vp9_short_idct4x4_add_neon| + + END diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh index 3400aec5b..9e8714cb2 100644 --- a/vp9/common/vp9_rtcd_defs.sh +++ b/vp9/common/vp9_rtcd_defs.sh @@ -298,7 +298,7 @@ prototype void vp9_short_idct4x4_1_add "int16_t *input, uint8_t *dest, int dest_ specialize vp9_short_idct4x4_1_add sse2 prototype void vp9_short_idct4x4_add "int16_t *input, uint8_t *dest, int dest_stride" -specialize vp9_short_idct4x4_add sse2 +specialize vp9_short_idct4x4_add sse2 neon prototype void vp9_short_idct8x8_1_add "int16_t *input, uint8_t *dest, int dest_stride" specialize vp9_short_idct8x8_1_add sse2 diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk index 76021e828..f9ad4bedd 100644 --- a/vp9/vp9_common.mk +++ b/vp9/vp9_common.mk @@ -93,6 +93,7 @@ VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve8_neon$(ASM) VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve8_avg_neon$(ASM) VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_loopfilter_neon$(ASM) VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_dc_only_idct_add_neon$(ASM) +VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_short_idct4x4_add_neon$(ASM) VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_short_idct8x8_add_neon$(ASM) VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_mb_lpf_neon$(ASM)