michael@0: /* michael@0: * Copyright 2011 The LibYuv 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 "libyuv/row.h" michael@0: michael@0: #ifdef __cplusplus michael@0: namespace libyuv { michael@0: extern "C" { michael@0: #endif michael@0: michael@0: // This module is for GCC Neon michael@0: #if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) michael@0: michael@0: // Read 8 Y, 4 U and 4 V from 422 michael@0: #define READYUV422 \ michael@0: "vld1.8 {d0}, [%0]! \n" \ michael@0: "vld1.32 {d2[0]}, [%1]! \n" \ michael@0: "vld1.32 {d2[1]}, [%2]! \n" michael@0: michael@0: // Read 8 Y, 2 U and 2 V from 422 michael@0: #define READYUV411 \ michael@0: "vld1.8 {d0}, [%0]! \n" \ michael@0: "vld1.16 {d2[0]}, [%1]! \n" \ michael@0: "vld1.16 {d2[1]}, [%2]! \n" \ michael@0: "vmov.u8 d3, d2 \n" \ michael@0: "vzip.u8 d2, d3 \n" michael@0: michael@0: // Read 8 Y, 8 U and 8 V from 444 michael@0: #define READYUV444 \ michael@0: "vld1.8 {d0}, [%0]! \n" \ michael@0: "vld1.8 {d2}, [%1]! \n" \ michael@0: "vld1.8 {d3}, [%2]! \n" \ michael@0: "vpaddl.u8 q1, q1 \n" \ michael@0: "vrshrn.u16 d2, q1, #1 \n" michael@0: michael@0: // Read 8 Y, and set 4 U and 4 V to 128 michael@0: #define READYUV400 \ michael@0: "vld1.8 {d0}, [%0]! \n" \ michael@0: "vmov.u8 d2, #128 \n" michael@0: michael@0: // Read 8 Y and 4 UV from NV12 michael@0: #define READNV12 \ michael@0: "vld1.8 {d0}, [%0]! \n" \ michael@0: "vld1.8 {d2}, [%1]! \n" \ michael@0: "vmov.u8 d3, d2 \n"/* split odd/even uv apart */\ michael@0: "vuzp.u8 d2, d3 \n" \ michael@0: "vtrn.u32 d2, d3 \n" michael@0: michael@0: // Read 8 Y and 4 VU from NV21 michael@0: #define READNV21 \ michael@0: "vld1.8 {d0}, [%0]! \n" \ michael@0: "vld1.8 {d2}, [%1]! \n" \ michael@0: "vmov.u8 d3, d2 \n"/* split odd/even uv apart */\ michael@0: "vuzp.u8 d3, d2 \n" \ michael@0: "vtrn.u32 d2, d3 \n" michael@0: michael@0: // Read 8 YUY2 michael@0: #define READYUY2 \ michael@0: "vld2.8 {d0, d2}, [%0]! \n" \ michael@0: "vmov.u8 d3, d2 \n" \ michael@0: "vuzp.u8 d2, d3 \n" \ michael@0: "vtrn.u32 d2, d3 \n" michael@0: michael@0: // Read 8 UYVY michael@0: #define READUYVY \ michael@0: "vld2.8 {d2, d3}, [%0]! \n" \ michael@0: "vmov.u8 d0, d3 \n" \ michael@0: "vmov.u8 d3, d2 \n" \ michael@0: "vuzp.u8 d2, d3 \n" \ michael@0: "vtrn.u32 d2, d3 \n" michael@0: michael@0: #define YUV422TORGB \ michael@0: "veor.u8 d2, d26 \n"/*subtract 128 from u and v*/\ michael@0: "vmull.s8 q8, d2, d24 \n"/* u/v B/R component */\ michael@0: "vmull.s8 q9, d2, d25 \n"/* u/v G component */\ michael@0: "vmov.u8 d1, #0 \n"/* split odd/even y apart */\ michael@0: "vtrn.u8 d0, d1 \n" \ michael@0: "vsub.s16 q0, q0, q15 \n"/* offset y */\ michael@0: "vmul.s16 q0, q0, q14 \n" \ michael@0: "vadd.s16 d18, d19 \n" \ michael@0: "vqadd.s16 d20, d0, d16 \n" /* B */ \ michael@0: "vqadd.s16 d21, d1, d16 \n" \ michael@0: "vqadd.s16 d22, d0, d17 \n" /* R */ \ michael@0: "vqadd.s16 d23, d1, d17 \n" \ michael@0: "vqadd.s16 d16, d0, d18 \n" /* G */ \ michael@0: "vqadd.s16 d17, d1, d18 \n" \ michael@0: "vqshrun.s16 d0, q10, #6 \n" /* B */ \ michael@0: "vqshrun.s16 d1, q11, #6 \n" /* G */ \ michael@0: "vqshrun.s16 d2, q8, #6 \n" /* R */ \ michael@0: "vmovl.u8 q10, d0 \n"/* set up for reinterleave*/\ michael@0: "vmovl.u8 q11, d1 \n" \ michael@0: "vmovl.u8 q8, d2 \n" \ michael@0: "vtrn.u8 d20, d21 \n" \ michael@0: "vtrn.u8 d22, d23 \n" \ michael@0: "vtrn.u8 d16, d17 \n" \ michael@0: "vmov.u8 d21, d16 \n" michael@0: michael@0: static vec8 kUVToRB = { 127, 127, 127, 127, 102, 102, 102, 102, michael@0: 0, 0, 0, 0, 0, 0, 0, 0 }; michael@0: static vec8 kUVToG = { -25, -25, -25, -25, -52, -52, -52, -52, michael@0: 0, 0, 0, 0, 0, 0, 0, 0 }; michael@0: michael@0: void I444ToARGBRow_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_argb, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV444 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%3]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_argb), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void I422ToARGBRow_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_argb, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV422 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%3]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_argb), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void I411ToARGBRow_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_argb, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV411 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%3]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_argb), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void I422ToBGRARow_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_bgra, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV422 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vswp.u8 d20, d22 \n" michael@0: "vmov.u8 d19, #255 \n" michael@0: "vst4.8 {d19, d20, d21, d22}, [%3]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_bgra), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void I422ToABGRRow_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_abgr, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV422 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vswp.u8 d20, d22 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%3]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_abgr), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void I422ToRGBARow_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_rgba, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV422 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vmov.u8 d19, #255 \n" michael@0: "vst4.8 {d19, d20, d21, d22}, [%3]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_rgba), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void I422ToRGB24Row_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_rgb24, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV422 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vst3.8 {d20, d21, d22}, [%3]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_rgb24), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void I422ToRAWRow_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_raw, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV422 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vswp.u8 d20, d22 \n" michael@0: "vst3.8 {d20, d21, d22}, [%3]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_raw), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: #define ARGBTORGB565 \ michael@0: "vshr.u8 d20, d20, #3 \n" /* B */ \ michael@0: "vshr.u8 d21, d21, #2 \n" /* G */ \ michael@0: "vshr.u8 d22, d22, #3 \n" /* R */ \ michael@0: "vmovl.u8 q8, d20 \n" /* B */ \ michael@0: "vmovl.u8 q9, d21 \n" /* G */ \ michael@0: "vmovl.u8 q10, d22 \n" /* R */ \ michael@0: "vshl.u16 q9, q9, #5 \n" /* G */ \ michael@0: "vshl.u16 q10, q10, #11 \n" /* R */ \ michael@0: "vorr q0, q8, q9 \n" /* BG */ \ michael@0: "vorr q0, q0, q10 \n" /* BGR */ michael@0: michael@0: void I422ToRGB565Row_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_rgb565, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV422 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: ARGBTORGB565 michael@0: "vst1.8 {q0}, [%3]! \n" // store 8 pixels RGB565. michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_rgb565), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: #define ARGBTOARGB1555 \ michael@0: "vshr.u8 q10, q10, #3 \n" /* B */ \ michael@0: "vshr.u8 d22, d22, #3 \n" /* R */ \ michael@0: "vshr.u8 d23, d23, #7 \n" /* A */ \ michael@0: "vmovl.u8 q8, d20 \n" /* B */ \ michael@0: "vmovl.u8 q9, d21 \n" /* G */ \ michael@0: "vmovl.u8 q10, d22 \n" /* R */ \ michael@0: "vmovl.u8 q11, d23 \n" /* A */ \ michael@0: "vshl.u16 q9, q9, #5 \n" /* G */ \ michael@0: "vshl.u16 q10, q10, #10 \n" /* R */ \ michael@0: "vshl.u16 q11, q11, #15 \n" /* A */ \ michael@0: "vorr q0, q8, q9 \n" /* BG */ \ michael@0: "vorr q1, q10, q11 \n" /* RA */ \ michael@0: "vorr q0, q0, q1 \n" /* BGRA */ michael@0: michael@0: void I422ToARGB1555Row_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_argb1555, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV422 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: ARGBTOARGB1555 michael@0: "vst1.8 {q0}, [%3]! \n" // store 8 pixels ARGB1555. michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_argb1555), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: #define ARGBTOARGB4444 \ michael@0: "vshr.u8 d20, d20, #4 \n" /* B */ \ michael@0: "vbic.32 d21, d21, d4 \n" /* G */ \ michael@0: "vshr.u8 d22, d22, #4 \n" /* R */ \ michael@0: "vbic.32 d23, d23, d4 \n" /* A */ \ michael@0: "vorr d0, d20, d21 \n" /* BG */ \ michael@0: "vorr d1, d22, d23 \n" /* RA */ \ michael@0: "vzip.u8 d0, d1 \n" /* BGRA */ michael@0: michael@0: void I422ToARGB4444Row_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_argb4444, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%5] \n" michael@0: "vld1.8 {d25}, [%6] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: "vmov.u8 d4, #0x0f \n" // bits to clear with vbic. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV422 michael@0: YUV422TORGB michael@0: "subs %4, %4, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: ARGBTOARGB4444 michael@0: "vst1.8 {q0}, [%3]! \n" // store 8 pixels ARGB4444. michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_argb4444), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(&kUVToRB), // %5 michael@0: "r"(&kUVToG) // %6 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void YToARGBRow_NEON(const uint8* src_y, michael@0: uint8* dst_argb, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%3] \n" michael@0: "vld1.8 {d25}, [%4] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUV400 michael@0: YUV422TORGB michael@0: "subs %2, %2, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%1]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(width) // %2 michael@0: : "r"(&kUVToRB), // %3 michael@0: "r"(&kUVToG) // %4 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void I400ToARGBRow_NEON(const uint8* src_y, michael@0: uint8* dst_argb, michael@0: int width) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "1: \n" michael@0: "vld1.8 {d20}, [%0]! \n" michael@0: "vmov d21, d20 \n" michael@0: "vmov d22, d20 \n" michael@0: "subs %2, %2, #8 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%1]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(width) // %2 michael@0: : michael@0: : "cc", "memory", "d20", "d21", "d22", "d23" michael@0: ); michael@0: } michael@0: michael@0: void NV12ToARGBRow_NEON(const uint8* src_y, michael@0: const uint8* src_uv, michael@0: uint8* dst_argb, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%4] \n" michael@0: "vld1.8 {d25}, [%5] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READNV12 michael@0: YUV422TORGB michael@0: "subs %3, %3, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%2]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_uv), // %1 michael@0: "+r"(dst_argb), // %2 michael@0: "+r"(width) // %3 michael@0: : "r"(&kUVToRB), // %4 michael@0: "r"(&kUVToG) // %5 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void NV21ToARGBRow_NEON(const uint8* src_y, michael@0: const uint8* src_uv, michael@0: uint8* dst_argb, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%4] \n" michael@0: "vld1.8 {d25}, [%5] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READNV21 michael@0: YUV422TORGB michael@0: "subs %3, %3, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%2]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_uv), // %1 michael@0: "+r"(dst_argb), // %2 michael@0: "+r"(width) // %3 michael@0: : "r"(&kUVToRB), // %4 michael@0: "r"(&kUVToG) // %5 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void NV12ToRGB565Row_NEON(const uint8* src_y, michael@0: const uint8* src_uv, michael@0: uint8* dst_rgb565, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%4] \n" michael@0: "vld1.8 {d25}, [%5] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READNV12 michael@0: YUV422TORGB michael@0: "subs %3, %3, #8 \n" michael@0: ARGBTORGB565 michael@0: "vst1.8 {q0}, [%2]! \n" // store 8 pixels RGB565. michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_uv), // %1 michael@0: "+r"(dst_rgb565), // %2 michael@0: "+r"(width) // %3 michael@0: : "r"(&kUVToRB), // %4 michael@0: "r"(&kUVToG) // %5 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void NV21ToRGB565Row_NEON(const uint8* src_y, michael@0: const uint8* src_uv, michael@0: uint8* dst_rgb565, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%4] \n" michael@0: "vld1.8 {d25}, [%5] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READNV21 michael@0: YUV422TORGB michael@0: "subs %3, %3, #8 \n" michael@0: ARGBTORGB565 michael@0: "vst1.8 {q0}, [%2]! \n" // store 8 pixels RGB565. michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_uv), // %1 michael@0: "+r"(dst_rgb565), // %2 michael@0: "+r"(width) // %3 michael@0: : "r"(&kUVToRB), // %4 michael@0: "r"(&kUVToG) // %5 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void YUY2ToARGBRow_NEON(const uint8* src_yuy2, michael@0: uint8* dst_argb, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%3] \n" michael@0: "vld1.8 {d25}, [%4] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READYUY2 michael@0: YUV422TORGB michael@0: "subs %2, %2, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%1]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_yuy2), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(width) // %2 michael@0: : "r"(&kUVToRB), // %3 michael@0: "r"(&kUVToG) // %4 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void UYVYToARGBRow_NEON(const uint8* src_uyvy, michael@0: uint8* dst_argb, michael@0: int width) { michael@0: asm volatile ( michael@0: "vld1.8 {d24}, [%3] \n" michael@0: "vld1.8 {d25}, [%4] \n" michael@0: "vmov.u8 d26, #128 \n" michael@0: "vmov.u16 q14, #74 \n" michael@0: "vmov.u16 q15, #16 \n" michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: READUYVY michael@0: YUV422TORGB michael@0: "subs %2, %2, #8 \n" michael@0: "vmov.u8 d23, #255 \n" michael@0: "vst4.8 {d20, d21, d22, d23}, [%1]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_uyvy), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(width) // %2 michael@0: : "r"(&kUVToRB), // %3 michael@0: "r"(&kUVToG) // %4 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // Reads 16 pairs of UV and write even values to dst_u and odd to dst_v. michael@0: void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, michael@0: int width) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld2.8 {q0, q1}, [%0]! \n" // load 16 pairs of UV michael@0: "subs %3, %3, #16 \n" // 16 processed per loop michael@0: "vst1.8 {q0}, [%1]! \n" // store U michael@0: "vst1.8 {q1}, [%2]! \n" // store V michael@0: "bgt 1b \n" michael@0: : "+r"(src_uv), // %0 michael@0: "+r"(dst_u), // %1 michael@0: "+r"(dst_v), // %2 michael@0: "+r"(width) // %3 // Output registers michael@0: : // Input registers michael@0: : "cc", "memory", "q0", "q1" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: // Reads 16 U's and V's and writes out 16 pairs of UV. michael@0: void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv, michael@0: int width) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load U michael@0: "vld1.8 {q1}, [%1]! \n" // load V michael@0: "subs %3, %3, #16 \n" // 16 processed per loop michael@0: "vst2.u8 {q0, q1}, [%2]! \n" // store 16 pairs of UV michael@0: "bgt 1b \n" michael@0: : michael@0: "+r"(src_u), // %0 michael@0: "+r"(src_v), // %1 michael@0: "+r"(dst_uv), // %2 michael@0: "+r"(width) // %3 // Output registers michael@0: : // Input registers michael@0: : "cc", "memory", "q0", "q1" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: // Copy multiple of 32. vld4.8 allow unaligned and is fastest on a15. michael@0: void CopyRow_NEON(const uint8* src, uint8* dst, int count) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {d0, d1, d2, d3}, [%0]! \n" // load 32 michael@0: "subs %2, %2, #32 \n" // 32 processed per loop michael@0: "vst1.8 {d0, d1, d2, d3}, [%1]! \n" // store 32 michael@0: "bgt 1b \n" michael@0: : "+r"(src), // %0 michael@0: "+r"(dst), // %1 michael@0: "+r"(count) // %2 // Output registers michael@0: : // Input registers michael@0: : "cc", "memory", "q0", "q1" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: // SetRow8 writes 'count' bytes using a 32 bit value repeated. michael@0: void SetRow_NEON(uint8* dst, uint32 v32, int count) { michael@0: asm volatile ( michael@0: "vdup.u32 q0, %2 \n" // duplicate 4 ints michael@0: "1: \n" michael@0: "subs %1, %1, #16 \n" // 16 bytes per loop michael@0: "vst1.8 {q0}, [%0]! \n" // store michael@0: "bgt 1b \n" michael@0: : "+r"(dst), // %0 michael@0: "+r"(count) // %1 michael@0: : "r"(v32) // %2 michael@0: : "cc", "memory", "q0" michael@0: ); michael@0: } michael@0: michael@0: // TODO(fbarchard): Make fully assembler michael@0: // SetRow32 writes 'count' words using a 32 bit value repeated. michael@0: void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width, michael@0: int dst_stride, int height) { michael@0: for (int y = 0; y < height; ++y) { michael@0: SetRow_NEON(dst, v32, width << 2); michael@0: dst += dst_stride; michael@0: } michael@0: } michael@0: michael@0: void MirrorRow_NEON(const uint8* src, uint8* dst, int width) { michael@0: asm volatile ( michael@0: // Start at end of source row. michael@0: "mov r3, #-16 \n" michael@0: "add %0, %0, %2 \n" michael@0: "sub %0, #16 \n" michael@0: michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0], r3 \n" // src -= 16 michael@0: "subs %2, #16 \n" // 16 pixels per loop. michael@0: "vrev64.8 q0, q0 \n" michael@0: "vst1.8 {d1}, [%1]! \n" // dst += 16 michael@0: "vst1.8 {d0}, [%1]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src), // %0 michael@0: "+r"(dst), // %1 michael@0: "+r"(width) // %2 michael@0: : michael@0: : "cc", "memory", "r3", "q0" michael@0: ); michael@0: } michael@0: michael@0: void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, michael@0: int width) { michael@0: asm volatile ( michael@0: // Start at end of source row. michael@0: "mov r12, #-16 \n" michael@0: "add %0, %0, %3, lsl #1 \n" michael@0: "sub %0, #16 \n" michael@0: michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld2.8 {d0, d1}, [%0], r12 \n" // src -= 16 michael@0: "subs %3, #8 \n" // 8 pixels per loop. michael@0: "vrev64.8 q0, q0 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // dst += 8 michael@0: "vst1.8 {d1}, [%2]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_uv), // %0 michael@0: "+r"(dst_u), // %1 michael@0: "+r"(dst_v), // %2 michael@0: "+r"(width) // %3 michael@0: : michael@0: : "cc", "memory", "r12", "q0" michael@0: ); michael@0: } michael@0: michael@0: void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) { michael@0: asm volatile ( michael@0: // Start at end of source row. michael@0: "mov r3, #-16 \n" michael@0: "add %0, %0, %2, lsl #2 \n" michael@0: "sub %0, #16 \n" michael@0: michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0], r3 \n" // src -= 16 michael@0: "subs %2, #4 \n" // 4 pixels per loop. michael@0: "vrev64.32 q0, q0 \n" michael@0: "vst1.8 {d1}, [%1]! \n" // dst += 16 michael@0: "vst1.8 {d0}, [%1]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src), // %0 michael@0: "+r"(dst), // %1 michael@0: "+r"(width) // %2 michael@0: : michael@0: : "cc", "memory", "r3", "q0" michael@0: ); michael@0: } michael@0: michael@0: void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d4, #255 \n" // Alpha michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld3.8 {d1, d2, d3}, [%0]! \n" // load 8 pixels of RGB24. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vst4.8 {d1, d2, d3, d4}, [%1]! \n" // store 8 pixels of ARGB. michael@0: "bgt 1b \n" michael@0: : "+r"(src_rgb24), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "d1", "d2", "d3", "d4" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d4, #255 \n" // Alpha michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld3.8 {d1, d2, d3}, [%0]! \n" // load 8 pixels of RAW. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vswp.u8 d1, d3 \n" // swap R, B michael@0: "vst4.8 {d1, d2, d3, d4}, [%1]! \n" // store 8 pixels of ARGB. michael@0: "bgt 1b \n" michael@0: : "+r"(src_raw), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "d1", "d2", "d3", "d4" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: #define RGB565TOARGB \ michael@0: "vshrn.u16 d6, q0, #5 \n" /* G xxGGGGGG */ \ michael@0: "vuzp.u8 d0, d1 \n" /* d0 xxxBBBBB RRRRRxxx */ \ michael@0: "vshl.u8 d6, d6, #2 \n" /* G GGGGGG00 upper 6 */ \ michael@0: "vshr.u8 d1, d1, #3 \n" /* R 000RRRRR lower 5 */ \ michael@0: "vshl.u8 q0, q0, #3 \n" /* B,R BBBBB000 upper 5 */ \ michael@0: "vshr.u8 q2, q0, #5 \n" /* B,R 00000BBB lower 3 */ \ michael@0: "vorr.u8 d0, d0, d4 \n" /* B */ \ michael@0: "vshr.u8 d4, d6, #6 \n" /* G 000000GG lower 2 */ \ michael@0: "vorr.u8 d2, d1, d5 \n" /* R */ \ michael@0: "vorr.u8 d1, d4, d6 \n" /* G */ michael@0: michael@0: void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d3, #255 \n" // Alpha michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 8 RGB565 pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: RGB565TOARGB michael@0: "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 pixels of ARGB. michael@0: "bgt 1b \n" michael@0: : "+r"(src_rgb565), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: #define ARGB1555TOARGB \ michael@0: "vshrn.u16 d7, q0, #8 \n" /* A Arrrrrxx */ \ michael@0: "vshr.u8 d6, d7, #2 \n" /* R xxxRRRRR */ \ michael@0: "vshrn.u16 d5, q0, #5 \n" /* G xxxGGGGG */ \ michael@0: "vmovn.u16 d4, q0 \n" /* B xxxBBBBB */ \ michael@0: "vshr.u8 d7, d7, #7 \n" /* A 0000000A */ \ michael@0: "vneg.s8 d7, d7 \n" /* A AAAAAAAA upper 8 */ \ michael@0: "vshl.u8 d6, d6, #3 \n" /* R RRRRR000 upper 5 */ \ michael@0: "vshr.u8 q1, q3, #5 \n" /* R,A 00000RRR lower 3 */ \ michael@0: "vshl.u8 q0, q2, #3 \n" /* B,G BBBBB000 upper 5 */ \ michael@0: "vshr.u8 q2, q0, #5 \n" /* B,G 00000BBB lower 3 */ \ michael@0: "vorr.u8 q1, q1, q3 \n" /* R,A */ \ michael@0: "vorr.u8 q0, q0, q2 \n" /* B,G */ \ michael@0: michael@0: // RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha. michael@0: #define RGB555TOARGB \ michael@0: "vshrn.u16 d6, q0, #5 \n" /* G xxxGGGGG */ \ michael@0: "vuzp.u8 d0, d1 \n" /* d0 xxxBBBBB xRRRRRxx */ \ michael@0: "vshl.u8 d6, d6, #3 \n" /* G GGGGG000 upper 5 */ \ michael@0: "vshr.u8 d1, d1, #2 \n" /* R 00xRRRRR lower 5 */ \ michael@0: "vshl.u8 q0, q0, #3 \n" /* B,R BBBBB000 upper 5 */ \ michael@0: "vshr.u8 q2, q0, #5 \n" /* B,R 00000BBB lower 3 */ \ michael@0: "vorr.u8 d0, d0, d4 \n" /* B */ \ michael@0: "vshr.u8 d4, d6, #5 \n" /* G 00000GGG lower 3 */ \ michael@0: "vorr.u8 d2, d1, d5 \n" /* R */ \ michael@0: "vorr.u8 d1, d4, d6 \n" /* G */ michael@0: michael@0: void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb, michael@0: int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d3, #255 \n" // Alpha michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 8 ARGB1555 pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: ARGB1555TOARGB michael@0: "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 pixels of ARGB. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb1555), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: #define ARGB4444TOARGB \ michael@0: "vuzp.u8 d0, d1 \n" /* d0 BG, d1 RA */ \ michael@0: "vshl.u8 q2, q0, #4 \n" /* B,R BBBB0000 */ \ michael@0: "vshr.u8 q1, q0, #4 \n" /* G,A 0000GGGG */ \ michael@0: "vshr.u8 q0, q2, #4 \n" /* B,R 0000BBBB */ \ michael@0: "vorr.u8 q0, q0, q2 \n" /* B,R BBBBBBBB */ \ michael@0: "vshl.u8 q2, q1, #4 \n" /* G,A GGGG0000 */ \ michael@0: "vorr.u8 q1, q1, q2 \n" /* G,A GGGGGGGG */ \ michael@0: "vswp.u8 d1, d2 \n" /* B,R,G,A -> B,G,R,A */ michael@0: michael@0: void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb, michael@0: int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d3, #255 \n" // Alpha michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 8 ARGB4444 pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: ARGB4444TOARGB michael@0: "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 pixels of ARGB. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb4444), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d1, d2, d3, d4}, [%0]! \n" // load 8 pixels of ARGB. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vst3.8 {d1, d2, d3}, [%1]! \n" // store 8 pixels of RGB24. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_rgb24), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "d1", "d2", "d3", "d4" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d1, d2, d3, d4}, [%0]! \n" // load 8 pixels of ARGB. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vswp.u8 d1, d3 \n" // swap R, B michael@0: "vst3.8 {d1, d2, d3}, [%1]! \n" // store 8 pixels of RAW. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_raw), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "d1", "d2", "d3", "d4" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld2.8 {q0, q1}, [%0]! \n" // load 16 pixels of YUY2. michael@0: "subs %2, %2, #16 \n" // 16 processed per loop. michael@0: "vst1.8 {q0}, [%1]! \n" // store 16 pixels of Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_yuy2), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld2.8 {q0, q1}, [%0]! \n" // load 16 pixels of UYVY. michael@0: "subs %2, %2, #16 \n" // 16 processed per loop. michael@0: "vst1.8 {q1}, [%1]! \n" // store 16 pixels of Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_uyvy), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v, michael@0: int pix) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 16 pixels of YUY2. michael@0: "subs %3, %3, #16 \n" // 16 pixels = 8 UVs. michael@0: "vst1.8 {d1}, [%1]! \n" // store 8 U. michael@0: "vst1.8 {d3}, [%2]! \n" // store 8 V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_yuy2), // %0 michael@0: "+r"(dst_u), // %1 michael@0: "+r"(dst_v), // %2 michael@0: "+r"(pix) // %3 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v, michael@0: int pix) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 16 pixels of UYVY. michael@0: "subs %3, %3, #16 \n" // 16 pixels = 8 UVs. michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 U. michael@0: "vst1.8 {d2}, [%2]! \n" // store 8 V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_uyvy), // %0 michael@0: "+r"(dst_u), // %1 michael@0: "+r"(dst_v), // %2 michael@0: "+r"(pix) // %3 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // stride + src_yuy2 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 16 pixels of YUY2. michael@0: "subs %4, %4, #16 \n" // 16 pixels = 8 UVs. michael@0: "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load next row YUY2. michael@0: "vrhadd.u8 d1, d1, d5 \n" // average rows of U michael@0: "vrhadd.u8 d3, d3, d7 \n" // average rows of V michael@0: "vst1.8 {d1}, [%2]! \n" // store 8 U. michael@0: "vst1.8 {d3}, [%3]! \n" // store 8 V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_yuy2), // %0 michael@0: "+r"(stride_yuy2), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // stride + src_uyvy michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 16 pixels of UYVY. michael@0: "subs %4, %4, #16 \n" // 16 pixels = 8 UVs. michael@0: "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load next row UYVY. michael@0: "vrhadd.u8 d0, d0, d4 \n" // average rows of U michael@0: "vrhadd.u8 d2, d2, d6 \n" // average rows of V michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 U. michael@0: "vst1.8 {d2}, [%3]! \n" // store 8 V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_uyvy), // %0 michael@0: "+r"(stride_uyvy), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void HalfRow_NEON(const uint8* src_uv, int src_uv_stride, michael@0: uint8* dst_uv, int pix) { michael@0: asm volatile ( michael@0: // change the stride to row 2 pointer michael@0: "add %1, %0 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load row 1 16 pixels. michael@0: "subs %3, %3, #16 \n" // 16 processed per loop michael@0: "vld1.8 {q1}, [%1]! \n" // load row 2 16 pixels. michael@0: "vrhadd.u8 q0, q1 \n" // average row 1 and 2 michael@0: "vst1.8 {q0}, [%2]! \n" michael@0: "bgt 1b \n" michael@0: : "+r"(src_uv), // %0 michael@0: "+r"(src_uv_stride), // %1 michael@0: "+r"(dst_uv), // %2 michael@0: "+r"(pix) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: // Select 2 channels from ARGB on alternating pixels. e.g. BGBGBGBG michael@0: void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer, michael@0: uint32 selector, int pix) { michael@0: asm volatile ( michael@0: "vmov.u32 d6[0], %3 \n" // selector michael@0: "1: \n" michael@0: "vld1.8 {q0, q1}, [%0]! \n" // load row 8 pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop michael@0: "vtbl.8 d4, {d0, d1}, d6 \n" // look up 4 pixels michael@0: "vtbl.8 d5, {d2, d3}, d6 \n" // look up 4 pixels michael@0: "vtrn.u32 d4, d5 \n" // combine 8 pixels michael@0: "vst1.8 {d4}, [%1]! \n" // store 8. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_bayer), // %1 michael@0: "+r"(pix) // %2 michael@0: : "r"(selector) // %3 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: // Select G channels from ARGB. e.g. GGGGGGGG michael@0: void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer, michael@0: uint32 /*selector*/, int pix) { michael@0: asm volatile ( michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load row 8 pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop michael@0: "vst1.8 {d1}, [%1]! \n" // store 8 G's. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_bayer), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: // For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. michael@0: void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb, michael@0: const uint8* shuffler, int pix) { michael@0: asm volatile ( michael@0: "vld1.8 {q2}, [%3] \n" // shuffler michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 4 pixels. michael@0: "subs %2, %2, #4 \n" // 4 processed per loop michael@0: "vtbl.8 d2, {d0, d1}, d4 \n" // look up 2 first pixels michael@0: "vtbl.8 d3, {d0, d1}, d5 \n" // look up 2 next pixels michael@0: "vst1.8 {q1}, [%1]! \n" // store 4. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(pix) // %2 michael@0: : "r"(shuffler) // %3 michael@0: : "cc", "memory", "q0", "q1", "q2" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: void I422ToYUY2Row_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_yuy2, int width) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld2.8 {d0, d2}, [%0]! \n" // load 16 Ys michael@0: "vld1.8 {d1}, [%1]! \n" // load 8 Us michael@0: "vld1.8 {d3}, [%2]! \n" // load 8 Vs michael@0: "subs %4, %4, #16 \n" // 16 pixels michael@0: "vst4.8 {d0, d1, d2, d3}, [%3]! \n" // Store 8 YUY2/16 pixels. michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_yuy2), // %3 michael@0: "+r"(width) // %4 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3" michael@0: ); michael@0: } michael@0: michael@0: void I422ToUYVYRow_NEON(const uint8* src_y, michael@0: const uint8* src_u, michael@0: const uint8* src_v, michael@0: uint8* dst_uyvy, int width) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld2.8 {d1, d3}, [%0]! \n" // load 16 Ys michael@0: "vld1.8 {d0}, [%1]! \n" // load 8 Us michael@0: "vld1.8 {d2}, [%2]! \n" // load 8 Vs michael@0: "subs %4, %4, #16 \n" // 16 pixels michael@0: "vst4.8 {d0, d1, d2, d3}, [%3]! \n" // Store 8 UYVY/16 pixels. michael@0: "bgt 1b \n" michael@0: : "+r"(src_y), // %0 michael@0: "+r"(src_u), // %1 michael@0: "+r"(src_v), // %2 michael@0: "+r"(dst_uyvy), // %3 michael@0: "+r"(width) // %4 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3" michael@0: ); michael@0: } michael@0: michael@0: void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d20, d21, d22, d23}, [%0]! \n" // load 8 pixels of ARGB. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: ARGBTORGB565 michael@0: "vst1.8 {q0}, [%1]! \n" // store 8 pixels RGB565. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_rgb565), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q8", "q9", "q10", "q11" michael@0: ); michael@0: } michael@0: michael@0: void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555, michael@0: int pix) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d20, d21, d22, d23}, [%0]! \n" // load 8 pixels of ARGB. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: ARGBTOARGB1555 michael@0: "vst1.8 {q0}, [%1]! \n" // store 8 pixels ARGB1555. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_argb1555), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q8", "q9", "q10", "q11" michael@0: ); michael@0: } michael@0: michael@0: void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444, michael@0: int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d4, #0x0f \n" // bits to clear with vbic. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d20, d21, d22, d23}, [%0]! \n" // load 8 pixels of ARGB. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: ARGBTOARGB4444 michael@0: "vst1.8 {q0}, [%1]! \n" // store 8 pixels ARGB4444. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_argb4444), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q8", "q9", "q10", "q11" michael@0: ); michael@0: } michael@0: michael@0: void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d24, #13 \n" // B * 0.1016 coefficient michael@0: "vmov.u8 d25, #65 \n" // G * 0.5078 coefficient michael@0: "vmov.u8 d26, #33 \n" // R * 0.2578 coefficient michael@0: "vmov.u8 d27, #16 \n" // Add 16 constant michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q2, d0, d24 \n" // B michael@0: "vmlal.u8 q2, d1, d25 \n" // G michael@0: "vmlal.u8 q2, d2, d26 \n" // R michael@0: "vqrshrun.s16 d0, q2, #7 \n" // 16 bit to 8 bit Y michael@0: "vqadd.u8 d0, d27 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q12", "q13" michael@0: ); michael@0: } michael@0: michael@0: void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d24, #15 \n" // B * 0.11400 coefficient michael@0: "vmov.u8 d25, #75 \n" // G * 0.58700 coefficient michael@0: "vmov.u8 d26, #38 \n" // R * 0.29900 coefficient michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q2, d0, d24 \n" // B michael@0: "vmlal.u8 q2, d1, d25 \n" // G michael@0: "vmlal.u8 q2, d2, d26 \n" // R michael@0: "vqrshrun.s16 d0, q2, #7 \n" // 15 bit to 8 bit Y michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q12", "q13" michael@0: ); michael@0: } michael@0: michael@0: // 8x1 pixels. michael@0: void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, michael@0: int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d24, #112 \n" // UB / VR 0.875 coefficient michael@0: "vmov.u8 d25, #74 \n" // UG -0.5781 coefficient michael@0: "vmov.u8 d26, #38 \n" // UR -0.2969 coefficient michael@0: "vmov.u8 d27, #18 \n" // VB -0.1406 coefficient michael@0: "vmov.u8 d28, #94 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. michael@0: "subs %3, %3, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q2, d0, d24 \n" // B michael@0: "vmlsl.u8 q2, d1, d25 \n" // G michael@0: "vmlsl.u8 q2, d2, d26 \n" // R michael@0: "vadd.u16 q2, q2, q15 \n" // +128 -> unsigned michael@0: michael@0: "vmull.u8 q3, d2, d24 \n" // R michael@0: "vmlsl.u8 q3, d1, d28 \n" // G michael@0: "vmlsl.u8 q3, d0, d27 \n" // B michael@0: "vadd.u16 q3, q3, q15 \n" // +128 -> unsigned michael@0: michael@0: "vqshrn.u16 d0, q2, #8 \n" // 16 bit to 8 bit U michael@0: "vqshrn.u16 d1, q3, #8 \n" // 16 bit to 8 bit V michael@0: michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%2]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_u), // %1 michael@0: "+r"(dst_v), // %2 michael@0: "+r"(pix) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // 16x1 pixels -> 8x1. pix is number of argb pixels. e.g. 16. michael@0: void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, michael@0: int pix) { michael@0: asm volatile ( michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. michael@0: "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. michael@0: michael@0: "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. michael@0: michael@0: "subs %3, %3, #16 \n" // 16 processed per loop. michael@0: "vmul.s16 q8, q0, q10 \n" // B michael@0: "vmls.s16 q8, q1, q11 \n" // G michael@0: "vmls.s16 q8, q2, q12 \n" // R michael@0: "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned michael@0: michael@0: "vmul.s16 q9, q2, q10 \n" // R michael@0: "vmls.s16 q9, q1, q14 \n" // G michael@0: "vmls.s16 q9, q0, q13 \n" // B michael@0: "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned michael@0: michael@0: "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U michael@0: "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V michael@0: michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%2]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_u), // %1 michael@0: "+r"(dst_v), // %2 michael@0: "+r"(pix) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // 32x1 pixels -> 8x1. pix is number of argb pixels. e.g. 32. michael@0: void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, michael@0: int pix) { michael@0: asm volatile ( michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. michael@0: "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. michael@0: "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. michael@0: "vld4.8 {d8, d10, d12, d14}, [%0]! \n" // load 8 more ARGB pixels. michael@0: "vld4.8 {d9, d11, d13, d15}, [%0]! \n" // load last 8 ARGB pixels. michael@0: "vpaddl.u8 q4, q4 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q5, q5 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q6, q6 \n" // R 16 bytes -> 8 shorts. michael@0: michael@0: "vpadd.u16 d0, d0, d1 \n" // B 16 shorts -> 8 shorts. michael@0: "vpadd.u16 d1, d8, d9 \n" // B michael@0: "vpadd.u16 d2, d2, d3 \n" // G 16 shorts -> 8 shorts. michael@0: "vpadd.u16 d3, d10, d11 \n" // G michael@0: "vpadd.u16 d4, d4, d5 \n" // R 16 shorts -> 8 shorts. michael@0: "vpadd.u16 d5, d12, d13 \n" // R michael@0: michael@0: "vrshr.u16 q0, q0, #1 \n" // 2x average michael@0: "vrshr.u16 q1, q1, #1 \n" michael@0: "vrshr.u16 q2, q2, #1 \n" michael@0: michael@0: "subs %3, %3, #32 \n" // 32 processed per loop. michael@0: "vmul.s16 q8, q0, q10 \n" // B michael@0: "vmls.s16 q8, q1, q11 \n" // G michael@0: "vmls.s16 q8, q2, q12 \n" // R michael@0: "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned michael@0: "vmul.s16 q9, q2, q10 \n" // R michael@0: "vmls.s16 q9, q1, q14 \n" // G michael@0: "vmls.s16 q9, q0, q13 \n" // B michael@0: "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned michael@0: "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U michael@0: "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%2]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_u), // %1 michael@0: "+r"(dst_v), // %2 michael@0: "+r"(pix) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. michael@0: #define RGBTOUV(QB, QG, QR) \ michael@0: "vmul.s16 q8, " #QB ", q10 \n" /* B */ \ michael@0: "vmls.s16 q8, " #QG ", q11 \n" /* G */ \ michael@0: "vmls.s16 q8, " #QR ", q12 \n" /* R */ \ michael@0: "vadd.u16 q8, q8, q15 \n" /* +128 -> unsigned */ \ michael@0: "vmul.s16 q9, " #QR ", q10 \n" /* R */ \ michael@0: "vmls.s16 q9, " #QG ", q14 \n" /* G */ \ michael@0: "vmls.s16 q9, " #QB ", q13 \n" /* B */ \ michael@0: "vadd.u16 q9, q9, q15 \n" /* +128 -> unsigned */ \ michael@0: "vqshrn.u16 d0, q8, #8 \n" /* 16 bit to 8 bit U */ \ michael@0: "vqshrn.u16 d1, q9, #8 \n" /* 16 bit to 8 bit V */ michael@0: michael@0: // TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr. michael@0: void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_argb michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. michael@0: "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. michael@0: "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. michael@0: "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more ARGB pixels. michael@0: "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 ARGB pixels. michael@0: "vpadal.u8 q0, q4 \n" // B 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q2, q6 \n" // R 16 bytes -> 8 shorts. michael@0: michael@0: "vrshr.u16 q0, q0, #1 \n" // 2x average michael@0: "vrshr.u16 q1, q1, #1 \n" michael@0: "vrshr.u16 q2, q2, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 32 processed per loop. michael@0: RGBTOUV(q0, q1, q2) michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(src_stride_argb), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // TODO(fbarchard): Subsample match C code. michael@0: void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_argb michael@0: "vmov.s16 q10, #127 / 2 \n" // UB / VR 0.500 coefficient michael@0: "vmov.s16 q11, #84 / 2 \n" // UG -0.33126 coefficient michael@0: "vmov.s16 q12, #43 / 2 \n" // UR -0.16874 coefficient michael@0: "vmov.s16 q13, #20 / 2 \n" // VB -0.08131 coefficient michael@0: "vmov.s16 q14, #107 / 2 \n" // VG -0.41869 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. michael@0: "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. michael@0: "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. michael@0: "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more ARGB pixels. michael@0: "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 ARGB pixels. michael@0: "vpadal.u8 q0, q4 \n" // B 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q2, q6 \n" // R 16 bytes -> 8 shorts. michael@0: michael@0: "vrshr.u16 q0, q0, #1 \n" // 2x average michael@0: "vrshr.u16 q1, q1, #1 \n" michael@0: "vrshr.u16 q2, q2, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 32 processed per loop. michael@0: RGBTOUV(q0, q1, q2) michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(src_stride_argb), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_bgra michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 BGRA pixels. michael@0: "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 BGRA pixels. michael@0: "vpaddl.u8 q3, q3 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q2, q2 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q1, q1 \n" // R 16 bytes -> 8 shorts. michael@0: "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more BGRA pixels. michael@0: "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 BGRA pixels. michael@0: "vpadal.u8 q3, q7 \n" // B 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q2, q6 \n" // G 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q1, q5 \n" // R 16 bytes -> 8 shorts. michael@0: michael@0: "vrshr.u16 q1, q1, #1 \n" // 2x average michael@0: "vrshr.u16 q2, q2, #1 \n" michael@0: "vrshr.u16 q3, q3, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 32 processed per loop. michael@0: RGBTOUV(q3, q2, q1) michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_bgra), // %0 michael@0: "+r"(src_stride_bgra), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_abgr michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ABGR pixels. michael@0: "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ABGR pixels. michael@0: "vpaddl.u8 q2, q2 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q0, q0 \n" // R 16 bytes -> 8 shorts. michael@0: "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more ABGR pixels. michael@0: "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 ABGR pixels. michael@0: "vpadal.u8 q2, q6 \n" // B 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q0, q4 \n" // R 16 bytes -> 8 shorts. michael@0: michael@0: "vrshr.u16 q0, q0, #1 \n" // 2x average michael@0: "vrshr.u16 q1, q1, #1 \n" michael@0: "vrshr.u16 q2, q2, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 32 processed per loop. michael@0: RGBTOUV(q2, q1, q0) michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_abgr), // %0 michael@0: "+r"(src_stride_abgr), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_rgba michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 RGBA pixels. michael@0: "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 RGBA pixels. michael@0: "vpaddl.u8 q0, q1 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q1, q2 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q2, q3 \n" // R 16 bytes -> 8 shorts. michael@0: "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more RGBA pixels. michael@0: "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 RGBA pixels. michael@0: "vpadal.u8 q0, q5 \n" // B 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q1, q6 \n" // G 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q2, q7 \n" // R 16 bytes -> 8 shorts. michael@0: michael@0: "vrshr.u16 q0, q0, #1 \n" // 2x average michael@0: "vrshr.u16 q1, q1, #1 \n" michael@0: "vrshr.u16 q2, q2, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 32 processed per loop. michael@0: RGBTOUV(q0, q1, q2) michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_rgba), // %0 michael@0: "+r"(src_stride_rgba), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_rgb24 michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld3.8 {d0, d2, d4}, [%0]! \n" // load 8 RGB24 pixels. michael@0: "vld3.8 {d1, d3, d5}, [%0]! \n" // load next 8 RGB24 pixels. michael@0: "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. michael@0: "vld3.8 {d8, d10, d12}, [%1]! \n" // load 8 more RGB24 pixels. michael@0: "vld3.8 {d9, d11, d13}, [%1]! \n" // load last 8 RGB24 pixels. michael@0: "vpadal.u8 q0, q4 \n" // B 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q2, q6 \n" // R 16 bytes -> 8 shorts. michael@0: michael@0: "vrshr.u16 q0, q0, #1 \n" // 2x average michael@0: "vrshr.u16 q1, q1, #1 \n" michael@0: "vrshr.u16 q2, q2, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 32 processed per loop. michael@0: RGBTOUV(q0, q1, q2) michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_rgb24), // %0 michael@0: "+r"(src_stride_rgb24), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_raw michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld3.8 {d0, d2, d4}, [%0]! \n" // load 8 RAW pixels. michael@0: "vld3.8 {d1, d3, d5}, [%0]! \n" // load next 8 RAW pixels. michael@0: "vpaddl.u8 q2, q2 \n" // B 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. michael@0: "vpaddl.u8 q0, q0 \n" // R 16 bytes -> 8 shorts. michael@0: "vld3.8 {d8, d10, d12}, [%1]! \n" // load 8 more RAW pixels. michael@0: "vld3.8 {d9, d11, d13}, [%1]! \n" // load last 8 RAW pixels. michael@0: "vpadal.u8 q2, q6 \n" // B 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. michael@0: "vpadal.u8 q0, q4 \n" // R 16 bytes -> 8 shorts. michael@0: michael@0: "vrshr.u16 q0, q0, #1 \n" // 2x average michael@0: "vrshr.u16 q1, q1, #1 \n" michael@0: "vrshr.u16 q2, q2, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 32 processed per loop. michael@0: RGBTOUV(q2, q1, q0) michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_raw), // %0 michael@0: "+r"(src_stride_raw), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. michael@0: void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_argb michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 8 RGB565 pixels. michael@0: RGB565TOARGB michael@0: "vpaddl.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. michael@0: "vld1.8 {q0}, [%0]! \n" // next 8 RGB565 pixels. michael@0: RGB565TOARGB michael@0: "vpaddl.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. michael@0: michael@0: "vld1.8 {q0}, [%1]! \n" // load 8 RGB565 pixels. michael@0: RGB565TOARGB michael@0: "vpadal.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. michael@0: "vld1.8 {q0}, [%1]! \n" // next 8 RGB565 pixels. michael@0: RGB565TOARGB michael@0: "vpadal.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. michael@0: michael@0: "vrshr.u16 q4, q4, #1 \n" // 2x average michael@0: "vrshr.u16 q5, q5, #1 \n" michael@0: "vrshr.u16 q6, q6, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 16 processed per loop. michael@0: "vmul.s16 q8, q4, q10 \n" // B michael@0: "vmls.s16 q8, q5, q11 \n" // G michael@0: "vmls.s16 q8, q6, q12 \n" // R michael@0: "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned michael@0: "vmul.s16 q9, q6, q10 \n" // R michael@0: "vmls.s16 q9, q5, q14 \n" // G michael@0: "vmls.s16 q9, q4, q13 \n" // B michael@0: "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned michael@0: "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U michael@0: "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_rgb565), // %0 michael@0: "+r"(src_stride_rgb565), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. michael@0: void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_argb michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 8 ARGB1555 pixels. michael@0: RGB555TOARGB michael@0: "vpaddl.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. michael@0: "vld1.8 {q0}, [%0]! \n" // next 8 ARGB1555 pixels. michael@0: RGB555TOARGB michael@0: "vpaddl.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. michael@0: michael@0: "vld1.8 {q0}, [%1]! \n" // load 8 ARGB1555 pixels. michael@0: RGB555TOARGB michael@0: "vpadal.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. michael@0: "vld1.8 {q0}, [%1]! \n" // next 8 ARGB1555 pixels. michael@0: RGB555TOARGB michael@0: "vpadal.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. michael@0: michael@0: "vrshr.u16 q4, q4, #1 \n" // 2x average michael@0: "vrshr.u16 q5, q5, #1 \n" michael@0: "vrshr.u16 q6, q6, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 16 processed per loop. michael@0: "vmul.s16 q8, q4, q10 \n" // B michael@0: "vmls.s16 q8, q5, q11 \n" // G michael@0: "vmls.s16 q8, q6, q12 \n" // R michael@0: "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned michael@0: "vmul.s16 q9, q6, q10 \n" // R michael@0: "vmls.s16 q9, q5, q14 \n" // G michael@0: "vmls.s16 q9, q4, q13 \n" // B michael@0: "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned michael@0: "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U michael@0: "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb1555), // %0 michael@0: "+r"(src_stride_argb1555), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. michael@0: void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444, michael@0: uint8* dst_u, uint8* dst_v, int pix) { michael@0: asm volatile ( michael@0: "add %1, %0, %1 \n" // src_stride + src_argb michael@0: "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient michael@0: "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient michael@0: "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient michael@0: "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient michael@0: "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient michael@0: "vmov.u16 q15, #0x8080 \n" // 128.5 michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 8 ARGB4444 pixels. michael@0: ARGB4444TOARGB michael@0: "vpaddl.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. michael@0: "vld1.8 {q0}, [%0]! \n" // next 8 ARGB4444 pixels. michael@0: ARGB4444TOARGB michael@0: "vpaddl.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpaddl.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. michael@0: michael@0: "vld1.8 {q0}, [%1]! \n" // load 8 ARGB4444 pixels. michael@0: ARGB4444TOARGB michael@0: "vpadal.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. michael@0: "vld1.8 {q0}, [%1]! \n" // next 8 ARGB4444 pixels. michael@0: ARGB4444TOARGB michael@0: "vpadal.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. michael@0: "vpadal.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. michael@0: michael@0: "vrshr.u16 q4, q4, #1 \n" // 2x average michael@0: "vrshr.u16 q5, q5, #1 \n" michael@0: "vrshr.u16 q6, q6, #1 \n" michael@0: michael@0: "subs %4, %4, #16 \n" // 16 processed per loop. michael@0: "vmul.s16 q8, q4, q10 \n" // B michael@0: "vmls.s16 q8, q5, q11 \n" // G michael@0: "vmls.s16 q8, q6, q12 \n" // R michael@0: "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned michael@0: "vmul.s16 q9, q6, q10 \n" // R michael@0: "vmls.s16 q9, q5, q14 \n" // G michael@0: "vmls.s16 q9, q4, q13 \n" // B michael@0: "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned michael@0: "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U michael@0: "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. michael@0: "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb4444), // %0 michael@0: "+r"(src_stride_argb4444), // %1 michael@0: "+r"(dst_u), // %2 michael@0: "+r"(dst_v), // %3 michael@0: "+r"(pix) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", michael@0: "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d24, #13 \n" // B * 0.1016 coefficient michael@0: "vmov.u8 d25, #65 \n" // G * 0.5078 coefficient michael@0: "vmov.u8 d26, #33 \n" // R * 0.2578 coefficient michael@0: "vmov.u8 d27, #16 \n" // Add 16 constant michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 8 RGB565 pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: RGB565TOARGB michael@0: "vmull.u8 q2, d0, d24 \n" // B michael@0: "vmlal.u8 q2, d1, d25 \n" // G michael@0: "vmlal.u8 q2, d2, d26 \n" // R michael@0: "vqrshrun.s16 d0, q2, #7 \n" // 16 bit to 8 bit Y michael@0: "vqadd.u8 d0, d27 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_rgb565), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13" michael@0: ); michael@0: } michael@0: michael@0: void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d24, #13 \n" // B * 0.1016 coefficient michael@0: "vmov.u8 d25, #65 \n" // G * 0.5078 coefficient michael@0: "vmov.u8 d26, #33 \n" // R * 0.2578 coefficient michael@0: "vmov.u8 d27, #16 \n" // Add 16 constant michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 8 ARGB1555 pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: ARGB1555TOARGB michael@0: "vmull.u8 q2, d0, d24 \n" // B michael@0: "vmlal.u8 q2, d1, d25 \n" // G michael@0: "vmlal.u8 q2, d2, d26 \n" // R michael@0: "vqrshrun.s16 d0, q2, #7 \n" // 16 bit to 8 bit Y michael@0: "vqadd.u8 d0, d27 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb1555), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13" michael@0: ); michael@0: } michael@0: michael@0: void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d24, #13 \n" // B * 0.1016 coefficient michael@0: "vmov.u8 d25, #65 \n" // G * 0.5078 coefficient michael@0: "vmov.u8 d26, #33 \n" // R * 0.2578 coefficient michael@0: "vmov.u8 d27, #16 \n" // Add 16 constant michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 8 ARGB4444 pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: ARGB4444TOARGB michael@0: "vmull.u8 q2, d0, d24 \n" // B michael@0: "vmlal.u8 q2, d1, d25 \n" // G michael@0: "vmlal.u8 q2, d2, d26 \n" // R michael@0: "vqrshrun.s16 d0, q2, #7 \n" // 16 bit to 8 bit Y michael@0: "vqadd.u8 d0, d27 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb4444), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13" michael@0: ); michael@0: } michael@0: michael@0: void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d4, #33 \n" // R * 0.2578 coefficient michael@0: "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient michael@0: "vmov.u8 d6, #13 \n" // B * 0.1016 coefficient michael@0: "vmov.u8 d7, #16 \n" // Add 16 constant michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of BGRA. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q8, d1, d4 \n" // R michael@0: "vmlal.u8 q8, d2, d5 \n" // G michael@0: "vmlal.u8 q8, d3, d6 \n" // B michael@0: "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y michael@0: "vqadd.u8 d0, d7 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_bgra), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" michael@0: ); michael@0: } michael@0: michael@0: void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d4, #33 \n" // R * 0.2578 coefficient michael@0: "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient michael@0: "vmov.u8 d6, #13 \n" // B * 0.1016 coefficient michael@0: "vmov.u8 d7, #16 \n" // Add 16 constant michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of ABGR. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q8, d0, d4 \n" // R michael@0: "vmlal.u8 q8, d1, d5 \n" // G michael@0: "vmlal.u8 q8, d2, d6 \n" // B michael@0: "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y michael@0: "vqadd.u8 d0, d7 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_abgr), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" michael@0: ); michael@0: } michael@0: michael@0: void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d4, #13 \n" // B * 0.1016 coefficient michael@0: "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient michael@0: "vmov.u8 d6, #33 \n" // R * 0.2578 coefficient michael@0: "vmov.u8 d7, #16 \n" // Add 16 constant michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of RGBA. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q8, d1, d4 \n" // B michael@0: "vmlal.u8 q8, d2, d5 \n" // G michael@0: "vmlal.u8 q8, d3, d6 \n" // R michael@0: "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y michael@0: "vqadd.u8 d0, d7 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_rgba), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" michael@0: ); michael@0: } michael@0: michael@0: void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d4, #13 \n" // B * 0.1016 coefficient michael@0: "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient michael@0: "vmov.u8 d6, #33 \n" // R * 0.2578 coefficient michael@0: "vmov.u8 d7, #16 \n" // Add 16 constant michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld3.8 {d0, d1, d2}, [%0]! \n" // load 8 pixels of RGB24. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q8, d0, d4 \n" // B michael@0: "vmlal.u8 q8, d1, d5 \n" // G michael@0: "vmlal.u8 q8, d2, d6 \n" // R michael@0: "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y michael@0: "vqadd.u8 d0, d7 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_rgb24), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" michael@0: ); michael@0: } michael@0: michael@0: void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) { michael@0: asm volatile ( michael@0: "vmov.u8 d4, #33 \n" // R * 0.2578 coefficient michael@0: "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient michael@0: "vmov.u8 d6, #13 \n" // B * 0.1016 coefficient michael@0: "vmov.u8 d7, #16 \n" // Add 16 constant michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld3.8 {d0, d1, d2}, [%0]! \n" // load 8 pixels of RAW. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q8, d0, d4 \n" // B michael@0: "vmlal.u8 q8, d1, d5 \n" // G michael@0: "vmlal.u8 q8, d2, d6 \n" // R michael@0: "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y michael@0: "vqadd.u8 d0, d7 \n" michael@0: "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. michael@0: "bgt 1b \n" michael@0: : "+r"(src_raw), // %0 michael@0: "+r"(dst_y), // %1 michael@0: "+r"(pix) // %2 michael@0: : michael@0: : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" michael@0: ); michael@0: } michael@0: michael@0: // Bilinear filter 16x2 -> 16x1 michael@0: void InterpolateRow_NEON(uint8* dst_ptr, michael@0: const uint8* src_ptr, ptrdiff_t src_stride, michael@0: int dst_width, int source_y_fraction) { michael@0: asm volatile ( michael@0: "cmp %4, #0 \n" michael@0: "beq 100f \n" michael@0: "add %2, %1 \n" michael@0: "cmp %4, #64 \n" michael@0: "beq 75f \n" michael@0: "cmp %4, #128 \n" michael@0: "beq 50f \n" michael@0: "cmp %4, #192 \n" michael@0: "beq 25f \n" michael@0: michael@0: "vdup.8 d5, %4 \n" michael@0: "rsb %4, #256 \n" michael@0: "vdup.8 d4, %4 \n" michael@0: // General purpose row blend. michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%1]! \n" michael@0: "vld1.8 {q1}, [%2]! \n" michael@0: "subs %3, %3, #16 \n" michael@0: "vmull.u8 q13, d0, d4 \n" michael@0: "vmull.u8 q14, d1, d4 \n" michael@0: "vmlal.u8 q13, d2, d5 \n" michael@0: "vmlal.u8 q14, d3, d5 \n" michael@0: "vrshrn.u16 d0, q13, #8 \n" michael@0: "vrshrn.u16 d1, q14, #8 \n" michael@0: "vst1.8 {q0}, [%0]! \n" michael@0: "bgt 1b \n" michael@0: "b 99f \n" michael@0: michael@0: // Blend 25 / 75. michael@0: "25: \n" michael@0: "vld1.8 {q0}, [%1]! \n" michael@0: "vld1.8 {q1}, [%2]! \n" michael@0: "subs %3, %3, #16 \n" michael@0: "vrhadd.u8 q0, q1 \n" michael@0: "vrhadd.u8 q0, q1 \n" michael@0: "vst1.8 {q0}, [%0]! \n" michael@0: "bgt 25b \n" michael@0: "b 99f \n" michael@0: michael@0: // Blend 50 / 50. michael@0: "50: \n" michael@0: "vld1.8 {q0}, [%1]! \n" michael@0: "vld1.8 {q1}, [%2]! \n" michael@0: "subs %3, %3, #16 \n" michael@0: "vrhadd.u8 q0, q1 \n" michael@0: "vst1.8 {q0}, [%0]! \n" michael@0: "bgt 50b \n" michael@0: "b 99f \n" michael@0: michael@0: // Blend 75 / 25. michael@0: "75: \n" michael@0: "vld1.8 {q1}, [%1]! \n" michael@0: "vld1.8 {q0}, [%2]! \n" michael@0: "subs %3, %3, #16 \n" michael@0: "vrhadd.u8 q0, q1 \n" michael@0: "vrhadd.u8 q0, q1 \n" michael@0: "vst1.8 {q0}, [%0]! \n" michael@0: "bgt 75b \n" michael@0: "b 99f \n" michael@0: michael@0: // Blend 100 / 0 - Copy row unchanged. michael@0: "100: \n" michael@0: "vld1.8 {q0}, [%1]! \n" michael@0: "subs %3, %3, #16 \n" michael@0: "vst1.8 {q0}, [%0]! \n" michael@0: "bgt 100b \n" michael@0: michael@0: "99: \n" michael@0: : "+r"(dst_ptr), // %0 michael@0: "+r"(src_ptr), // %1 michael@0: "+r"(src_stride), // %2 michael@0: "+r"(dst_width), // %3 michael@0: "+r"(source_y_fraction) // %4 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14" michael@0: ); michael@0: } michael@0: michael@0: // dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr michael@0: void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1, michael@0: uint8* dst_argb, int width) { michael@0: asm volatile ( michael@0: "subs %3, #8 \n" michael@0: "blt 89f \n" michael@0: // Blend 8 pixels. michael@0: "8: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of ARGB0. michael@0: "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load 8 pixels of ARGB1. michael@0: "subs %3, %3, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q10, d4, d3 \n" // db * a michael@0: "vmull.u8 q11, d5, d3 \n" // dg * a michael@0: "vmull.u8 q12, d6, d3 \n" // dr * a michael@0: "vqrshrn.u16 d20, q10, #8 \n" // db >>= 8 michael@0: "vqrshrn.u16 d21, q11, #8 \n" // dg >>= 8 michael@0: "vqrshrn.u16 d22, q12, #8 \n" // dr >>= 8 michael@0: "vqsub.u8 q2, q2, q10 \n" // dbg - dbg * a / 256 michael@0: "vqsub.u8 d6, d6, d22 \n" // dr - dr * a / 256 michael@0: "vqadd.u8 q0, q0, q2 \n" // + sbg michael@0: "vqadd.u8 d2, d2, d6 \n" // + sr michael@0: "vmov.u8 d3, #255 \n" // a = 255 michael@0: "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 pixels of ARGB. michael@0: "bge 8b \n" michael@0: michael@0: "89: \n" michael@0: "adds %3, #8-1 \n" michael@0: "blt 99f \n" michael@0: michael@0: // Blend 1 pixels. michael@0: "1: \n" michael@0: "vld4.8 {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n" // load 1 pixel ARGB0. michael@0: "vld4.8 {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n" // load 1 pixel ARGB1. michael@0: "subs %3, %3, #1 \n" // 1 processed per loop. michael@0: "vmull.u8 q10, d4, d3 \n" // db * a michael@0: "vmull.u8 q11, d5, d3 \n" // dg * a michael@0: "vmull.u8 q12, d6, d3 \n" // dr * a michael@0: "vqrshrn.u16 d20, q10, #8 \n" // db >>= 8 michael@0: "vqrshrn.u16 d21, q11, #8 \n" // dg >>= 8 michael@0: "vqrshrn.u16 d22, q12, #8 \n" // dr >>= 8 michael@0: "vqsub.u8 q2, q2, q10 \n" // dbg - dbg * a / 256 michael@0: "vqsub.u8 d6, d6, d22 \n" // dr - dr * a / 256 michael@0: "vqadd.u8 q0, q0, q2 \n" // + sbg michael@0: "vqadd.u8 d2, d2, d6 \n" // + sr michael@0: "vmov.u8 d3, #255 \n" // a = 255 michael@0: "vst4.8 {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n" // store 1 pixel. michael@0: "bge 1b \n" michael@0: michael@0: "99: \n" michael@0: michael@0: : "+r"(src_argb0), // %0 michael@0: "+r"(src_argb1), // %1 michael@0: "+r"(dst_argb), // %2 michael@0: "+r"(width) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12" michael@0: ); michael@0: } michael@0: michael@0: // Attenuate 8 pixels at a time. michael@0: void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) { michael@0: asm volatile ( michael@0: // Attenuate 8 pixels. michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of ARGB. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q10, d0, d3 \n" // b * a michael@0: "vmull.u8 q11, d1, d3 \n" // g * a michael@0: "vmull.u8 q12, d2, d3 \n" // r * a michael@0: "vqrshrn.u16 d0, q10, #8 \n" // b >>= 8 michael@0: "vqrshrn.u16 d1, q11, #8 \n" // g >>= 8 michael@0: "vqrshrn.u16 d2, q12, #8 \n" // r >>= 8 michael@0: "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 pixels of ARGB. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(width) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q10", "q11", "q12" michael@0: ); michael@0: } michael@0: michael@0: // Quantize 8 ARGB pixels (32 bytes). michael@0: // dst = (dst * scale >> 16) * interval_size + interval_offset; michael@0: void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size, michael@0: int interval_offset, int width) { michael@0: asm volatile ( michael@0: "vdup.u16 q8, %2 \n" michael@0: "vshr.u16 q8, q8, #1 \n" // scale >>= 1 michael@0: "vdup.u16 q9, %3 \n" // interval multiply. michael@0: "vdup.u16 q10, %4 \n" // interval add michael@0: michael@0: // 8 pixel loop. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d2, d4, d6}, [%0] \n" // load 8 pixels of ARGB. michael@0: "subs %1, %1, #8 \n" // 8 processed per loop. michael@0: "vmovl.u8 q0, d0 \n" // b (0 .. 255) michael@0: "vmovl.u8 q1, d2 \n" michael@0: "vmovl.u8 q2, d4 \n" michael@0: "vqdmulh.s16 q0, q0, q8 \n" // b * scale michael@0: "vqdmulh.s16 q1, q1, q8 \n" // g michael@0: "vqdmulh.s16 q2, q2, q8 \n" // r michael@0: "vmul.u16 q0, q0, q9 \n" // b * interval_size michael@0: "vmul.u16 q1, q1, q9 \n" // g michael@0: "vmul.u16 q2, q2, q9 \n" // r michael@0: "vadd.u16 q0, q0, q10 \n" // b + interval_offset michael@0: "vadd.u16 q1, q1, q10 \n" // g michael@0: "vadd.u16 q2, q2, q10 \n" // r michael@0: "vqmovn.u16 d0, q0 \n" michael@0: "vqmovn.u16 d2, q1 \n" michael@0: "vqmovn.u16 d4, q2 \n" michael@0: "vst4.8 {d0, d2, d4, d6}, [%0]! \n" // store 8 pixels of ARGB. michael@0: "bgt 1b \n" michael@0: : "+r"(dst_argb), // %0 michael@0: "+r"(width) // %1 michael@0: : "r"(scale), // %2 michael@0: "r"(interval_size), // %3 michael@0: "r"(interval_offset) // %4 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10" michael@0: ); michael@0: } michael@0: michael@0: // Shade 8 pixels at a time by specified value. michael@0: // NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8. michael@0: // Rounding in vqrdmulh does +1 to high if high bit of low s16 is set. michael@0: void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width, michael@0: uint32 value) { michael@0: asm volatile ( michael@0: "vdup.u32 q0, %3 \n" // duplicate scale value. michael@0: "vzip.u8 d0, d1 \n" // d0 aarrggbb. michael@0: "vshr.u16 q0, q0, #1 \n" // scale / 2. michael@0: michael@0: // 8 pixel loop. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d20, d22, d24, d26}, [%0]! \n" // load 8 pixels of ARGB. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmovl.u8 q10, d20 \n" // b (0 .. 255) michael@0: "vmovl.u8 q11, d22 \n" michael@0: "vmovl.u8 q12, d24 \n" michael@0: "vmovl.u8 q13, d26 \n" michael@0: "vqrdmulh.s16 q10, q10, d0[0] \n" // b * scale * 2 michael@0: "vqrdmulh.s16 q11, q11, d0[1] \n" // g michael@0: "vqrdmulh.s16 q12, q12, d0[2] \n" // r michael@0: "vqrdmulh.s16 q13, q13, d0[3] \n" // a michael@0: "vqmovn.u16 d20, q10 \n" michael@0: "vqmovn.u16 d22, q11 \n" michael@0: "vqmovn.u16 d24, q12 \n" michael@0: "vqmovn.u16 d26, q13 \n" michael@0: "vst4.8 {d20, d22, d24, d26}, [%1]! \n" // store 8 pixels of ARGB. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(width) // %2 michael@0: : "r"(value) // %3 michael@0: : "cc", "memory", "q0", "q10", "q11", "q12", "q13" michael@0: ); michael@0: } michael@0: michael@0: // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels michael@0: // Similar to ARGBToYJ but stores ARGB. michael@0: // C code is (15 * b + 75 * g + 38 * r + 64) >> 7; michael@0: void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) { michael@0: asm volatile ( michael@0: "vmov.u8 d24, #15 \n" // B * 0.11400 coefficient michael@0: "vmov.u8 d25, #75 \n" // G * 0.58700 coefficient michael@0: "vmov.u8 d26, #38 \n" // R * 0.29900 coefficient michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q2, d0, d24 \n" // B michael@0: "vmlal.u8 q2, d1, d25 \n" // G michael@0: "vmlal.u8 q2, d2, d26 \n" // R michael@0: "vqrshrun.s16 d0, q2, #7 \n" // 15 bit to 8 bit B michael@0: "vmov d1, d0 \n" // G michael@0: "vmov d2, d0 \n" // R michael@0: "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 ARGB pixels. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(width) // %2 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q12", "q13" michael@0: ); michael@0: } michael@0: michael@0: // Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels. michael@0: // b = (r * 35 + g * 68 + b * 17) >> 7 michael@0: // g = (r * 45 + g * 88 + b * 22) >> 7 michael@0: // r = (r * 50 + g * 98 + b * 24) >> 7 michael@0: void ARGBSepiaRow_NEON(uint8* dst_argb, int width) { michael@0: asm volatile ( michael@0: "vmov.u8 d20, #17 \n" // BB coefficient michael@0: "vmov.u8 d21, #68 \n" // BG coefficient michael@0: "vmov.u8 d22, #35 \n" // BR coefficient michael@0: "vmov.u8 d24, #22 \n" // GB coefficient michael@0: "vmov.u8 d25, #88 \n" // GG coefficient michael@0: "vmov.u8 d26, #45 \n" // GR coefficient michael@0: "vmov.u8 d28, #24 \n" // BB coefficient michael@0: "vmov.u8 d29, #98 \n" // BG coefficient michael@0: "vmov.u8 d30, #50 \n" // BR coefficient michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0] \n" // load 8 ARGB pixels. michael@0: "subs %1, %1, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q2, d0, d20 \n" // B to Sepia B michael@0: "vmlal.u8 q2, d1, d21 \n" // G michael@0: "vmlal.u8 q2, d2, d22 \n" // R michael@0: "vmull.u8 q3, d0, d24 \n" // B to Sepia G michael@0: "vmlal.u8 q3, d1, d25 \n" // G michael@0: "vmlal.u8 q3, d2, d26 \n" // R michael@0: "vmull.u8 q8, d0, d28 \n" // B to Sepia R michael@0: "vmlal.u8 q8, d1, d29 \n" // G michael@0: "vmlal.u8 q8, d2, d30 \n" // R michael@0: "vqshrn.u16 d0, q2, #7 \n" // 16 bit to 8 bit B michael@0: "vqshrn.u16 d1, q3, #7 \n" // 16 bit to 8 bit G michael@0: "vqshrn.u16 d2, q8, #7 \n" // 16 bit to 8 bit R michael@0: "vst4.8 {d0, d1, d2, d3}, [%0]! \n" // store 8 ARGB pixels. michael@0: "bgt 1b \n" michael@0: : "+r"(dst_argb), // %0 michael@0: "+r"(width) // %1 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", michael@0: "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // Tranform 8 ARGB pixels (32 bytes) with color matrix. michael@0: // TODO(fbarchard): Was same as Sepia except matrix is provided. This function michael@0: // needs to saturate. Consider doing a non-saturating version. michael@0: void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb, michael@0: const int8* matrix_argb, int width) { michael@0: asm volatile ( michael@0: "vld1.8 {q2}, [%3] \n" // load 3 ARGB vectors. michael@0: "vmovl.s8 q0, d4 \n" // B,G coefficients s16. michael@0: "vmovl.s8 q1, d5 \n" // R,A coefficients s16. michael@0: michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d16, d18, d20, d22}, [%0]! \n" // load 8 ARGB pixels. michael@0: "subs %2, %2, #8 \n" // 8 processed per loop. michael@0: "vmovl.u8 q8, d16 \n" // b (0 .. 255) 16 bit michael@0: "vmovl.u8 q9, d18 \n" // g michael@0: "vmovl.u8 q10, d20 \n" // r michael@0: "vmovl.u8 q15, d22 \n" // a michael@0: "vmul.s16 q12, q8, d0[0] \n" // B = B * Matrix B michael@0: "vmul.s16 q13, q8, d1[0] \n" // G = B * Matrix G michael@0: "vmul.s16 q14, q8, d2[0] \n" // R = B * Matrix R michael@0: "vmul.s16 q15, q8, d3[0] \n" // A = B * Matrix A michael@0: "vmul.s16 q4, q9, d0[1] \n" // B += G * Matrix B michael@0: "vmul.s16 q5, q9, d1[1] \n" // G += G * Matrix G michael@0: "vmul.s16 q6, q9, d2[1] \n" // R += G * Matrix R michael@0: "vmul.s16 q7, q9, d3[1] \n" // A += G * Matrix A michael@0: "vqadd.s16 q12, q12, q4 \n" // Accumulate B michael@0: "vqadd.s16 q13, q13, q5 \n" // Accumulate G michael@0: "vqadd.s16 q14, q14, q6 \n" // Accumulate R michael@0: "vqadd.s16 q15, q15, q7 \n" // Accumulate A michael@0: "vmul.s16 q4, q10, d0[2] \n" // B += R * Matrix B michael@0: "vmul.s16 q5, q10, d1[2] \n" // G += R * Matrix G michael@0: "vmul.s16 q6, q10, d2[2] \n" // R += R * Matrix R michael@0: "vmul.s16 q7, q10, d3[2] \n" // A += R * Matrix A michael@0: "vqadd.s16 q12, q12, q4 \n" // Accumulate B michael@0: "vqadd.s16 q13, q13, q5 \n" // Accumulate G michael@0: "vqadd.s16 q14, q14, q6 \n" // Accumulate R michael@0: "vqadd.s16 q15, q15, q7 \n" // Accumulate A michael@0: "vmul.s16 q4, q15, d0[3] \n" // B += A * Matrix B michael@0: "vmul.s16 q5, q15, d1[3] \n" // G += A * Matrix G michael@0: "vmul.s16 q6, q15, d2[3] \n" // R += A * Matrix R michael@0: "vmul.s16 q7, q15, d3[3] \n" // A += A * Matrix A michael@0: "vqadd.s16 q12, q12, q4 \n" // Accumulate B michael@0: "vqadd.s16 q13, q13, q5 \n" // Accumulate G michael@0: "vqadd.s16 q14, q14, q6 \n" // Accumulate R michael@0: "vqadd.s16 q15, q15, q7 \n" // Accumulate A michael@0: "vqshrun.s16 d16, q12, #6 \n" // 16 bit to 8 bit B michael@0: "vqshrun.s16 d18, q13, #6 \n" // 16 bit to 8 bit G michael@0: "vqshrun.s16 d20, q14, #6 \n" // 16 bit to 8 bit R michael@0: "vqshrun.s16 d22, q15, #6 \n" // 16 bit to 8 bit A michael@0: "vst4.8 {d16, d18, d20, d22}, [%1]! \n" // store 8 ARGB pixels. michael@0: "bgt 1b \n" michael@0: : "+r"(src_argb), // %0 michael@0: "+r"(dst_argb), // %1 michael@0: "+r"(width) // %2 michael@0: : "r"(matrix_argb) // %3 michael@0: : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", michael@0: "q10", "q11", "q12", "q13", "q14", "q15" michael@0: ); michael@0: } michael@0: michael@0: // TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable. michael@0: #ifdef HAS_ARGBMULTIPLYROW_NEON michael@0: // Multiply 2 rows of ARGB pixels together, 8 pixels at a time. michael@0: void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1, michael@0: uint8* dst_argb, int width) { michael@0: asm volatile ( michael@0: // 8 pixel loop. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. michael@0: "vld4.8 {d1, d3, d5, d7}, [%1]! \n" // load 8 more ARGB pixels. michael@0: "subs %3, %3, #8 \n" // 8 processed per loop. michael@0: "vmull.u8 q0, d0, d1 \n" // multiply B michael@0: "vmull.u8 q1, d2, d3 \n" // multiply G michael@0: "vmull.u8 q2, d4, d5 \n" // multiply R michael@0: "vmull.u8 q3, d6, d7 \n" // multiply A michael@0: "vrshrn.u16 d0, q0, #8 \n" // 16 bit to 8 bit B michael@0: "vrshrn.u16 d1, q1, #8 \n" // 16 bit to 8 bit G michael@0: "vrshrn.u16 d2, q2, #8 \n" // 16 bit to 8 bit R michael@0: "vrshrn.u16 d3, q3, #8 \n" // 16 bit to 8 bit A michael@0: "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. michael@0: "bgt 1b \n" michael@0: michael@0: : "+r"(src_argb0), // %0 michael@0: "+r"(src_argb1), // %1 michael@0: "+r"(dst_argb), // %2 michael@0: "+r"(width) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3" michael@0: ); michael@0: } michael@0: #endif // HAS_ARGBMULTIPLYROW_NEON michael@0: michael@0: // Add 2 rows of ARGB pixels together, 8 pixels at a time. michael@0: void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1, michael@0: uint8* dst_argb, int width) { michael@0: asm volatile ( michael@0: // 8 pixel loop. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. michael@0: "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load 8 more ARGB pixels. michael@0: "subs %3, %3, #8 \n" // 8 processed per loop. michael@0: "vqadd.u8 q0, q0, q2 \n" // add B, G michael@0: "vqadd.u8 q1, q1, q3 \n" // add R, A michael@0: "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. michael@0: "bgt 1b \n" michael@0: michael@0: : "+r"(src_argb0), // %0 michael@0: "+r"(src_argb1), // %1 michael@0: "+r"(dst_argb), // %2 michael@0: "+r"(width) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3" michael@0: ); michael@0: } michael@0: michael@0: // Subtract 2 rows of ARGB pixels, 8 pixels at a time. michael@0: void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1, michael@0: uint8* dst_argb, int width) { michael@0: asm volatile ( michael@0: // 8 pixel loop. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. michael@0: "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load 8 more ARGB pixels. michael@0: "subs %3, %3, #8 \n" // 8 processed per loop. michael@0: "vqsub.u8 q0, q0, q2 \n" // subtract B, G michael@0: "vqsub.u8 q1, q1, q3 \n" // subtract R, A michael@0: "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. michael@0: "bgt 1b \n" michael@0: michael@0: : "+r"(src_argb0), // %0 michael@0: "+r"(src_argb1), // %1 michael@0: "+r"(dst_argb), // %2 michael@0: "+r"(width) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1", "q2", "q3" michael@0: ); michael@0: } michael@0: michael@0: // Adds Sobel X and Sobel Y and stores Sobel into ARGB. michael@0: // A = 255 michael@0: // R = Sobel michael@0: // G = Sobel michael@0: // B = Sobel michael@0: void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, michael@0: uint8* dst_argb, int width) { michael@0: asm volatile ( michael@0: "vmov.u8 d3, #255 \n" // alpha michael@0: // 8 pixel loop. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {d0}, [%0]! \n" // load 8 sobelx. michael@0: "vld1.8 {d1}, [%1]! \n" // load 8 sobely. michael@0: "subs %3, %3, #8 \n" // 8 processed per loop. michael@0: "vqadd.u8 d0, d0, d1 \n" // add michael@0: "vmov.u8 d1, d0 \n" michael@0: "vmov.u8 d2, d0 \n" michael@0: "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. michael@0: "bgt 1b \n" michael@0: : "+r"(src_sobelx), // %0 michael@0: "+r"(src_sobely), // %1 michael@0: "+r"(dst_argb), // %2 michael@0: "+r"(width) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1" michael@0: ); michael@0: } michael@0: michael@0: // Adds Sobel X and Sobel Y and stores Sobel into plane. michael@0: void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, michael@0: uint8* dst_y, int width) { michael@0: asm volatile ( michael@0: // 16 pixel loop. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {q0}, [%0]! \n" // load 16 sobelx. michael@0: "vld1.8 {q1}, [%1]! \n" // load 16 sobely. michael@0: "subs %3, %3, #16 \n" // 16 processed per loop. michael@0: "vqadd.u8 q0, q0, q1 \n" // add michael@0: "vst1.8 {q0}, [%2]! \n" // store 16 pixels. michael@0: "bgt 1b \n" michael@0: : "+r"(src_sobelx), // %0 michael@0: "+r"(src_sobely), // %1 michael@0: "+r"(dst_y), // %2 michael@0: "+r"(width) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1" michael@0: ); michael@0: } michael@0: michael@0: // Mixes Sobel X, Sobel Y and Sobel into ARGB. michael@0: // A = 255 michael@0: // R = Sobel X michael@0: // G = Sobel michael@0: // B = Sobel Y michael@0: void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, michael@0: uint8* dst_argb, int width) { michael@0: asm volatile ( michael@0: "vmov.u8 d3, #255 \n" // alpha michael@0: // 8 pixel loop. michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {d2}, [%0]! \n" // load 8 sobelx. michael@0: "vld1.8 {d0}, [%1]! \n" // load 8 sobely. michael@0: "subs %3, %3, #8 \n" // 8 processed per loop. michael@0: "vqadd.u8 d1, d0, d2 \n" // add michael@0: "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. michael@0: "bgt 1b \n" michael@0: : "+r"(src_sobelx), // %0 michael@0: "+r"(src_sobely), // %1 michael@0: "+r"(dst_argb), // %2 michael@0: "+r"(width) // %3 michael@0: : michael@0: : "cc", "memory", "q0", "q1" michael@0: ); michael@0: } michael@0: michael@0: // SobelX as a matrix is michael@0: // -1 0 1 michael@0: // -2 0 2 michael@0: // -1 0 1 michael@0: void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1, michael@0: const uint8* src_y2, uint8* dst_sobelx, int width) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {d0}, [%0],%5 \n" // top michael@0: "vld1.8 {d1}, [%0],%6 \n" michael@0: "vsubl.u8 q0, d0, d1 \n" michael@0: "vld1.8 {d2}, [%1],%5 \n" // center * 2 michael@0: "vld1.8 {d3}, [%1],%6 \n" michael@0: "vsubl.u8 q1, d2, d3 \n" michael@0: "vadd.s16 q0, q0, q1 \n" michael@0: "vadd.s16 q0, q0, q1 \n" michael@0: "vld1.8 {d2}, [%2],%5 \n" // bottom michael@0: "vld1.8 {d3}, [%2],%6 \n" michael@0: "subs %4, %4, #8 \n" // 8 pixels michael@0: "vsubl.u8 q1, d2, d3 \n" michael@0: "vadd.s16 q0, q0, q1 \n" michael@0: "vabs.s16 q0, q0 \n" michael@0: "vqmovn.u16 d0, q0 \n" michael@0: "vst1.8 {d0}, [%3]! \n" // store 8 sobelx michael@0: "bgt 1b \n" michael@0: : "+r"(src_y0), // %0 michael@0: "+r"(src_y1), // %1 michael@0: "+r"(src_y2), // %2 michael@0: "+r"(dst_sobelx), // %3 michael@0: "+r"(width) // %4 michael@0: : "r"(2), // %5 michael@0: "r"(6) // %6 michael@0: : "cc", "memory", "q0", "q1" // Clobber List michael@0: ); michael@0: } michael@0: michael@0: // SobelY as a matrix is michael@0: // -1 -2 -1 michael@0: // 0 0 0 michael@0: // 1 2 1 michael@0: void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1, michael@0: uint8* dst_sobely, int width) { michael@0: asm volatile ( michael@0: ".p2align 2 \n" michael@0: "1: \n" michael@0: "vld1.8 {d0}, [%0],%4 \n" // left michael@0: "vld1.8 {d1}, [%1],%4 \n" michael@0: "vsubl.u8 q0, d0, d1 \n" michael@0: "vld1.8 {d2}, [%0],%4 \n" // center * 2 michael@0: "vld1.8 {d3}, [%1],%4 \n" michael@0: "vsubl.u8 q1, d2, d3 \n" michael@0: "vadd.s16 q0, q0, q1 \n" michael@0: "vadd.s16 q0, q0, q1 \n" michael@0: "vld1.8 {d2}, [%0],%5 \n" // right michael@0: "vld1.8 {d3}, [%1],%5 \n" michael@0: "subs %3, %3, #8 \n" // 8 pixels michael@0: "vsubl.u8 q1, d2, d3 \n" michael@0: "vadd.s16 q0, q0, q1 \n" michael@0: "vabs.s16 q0, q0 \n" michael@0: "vqmovn.u16 d0, q0 \n" michael@0: "vst1.8 {d0}, [%2]! \n" // store 8 sobely michael@0: "bgt 1b \n" michael@0: : "+r"(src_y0), // %0 michael@0: "+r"(src_y1), // %1 michael@0: "+r"(dst_sobely), // %2 michael@0: "+r"(width) // %3 michael@0: : "r"(1), // %4 michael@0: "r"(6) // %5 michael@0: : "cc", "memory", "q0", "q1" // Clobber List michael@0: ); michael@0: } michael@0: #endif // __ARM_NEON__ michael@0: michael@0: #ifdef __cplusplus michael@0: } // extern "C" michael@0: } // namespace libyuv michael@0: #endif