From 4e5e44e21e511f20d1fa326ca369ba2458469d57 Mon Sep 17 00:00:00 2001 From: "fbarchard@google.com" Date: Mon, 16 Jun 2014 17:09:48 +0000 Subject: [PATCH] scale neon nacl port BUG=333 TESTED=ncval R=tpsiaki@google.com Review URL: https://webrtc-codereview.appspot.com/18549004 git-svn-id: http://libyuv.googlecode.com/svn/trunk@1015 16f28f9a-4ce2-e073-06de-1de4eb20be90 --- README.chromium | 2 +- include/libyuv/version.h | 2 +- source/scale_neon.cc | 70 ++++++++++++++++++++++++++++++++++++++++ 3 files changed, 72 insertions(+), 2 deletions(-) diff --git a/README.chromium b/README.chromium index 956a8c02b..fffa1cc7c 100644 --- a/README.chromium +++ b/README.chromium @@ -1,6 +1,6 @@ Name: libyuv URL: http://code.google.com/p/libyuv/ -Version: 1014 +Version: 1015 License: BSD License File: LICENSE diff --git a/include/libyuv/version.h b/include/libyuv/version.h index 652ec3def..6f79f68e8 100644 --- a/include/libyuv/version.h +++ b/include/libyuv/version.h @@ -11,6 +11,6 @@ #ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT #define INCLUDE_LIBYUV_VERSION_H_ -#define LIBYUV_VERSION 1014 +#define LIBYUV_VERSION 1015 #endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT diff --git a/source/scale_neon.cc b/source/scale_neon.cc index 410364a1b..a094424b8 100644 --- a/source/scale_neon.cc +++ b/source/scale_neon.cc @@ -29,8 +29,10 @@ void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride, ".p2align 2 \n" "1: \n" // load even pixels into q0, odd into q1 + MEMACCESS(0) "vld2.8 {q0, q1}, [%0]! \n" "subs %2, %2, #16 \n" // 16 processed per loop + MEMACCESS(1) "vst1.8 {q1}, [%1]! \n" // store odd pixels "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -49,7 +51,9 @@ void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, "add %1, %0 \n" ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld1.8 {q0, q1}, [%0]! \n" // load row 1 and post inc + MEMACCESS(1) "vld1.8 {q2, q3}, [%1]! \n" // load row 2 and post inc "subs %3, %3, #16 \n" // 16 processed per loop "vpaddl.u8 q0, q0 \n" // row 1 add adjacent @@ -58,6 +62,7 @@ void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, "vpadal.u8 q1, q3 \n" "vrshrn.u16 d0, q0, #2 \n" // downshift, round and pack "vrshrn.u16 d1, q1, #2 \n" + MEMACCESS(2) "vst1.8 {q0}, [%2]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -74,8 +79,10 @@ void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride, asm volatile ( ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 "subs %2, %2, #8 \n" // 8 processed per loop + MEMACCESS(1) "vst1.8 {d2}, [%1]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -94,9 +101,11 @@ void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, "add %3, r5, %3 \n" ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld1.8 {q0}, [%0]! \n" // load up 16x4 "vld1.8 {q1}, [r4]! \n" "vld1.8 {q2}, [r5]! \n" + MEMACCESS(3) "vld1.8 {q3}, [%3]! \n" "subs %2, %2, #4 \n" "vpaddl.u8 q0, q0 \n" @@ -106,6 +115,7 @@ void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, "vpaddl.u16 q0, q0 \n" "vrshrn.u32 d0, q0, #4 \n" // divide by 16 w/rounding "vmovn.u16 d0, q0 \n" + MEMACCESS(1) "vst1.32 {d0[0]}, [%1]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -125,9 +135,11 @@ void ScaleRowDown34_NEON(const uint8* src_ptr, asm volatile ( ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 "subs %2, %2, #24 \n" "vmov d2, d3 \n" // order d0, d1, d2 + MEMACCESS(1) "vst3.8 {d0, d1, d2}, [%1]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -146,7 +158,9 @@ void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr, "add %3, %0 \n" ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 + MEMACCESS(3) "vld4.8 {d4, d5, d6, d7}, [%3]! \n" // src line 1 "subs %2, %2, #24 \n" @@ -183,6 +197,7 @@ void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr, "vmlal.u8 q8, d3, d24 \n" "vqrshrn.u16 d2, q8, #2 \n" + MEMACCESS(1) "vst3.8 {d0, d1, d2}, [%1]! \n" "bgt 1b \n" @@ -203,7 +218,9 @@ void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr, "add %3, %0 \n" ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 + MEMACCESS(3) "vld4.8 {d4, d5, d6, d7}, [%3]! \n" // src line 1 "subs %2, %2, #24 \n" // average src line 0 with src line 1 @@ -223,6 +240,7 @@ void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr, "vmlal.u8 q3, d3, d24 \n" "vqrshrn.u16 d2, q3, #2 \n" + MEMACCESS(1) "vst3.8 {d0, d1, d2}, [%1]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -251,14 +269,18 @@ void ScaleRowDown38_NEON(const uint8* src_ptr, ptrdiff_t src_stride, uint8* dst_ptr, int dst_width) { asm volatile ( + MEMACCESS(3) "vld1.8 {q3}, [%3] \n" ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld1.8 {d0, d1, d2, d3}, [%0]! \n" "subs %2, %2, #12 \n" "vtbl.u8 d4, {d0, d1, d2, d3}, d6 \n" "vtbl.u8 d5, {d0, d1, d2, d3}, d7 \n" + MEMACCESS(1) "vst1.8 {d4}, [%1]! \n" + MEMACCESS(1) "vst1.32 {d5[0]}, [%1]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -286,7 +308,9 @@ void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr, // d1 = 10 50 11 51 12 52 13 53 // d2 = 20 60 21 61 22 62 23 63 // d3 = 30 70 31 71 32 72 33 73 + MEMACCESS(0) "vld4.8 {d0, d1, d2, d3}, [%0]! \n" + MEMACCESS(3) "vld4.8 {d4, d5, d6, d7}, [%3]! \n" "vld4.8 {d16, d17, d18, d19}, [r4]! \n" "subs %2, %2, #12 \n" @@ -365,7 +389,9 @@ void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr, "vtbl.u8 d3, {d0, d1, d2}, d28 \n" "vtbl.u8 d4, {d0, d1, d2}, d29 \n" + MEMACCESS(1) "vst1.8 {d3}, [%1]! \n" + MEMACCESS(1) "vst1.32 {d4[0]}, [%1]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -385,7 +411,9 @@ void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, uint8* dst_ptr, int dst_width) { asm volatile ( + MEMACCESS(4) "vld1.16 {q13}, [%4] \n" + MEMACCESS(5) "vld1.8 {q14}, [%5] \n" "add %3, %0 \n" ".p2align 2 \n" @@ -395,7 +423,9 @@ void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr, // d1 = 10 50 11 51 12 52 13 53 // d2 = 20 60 21 61 22 62 23 63 // d3 = 30 70 31 71 32 72 33 73 + MEMACCESS(0) "vld4.8 {d0, d1, d2, d3}, [%0]! \n" + MEMACCESS(3) "vld4.8 {d4, d5, d6, d7}, [%3]! \n" "subs %2, %2, #12 \n" @@ -462,7 +492,9 @@ void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr, "vtbl.u8 d3, {d0, d1, d2}, d28 \n" "vtbl.u8 d4, {d0, d1, d2}, d29 \n" + MEMACCESS(1) "vst1.8 {d3}, [%1]! \n" + MEMACCESS(1) "vst1.32 {d4[0]}, [%1]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -495,7 +527,9 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "vdup.8 d4, %4 \n" // General purpose row blend. "1: \n" + MEMACCESS(1) "vld1.8 {q0}, [%1]! \n" + MEMACCESS(2) "vld1.8 {q1}, [%2]! \n" "subs %3, %3, #16 \n" "vmull.u8 q13, d0, d4 \n" @@ -504,50 +538,63 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "vmlal.u8 q14, d3, d5 \n" "vrshrn.u16 d0, q13, #8 \n" "vrshrn.u16 d1, q14, #8 \n" + MEMACCESS(0) "vst1.8 {q0}, [%0]! \n" "bgt 1b \n" "b 99f \n" // Blend 25 / 75. "25: \n" + MEMACCESS(1) "vld1.8 {q0}, [%1]! \n" + MEMACCESS(2) "vld1.8 {q1}, [%2]! \n" "subs %3, %3, #16 \n" "vrhadd.u8 q0, q1 \n" "vrhadd.u8 q0, q1 \n" + MEMACCESS(0) "vst1.8 {q0}, [%0]! \n" "bgt 25b \n" "b 99f \n" // Blend 50 / 50. "50: \n" + MEMACCESS(1) "vld1.8 {q0}, [%1]! \n" + MEMACCESS(2) "vld1.8 {q1}, [%2]! \n" "subs %3, %3, #16 \n" "vrhadd.u8 q0, q1 \n" + MEMACCESS(0) "vst1.8 {q0}, [%0]! \n" "bgt 50b \n" "b 99f \n" // Blend 75 / 25. "75: \n" + MEMACCESS(1) "vld1.8 {q1}, [%1]! \n" + MEMACCESS(2) "vld1.8 {q0}, [%2]! \n" "subs %3, %3, #16 \n" "vrhadd.u8 q0, q1 \n" "vrhadd.u8 q0, q1 \n" + MEMACCESS(0) "vst1.8 {q0}, [%0]! \n" "bgt 75b \n" "b 99f \n" // Blend 100 / 0 - Copy row unchanged. "100: \n" + MEMACCESS(1) "vld1.8 {q0}, [%1]! \n" "subs %3, %3, #16 \n" + MEMACCESS(0) "vst1.8 {q0}, [%0]! \n" "bgt 100b \n" "99: \n" + MEMACCESS(0) "vst1.8 {d1[7]}, [%0] \n" : "+r"(dst_ptr), // %0 "+r"(src_ptr), // %1 @@ -565,10 +612,14 @@ void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride, ".p2align 2 \n" "1: \n" // load even pixels into q0, odd into q1 + MEMACCESS(0) "vld2.32 {q0, q1}, [%0]! \n" + MEMACCESS(0) "vld2.32 {q2, q3}, [%0]! \n" "subs %2, %2, #8 \n" // 8 processed per loop + MEMACCESS(1) "vst1.8 {q1}, [%1]! \n" // store odd pixels + MEMACCESS(1) "vst1.8 {q3}, [%1]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -586,14 +637,18 @@ void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, "add %1, %1, %0 \n" ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(0) "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. "subs %3, %3, #8 \n" // 8 processed per loop. "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. "vpaddl.u8 q3, q3 \n" // A 16 bytes -> 8 shorts. + MEMACCESS(1) "vld4.8 {d16, d18, d20, d22}, [%1]! \n" // load 8 more ARGB pixels. + MEMACCESS(1) "vld4.8 {d17, d19, d21, d23}, [%1]! \n" // load last 8 ARGB pixels. "vpadal.u8 q0, q8 \n" // B 16 bytes -> 8 shorts. "vpadal.u8 q1, q9 \n" // G 16 bytes -> 8 shorts. @@ -603,6 +658,7 @@ void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, "vrshrn.u16 d1, q1, #2 \n" "vrshrn.u16 d2, q2, #2 \n" "vrshrn.u16 d3, q3, #2 \n" + MEMACCESS(2) "vst4.8 {d0, d1, d2, d3}, [%2]! \n" "bgt 1b \n" : "+r"(src_ptr), // %0 @@ -622,11 +678,16 @@ void ScaleARGBRowDownEven_NEON(const uint8* src_argb, ptrdiff_t src_stride, "mov r12, %3, lsl #2 \n" ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld1.32 {d0[0]}, [%0], r12 \n" + MEMACCESS(0) "vld1.32 {d0[1]}, [%0], r12 \n" + MEMACCESS(0) "vld1.32 {d1[0]}, [%0], r12 \n" + MEMACCESS(0) "vld1.32 {d1[1]}, [%0], r12 \n" "subs %2, %2, #4 \n" // 4 pixels per loop. + MEMACCESS(1) "vst1.8 {q0}, [%1]! \n" "bgt 1b \n" : "+r"(src_argb), // %0 @@ -647,13 +708,21 @@ void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride, "add %1, %1, %0 \n" ".p2align 2 \n" "1: \n" + MEMACCESS(0) "vld1.8 {d0}, [%0], r12 \n" // Read 4 2x2 blocks -> 2x1 + MEMACCESS(1) "vld1.8 {d1}, [%1], r12 \n" + MEMACCESS(0) "vld1.8 {d2}, [%0], r12 \n" + MEMACCESS(1) "vld1.8 {d3}, [%1], r12 \n" + MEMACCESS(0) "vld1.8 {d4}, [%0], r12 \n" + MEMACCESS(1) "vld1.8 {d5}, [%1], r12 \n" + MEMACCESS(0) "vld1.8 {d6}, [%0], r12 \n" + MEMACCESS(1) "vld1.8 {d7}, [%1], r12 \n" "vaddl.u8 q0, d0, d1 \n" "vaddl.u8 q1, d2, d3 \n" @@ -666,6 +735,7 @@ void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride, "vrshrn.u16 d0, q0, #2 \n" // first 2 pixels. "vrshrn.u16 d1, q2, #2 \n" // next 2 pixels. "subs %3, %3, #4 \n" // 4 pixels per loop. + MEMACCESS(2) "vst1.8 {q0}, [%2]! \n" "bgt 1b \n" : "+r"(src_argb), // %0