diff --git a/README.chromium b/README.chromium index 757e86d0b..c8f85b208 100644 --- a/README.chromium +++ b/README.chromium @@ -1,6 +1,6 @@ Name: libyuv URL: http://code.google.com/p/libyuv/ -Version: 1666 +Version: 1667 License: BSD License File: LICENSE diff --git a/include/libyuv/version.h b/include/libyuv/version.h index b9f3d6522..eca7f0334 100644 --- a/include/libyuv/version.h +++ b/include/libyuv/version.h @@ -11,6 +11,6 @@ #ifndef INCLUDE_LIBYUV_VERSION_H_ #define INCLUDE_LIBYUV_VERSION_H_ -#define LIBYUV_VERSION 1666 +#define LIBYUV_VERSION 1667 #endif // INCLUDE_LIBYUV_VERSION_H_ diff --git a/source/scale_neon.cc b/source/scale_neon.cc index c45c6409a..b03a82821 100644 --- a/source/scale_neon.cc +++ b/source/scale_neon.cc @@ -29,7 +29,7 @@ void ScaleRowDown2_NEON(const uint8* src_ptr, int dst_width) { (void)src_stride; asm volatile( - "1: \n" + "1: \n" // load even pixels into q0, odd into q1 "vld2.8 {q0, q1}, [%0]! \n" "subs %2, %2, #16 \n" // 16 processed per loop @@ -50,7 +50,7 @@ void ScaleRowDown2Linear_NEON(const uint8* src_ptr, int dst_width) { (void)src_stride; asm volatile( - "1: \n" + "1: \n" "vld2.8 {q0, q1}, [%0]! \n" // load 32 pixels "subs %2, %2, #16 \n" // 16 processed per loop "vrhadd.u8 q0, q0, q1 \n" // rounding half add @@ -72,7 +72,7 @@ void ScaleRowDown2Box_NEON(const uint8* src_ptr, asm volatile( // change the stride to row 2 pointer "add %1, %0 \n" - "1: \n" + "1: \n" "vld1.8 {q0, q1}, [%0]! \n" // load row 1 and post inc "vld1.8 {q2, q3}, [%1]! \n" // load row 2 and post inc "subs %3, %3, #16 \n" // 16 processed per loop @@ -101,7 +101,7 @@ void ScaleRowDown4_NEON(const uint8* src_ptr, int dst_width) { (void)src_stride; asm volatile( - "1: \n" + "1: \n" "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 "subs %2, %2, #8 \n" // 8 processed per loop "vst1.8 {d2}, [%1]! \n" @@ -121,7 +121,7 @@ void ScaleRowDown4Box_NEON(const uint8* src_ptr, const uint8* src_ptr2 = src_ptr + src_stride * 2; const uint8* src_ptr3 = src_ptr + src_stride * 3; asm volatile( - "1: \n" + "1: \n" "vld1.8 {q0}, [%0]! \n" // load up 16x4 "vld1.8 {q1}, [%3]! \n" "vld1.8 {q2}, [%4]! \n" @@ -155,12 +155,12 @@ void ScaleRowDown34_NEON(const uint8* src_ptr, int dst_width) { (void)src_stride; asm volatile( - "1: \n" - "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 - "subs %2, %2, #24 \n" - "vmov d2, d3 \n" // order d0, d1, d2 - "vst3.8 {d0, d1, d2}, [%1]! \n" - "bgt 1b \n" + "1: \n" + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 + "subs %2, %2, #24 \n" + "vmov d2, d3 \n" // order d0, d1, d2 + "vst3.8 {d0, d1, d2}, [%1]! \n" + "bgt 1b \n" : "+r"(src_ptr), // %0 "+r"(dst_ptr), // %1 "+r"(dst_width) // %2 @@ -175,7 +175,7 @@ void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr, asm volatile( "vmov.u8 d24, #3 \n" "add %3, %0 \n" - "1: \n" + "1: \n" "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 "vld4.8 {d4, d5, d6, d7}, [%3]! \n" // src line 1 "subs %2, %2, #24 \n" @@ -232,7 +232,7 @@ void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr, asm volatile( "vmov.u8 d24, #3 \n" "add %3, %0 \n" - "1: \n" + "1: \n" "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 "vld4.8 {d4, d5, d6, d7}, [%3]! \n" // src line 1 "subs %2, %2, #24 \n" @@ -280,7 +280,7 @@ void ScaleRowDown38_NEON(const uint8* src_ptr, (void)src_stride; asm volatile( "vld1.8 {q3}, [%3] \n" - "1: \n" + "1: \n" "vld1.8 {d0, d1, d2, d3}, [%0]! \n" "subs %2, %2, #12 \n" "vtbl.u8 d4, {d0, d1, d2, d3}, d6 \n" @@ -307,7 +307,7 @@ void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr, "vld1.8 {q14}, [%6] \n" "vld1.8 {q15}, [%7] \n" "add %3, %0 \n" - "1: \n" + "1: \n" // d0 = 00 40 01 41 02 42 03 43 // d1 = 10 50 11 51 12 52 13 53 @@ -416,7 +416,7 @@ void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr, "vld1.16 {q13}, [%4] \n" "vld1.8 {q14}, [%5] \n" "add %3, %0 \n" - "1: \n" + "1: \n" // d0 = 00 40 01 41 02 42 03 43 // d1 = 10 50 11 51 12 52 13 53 @@ -508,12 +508,12 @@ void ScaleAddRows_NEON(const uint8* src_ptr, int src_height) { const uint8* src_tmp; asm volatile( - "1: \n" + "1: \n" "mov %0, %1 \n" "mov r12, %5 \n" "veor q2, q2, q2 \n" "veor q3, q3, q3 \n" - "2: \n" + "2: \n" // load 16 pixels into q0 "vld1.8 {q0}, [%0], %3 \n" "vaddw.u8 q3, q3, d1 \n" @@ -535,15 +535,13 @@ void ScaleAddRows_NEON(const uint8* src_ptr, ); } -// clang-format off // TODO(Yang Zhang): Investigate less load instructions for // the x/dx stepping -#define LOAD2_DATA8_LANE(n) \ - "lsr %5, %3, #16 \n" \ - "add %6, %1, %5 \n" \ - "add %3, %3, %4 \n" \ - "vld2.8 {d6["#n"], d7["#n"]}, [%6] \n" -// clang-format on +#define LOAD2_DATA8_LANE(n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5 \n" \ + "add %3, %3, %4 \n" \ + "vld2.8 {d6[" #n "], d7[" #n "]}, [%6] \n" // The NEON version mimics this formula (from row_common.cc): // #define BLENDER(a, b, f) (uint8)((int)(a) + @@ -634,7 +632,7 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "rsb %4, #256 \n" "vdup.8 d4, %4 \n" // General purpose row blend. - "1: \n" + "1: \n" "vld1.8 {q0}, [%1]! \n" "vld1.8 {q1}, [%2]! \n" "subs %3, %3, #16 \n" @@ -649,7 +647,7 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "b 99f \n" // Blend 25 / 75. - "25: \n" + "25: \n" "vld1.8 {q0}, [%1]! \n" "vld1.8 {q1}, [%2]! \n" "subs %3, %3, #16 \n" @@ -660,7 +658,7 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "b 99f \n" // Blend 50 / 50. - "50: \n" + "50: \n" "vld1.8 {q0}, [%1]! \n" "vld1.8 {q1}, [%2]! \n" "subs %3, %3, #16 \n" @@ -670,7 +668,7 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "b 99f \n" // Blend 75 / 25. - "75: \n" + "75: \n" "vld1.8 {q1}, [%1]! \n" "vld1.8 {q0}, [%2]! \n" "subs %3, %3, #16 \n" @@ -681,13 +679,13 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "b 99f \n" // Blend 100 / 0 - Copy row unchanged. - "100: \n" + "100: \n" "vld1.8 {q0}, [%1]! \n" "subs %3, %3, #16 \n" "vst1.8 {q0}, [%0]! \n" "bgt 100b \n" - "99: \n" + "99: \n" "vst1.8 {d1[7]}, [%0] \n" : "+r"(dst_ptr), // %0 "+r"(src_ptr), // %1 @@ -719,7 +717,6 @@ void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ); } - // 46: f964 018d vld4.32 {d16,d18,d20,d22}, [r4]! // 4a: 3e04 subs r6, #4 // 4c: f964 118d vld4.32 {d17,d19,d21,d23}, [r4]! @@ -727,14 +724,13 @@ void ScaleARGBRowDown2_NEON(const uint8* src_ptr, // 54: f942 038d vst2.32 {d16-d19}, [r2]! // 58: d1f5 bne.n 46 - void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride, uint8* dst_argb, int dst_width) { (void)src_stride; asm volatile( - "1: \n" + "1: \n" "vld4.32 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. "vld4.32 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB "subs %2, %2, #8 \n" // 8 processed per loop @@ -757,25 +753,21 @@ void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, asm volatile( // change the stride to row 2 pointer "add %1, %1, %0 \n" - "1: \n" + "1: \n" "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. "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. "vld4.8 {d16, d18, d20, d22}, [%1]! \n" // load 8 more ARGB - // pixels. "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. "vpadal.u8 q2, q10 \n" // R 16 bytes -> 8 shorts. "vpadal.u8 q3, q11 \n" // A 16 bytes -> 8 shorts. - "vrshrn.u16 d0, q0, #2 \n" // downshift, round and - // pack + "vrshrn.u16 d0, q0, #2 \n" // round and pack to bytes "vrshrn.u16 d1, q1, #2 \n" "vrshrn.u16 d2, q2, #2 \n" "vrshrn.u16 d3, q3, #2 \n" @@ -799,7 +791,7 @@ void ScaleARGBRowDownEven_NEON(const uint8* src_argb, (void)src_stride; asm volatile( "mov r12, %3, lsl #2 \n" - "1: \n" + "1: \n" "vld1.32 {d0[0]}, [%0], r12 \n" "vld1.32 {d0[1]}, [%0], r12 \n" "vld1.32 {d1[0]}, [%0], r12 \n" @@ -824,9 +816,8 @@ void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, asm volatile( "mov r12, %4, lsl #2 \n" "add %1, %1, %0 \n" - "1: \n" - "vld1.8 {d0}, [%0], r12 \n" // Read 4 2x2 blocks -> - // 2x1 + "1: \n" + "vld1.8 {d0}, [%0], r12 \n" // 4 2x2 blocks -> 2x1 "vld1.8 {d1}, [%1], r12 \n" "vld1.8 {d2}, [%0], r12 \n" "vld1.8 {d3}, [%1], r12 \n" @@ -855,15 +846,13 @@ void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, : "memory", "cc", "r12", "q0", "q1", "q2", "q3"); } -// clang-format off // TODO(Yang Zhang): Investigate less load instructions for // the x/dx stepping -#define LOAD1_DATA32_LANE(dn, n) \ - "lsr %5, %3, #16 \n" \ - "add %6, %1, %5, lsl #2 \n" \ - "add %3, %3, %4 \n" \ +#define LOAD1_DATA32_LANE(dn, n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5, lsl #2 \n" \ + "add %3, %3, %4 \n" \ "vld1.32 {" #dn "[" #n "]}, [%6] \n" -// clang-format on void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb, @@ -873,15 +862,20 @@ void ScaleARGBCols_NEON(uint8* dst_argb, int tmp; const uint8* src_tmp = src_argb; asm volatile( - "1: \n" LOAD1_DATA32_LANE( - d0, 0) LOAD1_DATA32_LANE(d0, 1) LOAD1_DATA32_LANE(d1, 0) - LOAD1_DATA32_LANE(d1, 1) LOAD1_DATA32_LANE(d2, 0) LOAD1_DATA32_LANE( - d2, 1) LOAD1_DATA32_LANE(d3, 0) LOAD1_DATA32_LANE(d3, 1) - - "vst1.32 {q0, q1}, [%0]! \n" // store pixels - "subs %2, %2, #8 \n" // 8 processed per - // loop - "bgt 1b \n" + "1: \n" + // clang-format off + LOAD1_DATA32_LANE(d0, 0) + LOAD1_DATA32_LANE(d0, 1) + LOAD1_DATA32_LANE(d1, 0) + LOAD1_DATA32_LANE(d1, 1) + LOAD1_DATA32_LANE(d2, 0) + LOAD1_DATA32_LANE(d2, 1) + LOAD1_DATA32_LANE(d3, 0) + LOAD1_DATA32_LANE(d3, 1) + // clang-format on + "vst1.32 {q0, q1}, [%0]! \n" // store pixels + "subs %2, %2, #8 \n" // 8 processed per loop + "bgt 1b \n" : "+r"(dst_argb), // %0 "+r"(src_argb), // %1 "+r"(dst_width), // %2 @@ -895,15 +889,13 @@ void ScaleARGBCols_NEON(uint8* dst_argb, #undef LOAD1_DATA32_LANE -// clang-format off // TODO(Yang Zhang): Investigate less load instructions for // the x/dx stepping -#define LOAD2_DATA32_LANE(dn1, dn2, n) \ - "lsr %5, %3, #16 \n" \ - "add %6, %1, %5, lsl #2 \n" \ - "add %3, %3, %4 \n" \ - "vld2.32 {" #dn1 "[" #n "], " #dn2 "[" #n "]}, [%6] \n" -// clang-format on +#define LOAD2_DATA32_LANE(dn1, dn2, n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5, lsl #2 \n" \ + "add %3, %3, %4 \n" \ + "vld2.32 {" #dn1 "[" #n "], " #dn2 "[" #n "]}, [%6] \n" void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb, diff --git a/source/scale_neon64.cc b/source/scale_neon64.cc index 47c883eb7..c8e576883 100644 --- a/source/scale_neon64.cc +++ b/source/scale_neon64.cc @@ -27,7 +27,7 @@ void ScaleRowDown2_NEON(const uint8* src_ptr, int dst_width) { (void)src_stride; asm volatile( - "1: \n" + "1: \n" // load even pixels into v0, odd into v1 "ld2 {v0.16b,v1.16b}, [%0], #32 \n" "subs %w2, %w2, #16 \n" // 16 processed per loop @@ -48,11 +48,11 @@ void ScaleRowDown2Linear_NEON(const uint8* src_ptr, int dst_width) { (void)src_stride; asm volatile( - "1: \n" + "1: \n" // load even pixels into v0, odd into v1 "ld2 {v0.16b,v1.16b}, [%0], #32 \n" - "urhadd v0.16b, v0.16b, v1.16b \n" // rounding half add "subs %w2, %w2, #16 \n" // 16 processed per loop + "urhadd v0.16b, v0.16b, v1.16b \n" // rounding half add "st1 {v0.16b}, [%1], #16 \n" "b.gt 1b \n" : "+r"(src_ptr), // %0 @@ -71,7 +71,7 @@ void ScaleRowDown2Box_NEON(const uint8* src_ptr, asm volatile( // change the stride to row 2 pointer "add %1, %1, %0 \n" - "1: \n" + "1: \n" "ld1 {v0.16b, v1.16b}, [%0], #32 \n" // load row 1 and post inc "ld1 {v2.16b, v3.16b}, [%1], #32 \n" // load row 2 and post inc "subs %w3, %w3, #16 \n" // 16 processed per loop @@ -100,7 +100,7 @@ void ScaleRowDown4_NEON(const uint8* src_ptr, int dst_width) { (void)src_stride; asm volatile( - "1: \n" + "1: \n" "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // src line 0 "subs %w2, %w2, #8 \n" // 8 processed per loop "st1 {v2.8b}, [%1], #8 \n" @@ -120,7 +120,7 @@ void ScaleRowDown4Box_NEON(const uint8* src_ptr, const uint8* src_ptr2 = src_ptr + src_stride * 2; const uint8* src_ptr3 = src_ptr + src_stride * 3; asm volatile( - "1: \n" + "1: \n" "ld1 {v0.16b}, [%0], #16 \n" // load up 16x4 "ld1 {v1.16b}, [%2], #16 \n" "ld1 {v2.16b}, [%3], #16 \n" @@ -153,11 +153,10 @@ void ScaleRowDown34_NEON(const uint8* src_ptr, int dst_width) { (void)src_stride; asm volatile( - "1: \n" + "1: \n" "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // src line 0 "subs %w2, %w2, #24 \n" - "orr v2.16b, v3.16b, v3.16b \n" // order v0, v1, - // v2 + "orr v2.16b, v3.16b, v3.16b \n" // order v0,v1,v2 "st3 {v0.8b,v1.8b,v2.8b}, [%1], #24 \n" "b.gt 1b \n" : "+r"(src_ptr), // %0 @@ -174,7 +173,7 @@ void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr, asm volatile( "movi v20.8b, #3 \n" "add %3, %3, %0 \n" - "1: \n" + "1: \n" "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // src line 0 "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32 \n" // src line 1 "subs %w2, %w2, #24 \n" @@ -307,14 +306,14 @@ void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr, "ld1 {v30.16b}, [%6] \n" "ld1 {v31.8h}, [%7] \n" "add %2, %2, %0 \n" - "1: \n" + "1: \n" // 00 40 01 41 02 42 03 43 // 10 50 11 51 12 52 13 53 // 20 60 21 61 22 62 23 63 // 30 70 31 71 32 72 33 73 - "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" - "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n" + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n" "ld4 {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32 \n" "subs %w4, %w4, #12 \n" @@ -426,7 +425,7 @@ void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr, "ld1 {v30.8h}, [%4] \n" "ld1 {v31.16b}, [%5] \n" "add %2, %2, %0 \n" - "1: \n" + "1: \n" // 00 40 01 41 02 42 03 43 // 10 50 11 51 12 52 13 53 @@ -523,12 +522,12 @@ void ScaleAddRows_NEON(const uint8* src_ptr, int src_height) { const uint8* src_tmp; asm volatile( - "1: \n" + "1: \n" "mov %0, %1 \n" "mov w12, %w5 \n" "eor v2.16b, v2.16b, v2.16b \n" "eor v3.16b, v3.16b, v3.16b \n" - "2: \n" + "2: \n" // load 16 pixels into q0 "ld1 {v0.16b}, [%0], %3 \n" "uaddw2 v3.8h, v3.8h, v0.16b \n" @@ -550,15 +549,13 @@ void ScaleAddRows_NEON(const uint8* src_ptr, ); } -// clang-format off // TODO(Yang Zhang): Investigate less load instructions for // the x/dx stepping -#define LOAD2_DATA8_LANE(n) \ - "lsr %5, %3, #16 \n" \ - "add %6, %1, %5 \n" \ - "add %3, %3, %4 \n" \ +#define LOAD2_DATA8_LANE(n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5 \n" \ + "add %3, %3, %4 \n" \ "ld2 {v4.b, v5.b}[" #n "], [%6] \n" -// clang-format on // The NEON version mimics this formula (from row_common.cc): // #define BLENDER(a, b, f) (uint8)((int)(a) + @@ -572,8 +569,8 @@ void ScaleFilterCols_NEON(uint8* dst_ptr, int dx_offset[4] = {0, 1, 2, 3}; int* tmp = dx_offset; const uint8* src_tmp = src_ptr; - int64 x64 = (int64)x; - int64 dx64 = (int64)dx; + int64 x64 = (int64)x; // NOLINT + int64 dx64 = (int64)dx; // NOLINT asm volatile ( "dup v0.4s, %w3 \n" // x "dup v1.4s, %w4 \n" // dx @@ -651,7 +648,7 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "dup v5.8b, %w4 \n" "dup v4.8b, %w5 \n" // General purpose row blend. - "1: \n" + "1: \n" "ld1 {v0.16b}, [%1], #16 \n" "ld1 {v1.16b}, [%2], #16 \n" "subs %w3, %w3, #16 \n" @@ -666,7 +663,7 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "b 99f \n" // Blend 25 / 75. - "25: \n" + "25: \n" "ld1 {v0.16b}, [%1], #16 \n" "ld1 {v1.16b}, [%2], #16 \n" "subs %w3, %w3, #16 \n" @@ -677,7 +674,7 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "b 99f \n" // Blend 50 / 50. - "50: \n" + "50: \n" "ld1 {v0.16b}, [%1], #16 \n" "ld1 {v1.16b}, [%2], #16 \n" "subs %w3, %w3, #16 \n" @@ -687,7 +684,7 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "b 99f \n" // Blend 75 / 25. - "75: \n" + "75: \n" "ld1 {v1.16b}, [%1], #16 \n" "ld1 {v0.16b}, [%2], #16 \n" "subs %w3, %w3, #16 \n" @@ -698,13 +695,13 @@ void ScaleFilterRows_NEON(uint8* dst_ptr, "b 99f \n" // Blend 100 / 0 - Copy row unchanged. - "100: \n" + "100: \n" "ld1 {v0.16b}, [%1], #16 \n" "subs %w3, %w3, #16 \n" "st1 {v0.16b}, [%0], #16 \n" "b.gt 100b \n" - "99: \n" + "99: \n" "st1 {v0.b}[15], [%0] \n" : "+r"(dst_ptr), // %0 "+r"(src_ptr), // %1 @@ -724,7 +721,7 @@ void ScaleARGBRowDown2_NEON(const uint8* src_ptr, asm volatile( "1: \n" // load 16 ARGB pixels with even pixels into q0/q2, odd into q1/q3 - "ld4 {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n" + "ld4 {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n" "subs %w2, %w2, #8 \n" // 8 processed per loop "mov v2.16b, v3.16b \n" "st2 {v1.4s,v2.4s}, [%1], #32 \n" // store 8 odd pixels @@ -745,7 +742,7 @@ void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, asm volatile( "1: \n" // load 16 ARGB pixels with even pixels into q0/q2, odd into q1/q3 - "ld4 {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n" + "ld4 {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n" "subs %w2, %w2, #8 \n" // 8 processed per loop "urhadd v0.16b, v0.16b, v1.16b \n" // rounding half add @@ -767,17 +764,14 @@ void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, asm volatile( // change the stride to row 2 pointer "add %1, %1, %0 \n" - "1: \n" - "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 8 ARGB - // pixels. + "1: \n" + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 8 ARGB "subs %w3, %w3, #8 \n" // 8 processed per loop. "uaddlp v0.8h, v0.16b \n" // B 16 bytes -> 8 shorts. "uaddlp v1.8h, v1.16b \n" // G 16 bytes -> 8 shorts. "uaddlp v2.8h, v2.16b \n" // R 16 bytes -> 8 shorts. "uaddlp v3.8h, v3.16b \n" // A 16 bytes -> 8 shorts. "ld4 {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n" // load 8 - // more ARGB - // pixels. "uadalp v0.8h, v16.16b \n" // B 16 bytes -> 8 shorts. "uadalp v1.8h, v17.16b \n" // G 16 bytes -> 8 shorts. "uadalp v2.8h, v18.16b \n" // R 16 bytes -> 8 shorts. @@ -806,7 +800,7 @@ void ScaleARGBRowDownEven_NEON(const uint8* src_argb, int dst_width) { (void)src_stride; asm volatile( - "1: \n" + "1: \n" "ld1 {v0.s}[0], [%0], %3 \n" "ld1 {v0.s}[1], [%0], %3 \n" "ld1 {v0.s}[2], [%0], %3 \n" @@ -832,7 +826,7 @@ void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, int dst_width) { asm volatile( "add %1, %1, %0 \n" - "1: \n" + "1: \n" "ld1 {v0.8b}, [%0], %4 \n" // Read 4 2x2 blocks -> // 2x1 "ld1 {v1.8b}, [%1], %4 \n" @@ -867,15 +861,13 @@ void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"); } -// clang-format off // TODO(Yang Zhang): Investigate less load instructions for // the x/dx stepping -#define LOAD1_DATA32_LANE(vn, n) \ - "lsr %5, %3, #16 \n" \ - "add %6, %1, %5, lsl #2 \n" \ - "add %3, %3, %4 \n" \ +#define LOAD1_DATA32_LANE(vn, n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5, lsl #2 \n" \ + "add %3, %3, %4 \n" \ "ld1 {" #vn ".s}[" #n "], [%6] \n" -// clang-format on void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb, @@ -883,19 +875,25 @@ void ScaleARGBCols_NEON(uint8* dst_argb, int x, int dx) { const uint8* src_tmp = src_argb; - int64 x64 = (int64)x; - int64 dx64 = (int64)dx; + int64 x64 = (int64)x; // NOLINT + int64 dx64 = (int64)dx; // NOLINT int64 tmp64; asm volatile( - "1: \n" LOAD1_DATA32_LANE( - v0, 0) LOAD1_DATA32_LANE(v0, 1) LOAD1_DATA32_LANE(v0, 2) - LOAD1_DATA32_LANE(v0, 3) LOAD1_DATA32_LANE(v1, 0) LOAD1_DATA32_LANE( - v1, 1) LOAD1_DATA32_LANE(v1, 2) LOAD1_DATA32_LANE(v1, 3) - - "st1 {v0.4s, v1.4s}, [%0], #32 \n" // store pixels - "subs %w2, %w2, #8 \n" // 8 processed per - // loop - "b.gt 1b \n" + "1: \n" + // clang-format off + LOAD1_DATA32_LANE(v0, 0) + LOAD1_DATA32_LANE(v0, 1) + LOAD1_DATA32_LANE(v0, 2) + LOAD1_DATA32_LANE(v0, 3) + LOAD1_DATA32_LANE(v1, 0) + LOAD1_DATA32_LANE(v1, 1) + LOAD1_DATA32_LANE(v1, 2) + LOAD1_DATA32_LANE(v1, 3) + // clang-format on + "st1 {v0.4s, v1.4s}, [%0], #32 \n" // store pixels + "subs %w2, %w2, #8 \n" // 8 processed per + // loop + "b.gt 1b \n" : "+r"(dst_argb), // %0 "+r"(src_argb), // %1 "+r"(dst_width), // %2 @@ -909,15 +907,13 @@ void ScaleARGBCols_NEON(uint8* dst_argb, #undef LOAD1_DATA32_LANE -// clang-format off // TODO(Yang Zhang): Investigate less load instructions for // the x/dx stepping -#define LOAD2_DATA32_LANE(vn1, vn2, n) \ - "lsr %5, %3, #16 \n" \ - "add %6, %1, %5, lsl #2 \n" \ - "add %3, %3, %4 \n" \ +#define LOAD2_DATA32_LANE(vn1, vn2, n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5, lsl #2 \n" \ + "add %3, %3, %4 \n" \ "ld2 {" #vn1 ".s, " #vn2 ".s}[" #n "], [%6] \n" -// clang-format on void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb, @@ -927,8 +923,8 @@ void ScaleARGBFilterCols_NEON(uint8* dst_argb, int dx_offset[4] = {0, 1, 2, 3}; int* tmp = dx_offset; const uint8* src_tmp = src_argb; - int64 x64 = (int64)x; - int64 dx64 = (int64)dx; + int64 x64 = (int64)x; // NOLINT + int64 dx64 = (int64)dx; // NOLINT asm volatile ( "dup v0.4s, %w3 \n" // x "dup v1.4s, %w4 \n" // dx