diff options
Diffstat (limited to 'arm')
-rw-r--r-- | arm/arm_init.c | 21 | ||||
-rw-r--r-- | arm/filter_neon.S | 10 | ||||
-rw-r--r-- | arm/filter_neon_intrinsics.c | 372 |
3 files changed, 396 insertions, 7 deletions
diff --git a/arm/arm_init.c b/arm/arm_init.c index d26b8843c..098771781 100644 --- a/arm/arm_init.c +++ b/arm/arm_init.c @@ -3,7 +3,7 @@ * * Copyright (c) 2013 Glenn Randers-Pehrson * Written by Mans Rullgard, 2011. - * Last changed in libpng 1.6.5 [September 14, 2013] + * Last changed in libpng 1.6.6 [September 16, 2013] * * This code is released under the libpng license. * For conditions of distribution and use, see the disclaimer @@ -17,8 +17,6 @@ #include "../pngpriv.h" #ifdef PNG_READ_SUPPORTED - if (((png_ptr->options >> PNG_MAXIMUM_INFLATE_WINDOW) & 3) == - PNG_OPTION_ON) #if PNG_ARM_NEON_OPT > 0 #ifdef PNG_ARM_NEON_CHECK_SUPPORTED /* Do run-time checks */ #include <signal.h> /* for sig_atomic_t */ @@ -156,6 +154,16 @@ png_have_neon(png_structp png_ptr) void png_init_filter_functions_neon(png_structp pp, unsigned int bpp) { + /* The switch statement is compiled in for ARM_NEON_API, the call to + * png_have_neon is compiled in for ARM_NEON_CHECK. If both are defined + * the check is only performed if the API has not set the NEON option on + * or off explicitly. In this case the check controls what happens. + * + * If the CHECK is not compiled in and the option is UNSET the behavior prior + * to 1.6.7 was to use the NEON code - this was a bug caused by having the + * wrong order of the 'ON' and 'default' cases. UNSET now defaults to OFF, + * as documented in png.h + */ #ifdef PNG_ARM_NEON_API_SUPPORTED switch ((pp->options >> PNG_ARM_NEON) & 3) { @@ -180,13 +188,14 @@ png_init_filter_functions_neon(png_structp pp, unsigned int bpp) break; #endif #endif /* PNG_ARM_NEON_CHECK_SUPPORTED */ + #ifdef PNG_ARM_NEON_API_SUPPORTED + default: /* OFF or INVALID */ + return; + case PNG_OPTION_ON: /* Option turned on */ break; - - default: /* OFF or INVALID */ - return; } #endif diff --git a/arm/filter_neon.S b/arm/filter_neon.S index b8aef1053..3d1ccf505 100644 --- a/arm/filter_neon.S +++ b/arm/filter_neon.S @@ -3,7 +3,7 @@ * * Copyright (c) 2013 Glenn Randers-Pehrson * Written by Mans Rullgard, 2011. - * Last changed in libpng 1.5.17 [July 18, 2013] + * Last changed in libpng 1.6.7 [November 14, 2013] * * This code is released under the libpng license. * For conditions of distribution and use, see the disclaimer @@ -20,6 +20,13 @@ .section .note.GNU-stack,"",%progbits /* mark stack as non-executable */ #endif +/* Assembler NEON support - only works for 32-bit ARM (i.e. it does not work for + * ARM64). The code in arm/filter_neon_intrinsics.c supports ARM64, however it + * only works if -mfpu=neon is specified on the GCC command line. See pngpriv.h + * for the logic which sets PNG_USE_ARM_NEON_ASM: + */ +#if PNG_ARM_NEON_IMPLEMENTATION == 2 /* hand-coded assembler */ + #ifdef PNG_READ_SUPPORTED #if PNG_ARM_NEON_OPT > 0 @@ -235,3 +242,4 @@ func png_read_filter_row_paeth3_neon, export=1 endfunc #endif /* PNG_ARM_NEON_OPT > 0 */ #endif /* PNG_READ_SUPPORTED */ +#endif /* PNG_ARM_NEON_IMPLEMENTATION == 2 (assembler) */ diff --git a/arm/filter_neon_intrinsics.c b/arm/filter_neon_intrinsics.c new file mode 100644 index 000000000..e6a0217ab --- /dev/null +++ b/arm/filter_neon_intrinsics.c @@ -0,0 +1,372 @@ + +/* filter_neon_intrinsics.c - NEON optimised filter functions + * + * Copyright (c) 2013 Glenn Randers-Pehrson + * Written by James Yu <james.yu at linaro.org>, October 2013. + * Based on filter_neon.S, written by Mans Rullgard, 2011. + * + * Last changed in libpng 1.6.7 [November 14, 2013] + * + * This code is released under the libpng license. + * For conditions of distribution and use, see the disclaimer + * and license in png.h + */ + +#include "../pngpriv.h" + +/* This code requires -mfpu=neon on the command line: */ +#if PNG_ARM_NEON_IMPLEMENTATION == 1 /* intrinsics code */ + +#include <arm_neon.h> + +/* libpng row pointers are not necessarily aligned to any particular boundary, + * however this code will only work with appropriate alignment. arm/arm_init.c + * checks for this (and will not compile unless it is done), this code uses + * variants of png_aligncast to avoid compiler warnings. + */ +#define png_ptr(type,pointer) png_aligncast(type *,pointer) +#define png_ptrc(type,pointer) png_aligncastconst(const type *,pointer) + +/* The following relies on a variable 'temp_pointer' being declared with type + * 'type'. This is written this way just to hide the GCC strict aliasing + * warning; note that the code is safe because there never is an alias between + * the input and output pointers. + */ +#define png_ldr(type,pointer)\ + (temp_pointer = png_ptr(type,pointer), *temp_pointer) + +#ifdef PNG_READ_SUPPORTED +#if PNG_ARM_NEON_OPT > 0 + +void +png_read_filter_row_up_neon(png_row_infop row_info, png_bytep row, + png_const_bytep prev_row) +{ + png_bytep rp = row; + png_bytep rp_stop = row + row_info->rowbytes; + png_const_bytep pp = prev_row; + + for (; rp < rp_stop; rp += 16, pp += 16) + { + uint8x16_t qrp, qpp; + + qrp = vld1q_u8(rp); + qpp = vld1q_u8(pp); + qrp = vaddq_u8(qrp, qpp); + vst1q_u8(rp, qrp); + } +} + +void +png_read_filter_row_sub3_neon(png_row_infop row_info, png_bytep row, + png_const_bytep prev_row) +{ + png_bytep rp = row; + png_bytep rp_stop = row + row_info->rowbytes; + + uint8x16_t vtmp = vld1q_u8(rp); + uint8x8x2_t *vrpt = png_ptr(uint8x8x2_t, &vtmp); + uint8x8x2_t vrp = *vrpt; + + uint8x8x4_t vdest; + vdest.val[3] = vdup_n_u8(0); + + for (; rp < rp_stop;) + { + uint8x8_t vtmp1, vtmp2; + uint32x2_t *temp_pointer; + + vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3); + vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]); + vtmp2 = vext_u8(vrp.val[0], vrp.val[1], 6); + vdest.val[1] = vadd_u8(vdest.val[0], vtmp1); + + vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1); + vdest.val[2] = vadd_u8(vdest.val[1], vtmp2); + vdest.val[3] = vadd_u8(vdest.val[2], vtmp1); + + vtmp = vld1q_u8(rp + 12); + vrpt = png_ptr(uint8x8x2_t, &vtmp); + vrp = *vrpt; + + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0); + rp += 3; + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0); + rp += 3; + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0); + rp += 3; + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0); + rp += 3; + } + + PNG_UNUSED(prev_row) +} + +void +png_read_filter_row_sub4_neon(png_row_infop row_info, png_bytep row, + png_const_bytep prev_row) +{ + png_bytep rp = row; + png_bytep rp_stop = row + row_info->rowbytes; + + uint8x8x4_t vdest; + vdest.val[3] = vdup_n_u8(0); + + for (; rp < rp_stop; rp += 16) + { + uint32x2x4_t vtmp = vld4_u32(png_ptr(uint32_t,rp)); + uint8x8x4_t *vrpt = png_ptr(uint8x8x4_t,&vtmp); + uint8x8x4_t vrp = *vrpt; + uint32x2x4_t *temp_pointer; + + vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]); + vdest.val[1] = vadd_u8(vdest.val[0], vrp.val[1]); + vdest.val[2] = vadd_u8(vdest.val[1], vrp.val[2]); + vdest.val[3] = vadd_u8(vdest.val[2], vrp.val[3]); + vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0); + } + + PNG_UNUSED(prev_row) +} + +void +png_read_filter_row_avg3_neon(png_row_infop row_info, png_bytep row, + png_const_bytep prev_row) +{ + png_bytep rp = row; + png_const_bytep pp = prev_row; + png_bytep rp_stop = row + row_info->rowbytes; + + uint8x16_t vtmp; + uint8x8x2_t *vrpt; + uint8x8x2_t vrp; + uint8x8x4_t vdest; + vdest.val[3] = vdup_n_u8(0); + + vtmp = vld1q_u8(rp); + vrpt = png_ptr(uint8x8x2_t,&vtmp); + vrp = *vrpt; + + for (; rp < rp_stop; pp += 12) + { + uint8x8_t vtmp1, vtmp2, vtmp3; + + uint8x8x2_t *vppt; + uint8x8x2_t vpp; + + uint32x2_t *temp_pointer; + + vtmp = vld1q_u8(pp); + vppt = png_ptr(uint8x8x2_t,&vtmp); + vpp = *vppt; + + vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3); + vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]); + vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]); + + vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3); + vtmp3 = vext_u8(vrp.val[0], vrp.val[1], 6); + vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2); + vdest.val[1] = vadd_u8(vdest.val[1], vtmp1); + + vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 6); + vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1); + + vtmp = vld1q_u8(rp + 12); + vrpt = png_ptr(uint8x8x2_t,&vtmp); + vrp = *vrpt; + + vdest.val[2] = vhadd_u8(vdest.val[1], vtmp2); + vdest.val[2] = vadd_u8(vdest.val[2], vtmp3); + + vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1); + + vdest.val[3] = vhadd_u8(vdest.val[2], vtmp2); + vdest.val[3] = vadd_u8(vdest.val[3], vtmp1); + + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0); + rp += 3; + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0); + rp += 3; + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0); + rp += 3; + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0); + rp += 3; + } +} + +void +png_read_filter_row_avg4_neon(png_row_infop row_info, png_bytep row, + png_const_bytep prev_row) +{ + png_bytep rp = row; + png_bytep rp_stop = row + row_info->rowbytes; + png_const_bytep pp = prev_row; + + uint8x8x4_t vdest; + vdest.val[3] = vdup_n_u8(0); + + for (; rp < rp_stop; rp += 16, pp += 16) + { + uint32x2x4_t vtmp; + uint8x8x4_t *vrpt, *vppt; + uint8x8x4_t vrp, vpp; + uint32x2x4_t *temp_pointer; + + vtmp = vld4_u32(png_ptr(uint32_t,rp)); + vrpt = png_ptr(uint8x8x4_t,&vtmp); + vrp = *vrpt; + vtmp = vld4_u32(png_ptrc(uint32_t,pp)); + vppt = png_ptr(uint8x8x4_t,&vtmp); + vpp = *vppt; + + vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]); + vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]); + vdest.val[1] = vhadd_u8(vdest.val[0], vpp.val[1]); + vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]); + vdest.val[2] = vhadd_u8(vdest.val[1], vpp.val[2]); + vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]); + vdest.val[3] = vhadd_u8(vdest.val[2], vpp.val[3]); + vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]); + + vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0); + } +} + +static uint8x8_t +paeth(uint8x8_t a, uint8x8_t b, uint8x8_t c) +{ + uint8x8_t d, e; + uint16x8_t p1, pa, pb, pc; + + p1 = vaddl_u8(a, b); /* a + b */ + pc = vaddl_u8(c, c); /* c * 2 */ + pa = vabdl_u8(b, c); /* pa */ + pb = vabdl_u8(a, c); /* pb */ + pc = vabdq_u16(p1, pc); /* pc */ + + p1 = vcleq_u16(pa, pb); /* pa <= pb */ + pa = vcleq_u16(pa, pc); /* pa <= pc */ + pb = vcleq_u16(pb, pc); /* pb <= pc */ + + p1 = vandq_u16(p1, pa); /* pa <= pb && pa <= pc */ + + d = vmovn_u16(pb); + e = vmovn_u16(p1); + + d = vbsl_u8(d, b, c); + e = vbsl_u8(e, a, d); + + return e; +} + +void +png_read_filter_row_paeth3_neon(png_row_infop row_info, png_bytep row, + png_const_bytep prev_row) +{ + png_bytep rp = row; + png_const_bytep pp = prev_row; + png_bytep rp_stop = row + row_info->rowbytes; + + uint8x16_t vtmp; + uint8x8x2_t *vrpt; + uint8x8x2_t vrp; + uint8x8_t vlast = vdup_n_u8(0); + uint8x8x4_t vdest; + vdest.val[3] = vdup_n_u8(0); + + vtmp = vld1q_u8(rp); + vrpt = png_ptr(uint8x8x2_t,&vtmp); + vrp = *vrpt; + + for (; rp < rp_stop; pp += 12) + { + uint8x8x2_t *vppt; + uint8x8x2_t vpp; + uint8x8_t vtmp1, vtmp2, vtmp3; + uint32x2_t *temp_pointer; + + vtmp = vld1q_u8(pp); + vppt = png_ptr(uint8x8x2_t,&vtmp); + vpp = *vppt; + + vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast); + vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]); + + vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3); + vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3); + vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]); + vdest.val[1] = vadd_u8(vdest.val[1], vtmp1); + + vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 6); + vtmp3 = vext_u8(vpp.val[0], vpp.val[1], 6); + vdest.val[2] = paeth(vdest.val[1], vtmp3, vtmp2); + vdest.val[2] = vadd_u8(vdest.val[2], vtmp1); + + vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1); + vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1); + + vtmp = vld1q_u8(rp + 12); + vrpt = png_ptr(uint8x8x2_t,&vtmp); + vrp = *vrpt; + + vdest.val[3] = paeth(vdest.val[2], vtmp2, vtmp3); + vdest.val[3] = vadd_u8(vdest.val[3], vtmp1); + + vlast = vtmp2; + + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0); + rp += 3; + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0); + rp += 3; + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0); + rp += 3; + vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0); + rp += 3; + } +} + +void +png_read_filter_row_paeth4_neon(png_row_infop row_info, png_bytep row, + png_const_bytep prev_row) +{ + png_bytep rp = row; + png_bytep rp_stop = row + row_info->rowbytes; + png_const_bytep pp = prev_row; + + uint8x8_t vlast = vdup_n_u8(0); + uint8x8x4_t vdest; + vdest.val[3] = vdup_n_u8(0); + + for (; rp < rp_stop; rp += 16, pp += 16) + { + uint32x2x4_t vtmp; + uint8x8x4_t *vrpt, *vppt; + uint8x8x4_t vrp, vpp; + uint32x2x4_t *temp_pointer; + + vtmp = vld4_u32(png_ptr(uint32_t,rp)); + vrpt = png_ptr(uint8x8x4_t,&vtmp); + vrp = *vrpt; + vtmp = vld4_u32(png_ptrc(uint32_t,pp)); + vppt = png_ptr(uint8x8x4_t,&vtmp); + vpp = *vppt; + + vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast); + vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]); + vdest.val[1] = paeth(vdest.val[0], vpp.val[1], vpp.val[0]); + vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]); + vdest.val[2] = paeth(vdest.val[1], vpp.val[2], vpp.val[1]); + vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]); + vdest.val[3] = paeth(vdest.val[2], vpp.val[3], vpp.val[2]); + vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]); + + vlast = vpp.val[3]; + + vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0); + } +} + +#endif /* PNG_ARM_NEON_OPT > 0 */ +#endif /* PNG_READ_SUPPORTED */ +#endif /* PNG_ARM_NEON_IMPLEMENTATION == 1 (intrinsics) */ |