diff options
| author | Simon Hosie <simon.hosie@arm.com> | 2014-05-01 23:28:45 -0700 |
|---|---|---|
| committer | Simon Hosie <simon.hosie@arm.com> | 2014-05-07 15:03:01 -0700 |
| commit | 1d9c887c58d115975e01c9d500595f503803dc8c (patch) | |
| tree | 7238184201d2cca382d88f93fd26c68c9ed3f817 | |
| parent | 9bd08e84ba2c9bd0708de5600877162126c4467c (diff) | |
| download | rs-1d9c887c58d115975e01c9d500595f503803dc8c.tar.gz | |
YuvToRGB sub-rectangle handling.
Fix some difficult edge cases when processing only a portion of the image.
Also fix a register-marshalling bug in AArch64 assembly.
Change-Id: I8cd67f394fb42b216b2c3c7401e90eb2b86fca3d
| -rw-r--r-- | cpu_ref/rsCpuIntrinsicYuvToRGB.cpp | 17 | ||||
| -rw-r--r-- | cpu_ref/rsCpuIntrinsics_advsimd_YuvToRGB.S | 6 | ||||
| -rw-r--r-- | cpu_ref/rsCpuIntrinsics_neon_YuvToRGB.S | 4 |
3 files changed, 18 insertions, 9 deletions
diff --git a/cpu_ref/rsCpuIntrinsicYuvToRGB.cpp b/cpu_ref/rsCpuIntrinsicYuvToRGB.cpp index 2d905de3..f65d6aa2 100644 --- a/cpu_ref/rsCpuIntrinsicYuvToRGB.cpp +++ b/cpu_ref/rsCpuIntrinsicYuvToRGB.cpp @@ -123,7 +123,7 @@ void RsdCpuScriptIntrinsicYuvToRGB::kernel(const RsForEachStubParamStruct *p, } const uchar *Y = pinY + (p->y * strideY); - uchar4 *out = (uchar4 *)p->out; + uchar4 *out = (uchar4 *)p->out + xstart; uint32_t x1 = xstart; uint32_t x2 = xend; @@ -152,11 +152,20 @@ void RsdCpuScriptIntrinsicYuvToRGB::kernel(const RsForEachStubParamStruct *p, cstep = 2; } + /* If we start on an odd pixel then deal with it here and bump things along + * so that subsequent code can carry on with even-odd pairing assumptions. + */ + if((x1 & 1) && (x2 > x1)) { + int cx = (x1 >> 1) * cstep; + *out = rsYuvToRGBA_uchar4(Y[x1], u[cx], v[cx]); + out++; + x1++; + } #if defined(ARCH_ARM_HAVE_VFP) if((x2 > x1) && gArchUseSIMD) { int32_t len = x2 - x1; if (cstep == 1) { - rsdIntrinsicYuv2_K(out, Y, u, v, x1, x2); + rsdIntrinsicYuv2_K(p->out, Y, u, v, x1, x2); x1 += len; out += len; } else if (cstep == 2) { @@ -165,11 +174,11 @@ void RsdCpuScriptIntrinsicYuvToRGB::kernel(const RsForEachStubParamStruct *p, intptr_t ipv = (intptr_t)v; if (ipu == (ipv + 1)) { - rsdIntrinsicYuv_K(out, Y, v, x1, x2); + rsdIntrinsicYuv_K(p->out, Y, v, x1, x2); x1 += len; out += len; } else if (ipu == (ipv - 1)) { - rsdIntrinsicYuvR_K(out, Y, u, x1, x2); + rsdIntrinsicYuvR_K(p->out, Y, u, x1, x2); x1 += len; out += len; } diff --git a/cpu_ref/rsCpuIntrinsics_advsimd_YuvToRGB.S b/cpu_ref/rsCpuIntrinsics_advsimd_YuvToRGB.S index 9232a796..632ef7a4 100644 --- a/cpu_ref/rsCpuIntrinsics_advsimd_YuvToRGB.S +++ b/cpu_ref/rsCpuIntrinsics_advsimd_YuvToRGB.S @@ -165,7 +165,7 @@ 1: tbz x2, #0, 1f ld1 {v8.b}[1], [x1], #1 .if \interleaved - ld1 {v16.b}[1], [x3], #1 + ld1 {v16.h}[0], [x3], #2 .else ld1 {v16.b}[0], [x3], #1 ld1 {v17.b}[0], [x4], #1 @@ -247,7 +247,7 @@ END(rsdIntrinsicYuv2_K) * size_t xend); // x4 */ ENTRY(rsdIntrinsicYuv_K) - bic x5, x4, #1 + bic x5, x3, #1 add x0, x0, x5, LSL #2 add x1, x1, x5 add x3, x2, x5 @@ -273,7 +273,7 @@ END(rsdIntrinsicYuv_K) * size_t xend); // x4 */ ENTRY(rsdIntrinsicYuvR_K) - bic x5, x4, #1 + bic x5, x3, #1 add x0, x0, x5, LSL #2 add x1, x1, x5 add x3, x2, x5 diff --git a/cpu_ref/rsCpuIntrinsics_neon_YuvToRGB.S b/cpu_ref/rsCpuIntrinsics_neon_YuvToRGB.S index da4cdedb..5c3bce41 100644 --- a/cpu_ref/rsCpuIntrinsics_neon_YuvToRGB.S +++ b/cpu_ref/rsCpuIntrinsics_neon_YuvToRGB.S @@ -117,7 +117,7 @@ \kernel - subs r2, #16 + subs r2, #16 vst4.u8 {d0,d2,d4,d6}, [r0]! vst4.u8 {d1,d3,d5,d7}, [r0]! @@ -169,7 +169,7 @@ beq 1f vld1.u8 d16[1], [r1]! .if \interleaved - vld1.u8 d20[1], [r3]! + vld1.u16 d20[0], [r3]! .else vld1.u8 d20[0], [r3]! vld1.u8 d21[0], [r4]! |