diff options
Diffstat (limited to 'source/row_neon64.cc')
| -rw-r--r-- | source/row_neon64.cc | 120 |
1 files changed, 120 insertions, 0 deletions
diff --git a/source/row_neon64.cc b/source/row_neon64.cc index acefd96d..941c9b98 100644 --- a/source/row_neon64.cc +++ b/source/row_neon64.cc @@ -3526,6 +3526,126 @@ void HalfMergeUVRow_NEON(const uint8_t* src_u, : "cc", "memory", "v0", "v1", "v2", "v3"); } +void SplitUVRow_16_NEON(const uint16_t* src_uv, + uint16_t* dst_u, + uint16_t* dst_v, + int depth, + int width) { + asm volatile( + "dup v0.4s, %w3 \n" + "1: \n" + "ld2 {v1.8h, v2.8h}, [%0], #32 \n" // load 8 UV + "prfm pldl1keep, [%0, 448] \n" + "ushll v3.4s, v1.4h, #0 \n" + "ushll2 v4.4s, v1.8h, #0 \n" + "ushl v3.4s, v3.4s, v0.4s \n" + "ushl v4.4s, v4.4s, v0.4s \n" + "xtn v1.4h, v3.4s \n" + "xtn2 v1.8h, v4.4s \n" + "ushll v3.4s, v2.4h, #0 \n" + "ushll2 v4.4s, v2.8h, #0 \n" + "ushl v3.4s, v3.4s, v0.4s \n" + "ushl v4.4s, v4.4s, v0.4s \n" + "xtn v2.4h, v3.4s \n" + "xtn2 v2.8h, v4.4s \n" + "subs %w4, %w4, #8 \n" // 8 src pixels per loop + "st1 {v1.8h}, [%1], #16 \n" // store 8 U pixels + "st1 {v2.8h}, [%2], #16 \n" // store 8 V pixels + "b.gt 1b \n" + : "+r"(src_uv), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(depth), // %3 + "+r"(width) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4"); +} + +void MergeUVRow_16_NEON(const uint16_t* src_u, + const uint16_t* src_v, + uint16_t* dst_uv, + int depth, + int width) { + int shift = 16 - depth; + asm volatile( + "dup v2.8h, %w3 \n" + "1: \n" + "ld1 {v0.8h}, [%0], #16 \n" // load 8 U + "prfm pldl1keep, [%0, 448] \n" + "ld1 {v1.8h}, [%1], #16 \n" // load 8 V + "prfm pldl1keep, [%1, 448] \n" + "ushl v0.8h, v0.8h, v2.8h \n" + "ushl v1.8h, v1.8h, v2.8h \n" + "subs %w4, %w4, #8 \n" // 8 src pixels per loop + "st2 {v0.8h, v1.8h}, [%2], #32 \n" // store 8 UV pixels + "b.gt 1b \n" + : "+r"(src_u), // %0 + "+r"(src_v), // %1 + "+r"(dst_uv), // %2 + "+r"(shift), // %3 + "+r"(width) // %4 + : + : "cc", "memory", "v0", "v1", "v2"); +} + +void MultiplyRow_16_NEON(const uint16_t* src_y, + uint16_t* dst_y, + int scale, + int width) { + asm volatile( + "dup v2.8h, %w2 \n" + "1: \n" + "ldp q0, q1, [%0] \n" + "add %0, %0, #32 \n" + "prfm pldl1keep, [%0, 448] \n" + "mul v0.8h, v0.8h, v2.8h \n" + "mul v1.8h, v1.8h, v2.8h \n" + "stp q0, q1, [%1] \n" // store 16 pixels + "add %1, %1, #32 \n" + "subs %w3, %w3, #16 \n" // 16 src pixels per loop + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(dst_y), // %1 + "+r"(scale), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "v0", "v1", "v2"); +} + +void DivideRow_16_NEON(const uint16_t* src_y, + uint16_t* dst_y, + int scale, + int width) { + asm volatile( + "dup v0.8h, %w2 \n" + "1: \n" + "ldp q1, q2, [%0] \n" + "add %0, %0, #32 \n" + "prfm pldl1keep, [%0, 448] \n" + "ushll v3.4s, v1.4h, #0 \n" + "ushll v4.4s, v2.4h, #0 \n" + "ushll2 v1.4s, v1.8h, #0 \n" + "ushll2 v2.4s, v2.8h, #0 \n" + "mul v3.4s, v0.4s, v3.4s \n" + "mul v4.4s, v0.4s, v4.4s \n" + "mul v1.4s, v0.4s, v1.4s \n" + "mul v2.4s, v0.4s, v2.4s \n" + "shrn v3.4h, v3.4s, #16 \n" + "shrn v4.4h, v4.4s, #16 \n" + "shrn2 v3.8h, v1.4s, #16 \n" + "shrn2 v4.8h, v2.4s, #16 \n" + "stp q3, q3, [%1] \n" // store 16 pixels + "add %1, %1, #32 \n" + "subs %w3, %w3, #16 \n" // 16 src pixels per loop + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(dst_y), // %1 + "+r"(scale), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4"); +} + #endif // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) #ifdef __cplusplus |