diff options
| author | Jukka Ojanen <jukka.ojanen@linkotec.net> | 2016-03-28 13:57:47 +0300 |
|---|---|---|
| committer | Jukka Ojanen <jukka.ojanen@linkotec.net> | 2016-03-28 13:57:47 +0300 |
| commit | 2cf68165a461a9faf7069e094436d16c22990aff (patch) | |
| tree | f91cf852ae9ac06d12ae84f64c66df6d5cb22385 /src/ffts_nd.c | |
| parent | a62bc084cad3ea82128c8060b17a488f2fce2587 (diff) | |
| download | ffts-2cf68165a461a9faf7069e094436d16c22990aff.zip ffts-2cf68165a461a9faf7069e094436d16c22990aff.tar.gz | |
Improve performance of small complex 2D Neon transform by 15%
Diffstat (limited to 'src/ffts_nd.c')
| -rw-r--r-- | src/ffts_nd.c | 69 |
1 files changed, 6 insertions, 63 deletions
diff --git a/src/ffts_nd.c b/src/ffts_nd.c index 72e21e7..2bde9c4 100644 --- a/src/ffts_nd.c +++ b/src/ffts_nd.c @@ -92,74 +92,17 @@ static void ffts_free_nd(ffts_plan_t *p) static void ffts_transpose(uint64_t *in, uint64_t *out, int w, int h, uint64_t *buf) { #ifdef HAVE_NEON - size_t i, j, k; - int linebytes = 8 * w; +#if 0 + neon_transpose(in, out, w, h); +#else + size_t i, j; for (j = 0; j < h; j += 8) { for (i = 0; i < w; i += 8) { - neon_transpose_to_buf(in + j*w + i, buf, w); - - uint64_t *p = out + i*h + j; - uint64_t *pbuf = buf; - uint64_t *ptemp; - - __asm__ __volatile__( - "mov %[ptemp], %[p]\n\t" - "add %[p], %[p], %[w], lsl #3\n\t" - "vld1.32 {q8,q9}, [%[pbuf], :128]!\n\t" - "vld1.32 {q10,q11}, [%[pbuf], :128]!\n\t" - "vld1.32 {q12,q13}, [%[pbuf], :128]!\n\t" - "vld1.32 {q14,q15}, [%[pbuf], :128]!\n\t" - "vst1.32 {q8,q9}, [%[ptemp], :128]!\n\t" - "vst1.32 {q10,q11}, [%[ptemp], :128]!\n\t" - "mov %[ptemp], %[p]\n\t" - "add %[p], %[p], %[w], lsl #3\n\t" - "vst1.32 {q12,q13}, [%[ptemp], :128]!\n\t" - "vst1.32 {q14,q15}, [%[ptemp], :128]!\n\t" - "mov %[ptemp], %[p]\n\t" - "add %[p], %[p], %[w], lsl #3\n\t" - "vld1.32 {q8,q9}, [%[pbuf], :128]!\n\t" - "vld1.32 {q10,q11}, [%[pbuf], :128]!\n\t" - "vld1.32 {q12,q13}, [%[pbuf], :128]!\n\t" - "vld1.32 {q14,q15}, [%[pbuf], :128]!\n\t" - "vst1.32 {q8,q9}, [%[ptemp], :128]!\n\t" - "vst1.32 {q10,q11}, [%[ptemp], :128]!\n\t" - "mov %[ptemp], %[p]\n\t" - "add %[p], %[p], %[w], lsl #3\n\t" - "vst1.32 {q12,q13}, [%[ptemp], :128]!\n\t" - "vst1.32 {q14,q15}, [%[ptemp], :128]!\n\t" - "mov %[ptemp], %[p]\n\t" - "add %[p], %[p], %[w], lsl #3\n\t" - "vld1.32 {q8,q9}, [%[pbuf], :128]!\n\t" - "vld1.32 {q10,q11}, [%[pbuf], :128]!\n\t" - "vld1.32 {q12,q13}, [%[pbuf], :128]!\n\t" - "vld1.32 {q14,q15}, [%[pbuf], :128]!\n\t" - "vst1.32 {q8,q9}, [%[ptemp], :128]!\n\t" - "vst1.32 {q10,q11}, [%[ptemp], :128]!\n\t" - "mov %[ptemp], %[p]\n\t" - "add %[p], %[p], %[w], lsl #3\n\t" - "vst1.32 {q12,q13}, [%[ptemp], :128]!\n\t" - "vst1.32 {q14,q15}, [%[ptemp], :128]!\n\t" - "mov %[ptemp], %[p]\n\t" - "add %[p], %[p], %[w], lsl #3\n\t" - "vld1.32 {q8,q9}, [%[pbuf], :128]!\n\t" - "vld1.32 {q10,q11}, [%[pbuf], :128]!\n\t" - "vld1.32 {q12,q13}, [%[pbuf], :128]!\n\t" - "vld1.32 {q14,q15}, [%[pbuf], :128]!\n\t" - "vst1.32 {q8,q9}, [%[ptemp], :128]!\n\t" - "vst1.32 {q10,q11}, [%[ptemp], :128]!\n\t" - "mov %[ptemp], %[p]\n\t" - "vst1.32 {q12,q13}, [%[ptemp], :128]!\n\t" - "vst1.32 {q14,q15}, [%[ptemp], :128]!\n\t" - - : [p] "+r" (p), [pbuf] "+r" (pbuf), [ptemp] "+r" (ptemp) - : [w] "r" (w) - : "memory", "q8", "q9", "q10", "q11" - ); - - /* out[i*h + j] = in[j*w + i]; */ + neon_transpose_to_buf(in + j*w + i, out + i*h + j, w); } } +#endif #else #ifdef HAVE_SSE uint64_t FFTS_ALIGN(64) tmp[TSIZE*TSIZE]; |
