[FFmpeg-devel] [PATCH 6/9] x86: simple_idct: 12bits versions

[Christophe Gisquet]

On 12 frames of a 444p 12 bits DNxHR sequence, _put function:
C:         78902 decicycles in idct,  262071 runs,     73 skips
avx:       32478 decicycles in idct,  262045 runs,     99 skips

Difference between the 2:
stddev:    0.39 PSNR:104.47 MAXDIFF:    2

This is unavoidable and due to the scale factors used in the x86
version, which cannot match the C ones, as this would cause
overflows (there's one less 1bit of precision). In particular,
the trick to merge an addition into a multiplication of the first
butterfly of the pass can cause an overflow (15bits coeff now
needing 16).
---
 libavcodec/x86/idctdsp_init.c    | 21 +++++++++++++++++++--
 libavcodec/x86/simple_idct.h     |  6 ++++++
 libavcodec/x86/simple_idct10.asm | 17 +++++++++++++++++
 3 files changed, 42 insertions(+), 2 deletions(-)

diff --git a/libavcodec/x86/idctdsp_init.c b/libavcodec/x86/idctdsp_init.c
index 4fc9b0d..be563c2 100644
--- a/libavcodec/x86/idctdsp_init.c
+++ b/libavcodec/x86/idctdsp_init.c
@@ -86,8 +86,8 @@ av_cold void ff_idctdsp_init_x86(IDCTDSPContext *c, AVCodecContext *avctx,
         c->add_pixels_clamped        = ff_add_pixels_clamped_sse2;
     }
 
-    if (ARCH_X86_64 &&
-        avctx->bits_per_raw_sample == 10 && avctx->lowres == 0 &&
+    if (ARCH_X86_64 && avctx->lowres == 0) {
+        if (avctx->bits_per_raw_sample == 10 &&
         (avctx->idct_algo == FF_IDCT_AUTO ||
          avctx->idct_algo == FF_IDCT_SIMPLEAUTO ||
          avctx->idct_algo == FF_IDCT_SIMPLE)) {
@@ -102,5 +102,22 @@ av_cold void ff_idctdsp_init_x86(IDCTDSPContext *c, AVCodecContext *avctx,
             c->idct_put  = ff_simple_idct10_put_avx;
             c->perm_type = FF_IDCT_PERM_TRANSPOSE;
         }
+        }
+
+        if (avctx->bits_per_raw_sample == 12 &&
+            (avctx->idct_algo == FF_IDCT_AUTO ||
+             avctx->idct_algo == FF_IDCT_SIMPLEMMX)) {
+            if (EXTERNAL_SSE2(cpu_flags)) {
+                c->idct      = ff_simple_idct12_sse2;
+                c->idct_put  = ff_simple_idct12_put_sse2;
+                c->perm_type = FF_IDCT_PERM_TRANSPOSE;
+
+            }
+            if (EXTERNAL_AVX(cpu_flags)) {
+                c->idct      = ff_simple_idct12_avx;
+                c->idct_put  = ff_simple_idct12_put_avx;
+                c->perm_type = FF_IDCT_PERM_TRANSPOSE;
+            }
+        }
     }
 }
diff --git a/libavcodec/x86/simple_idct.h b/libavcodec/x86/simple_idct.h
index e8f59c1..8eeb31e 100644
--- a/libavcodec/x86/simple_idct.h
+++ b/libavcodec/x86/simple_idct.h
@@ -31,4 +31,10 @@ void ff_simple_idct10_avx(int16_t *block);
 void ff_simple_idct10_put_sse2(uint8_t *dest, int line_size, int16_t *block);
 void ff_simple_idct10_put_avx(uint8_t *dest, int line_size, int16_t *block);
 
+void ff_simple_idct12_sse2(int16_t *block);
+void ff_simple_idct12_avx(int16_t *block);
+
+void ff_simple_idct12_put_sse2(uint8_t *dest, int line_size, int16_t *block);
+void ff_simple_idct12_put_avx(uint8_t *dest, int line_size, int16_t *block);
+
 #endif /* AVCODEC_X86_SIMPLE_IDCT_H */
diff --git a/libavcodec/x86/simple_idct10.asm b/libavcodec/x86/simple_idct10.asm
index b1f45ea..a410191 100644
--- a/libavcodec/x86/simple_idct10.asm
+++ b/libavcodec/x86/simple_idct10.asm
@@ -29,9 +29,13 @@
 
 SECTION_RODATA
 
+cextern pw_1
+cextern pw_2
 cextern pw_8
 cextern pw_1023
+cextern pw_4095
 pd_round: times 4 dd 1<<(13-1)
+pd_round2: times 4 dd 1<<(15-1)
 
 %include "libavcodec/x86/simple_idct10_template.asm"
 
@@ -47,6 +51,19 @@ cglobal simple_idct10_put, 3, 3, 16
     mova          m15, [pd_round]
     IDCT_FN    "", 13, pw_8, 18, 0, pw_1023
     RET
+
+cglobal simple_idct12, 1, 1, 16
+    mova          m15, [pd_round2]
+    IDCT_FN    "", 15, pw_2, 16
+    RET
+
+cglobal simple_idct12_put, 3, 3, 16
+    ; range isn't known, so the C simple_idct range is used
+    ; Also, using a bias on input overflows, so use the bias
+    ; on output of the first butterfly instead
+    mova          m15, [pd_round2]
+    IDCT_FN    "", 15, pw_2, 16, 0, pw_4095
+    RET
 %endmacro
 
 INIT_XMM sse2
-- 
2.6.0