[x265] [PATCH 2/3] AArch64: Refactor loop unroll pragmas in dct primitives

[chen]

In my view, compiler option is not good idea, unroll these loop in manual are more better, it helpful us find out optimzie point and improve algorithm future.

For example, split DCT into deep Even part (EE/EEE, etc) does not get good performance




At 2024-08-21 01:42:12, "Hari Limaye" <hari.limaye at arm.com> wrote:

>Make #pragma unroll directives portable for Clang and GCC, as currently
>GCC will simply ignore the unsupported directives.
>---
> source/common/aarch64/dct-prim.cpp | 36 ++++++++++++++++++------------
> 1 file changed, 22 insertions(+), 14 deletions(-)
>
>diff --git a/source/common/aarch64/dct-prim.cpp b/source/common/aarch64/dct-prim.cpp
>index 416532e54..acc50d4f4 100644
>--- a/source/common/aarch64/dct-prim.cpp
>+++ b/source/common/aarch64/dct-prim.cpp
>@@ -5,6 +5,14 @@
> 
> #include <arm_neon.h>
> 
>+#define X265_PRAGMA(text)       _Pragma(#text)
>+#if defined(__clang__)
>+#define X265_PRAGMA_UNROLL(n)   X265_PRAGMA(unroll(n))
>+#elif defined(__GNUC__)
>+#define X265_PRAGMA_UNROLL(n)   X265_PRAGMA(GCC unroll (n))
>+#else
>+#define X265_PRAGMA_UNROLL(n)
>+#endif
> 
> namespace
> {
>@@ -472,12 +480,12 @@ static void partialButterflyInverse16_neon(const int16_t *src, int16_t *orig_dst
>     const int add = 1 << (shift - 1);
> 
> 
>-#pragma unroll(4)
>+X265_PRAGMA_UNROLL(4)
>     for (j = 0; j < line; j += 4)
>     {
>         /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */
> 
>-#pragma unroll(2)
>+X265_PRAGMA_UNROLL(2)
>         for (k = 0; k < 2; k++)
>         {
>             int32x4_t s;
>@@ -496,7 +504,7 @@ static void partialButterflyInverse16_neon(const int16_t *src, int16_t *orig_dst
>         EE[3] = vsubq_s32(EEE[0] , EEO[0]);
> 
> 
>-#pragma unroll(1)
>+X265_PRAGMA_UNROLL(1)
>         for (k = 0; k < 4; k += 4)
>         {
>             int32x4_t s[4];
>@@ -522,14 +530,14 @@ static void partialButterflyInverse16_neon(const int16_t *src, int16_t *orig_dst
>         static const int32x4_t max = vdupq_n_s32(32767);
>         const int32x4_t minus_shift = vdupq_n_s32(-shift);
> 
>-#pragma unroll(4)
>+X265_PRAGMA_UNROLL(4)
>         for (k = 0; k < 4; k++)
>         {
>             E[k] = vaddq_s32(EE[k] , EO[k]);
>             E[k + 4] = vsubq_s32(EE[3 - k] , EO[3 - k]);
>         }
> 
>-#pragma unroll(2)
>+X265_PRAGMA_UNROLL(2)
>         for (k = 0; k < 8; k += 4)
>         {
>             int32x4_t s[4];
>@@ -584,7 +592,7 @@ static void partialButterflyInverse16_neon(const int16_t *src, int16_t *orig_dst
>         }
> 
> 
>-#pragma unroll(2)
>+X265_PRAGMA_UNROLL(2)
>         for (k = 0; k < 8; k += 4)
>         {
>             int32x4_t t;
>@@ -657,10 +665,10 @@ static void partialButterflyInverse32_neon(const int16_t *src, int16_t *orig_dst
>     int16x4_t dst[32];
>     int add = 1 << (shift - 1);
> 
>-#pragma unroll (8)
>+X265_PRAGMA_UNROLL(8)
>     for (j = 0; j < line; j += 4)
>     {
>-#pragma unroll (4)
>+X265_PRAGMA_UNROLL(4)
>         for (k = 0; k < 16; k += 4)
>         {
>             int32x4_t s[4];
>@@ -681,7 +689,7 @@ static void partialButterflyInverse32_neon(const int16_t *src, int16_t *orig_dst
>         }
> 
> 
>-#pragma unroll (2)
>+X265_PRAGMA_UNROLL(2)
>         for (k = 0; k < 8; k += 4)
>         {
>             int32x4_t s[4];
>@@ -721,7 +729,7 @@ static void partialButterflyInverse32_neon(const int16_t *src, int16_t *orig_dst
>             EEO[k + 3] = s[3];
>         }
> 
>-#pragma unroll (2)
>+X265_PRAGMA_UNROLL(2)
>         for (k = 0; k < 2; k++)
>         {
>             int32x4_t s;
>@@ -736,14 +744,14 @@ static void partialButterflyInverse32_neon(const int16_t *src, int16_t *orig_dst
>         EEE[1] = vaddq_s32(EEEE[1], EEEO[1]);
>         EEE[2] = vsubq_s32(EEEE[1], EEEO[1]);
> 
>-#pragma unroll (4)
>+X265_PRAGMA_UNROLL(4)
>         for (k = 0; k < 4; k++)
>         {
>             EE[k] = vaddq_s32(EEE[k], EEO[k]);
>             EE[k + 4] = vsubq_s32((EEE[3 - k]), (EEO[3 - k]));
>         }
> 
>-#pragma unroll (8)
>+X265_PRAGMA_UNROLL(8)
>         for (k = 0; k < 8; k++)
>         {
>             E[k] = vaddq_s32(EE[k], EO[k]);
>@@ -755,7 +763,7 @@ static void partialButterflyInverse32_neon(const int16_t *src, int16_t *orig_dst
> 
> 
> 
>-#pragma unroll (16)
>+X265_PRAGMA_UNROLL(16)
>         for (k = 0; k < 16; k++)
>         {
>             int32x4_t adde = vaddq_s32(vdupq_n_s32(add), E[k]);
>@@ -777,7 +785,7 @@ static void partialButterflyInverse32_neon(const int16_t *src, int16_t *orig_dst
>         }
> 
> 
>-#pragma unroll (8)
>+X265_PRAGMA_UNROLL(8)
>         for (k = 0; k < 32; k += 4)
>         {
>             int16x4_t x0 = dst[k + 0];
>-- 
>2.42.1
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20240821/7c0f53af/attachment-0001.htm>