Re: [E-devel] [PATCH 2/8] evas: add common SSE3 blending functions/macro

[Vincent T. <vt...@un...>]

On Thu, 29 Sep 2011, Jim Kukunas wrote:

> mul_256_sse3
> sub4_alpha_sse3
> interp4_256_sse3
> mul_sym_sse3
> mul4_sym_sse3
> mul3_sym_sse3
>
> LOOP_ALIGNED_U1_A48_SSE3
>
> __attribute__((always_inline)) is needed to coax GCC (< 4.6.0)
> into inlining the common blend ops. Not inlining these functions
> causes a steep performance penalty.
> ---
> evas/src/lib/include/evas_blend_ops.h |  200 +++++++++++++++++++++++++++++++++
> 1 files changed, 200 insertions(+), 0 deletions(-)
>
> diff --git a/evas/src/lib/include/evas_blend_ops.h b/evas/src/lib/include/evas_blend_ops.h
> index 9647800..774f881 100644
> --- a/evas/src/lib/include/evas_blend_ops.h
> +++ b/evas/src/lib/include/evas_blend_ops.h
> @@ -5,6 +5,19 @@
> #include "evas_mmx.h"
> #endif
>
> +#if defined BUILD_SSE3
> +#include <immintrin.h>
> +#endif
> +
> +#ifndef always_inline
> +#if defined(__GNUC__) && (__GNUC__ > 3 || __GNUC__ == 3 && __GNUC_MINOR__ > 0)
> +#    define always_inline __attribute__((always_inline)) inline
> +#else
> +#    define always_inline inline
> +#endif
> +#endif

instead of that code, i would prefer a check in m4/efl_attribute.m4, by 
compiling a code that would fail if __attribute__((always_inline)) is not 
supported, and not using macro like __GNUC__

Vincent

> +
> +
> /* src pixel flags: */
>
> /* pixels none */
> @@ -178,4 +191,191 @@ extern const DATA32 ALPHA_256;
>
> #endif
>
> +
> +/* some useful SSE3 inline functions */
> +
> +#ifdef BUILD_SSE3
> +
> +static __m128i GA_MASK_SSE3;
> +static __m128i RB_MASK_SSE3;
> +static __m128i SYM4_MASK_SSE3;
> +static __m128i RGB_MASK_SSE3;
> +static __m128i A_MASK_SSE3;
> +
> +static __m128i ALPHA_SSE3;
> +
> +static always_inline __m128i
> +mul_256_sse3(__m128i a, __m128i c) {
> +
> +   /* prepare alpha for word multiplication */
> +   __m128i a_l = a;
> +   __m128i a_h = a;
> +   a_l = _mm_unpacklo_epi16(a_l, a_l);
> +   a_h = _mm_unpackhi_epi16(a_h, a_h);
> +   __m128i a0 = (__m128i) _mm_shuffle_ps( (__m128)a_l, (__m128)a_h, 0x88);
> +
> +   /* first half of calc */
> +   __m128i c0 = c;
> +   c0 = _mm_srli_epi32(c0, 8);
> +   c0 = _mm_and_si128(GA_MASK_SSE3, c0);
> +   c0 = _mm_mullo_epi16(a0, c0);
> +   c0 = _mm_and_si128(RB_MASK_SSE3, c0);
> +
> +   /* second half of calc */
> +   __m128i c1 = c;
> +   c1 = _mm_and_si128(GA_MASK_SSE3, c1);
> +   c1 = _mm_mullo_epi16(a0, c1);
> +   c1 = _mm_srli_epi32(c1, 8);
> +   c1 = _mm_and_si128(GA_MASK_SSE3, c1);
> +
> +   /* combine */
> +   return _mm_add_epi32(c0, c1);
> +}
> +
> +static always_inline __m128i
> +sub4_alpha_sse3(__m128i c) {
> +
> +   __m128i c0 = c;
> +
> +   c0 = _mm_srli_epi32(c0, 24);
> +   return _mm_sub_epi32(ALPHA_SSE3, c0);
> +}
> +
> +static always_inline __m128i
> +interp4_256_sse3(__m128i a, __m128i c0, __m128i c1)
> +{
> +   const __m128i zero = _mm_setzero_si128();
> +
> +   __m128i a_l = a;
> +   __m128i a_h = a;
> +   a_l = _mm_unpacklo_epi16(a_l, a_l);
> +   a_h = _mm_unpackhi_epi16(a_h, a_h);
> +
> +   __m128i a_t = _mm_slli_epi64(a_l, 32);
> +   __m128i a_t0 = _mm_slli_epi64(a_h, 32);
> +
> +   a_l = _mm_add_epi32(a_l, a_t);
> +   a_h = _mm_add_epi32(a_h, a_t0);
> +
> +   __m128i c0_l = c0;
> +   __m128i c0_h = c0;
> +
> +   c0_l = _mm_unpacklo_epi8(c0_l, zero);
> +   c0_h = _mm_unpackhi_epi8(c0_h, zero);
> +
> +   __m128i c1_l = c1;
> +   __m128i c1_h = c1;
> +
> +   c1_l = _mm_unpacklo_epi8(c1_l, zero);
> +   c1_h = _mm_unpackhi_epi8(c1_h, zero);
> +
> +   __m128i cl_sub = _mm_sub_epi16(c0_l, c1_l);
> +   __m128i ch_sub = _mm_sub_epi16(c0_h, c1_h);
> +
> +   cl_sub = _mm_mullo_epi16(cl_sub, a_l);
> +   ch_sub = _mm_mullo_epi16(ch_sub, a_h);
> +
> +   __m128i c1ls = _mm_slli_epi16(c1_l, 8);
> +   __m128i c1hs = _mm_slli_epi16(c1_h, 8);
> +
> +   cl_sub = _mm_add_epi16(cl_sub, c1ls);
> +   ch_sub = _mm_add_epi16(ch_sub, c1hs);
> +
> +   cl_sub = _mm_and_si128(cl_sub, RB_MASK_SSE3);
> +   ch_sub = _mm_and_si128(ch_sub, RB_MASK_SSE3);
> +
> +   cl_sub = _mm_srli_epi64(cl_sub, 8);
> +   ch_sub = _mm_srli_epi64(ch_sub, 8);
> +
> +   cl_sub = _mm_packus_epi16(cl_sub, cl_sub);
> +   ch_sub = _mm_packus_epi16(ch_sub, ch_sub);
> +
> +   return  (__m128i) _mm_shuffle_ps( (__m128)cl_sub, (__m128)ch_sub, 0x44);
> +}
> +
> +static always_inline __m128i
> +mul_sym_sse3(__m128i a, __m128i c) {
> +
> +      /* Prepare alpha for word mult */
> +      __m128i a_l = a;
> +      __m128i a_h = a;
> +      a_l = _mm_unpacklo_epi16(a_l, a_l);
> +      a_h = _mm_unpackhi_epi16(a_h, a_h);
> +      __m128i a0 = (__m128i) _mm_shuffle_ps( (__m128)a_l, (__m128)a_h, 0x88);
> +
> +      /* first part */
> +      __m128i c0 = c;
> +      c0 = _mm_srli_epi32(c0, 8);
> +      c0 = _mm_and_si128(GA_MASK_SSE3, c0);
> +      c0 = _mm_mullo_epi16(a0, c0);
> +      c0 = _mm_add_epi32(c0, GA_MASK_SSE3);
> +      c0 = _mm_and_si128(RB_MASK_SSE3, c0);
> +
> +      /* second part */
> +      __m128i c1 = c;
> +      c1 = _mm_and_si128(GA_MASK_SSE3, c1);
> +      c1 = _mm_mullo_epi16(a0, c1);
> +      c1 = _mm_add_epi32(c1, GA_MASK_SSE3);
> +      c1 = _mm_srli_epi32(c1, 8);
> +      c1 = _mm_and_si128(GA_MASK_SSE3, c1);
> +
> +      return _mm_add_epi32(c0, c1);
> +}
> +
> +static always_inline __m128i
> +mul4_sym_sse3(__m128i x, __m128i y) {
> +
> +   const __m128i zero = _mm_setzero_si128();
> +
> +   __m128i x_l = _mm_unpacklo_epi8(x, zero);
> +   __m128i x_h = _mm_unpackhi_epi8(x, zero);
> +
> +   __m128i y_l = _mm_unpacklo_epi8(y, zero);
> +   __m128i y_h = _mm_unpackhi_epi8(y, zero);
> +
> +   __m128i r_l = _mm_mullo_epi16(x_l, y_l);
> +   __m128i r_h = _mm_mullo_epi16(x_h, y_h);
> +
> +   r_l = _mm_add_epi16(r_l, SYM4_MASK_SSE3);
> +   r_h = _mm_add_epi16(r_h, SYM4_MASK_SSE3);
> +
> +   r_l = _mm_srli_epi16(r_l, 8);
> +   r_h = _mm_srli_epi16(r_h, 8);
> +
> +   return  _mm_packus_epi16(r_l, r_h);
> +}
> +
> +static always_inline __m128i
> +mul3_sym_sse3(__m128i x, __m128i y) {
> +
> +   __m128i res = mul4_sym_sse3(x, y);
> +   return  _mm_and_si128(res, RGB_MASK_SSE3);
> +}
> +
> +#define LOOP_ALIGNED_U1_A48_SSE3(D, LENGTH, UOP,A4OP, A8OP) \
> +   { \
> +      while((uintptr_t)d & 0xF && l) UOP \
> +   \
> +      while(l) { \
> +         switch(l) { \
> +            case 3: UOP \
> +            case 2: UOP \
> +            case 1: UOP \
> +               break; \
> +            case 7: \
> +            case 6: \
> +            case 5: \
> +            case 4: \
> +               A4OP \
> +               break; \
> +            default: \
> +               A8OP \
> +               break; \
> +         } \
> +      } \
> +   }
> +
> +
> +#endif
> +
> #endif
> -- 
> 1.7.3.4
>
>
> ------------------------------------------------------------------------------
> All the data continuously generated in your IT infrastructure contains a
> definitive record of customers, application performance, security
> threats, fraudulent activity and more. Splunk takes this data and makes
> sense of it. Business sense. IT sense. Common sense.
> http://p.sf.net/sfu/splunk-d2dcopy1
> _______________________________________________
> enlightenment-devel mailing list
> enl...@li...
> https://lists.sourceforge.net/lists/listinfo/enlightenment-devel
>
>