Attachment 23964 Details for Bug 20709 – Use SSE2/MMX to copy ARGB pixels, setting alpha to 0xFF

[patch] Use SSE2/MMX to copy ARGB pixels, setting alpha to 0xFF

pixman-fetch-fdt.patch (text/plain), 4.90 KB, created by Steve Snyder on 2009-03-17 07:04:07 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Steve Snyder

Created: 2009-03-17 07:04:07 UTC

Size: 4.90 KB

patch

obsolete

>diff -Nr -U5 libpixman.original/src/pixman-access.c libpixman/src/pixman-access.c
>--- libpixman.original/src/pixman-access.c	2009-02-23 16:41:44.000000000 -0500
>+++ libpixman/src/pixman-access.c	2009-03-17 09:40:53.000000000 -0400
>@@ -66,10 +66,20 @@
> 
> #define YV12_V(line)	      \
>     ((uint8_t *) ((bits) + offset0 + \
> 		((stride) >> 1) * ((line) >> 1)))
> 
>+#if (defined(_MSC_VER) && defined(_M_IX86) && (_M_IX86_FP == 2)) || \
>+    (defined(__GNUC__) && defined(__SSE2__))
>+# include "emmintrin.h"
>+# define _USE_ALPHA_COPY_SSE2_
>+#elif (defined(_MSC_VER) && defined(_M_IX86) && (_M_IX86_FP == 1)) || \
>+    (defined(__GNUC__) && defined(__SSE__))
>+# include "xmmintrin.h"
>+# define _USE_ALPHA_COPY_SSE_
>+#endif
>+
> /*********************************** Fetch ************************************/
> 
> static FASTCALL void
> fbFetch_a8r8g8b8 (bits_image_t *pict, int x, int y, int width, uint32_t *buffer)
> {
>@@ -83,10 +93,70 @@
> fbFetch_x8r8g8b8 (bits_image_t *pict, int x, int y, int width, uint32_t *buffer)
> {
>     const uint32_t *bits = pict->bits + y*pict->rowstride;
>     const uint32_t *pixel = (const uint32_t *)bits + x;
>     const uint32_t *end = pixel + width;
>+
>+#if defined(_USE_ALPHA_COPY_SSE2_) || defined(_USE_ALPHA_COPY_SSE_)
>+    ALIGN16 static const unsigned int alphas[4] = 
>+        {0xFF000000, 0xFF000000, 0xFF000000, 0xFF000000};
>+    __m64 alpha2, tmp8a, tmp8b, tmp8c, tmp8d;
>+
>+#ifdef _USE_ALPHA_COPY_SSE2_
>+    __m128i alpha4, tmp16a, tmp16b;
>+
>+    // if both source and dest pointers are or can be aligned...
>+    if ( ((uint32_t)pixel & 0x0F) == ((uint32_t)buffer & 0x0F) )
>+    {
>+        // if not already aligned, align pointers with single-pixel copies
>+        while ( ((uint32_t)pixel & 0x0F) && width ) {
>+            *buffer++ = *pixel++ | 0xff000000;
>+            width--;
>+        }
>+
>+        // use of tmp0 and tmp1 variables are to avoid
>+        //   back-to-back register operations
>+        // also: MSVC8 code gen for MMX sucks; use asm instead?
>+
>+        alpha4 = *((__m128i *)&alphas);
>+        while (width >= 8)
>+        {
>+            tmp16a = *((__m128i *)&pixel[0]);
>+            tmp16b = *((__m128i *)&pixel[4]);
>+            *((__m128i *)&buffer[0]) = _mm_or_si128(tmp16a, alpha4);
>+            *((__m128i *)&buffer[4]) = _mm_or_si128(tmp16b, alpha4);
>+            buffer+=8; pixel+=8; width-=8;
>+        }
>+    }
>+#endif // _USE_ALPHA_COPY_SSE2_
>+
>+    // if both source and dest pointers are or can be aligned...
>+    if ( (width >= 8) && (((uint32_t)pixel & 0x07) == ((uint32_t)buffer & 0x07)) )
>+    {
>+        // if not already aligned, align pointers with single-pixel copies
>+        while ( ((uint32_t)pixel & 0x07) && width ) {
>+            *buffer++ = *pixel++ | 0xff000000;
>+            width--;
>+        }
>+
>+        alpha2 = *((__m64 *)&alphas);
>+        while (width >= 8)
>+        {
>+            tmp8a = *((__m64 *)&pixel[0]);
>+            tmp8b = *((__m64 *)&pixel[2]);
>+            tmp8c = *((__m64 *)&pixel[4]);
>+            tmp8d = *((__m64 *)&pixel[6]);
>+            *((__m64 *)&buffer[0]) = _m_por(tmp8a, alpha2);
>+            *((__m64 *)&buffer[2]) = _m_por(tmp8b, alpha2);
>+            *((__m64 *)&buffer[4]) = _m_por(tmp8c, alpha2);
>+            *((__m64 *)&buffer[6]) = _m_por(tmp8d, alpha2);
>+            buffer+=8; pixel+=8; width-=8;
>+        }
>+        _m_empty();
>+    }
>+#endif // defined(_USE_ALPHA_COPY_SSE2_) || defined(_USE_ALPHA_COPY_SSE_)
>+
>     while (pixel < end) {
> 	*buffer++ = READ(pict, pixel++) | 0xff000000;
>     }
> }
> 
>diff -Nr -U5 libpixman.original/src/pixman-compose.c libpixman/src/pixman-compose.c
>--- libpixman.original/src/pixman-compose.c	2009-02-23 16:41:44.000000000 -0500
>+++ libpixman/src/pixman-compose.c	2009-03-17 09:40:53.000000000 -0400
>@@ -546,11 +546,11 @@
> #define SCANLINE_BUFFER_LENGTH 2048
> 
> void
> pixman_composite_rect_general (const FbComposeData *data)
> {
>-    uint32_t _scanline_buffer[SCANLINE_BUFFER_LENGTH * 3];
>+    ALIGN16 uint32_t _scanline_buffer[SCANLINE_BUFFER_LENGTH * 3];
>     const pixman_format_code_t srcFormat = data->src->type == BITS ? data->src->bits.format : 0;
>     const pixman_format_code_t maskFormat = data->mask && data->mask->type == BITS ? data->mask->bits.format : 0;
>     const pixman_format_code_t destFormat = data->dest->type == BITS ? data->dest->bits.format : 0;
>     const int srcWide = PIXMAN_FORMAT_16BPC(srcFormat);
>     const int maskWide = data->mask && PIXMAN_FORMAT_16BPC(maskFormat);
>diff -Nr -U5 libpixman.original/src/pixman-private.h libpixman/src/pixman-private.h
>--- libpixman.original/src/pixman-private.h	2009-02-23 16:41:44.000000000 -0500
>+++ libpixman/src/pixman-private.h	2009-03-17 09:40:53.000000000 -0400
>@@ -8,10 +8,16 @@
> #include "cairo-platform.h"
> 
> #include "pixman.h"
> #include <time.h>
> 
>+#if defined(_MSC_VER)
>+# define ALIGN16 __declspec(align(16))
>+#elif defined(__GNUC__)
>+# define ALIGN16 __attribute__ ((aligned (16)))
>+#endif // _MSC_VER || __GNUC__
>+
> #ifndef FALSE
> #define FALSE 0
> #endif
> 
> #ifndef TRUE

Actions: View | Splinter Review

Attachments on bug 20709: 23964

Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.