Michael,
Remember me, I'm back! :) I recall you stating the last time
you saw me: "Speaking of someone that likes to fuck with shit..." Well
I aim to please. I have a staggered set of patches that are coming that
rework the entire SIMD engine. If you're still reading and haven't
opened an email reply for a point by point shoot-down-in-flames then I
might have a chance. :)
I'm getting ready to explore that alignment issue that came up a
few months ago within imlib2. (Do we have a mail archive that stores
stuff older than a few months? Sourceforge is disappointing in that
regard.) On that note I've been digging into the various code bases as
much as the processor's internals to figure it out
The 15/16 bpp and the 32 bpp shading routines work fundamentally
different. The 15/16 bpp routines read in a pixel and make two more
copies for the three colors: red, green, and blue. They then mask out
the other colors in the copy shade tint each color with its modifier and
then blend the colors back into a single pixel. The 32 bpp routines
read in a pixel of three 8 bit colors and expand them to 16 bit values
in place, do the math and then truncate them back. That is why the
32bpp routines only read in 64bits at a time in SSE mode. Those two 32
bit pixels, with three 8 bit colors and maybe an 8 bit alpha channel
each expand to a 64 bit pixel with 16 bit values internally. After it
is tinted or shaded it is truncated back to four 8 bit values per pixel
and the two 32 bit pixels are then written back as a 64 bit chuck.
Therefore there is no need for using aligned/unaligned 128 bit memory
moves in 32bpp mode. This is what also creates rounding error
differences between the C routines. C does the temporary work in 32 bit
integers instead of 16 bits inside the SIMD engine.
Anyway, this presumes that you've attached the first set of patches
from the other day. This should be added to the "needs lots of testing"
branch of the code. I've tested it here and all is good. The only
thing that would be better than this patch is one that gauranteed the
data was aligned. That involves digging into X, which I'm doing, but
with great trepidation as that monster is a bit intimidating. Anyway,
patches for guaranteed alignment will go to X for XImage data. The
patches for x86/SSE are going to first get released as a test program to
see if I can get some widespread testing as I don't have the hardware.
I'm sure you'll see that on the list prior to me incorporating them into
Eterm and submitting them to you. Feel free to beep me on IRC if there
is anything else.
Cheers,
--
Tres Melton
IRC & Gentoo: RiverRat
Index: eterm/Eterm/src/pixmap.c
===================================================================
RCS file: /cvsroot/enlightenment/eterm/Eterm/src/pixmap.c,v
retrieving revision 1.115
diff -u -b -B -u -r1.115 pixmap.c
--- eterm/Eterm/src/pixmap.c 22 Dec 2005 23:31:33 -0000 1.115
+++ eterm/Eterm/src/pixmap.c 3 Jan 2006 22:34:31 -0000
@@ -66,9 +66,25 @@
extern void shade_ximage_32_mmx(void *data, int bpl, int w, int h, int rm, int gm, int bm);
/* Assembler routines for 64 bit cpu with sse2 */
-extern void shade_ximage_15_sse2(void *data, int bpl, int w, int h, int rm, int gm, int bm);
-extern void shade_ximage_16_sse2(void *data, int bpl, int w, int h, int rm, int gm, int bm);
-extern void shade_ximage_32_sse2(void *data, int bpl, int w, int h, int rm, int gm, int bm);
+extern void shade_ximage_15_sse2_A(void *data, int bpl, int w, int h, int rm, int gm, int bm );
+extern void shade_ximage_15_sse2_U(void *data, int bpl, int w, int h, int rm, int gm, int bm );
+extern void shade_ximage_16_sse2_A(void *data, int bpl, int w, int h, int rm, int gm, int bm );
+extern void shade_ximage_16_sse2_U(void *data, int bpl, int w, int h, int rm, int gm, int bm );
+
+#define shade_ximage_15_sse2( data, bpl, w, h, rm, gm, bm ) \
+{ \
+ (((long) ( data )) & ((long) ( bpl )) & ((long) ( ETERM_ALIGNMENT - 1 ))) ? \
+ shade_ximage_15_sse2_U((data), (bpl), (w), (h), (rm), (gm), (bm)) : \
+ shade_ximage_15_sse2_A((data), (bpl), (w), (h), (rm), (gm), (bm)); \
+}
+
+#define shade_ximage_16_sse2( data, bpl, w, h, rm, gm, bm ) \
+{ \
+ (((long) ( data )) & ((long) ( bpl )) & ((long) ( ETERM_ALIGNMENT - 1 ))) ? \
+ shade_ximage_16_sse2_U((data), (bpl), (w), (h), (rm), (gm), (bm)) : \
+ shade_ximage_16_sse2_A((data), (bpl), (w), (h), (rm), (gm), (bm)); \
+}
+
#ifdef PIXMAP_SUPPORT
static Imlib_Border bord_none = { 0, 0, 0, 0 };
Index: eterm/Eterm/src/sse2_cmod.c
===================================================================
RCS file: /cvsroot/enlightenment/eterm/Eterm/src/sse2_cmod.c,v
retrieving revision 1.1
diff -u -b -B -u -r1.1 sse2_cmod.c
--- eterm/Eterm/src/sse2_cmod.c 14 Jun 2005 19:39:01 -0000 1.1
+++ eterm/Eterm/src/sse2_cmod.c 3 Jan 2006 22:34:33 -0000
@@ -94,7 +88,7 @@
#ifdef HAVE_SSE2
-void shade_ximage_15_sse2( volatile void *data, volatile int bpl, volatile int w, volatile int h, volatile int rm, volatile int gm, volatile int bm )
+void shade_ximage_15_sse2_U( volatile void *data, volatile int bpl, volatile int w, volatile int h, volatile int rm, volatile int gm, volatile int bm )
{
__asm__ __volatile__ (
".align 16 \n\t" /* SIMD instructions should be aligned on 16 byte (128 bit) boundraries for performance reasons.*/
@@ -269,7 +263,7 @@
}
-void shade_ximage_16_sse2( volatile void *data, volatile int bpl, volatile int w, volatile int h, volatile int rm, volatile int gm, volatile int bm )
+void shade_ximage_16_sse2_U( volatile void *data, volatile int bpl, volatile int w, volatile int h, volatile int rm, volatile int gm, volatile int bm )
{
__asm__ __volatile__ (
".align 16 \n\t" /* SIMD instructions should be aligned on 16 byte (128 bit) boundraries for performance reasons.*/
@@ -447,6 +441,359 @@
); /* End of Assembly */
}
+void shade_ximage_15_sse2_A( volatile void *data, volatile int bpl, volatile int w, volatile int h, volatile int rm, volatile int gm, volatile int bm )
+{
+ __asm__ __volatile__ (
+ ".align 16 \n\t" /* SIMD instructions should be aligned on 16 byte (128 bit) boundraries for performance reasons.*/
+ "leaq -14(%%rsi, %%rbx, 2), %%rsi\n\t" /* Load the stack index register with a pointer to data + ( width * bytes/pixel ) -6 */
+ "negq %%rbx \n\t" /* Negate the width to that we can increment the counter */
+ "jz 10f \n\t" /* Jump to end if the line count is zero */
+ "movd %[red_mod], %%xmm5 \n\t" /* Load the color modifiers into mmx registers */
+ "movd %[green_mod], %%xmm6 \n\t" /* " " */
+ "movd %[blue_mod], %%xmm7 \n\t" /* " " */
+ "punpcklwd %%xmm5, %%xmm5 \n\t" /* Unpack and Interleave low words. From A64_128bit_Media_Programming (p. 380) */
+ "punpcklwd %%xmm6, %%xmm6 \n\t" /* Duplicate the bottom 16 bits into the next 16 bits (both operands are the same) */
+ "punpcklwd %%xmm7, %%xmm7 \n\t"
+ "punpckldq %%xmm5, %%xmm5 \n\t" /* Unpack and Interleave low double words. From A64_128bit_Media_Programming (p. 376) */
+ "punpckldq %%xmm6, %%xmm6 \n\t" /* Duplicate the bottom 32 bits into the next 32 bits (both operands are the same) */
+ "punpckldq %%xmm7, %%xmm7 \n\t"
+ "punpcklqdq %%xmm5, %%xmm5 \n\t" /* Unpack and Interleave low quad words. From A64_128bit_Media_Programming (p. 378) */
+ "punpcklqdq %%xmm6, %%xmm6 \n\t" /* Duplicate the bottom 64 bits into the next 64 bits (both operands are the same) */
+ "punpcklqdq %%xmm7, %%xmm7 \n\t"
+ "or %[red_mod], %[green_mod] \n\t" /* This, and the following 4 instructions, check to see if all three colormodifiers are */
+ "or %[blue_mod], %[green_mod] \n\t" /* less than 256. If any of the modifiers are > 256 then they will have the 9th, or higher, */
+ "sar $8, %[green_mod] \n\t" /* bit set. Then we shift off eight bits, leaving something set if a modifier > 256. */
+ "movq %%rax, %[blue_mod] \n\t" /* Use the register named blue_mod to now store bytes_per_line. */
+ "xor %[red_mod], %[red_mod] \n\t" /* zero red so we don't have to load an immediate value for the following compare. */
+ "cmp %[red_mod], %[green_mod] \n\t" /* Compare the left over bits to zero */
+ "jg 5f \n\t" /* If one of the colors (might) need saturated then jump to the secondary set of loops. */
+ "1: \n\t" /* Start of the outer loop (lines). */
+ "movq %%rbx, %%rcx \n\t" /* Move the width into the count register */
+ "addq $7, %%rcx \n\t"
+ "jns 3f \n\t"
+ "2: \n\t" /* Start of the inner loop (pixels 8 at a time --> 8 * 16 = 128bits/xmm register ) */
+ "movdqa (%%rsi, %%rcx, 2), %%xmm0\n\t" /* Load the 16 bits of the pixel (5 bits for red, 6 bits for green, 5 bits for blue) */
+ "movdqa %%xmm0, %%xmm1 \n\t" /* Create a copy of the pixel for the green color */
+ "movdqa %%xmm0, %%xmm2 \n\t" /* Create a copy of the pixel for the blue color */
+ "psrlw $5, %%xmm1 \n\t" /* Packed Shift Right Logical Words */
+ /* From A64_128bit_Media_Programming (p. 347) */
+ /* Shifts the blue off of the green color */
+ "psrlw $10, %%xmm0 \n\t" /* Shifts the blue & green off of the red color */
+ "psllw $11, %%xmm2 \n\t" /* Packed Shift Left Logical Words */
+ /* From A64_128bit_Media_Programming (p. 330) */
+ /* Shifts the red & green off of the blue color */
+ "psllw $11, %%xmm1 \n\t" /* Shifts the red off of the green color */
+ "psllw $8, %%xmm0 \n\t" /* Shifts the red color into position */
+ "psrlw $3, %%xmm1 \n\t" /* Shifts the green color into position */
+ "psrlw $3, %%xmm2 \n\t" /* Shifts the blue color into position */
+ "pmulhw %%xmm5, %%xmm0 \n\t" /* color *= modifier */
+ "pmulhw %%xmm6, %%xmm1 \n\t"
+ "pmulhw %%xmm7, %%xmm2 \n\t"
+ "psllw $10, %%xmm0 \n\t" /* Shift red back into its original position */
+ "psllw $5, %%xmm1 \n\t" /* Shift green back into its original position */
+ "por %%xmm2, %%xmm0 \n\t" /* Mesh the colors back together */
+ "por %%xmm1, %%xmm0 \n\t"
+ "movdqa %%xmm0, (%%rsi, %%rcx, 2)\n\t" /* Place the shaded 8 pixels back into the image map */
+ "addq $8, %%rcx \n\t"
+ "js 2b \n\t"
+ "jmp 4f \n\t"
+ "3: \n\t" /* Deal with pixels one at a time here. */
+ "movw (%%rsi, %%rcx, 2), %%ax \n\t"
+ "movd %%eax, %%xmm0 \n\t"
+ "movq %%xmm0, %%xmm1 \n\t"
+ "movq %%xmm0, %%xmm2 \n\t"
+ "psrlw $5, %%xmm1 \n\t"
+ "psrlw $10, %%xmm0 \n\t"
+ "psllw $11, %%xmm2 \n\t"
+ "psllw $11, %%xmm1 \n\t"
+ "psllw $8, %%xmm0 \n\t"
+ "psrlw $3, %%xmm1 \n\t"
+ "psrlw $3, %%xmm2 \n\t"
+ "pmulhw %%xmm5, %%xmm0 \n\t"
+ "pmulhw %%xmm6, %%xmm1 \n\t"
+ "pmulhw %%xmm7, %%xmm2 \n\t"
+ "psllw $10, %%xmm0 \n\t"
+ "psllw $5, %%xmm1 \n\t"
+ "por %%xmm2, %%xmm0 \n\t"
+ "por %%xmm1, %%xmm0 \n\t"
+ "movd %%xmm0, %%eax \n\t"
+ "movw %%ax, (%%rsi, %%rcx, 2) \n\t"
+ "incq %%rcx \n\t"
+ "4: \n\t"
+ "cmpq $6, %%rcx \n\t"
+ "jng 3b \n\t"
+ "addq %[blue_mod], %%rsi \n\t" /* Blue_mod is the name of a register that now contains bytes_per_line. */
+ "decq %%rdx \n\t"
+ "jnz 1b \n\t"
+ "jmp 10f \n\t" /* We're done! */
+
+ "5: \n\t" /* Saturation is required */
+ "pcmpeqw %%xmm3, %%xmm3 \n\t" /* Packed Compare Equal Words */
+ /* From A64_128bit_Media_Programming (p. 276) */
+ /* This sets xmm3 to 128 1's (since mm6 = mm6) */
+ "psllw $5, %%xmm3 \n\t" /* xmm3 = 8 copies of 1111 1111 1110 0000 */
+ "6: \n\t"
+ "movq %%rbx, %%rcx \n\t"
+ "addq $7, %%rcx \n\t"
+ "jns 8f \n\t"
+ "7: \n\t"
+ "movdqa (%%rsi, %%rcx, 2), %%xmm0\n\t"
+ "movdqa %%xmm0, %%xmm1 \n\t"
+ "movdqa %%xmm0, %%xmm2 \n\t"
+ "psrlw $5, %%xmm1 \n\t"
+ "psrlw $10, %%xmm0 \n\t"
+ "psllw $11, %%xmm2 \n\t"
+ "psllw $11, %%xmm1 \n\t"
+ "psllw $8, %%xmm0 \n\t"
+ "psrlw $3, %%xmm1 \n\t"
+ "psrlw $3, %%xmm2 \n\t"
+ "pmulhw %%xmm5, %%xmm0 \n\t"
+ "pmulhw %%xmm6, %%xmm1 \n\t"
+ "pmulhw %%xmm7, %%xmm2 \n\t"
+ "paddusw %%xmm3, %%xmm0 \n\t"
+ "paddusw %%xmm3, %%xmm1 \n\t"
+ "paddusw %%xmm3, %%xmm2 \n\t"
+ "psubw %%xmm3, %%xmm0 \n\t" /* FIXME: This line needs added to the original asm code */
+ "psubw %%xmm3, %%xmm1 \n\t"
+ "psubw %%xmm3, %%xmm2 \n\t"
+ "psllw $10, %%xmm0 \n\t"
+ "psllw $5, %%xmm1 \n\t"
+ "por %%xmm2, %%xmm0 \n\t"
+ "por %%xmm1, %%xmm0 \n\t"
+ "movdqa %%xmm0, (%%rsi, %%rcx, 2)\n\t"
+ "addq $8, %%rcx \n\t"
+ "js 7b \n\t"
+ "jmp 9f \n\t"
+ "8: \n\t"
+ "movw (%%rsi, %%rcx, 2), %%ax \n\t"
+ "movd %%eax, %%xmm0 \n\t"
+ "movq %%xmm0, %%xmm1 \n\t"
+ "movq %%xmm0, %%xmm2 \n\t"
+ "psrlw $5, %%xmm1 \n\t"
+ "psrlw $10, %%xmm0 \n\t"
+ "psllw $11, %%xmm2 \n\t"
+ "psllw $11, %%xmm1 \n\t"
+ "psllw $8, %%xmm0 \n\t"
+ "psrlw $3, %%xmm1 \n\t"
+ "psrlw $3, %%xmm2 \n\t"
+ "pmulhw %%xmm5, %%xmm0 \n\t"
+ "pmulhw %%xmm6, %%xmm1 \n\t"
+ "pmulhw %%xmm7, %%xmm2 \n\t"
+ "paddusw %%xmm3, %%xmm0 \n\t"
+ "paddusw %%xmm3, %%xmm1 \n\t"
+ "paddusw %%xmm3, %%xmm2 \n\t"
+ "psubw %%xmm3, %%xmm0 \n\t" /* FIXME: This line needs added to the original asm code */
+ "psubw %%xmm3, %%xmm1 \n\t"
+ "psubw %%xmm3, %%xmm2 \n\t"
+ "psllw $10, %%xmm0 \n\t"
+ "psllw $5, %%xmm1 \n\t"
+ "por %%xmm2, %%xmm0 \n\t"
+ "por %%xmm1, %%xmm0 \n\t"
+ "movd %%xmm0, %%eax \n\t"
+ "movw %%ax, (%%rsi, %%rcx, 2) \n\t"
+ "incq %%rcx \n\t"
+ "9: \n\t"
+ "cmpq $6, %%rcx \n\t"
+ "jng 8b \n\t"
+ "addq %[blue_mod], %%rsi \n\t" /* Blue_mod is the name of a register that now contains bytes_per_line. */
+ "decq %%rdx \n\t"
+ "jnz 6b \n\t"
+ "10: \n\t" /* This is the end. Jump here if the line count is zero. */
+ "emms \n\t" /* exit multi-media state (last asm instruction) */
+ : /* outputs: none */
+ /* inputs: (many operations cannot be performed with a mix of 32bit & 64bit operands directly) */
+ /* (however the compiler/assembler can preload 32bit values into 64bit registers) */
+ /* (that is why certain variables cannot be referenced by name -- use their register) */
+ : [data] "S" (data), /* put the pointer data into the rsi register */
+ [width] "b" (w), /* put the width in the %rbx register (cannot be referenced by name) */
+ [height] "d" (h), /* put the heigth in the %rdx register (cannot be referenced by name) */
+ [red_mod] "r" ((unsigned long)(rm)),/* put the red_modifier in a register (referenced by name) */
+ [green_mod] "r" ((unsigned long)(gm)),/* put the green_modifier in a register (referenced by name) */
+ [blue_mod] "r" ((unsigned long)(bm)),/* put the blue_modifier in a register (referenced by name) Later store the bytes_line here */
+ [bytes_line] "a" (bpl) /* put the bytes_per_line in the %rax register (cannot be referenced by name) */
+ : "memory" /* clobbers: (memory includes all the registers) */
+ ); /* End of Assembly */
+}
+
+
+void shade_ximage_16_sse2_A( volatile void *data, volatile int bpl, volatile int w, volatile int h, volatile int rm, volatile int gm, volatile int bm )
+{
+ __asm__ __volatile__ (
+ ".align 16 \n\t" /* SIMD instructions should be aligned on 16 byte (128 bit) boundraries for performance reasons.*/
+ "leaq -14(%%rsi, %%rbx, 2), %%rsi\n\t" /* Load the stack index register with a pointer to data + ( width * bytes/pixel ) -6 */
+ "negq %%rbx \n\t" /* Negate the width to that we can increment the counter */
+ "jz 10f \n\t" /* Jump to end if the line count is zero */
+ "movd %[red_mod], %%xmm5 \n\t" /* Load the color modifiers into mmx registers */
+ "movd %[green_mod], %%xmm6 \n\t" /* " " */
+ "movd %[blue_mod], %%xmm7 \n\t" /* " " */
+ "punpcklwd %%xmm5, %%xmm5 \n\t" /* Unpack and Interleave low words. From A64_128bit_Media_Programming (p. 380) */
+ "punpcklwd %%xmm6, %%xmm6 \n\t" /* Duplicate the bottom 16 bits into the next 16 bits (both operands are the same) */
+ "punpcklwd %%xmm7, %%xmm7 \n\t"
+ "punpckldq %%xmm5, %%xmm5 \n\t" /* Unpack and Interleave low double words. From A64_128bit_Media_Programming (p. 376) */
+ "punpckldq %%xmm6, %%xmm6 \n\t" /* Duplicate the bottom 32 bits into the next 32 bits (both operands are the same) */
+ "punpckldq %%xmm7, %%xmm7 \n\t"
+ "punpcklqdq %%xmm5, %%xmm5 \n\t" /* Unpack and Interleave low quad words. From A64_128bit_Media_Programming (p. 378) */
+ "punpcklqdq %%xmm6, %%xmm6 \n\t" /* Duplicate the bottom 64 bits into the next 64 bits (both operands are the same) */
+ "punpcklqdq %%xmm7, %%xmm7 \n\t"
+ "or %[red_mod], %[green_mod] \n\t" /* This, and the following 4 instructions, check to see if all three colormodifiers are */
+ "or %[blue_mod], %[green_mod] \n\t" /* less than 256. If any of the modifiers are > 256 then they will have the 9th, or higher, */
+ "sar $8, %[green_mod] \n\t" /* bit set. Then we shift off eight bits, leaving something set if a modifier > 256. */
+ "movq %%rax, %[blue_mod] \n\t" /* Use the register named blue_mod to now store bytes_per_line. */
+ "xor %[red_mod], %[red_mod] \n\t" /* zero red so we don't have to load an immediate value for the following compare. */
+ "cmp %[red_mod], %[green_mod] \n\t" /* Compare the left over bits to zero */
+ "jg 5f \n\t" /* If one of the colors (might) need saturated then jump to the secondary set of loops. */
+ "1: \n\t" /* Start of the outer loop (lines). */
+ "movq %%rbx, %%rcx \n\t" /* Move the width into the count register */
+ "addq $7, %%rcx \n\t"
+ "jns 3f \n\t"
+ "2: \n\t" /* Start of the inner loop (pixels 8 at a time --> 8 * 16 = 128bits/xmm register ) */
+ "movdqa (%%rsi, %%rcx, 2), %%xmm0\n\t" /* Load the 16 bits of the pixel (5 bits for red, 6 bits for green, 5 bits for blue) */
+ "movdqa %%xmm0, %%xmm1 \n\t" /* Create a copy of the pixel for the green color */
+ "movdqa %%xmm0, %%xmm2 \n\t" /* Create a copy of the pixel for the blue color */
+ "psrlw $5, %%xmm1 \n\t" /* Packed Shift Right Logical Words */
+ /* From A64_128bit_Media_Programming (p. 347) */
+ /* Shifts the blue off of the green color */
+ "psrlw $11, %%xmm0 \n\t" /* Shifts the blue & green off of the red color */
+ "psllw $11, %%xmm2 \n\t" /* Packed Shift Left Logical Words */
+ /* From A64_128bit_Media_Programming (p. 330) */
+ /* Shifts the red & green off of the blue color */
+ "psllw $10, %%xmm1 \n\t" /* Shifts the red off of the green color */
+ "psllw $8, %%xmm0 \n\t" /* Shifts the red color into position */
+ "psrlw $2, %%xmm1 \n\t" /* Shifts the green color into position */
+ "psrlw $3, %%xmm2 \n\t" /* Shifts the blue color into position */
+ "pmulhw %%xmm5, %%xmm0 \n\t" /* color *= modifier */
+ "pmulhw %%xmm6, %%xmm1 \n\t"
+ "pmulhw %%xmm7, %%xmm2 \n\t"
+ "psllw $11, %%xmm0 \n\t" /* Shift red back into its original position */
+ "psllw $5, %%xmm1 \n\t" /* Shift green back into its original position */
+ "por %%xmm2, %%xmm0 \n\t" /* Mesh the colors back together */
+ "por %%xmm1, %%xmm0 \n\t"
+ "movdqa %%xmm0, (%%rsi, %%rcx, 2)\n\t" /* Place the shaded 8 pixels back into the image map */
+ "addq $8, %%rcx \n\t"
+ "js 2b \n\t"
+ "jmp 4f \n\t"
+ "3: \n\t" /* Deal with pixels one at a time here. */
+ "movw (%%rsi, %%rcx, 2), %%ax \n\t"
+ "movd %%eax, %%xmm0 \n\t"
+ "movq %%xmm0, %%xmm1 \n\t"
+ "movq %%xmm0, %%xmm2 \n\t"
+ "psrlw $5, %%xmm1 \n\t"
+ "psrlw $11, %%xmm0 \n\t"
+ "psllw $11, %%xmm2 \n\t"
+ "psllw $10, %%xmm1 \n\t"
+ "psllw $8, %%xmm0 \n\t"
+ "psrlw $2, %%xmm1 \n\t"
+ "psrlw $3, %%xmm2 \n\t"
+ "pmulhw %%xmm5, %%xmm0 \n\t"
+ "pmulhw %%xmm6, %%xmm1 \n\t"
+ "pmulhw %%xmm7, %%xmm2 \n\t"
+ "psllw $11, %%xmm0 \n\t"
+ "psllw $5, %%xmm1 \n\t"
+ "por %%xmm2, %%xmm0 \n\t"
+ "por %%xmm1, %%xmm0 \n\t"
+ "movd %%xmm0, %%eax \n\t"
+ "movw %%ax, (%%rsi, %%rcx, 2) \n\t"
+ "incq %%rcx \n\t"
+ "4: \n\t"
+ "cmpq $6, %%rcx \n\t"
+ "jng 3b \n\t"
+ "addq %[blue_mod], %%rsi \n\t" /* Blue_mod is the name of a register that now contains bytes_per_line. */
+ "decq %%rdx \n\t"
+ "jnz 1b \n\t"
+ "jmp 10f \n\t" /* We're done! */
+
+ "5: \n\t" /* Saturation is required */
+ "pcmpeqw %%xmm3, %%xmm3 \n\t" /* Packed Compare Equal Words */
+ /* From A64_128bit_Media_Programming (p. 276) */
+ /* This sets xmm3 to 128 1's (since mm6 = mm6) */
+ "movdqa %%xmm3, %%xmm4 \n\t" /* Make copy of 128 ones */
+ "psllw $5, %%xmm3 \n\t" /* xmm3 = 8 copies of 1111 1111 1110 0000 */
+ "psllw $6, %%xmm4 \n\t" /* xmm4 = 8 copies of 1111 1111 1100 0000 */
+ "6: \n\t"
+ "movq %%rbx, %%rcx \n\t"
+ "addq $7, %%rcx \n\t"
+ "jns 8f \n\t"
+ "7: \n\t"
+ "movdqa (%%rsi, %%rcx, 2), %%xmm0\n\t"
+ "movdqa %%xmm0, %%xmm1 \n\t"
+ "movdqa %%xmm0, %%xmm2 \n\t"
+ "psrlw $5, %%xmm1 \n\t"
+ "psrlw $11, %%xmm0 \n\t"
+ "psllw $11, %%xmm2 \n\t"
+ "psllw $10, %%xmm1 \n\t"
+ "psllw $8, %%xmm0 \n\t"
+ "psrlw $2, %%xmm1 \n\t"
+ "psrlw $3, %%xmm2 \n\t"
+ "pmulhw %%xmm5, %%xmm0 \n\t"
+ "pmulhw %%xmm6, %%xmm1 \n\t"
+ "pmulhw %%xmm7, %%xmm2 \n\t"
+ "paddusw %%xmm3, %%xmm0 \n\t"
+ "paddusw %%xmm4, %%xmm1 \n\t"
+ "paddusw %%xmm3, %%xmm2 \n\t"
+ "psubw %%xmm4, %%xmm1 \n\t"
+ "psubw %%xmm3, %%xmm2 \n\t"
+ "psllw $11, %%xmm0 \n\t"
+ "psllw $5, %%xmm1 \n\t"
+ "por %%xmm2, %%xmm0 \n\t"
+ "por %%xmm1, %%xmm0 \n\t"
+ "movdqa %%xmm0, (%%rsi, %%rcx, 2)\n\t"
+ "addq $8, %%rcx \n\t"
+ "js 7b \n\t"
+ "jmp 9f \n\t"
+ "8: \n\t"
+ "movw (%%rsi, %%rcx, 2), %%ax \n\t"
+ "movd %%eax, %%xmm0 \n\t"
+ "movq %%xmm0, %%xmm1 \n\t"
+ "movq %%xmm0, %%xmm2 \n\t"
+ "psrlw $5, %%xmm1 \n\t"
+ "psrlw $11, %%xmm0 \n\t"
+ "psllw $11, %%xmm2 \n\t"
+ "psllw $10, %%xmm1 \n\t"
+ "psllw $8, %%xmm0 \n\t"
+ "psrlw $2, %%xmm1 \n\t"
+ "psrlw $3, %%xmm2 \n\t"
+ " \n\t"
+ "pmulhw %%xmm5, %%xmm0 \n\t"
+ "pmulhw %%xmm6, %%xmm1 \n\t"
+ "pmulhw %%xmm7, %%xmm2 \n\t"
+ " \n\t"
+ "paddusw %%xmm3, %%xmm0 \n\t"
+ "paddusw %%xmm4, %%xmm1 \n\t"
+ "paddusw %%xmm3, %%xmm2 \n\t"
+ " \n\t"
+ "psubw %%xmm4, %%xmm1 \n\t"
+ "psubw %%xmm3, %%xmm2 \n\t"
+ " \n\t"
+ "psllw $11, %%xmm0 \n\t"
+ "psllw $5, %%xmm1 \n\t"
+ "por %%xmm2, %%xmm0 \n\t"
+ "por %%xmm1, %%xmm0 \n\t"
+ "movd %%xmm0, %%eax \n\t"
+ "movw %%ax, (%%rsi, %%rcx, 2) \n\t"
+ "incq %%rcx \n\t"
+ "9: \n\t"
+ "cmpq $6, %%rcx \n\t"
+ "jng 8b \n\t"
+ "addq %[blue_mod], %%rsi \n\t" /* Blue_mod is the name of a register that now contains bytes_per_line. */
+ "decq %%rdx \n\t"
+ "jnz 6b \n\t"
+ "10: \n\t" /* This is the end. Jump here if the line count is zero. */
+ "emms \n\t" /* exit multi-media state (last asm instruction) */
+ : /* outputs: none */
+ /* inputs: (many operations cannot be performed with a mix of 32bit & 64bit operands directly) */
+ /* (however the compiler/assembler can preload 32bit values into 64bit registers) */
+ /* (that is why certain variables cannot be referenced by name -- use their register) */
+ : [data] "S" (data), /* put the pointer data into the rsi register */
+ [width] "b" (w), /* put the width in the %rbx register (cannot be referenced by name) */
+ [height] "d" (h), /* put the heigth in the %rdx register (cannot be referenced by name) */
+ [red_mod] "r" ((unsigned long)(rm)),/* put the red_modifier in a register (referenced by name) */
+ [green_mod] "r" ((unsigned long)(gm)),/* put the green_modifier in a register (referenced by name) */
+ [blue_mod] "r" ((unsigned long)(bm)),/* put the blue_modifier in a register (referenced by name) Later store the bytes_line here */
+ [bytes_line] "a" (bpl) /* put the bytes_per_line in the %rax register (cannot be referenced by name) */
+ : "memory" /* clobbers: (memory includes all the registers) */
+ ); /* End of Assembly */
+}
+
void shade_ximage_32_sse2( volatile void *data, volatile int bpl, volatile int w, volatile int h, volatile int rm, volatile int gm, volatile int bm )
{
signature.asc
Description: This is a digitally signed message part
