[fd5b1fe] | 1 | Submitted By: Ryan Oliver <ryan.oliver@pha.com.au>
|
---|
| 2 | Date: 2005-07-31
|
---|
| 3 | Initial Package Version: 0.9.21
|
---|
| 4 | Upstream Status: Unknown
|
---|
| 5 | Origin: Gentoo
|
---|
| 6 | Description: Supposedly fixes simd instructions on amd64. Didn't work here.
|
---|
| 7 |
|
---|
| 8 | diff -urN DirectFB-0.9.21.orig/src/gfx/generic/generic_mmx.h DirectFB-0.9.21/src/gfx/generic/generic_mmx.h
|
---|
| 9 | --- DirectFB-0.9.21.orig/src/gfx/generic/generic_mmx.h 2003-06-30 23:26:04.000000000 +0200
|
---|
| 10 | +++ DirectFB-0.9.21/src/gfx/generic/generic_mmx.h 2004-01-08 23:46:40.000000000 +0100
|
---|
| 11 | @@ -172,15 +172,19 @@
|
---|
| 12 | "movq %%mm1, (%1)\n\t"
|
---|
| 13 | "dec %2\n\t"
|
---|
| 14 | "jz 3f\n\t"
|
---|
| 15 | - "addl $8, %1\n\t"
|
---|
| 16 | - "addl %4, %0\n\t"
|
---|
| 17 | - "testl $0xFFFF0000, %0\n\t"
|
---|
| 18 | + "add $8, %1\n\t"
|
---|
| 19 | + "add %4, %0\n\t"
|
---|
| 20 | + "test $0xFFFF0000, %0\n\t"
|
---|
| 21 | "jz 2b\n\t"
|
---|
| 22 | - "movl %0, %%edx\n\t"
|
---|
| 23 | - "andl $0xFFFF0000, %%edx\n\t"
|
---|
| 24 | - "shrl $14, %%edx\n\t"
|
---|
| 25 | + "mov %0, %%edx\n\t"
|
---|
| 26 | + "and $0xFFFF0000, %%edx\n\t"
|
---|
| 27 | + "shr $14, %%edx\n\t"
|
---|
| 28 | +#ifdef __x86_64__
|
---|
| 29 | + "add %%rbx, %3\n\t"
|
---|
| 30 | +#else
|
---|
| 31 | "add %%edx, %3\n\t"
|
---|
| 32 | - "andl $0xFFFF, %0\n\t"
|
---|
| 33 | +#endif
|
---|
| 34 | + "and $0xFFFF, %0\n\t"
|
---|
| 35 | "jmp 1b\n"
|
---|
| 36 | "3:\n\t"
|
---|
| 37 | "emms"
|
---|
| 38 | @@ -201,8 +205,8 @@
|
---|
| 39 | "movd (%2), %%mm1\n\t"
|
---|
| 40 | "punpcklbw %%mm0, %%mm1\n\t"
|
---|
| 41 | "movq %%mm1, (%0)\n\t"
|
---|
| 42 | - "addl $4, %2\n\t"
|
---|
| 43 | - "addl $8, %0\n\t"
|
---|
| 44 | + "add $4, %2\n\t"
|
---|
| 45 | + "add $8, %0\n\t"
|
---|
| 46 | "dec %1\n\t"
|
---|
| 47 | "jnz 1b\n\t"
|
---|
| 48 | "emms"
|
---|
| 49 | @@ -238,7 +242,7 @@
|
---|
| 50 | "dec %1\n\t"
|
---|
| 51 | "jz 2f\n\t"
|
---|
| 52 | "psrlq $16, %%mm0\n\t"
|
---|
| 53 | - "addl $8, %0\n\t"
|
---|
| 54 | + "add $8, %0\n\t"
|
---|
| 55 | /* 2. Konvertierung nach 24 bit interleaved */
|
---|
| 56 | "movq %%mm0, %%mm3\n\t"
|
---|
| 57 | "punpcklwd %%mm3, %%mm3\n\t"
|
---|
| 58 | @@ -252,7 +256,7 @@
|
---|
| 59 | "dec %1\n\t"
|
---|
| 60 | "jz 2f\n\t"
|
---|
| 61 | "psrlq $16, %%mm0\n\t"
|
---|
| 62 | - "addl $8, %0\n\t"
|
---|
| 63 | + "add $8, %0\n\t"
|
---|
| 64 | /* 3. Konvertierung nach 24 bit interleaved */
|
---|
| 65 | "movq %%mm0, %%mm3\n\t"
|
---|
| 66 | "punpcklwd %%mm3, %%mm3\n\t"
|
---|
| 67 | @@ -266,7 +270,7 @@
|
---|
| 68 | "dec %1\n\t"
|
---|
| 69 | "jz 2f\n\t"
|
---|
| 70 | "psrlq $16, %%mm0\n\t"
|
---|
| 71 | - "addl $8, %0\n\t"
|
---|
| 72 | + "add $8, %0\n\t"
|
---|
| 73 | /* 4. Konvertierung nach 24 bit interleaved */
|
---|
| 74 | "movq %%mm0, %%mm3\n\t"
|
---|
| 75 | "punpcklwd %%mm3, %%mm3\n\t"
|
---|
| 76 | @@ -279,8 +283,8 @@
|
---|
| 77 | "movq %%mm3, (%0)\n\t"
|
---|
| 78 | "dec %1\n\t"
|
---|
| 79 | "jz 2f\n\t"
|
---|
| 80 | - "addl $8, %0\n\t"
|
---|
| 81 | - "addl $8, %2\n\t"
|
---|
| 82 | + "add $8, %0\n\t"
|
---|
| 83 | + "add $8, %2\n\t"
|
---|
| 84 | "jmp 1b\n"
|
---|
| 85 | "2:\n\t"
|
---|
| 86 | "emms"
|
---|
| 87 | @@ -304,8 +308,8 @@
|
---|
| 88 | "punpcklbw %%mm6, %%mm0\n\t"
|
---|
| 89 | "por %%mm7, %%mm0\n\t"
|
---|
| 90 | "movq %%mm0, (%0)\n\t"
|
---|
| 91 | - "addl $4, %2\n\t"
|
---|
| 92 | - "addl $8, %0\n\t"
|
---|
| 93 | + "add $4, %2\n\t"
|
---|
| 94 | + "add $8, %0\n\t"
|
---|
| 95 | "dec %1\n\t"
|
---|
| 96 | "jnz 1b\n\t"
|
---|
| 97 | "emms"
|
---|
| 98 | @@ -322,7 +326,7 @@
|
---|
| 99 |
|
---|
| 100 | __asm__ __volatile__ (
|
---|
| 101 | "movq %3, %%mm7\n\t"
|
---|
| 102 | - "cmpl $0, %2\n\t"
|
---|
| 103 | + "cmp $0, %2\n\t"
|
---|
| 104 | "jne 3f\n\t"
|
---|
| 105 | "movq %4, %%mm6\n\t"
|
---|
| 106 | "movd %5, %%mm0\n\t"
|
---|
| 107 | @@ -341,7 +345,7 @@
|
---|
| 108 | "psrlw $8, %%mm0\n\t"
|
---|
| 109 | "movq %%mm0, (%0)\n"
|
---|
| 110 | "1:\n\t"
|
---|
| 111 | - "addl $8, %0\n\t"
|
---|
| 112 | + "add $8, %0\n\t"
|
---|
| 113 | "dec %1\n\t"
|
---|
| 114 | "jnz 4b\n\t"
|
---|
| 115 | "jmp 2f\n\t"
|
---|
| 116 | @@ -360,8 +364,8 @@
|
---|
| 117 | "psrlw $8, %%mm0\n\t"
|
---|
| 118 | "movq %%mm0, (%0)\n"
|
---|
| 119 | "1:\n\t"
|
---|
| 120 | - "addl $8, %2\n\t"
|
---|
| 121 | - "addl $8, %0\n\t"
|
---|
| 122 | + "add $8, %2\n\t"
|
---|
| 123 | + "add $8, %0\n\t"
|
---|
| 124 | "dec %1\n\t"
|
---|
| 125 | "jnz 3b\n\t"
|
---|
| 126 | "2:\n\t"
|
---|
| 127 | @@ -379,7 +383,7 @@
|
---|
| 128 |
|
---|
| 129 | __asm__ __volatile__ (
|
---|
| 130 | "movq %3, %%mm7\n\t"
|
---|
| 131 | - "cmpl $0, %2\n\t"
|
---|
| 132 | + "cmp $0, %2\n\t"
|
---|
| 133 | "jne 3f\n\t"
|
---|
| 134 | "movq %4, %%mm6\n\t"
|
---|
| 135 | "movd %5, %%mm0\n\t"
|
---|
| 136 | @@ -397,7 +401,7 @@
|
---|
| 137 | "psrlw $8, %%mm1\n\t"
|
---|
| 138 | "movq %%mm1, (%0)\n"
|
---|
| 139 | "1:\n\t"
|
---|
| 140 | - "addl $8, %0\n\t"
|
---|
| 141 | + "add $8, %0\n\t"
|
---|
| 142 | "dec %1\n\t"
|
---|
| 143 | "jnz 4b\n\t"
|
---|
| 144 | "jmp 2f\n\t"
|
---|
| 145 | @@ -415,8 +419,8 @@
|
---|
| 146 | "psrlw $8, %%mm1\n\t"
|
---|
| 147 | "movq %%mm1, (%0)\n"
|
---|
| 148 | "1:\n\t"
|
---|
| 149 | - "addl $8, %2\n\t"
|
---|
| 150 | - "addl $8, %0\n\t"
|
---|
| 151 | + "add $8, %2\n\t"
|
---|
| 152 | + "add $8, %0\n\t"
|
---|
| 153 | "dec %1\n\t"
|
---|
| 154 | "jnz 3b\n\t"
|
---|
| 155 | "2:\n\t"
|
---|
| 156 | diff -urN DirectFB-0.9.21.orig/src/misc/cpu_accel.c DirectFB-0.9.21/src/misc/cpu_accel.c
|
---|
| 157 | --- DirectFB-0.9.21.orig/lib/direct/cpu_accel.c 2003-08-15 13:32:45.000000000 +0200
|
---|
| 158 | +++ DirectFB-0.9.21/lib/direct/cpu_accel.c 2004-01-08 23:48:05.000000000 +0100
|
---|
| 159 | @@ -57,6 +57,13 @@
|
---|
| 160 |
|
---|
| 161 | #include "cpu_accel.h"
|
---|
| 162 |
|
---|
| 163 | +#ifdef __x86_64__
|
---|
| 164 | +static __u32 arch_accel (void)
|
---|
| 165 | +{
|
---|
| 166 | + return MM_ACCEL_X86_MMX | MM_ACCEL_X86_SSE | MM_ACCEL_X86_MMXEXT | MM_ACCEL_X86_SSE2 | MM_ACCEL_X86_3DNOW;
|
---|
| 167 | +}
|
---|
| 168 | +#endif
|
---|
| 169 | +
|
---|
| 170 | #ifdef ARCH_X86
|
---|
| 171 | static __u32 arch_accel (void)
|
---|
| 172 | {
|
---|
| 173 | @@ -175,7 +182,7 @@
|
---|
| 174 |
|
---|
| 175 | __u32 dfb_mm_accel (void)
|
---|
| 176 | {
|
---|
| 177 | -#if defined (ARCH_X86) || (defined (ARCH_PPC) && defined (ENABLE_ALTIVEC))
|
---|
| 178 | +#if defined(__x86_64__) || defined (ARCH_X86) || (defined (ARCH_PPC) && defined (ENABLE_ALTIVEC))
|
---|
| 179 | static __u32 accel = ~0U;
|
---|
| 180 |
|
---|
| 181 | if (accel != ~0U)
|
---|
| 182 | @@ -183,7 +190,7 @@
|
---|
| 183 |
|
---|
| 184 | accel = arch_accel ();
|
---|
| 185 |
|
---|
| 186 | -#ifdef USE_SSE
|
---|
| 187 | +#if defined(USE_SSE) && !defined(__x86_64__)
|
---|
| 188 |
|
---|
| 189 | /* test OS support for SSE */
|
---|
| 190 | if (accel & MM_ACCEL_X86_SSE) {
|
---|