| 1 | ; $Id: vertex_3dnow.asm,v 1.1 2000-03-01 18:49:39 jeroen Exp $
|
|---|
| 2 | ;
|
|---|
| 3 | ; 3dNow! transforms
|
|---|
| 4 | ;
|
|---|
| 5 |
|
|---|
| 6 | .586p
|
|---|
| 7 |
|
|---|
| 8 | CODE32 SEGMENT PARA USE32 PUBLIC 'CODE'
|
|---|
| 9 | CODE32 ENDS
|
|---|
| 10 | DATA32 SEGMENT PARA USE32 PUBLIC 'DATA'
|
|---|
| 11 | ASSUME DS:FLAT, SS:FLAT, ES:FLAT
|
|---|
| 12 | DATA32 ENDS
|
|---|
| 13 |
|
|---|
| 14 | CODE32 SEGMENT
|
|---|
| 15 | PUBLIC _gl_3dnow_project_vertices
|
|---|
| 16 | _gl_3dnow_project_vertices:
|
|---|
| 17 |
|
|---|
| 18 | push EBP
|
|---|
| 19 | femms
|
|---|
| 20 | prefetch [ESP + 8]
|
|---|
| 21 |
|
|---|
| 22 | mov ECX, dword ptr [ESP + 8]
|
|---|
| 23 | mov EDX, dword ptr [ESP + 12]
|
|---|
| 24 | mov EBP, dword ptr [ESP + 16]
|
|---|
| 25 | mov EAX, dword ptr [ESP + 20]
|
|---|
| 26 |
|
|---|
| 27 | movd mm6, [ EBP + 48]
|
|---|
| 28 | punpckldq mm6, [ EBP + 52]
|
|---|
| 29 | movd mm5, [EBP]
|
|---|
| 30 | punpckldq mm5, [ EBP + 20]
|
|---|
| 31 | movd mm1, [ EBP + 40]
|
|---|
| 32 | sub EDX, ECX
|
|---|
| 33 |
|
|---|
| 34 | ALIGN 32
|
|---|
| 35 | v16_3dnow_pv_loop_start:
|
|---|
| 36 |
|
|---|
| 37 | prefetch [ECX + 64]
|
|---|
| 38 | movd mm0, [ECX + 12]
|
|---|
| 39 | pfrcp mm0, mm0
|
|---|
| 40 | movd mm7, [ECX + 12]
|
|---|
| 41 | pfrcpit1 mm7, mm0
|
|---|
| 42 | pfrcpit2 mm7, mm0
|
|---|
| 43 | punpckldq mm7, mm7
|
|---|
| 44 | movq mm2, [ECX]
|
|---|
| 45 | pfmul mm2, mm7
|
|---|
| 46 | movd mm3, [ECX + 8]
|
|---|
| 47 | pfmul mm3, mm7
|
|---|
| 48 | movd mm0, [ EBP + 56]
|
|---|
| 49 | pfmul mm3, mm1
|
|---|
| 50 | pfadd mm3, mm0
|
|---|
| 51 | pfmul mm2, mm5
|
|---|
| 52 | pfadd mm2, mm6
|
|---|
| 53 | punpckldq mm3, mm7
|
|---|
| 54 | movq [ECX + 0], mm2
|
|---|
| 55 | movq [ECX + 8], mm3
|
|---|
| 56 | add ECX, EAX
|
|---|
| 57 | sub EDX, EAX
|
|---|
| 58 | ja v16_3dnow_pv_loop_start
|
|---|
| 59 |
|
|---|
| 60 | femms
|
|---|
| 61 | pop EBP
|
|---|
| 62 | ret
|
|---|
| 63 |
|
|---|
| 64 |
|
|---|
| 65 |
|
|---|
| 66 |
|
|---|
| 67 |
|
|---|
| 68 |
|
|---|
| 69 |
|
|---|
| 70 |
|
|---|
| 71 | PUBLIC _gl_3dnow_project_clipped_vertices
|
|---|
| 72 | _gl_3dnow_project_clipped_vertices:
|
|---|
| 73 |
|
|---|
| 74 | push EBP
|
|---|
| 75 | push ESI
|
|---|
| 76 |
|
|---|
| 77 | femms
|
|---|
| 78 |
|
|---|
| 79 | prefetch [ESP + 12]
|
|---|
| 80 |
|
|---|
| 81 | mov ECX, dword ptr [ESP + 12]
|
|---|
| 82 | mov EDX, dword ptr [ESP + 16]
|
|---|
| 83 | mov EBP, dword ptr [ESP + 20]
|
|---|
| 84 | mov EAX, dword ptr [ESP + 24]
|
|---|
| 85 | mov ESI, dword ptr [ESP + 28]
|
|---|
| 86 |
|
|---|
| 87 |
|
|---|
| 88 | movd mm6, [ EBP + 48]
|
|---|
| 89 | punpckldq mm6, [ EBP + 52]
|
|---|
| 90 | movd mm5, [EBP]
|
|---|
| 91 | punpckldq mm5, [ EBP + 20]
|
|---|
| 92 | movd mm1, [ EBP + 40]
|
|---|
| 93 |
|
|---|
| 94 |
|
|---|
| 95 | ALIGN 32
|
|---|
| 96 | v16_3dnow_pcv_loop_start:
|
|---|
| 97 |
|
|---|
| 98 | cmp byte ptr [ESI], 0
|
|---|
| 99 | jne v16_3dnow_pcv_skip
|
|---|
| 100 |
|
|---|
| 101 | movd mm0, [ECX + 12]
|
|---|
| 102 | pfrcp mm0, mm0
|
|---|
| 103 | movd mm7, [ECX + 12]
|
|---|
| 104 | pfrcpit1 mm7, mm0
|
|---|
| 105 | pfrcpit2 mm7, mm0
|
|---|
| 106 | punpckldq mm7, mm7
|
|---|
| 107 | movq mm2, [ECX]
|
|---|
| 108 | pfmul mm2, mm7
|
|---|
| 109 | movd mm3, [ECX + 8]
|
|---|
| 110 | pfmul mm3, mm7
|
|---|
| 111 | movd mm0, [ EBP + 56]
|
|---|
| 112 | pfmul mm3, mm1
|
|---|
| 113 | pfadd mm3, mm0
|
|---|
| 114 | pfmul mm2, mm5
|
|---|
| 115 | pfadd mm2, mm6
|
|---|
| 116 | punpckldq mm3, mm7
|
|---|
| 117 | movq [ECX + 0], mm2
|
|---|
| 118 | movq [ECX + 8], mm3
|
|---|
| 119 |
|
|---|
| 120 | v16_3dnow_pcv_skip:
|
|---|
| 121 | add ECX, EAX
|
|---|
| 122 | inc ESI
|
|---|
| 123 |
|
|---|
| 124 | cmp EDX, ECX
|
|---|
| 125 | jne v16_3dnow_pcv_loop_start
|
|---|
| 126 |
|
|---|
| 127 | femms
|
|---|
| 128 |
|
|---|
| 129 | pop ESI
|
|---|
| 130 | pop EBP
|
|---|
| 131 | ret
|
|---|
| 132 |
|
|---|
| 133 |
|
|---|
| 134 |
|
|---|
| 135 |
|
|---|
| 136 |
|
|---|
| 137 |
|
|---|
| 138 | PUBLIC _gl_v16_3dnow_general_xform
|
|---|
| 139 | _gl_v16_3dnow_general_xform:
|
|---|
| 140 |
|
|---|
| 141 |
|
|---|
| 142 | push EDI
|
|---|
| 143 | push ESI
|
|---|
| 144 |
|
|---|
| 145 | mov EAX, dword ptr [ESP + 12]
|
|---|
| 146 | mov ESI, dword ptr [ESP + 16]
|
|---|
| 147 | mov EDX, dword ptr [ESP + 20]
|
|---|
| 148 | mov EDI, dword ptr [ESP + 24]
|
|---|
| 149 | mov ECX, dword ptr [ESP + 28]
|
|---|
| 150 |
|
|---|
| 151 | femms
|
|---|
| 152 |
|
|---|
| 153 | movq mm7, [ ESI + 48]
|
|---|
| 154 | movq mm3, [ ESI + 56]
|
|---|
| 155 |
|
|---|
| 156 | ALIGN 32
|
|---|
| 157 | v16_3dnow_general_loop:
|
|---|
| 158 | prefetch [EAX + 128]
|
|---|
| 159 | prefetch [EDX + 32]
|
|---|
| 160 |
|
|---|
| 161 | movq mm0, [EDX]
|
|---|
| 162 | movd mm1, [EDX + 8]
|
|---|
| 163 | movq mm4, [ESI]
|
|---|
| 164 | punpckhdq mm2, mm0
|
|---|
| 165 | movq mm5, [ESI + 16]
|
|---|
| 166 | punpckldq mm0, mm0
|
|---|
| 167 | movq mm6, [ESI + 32]
|
|---|
| 168 | pfmul mm4, mm0
|
|---|
| 169 | punpckhdq mm2, mm2
|
|---|
| 170 | pfmul mm5, mm2
|
|---|
| 171 | punpckldq mm1, mm1
|
|---|
| 172 | pfmul mm0, [ESI + 8]
|
|---|
| 173 | pfmul mm2, [ESI + 24]
|
|---|
| 174 | pfmul mm6, mm1
|
|---|
| 175 | pfadd mm5, mm4
|
|---|
| 176 | pfmul mm1, [ESI + 40]
|
|---|
| 177 | pfadd mm2, mm0
|
|---|
| 178 | pfadd mm6, mm5
|
|---|
| 179 | pfadd mm2, mm1
|
|---|
| 180 | pfadd mm6, mm7
|
|---|
| 181 | pfadd mm2, mm3
|
|---|
| 182 | add EDX, EDI
|
|---|
| 183 | movq [EAX], mm6
|
|---|
| 184 | movq [EAX + 8], mm2
|
|---|
| 185 | add EAX, 64
|
|---|
| 186 | dec ECX
|
|---|
| 187 | jne v16_3dnow_general_loop
|
|---|
| 188 |
|
|---|
| 189 | femms
|
|---|
| 190 |
|
|---|
| 191 | pop ESI
|
|---|
| 192 | pop EDI
|
|---|
| 193 | ret
|
|---|
| 194 |
|
|---|
| 195 | CODE32 ENDS
|
|---|
| 196 |
|
|---|
| 197 | end
|
|---|