| Line | |
|---|
| 1 | asm
|
|---|
| 2 | //vL := vLS- vP*LightDestFactor;
|
|---|
| 3 | movups xmm4, vP
|
|---|
| 4 | movups xmm6,LightDestFactor4
|
|---|
| 5 | mulps xmm6, xmm4 //keep xmm4 = vP
|
|---|
| 6 | movaps xmm0, xmm1
|
|---|
| 7 | subps xmm0, xmm6
|
|---|
| 8 |
|
|---|
| 9 | movups xmm2, vV //preload xmm2 (A)
|
|---|
| 10 |
|
|---|
| 11 | //normalize(vL);
|
|---|
| 12 | movaps xmm6, xmm0 //save
|
|---|
| 13 | mulps xmm0, xmm0
|
|---|
| 14 |
|
|---|
| 15 | {$IFDEF PARAM_USESSE3}
|
|---|
| 16 | haddps xmm0,xmm0
|
|---|
| 17 | haddps xmm0,xmm0
|
|---|
| 18 | {$ELSE}
|
|---|
| 19 | //mix1
|
|---|
| 20 | movaps xmm7, xmm0
|
|---|
| 21 | shufps xmm7, xmm7, $4e
|
|---|
| 22 | addps xmm0, xmm7
|
|---|
| 23 | //mix2
|
|---|
| 24 | movaps xmm7, xmm0
|
|---|
| 25 | shufps xmm7, xmm7, $11
|
|---|
| 26 | addps xmm0, xmm7
|
|---|
| 27 | {$ENDIF}
|
|---|
| 28 |
|
|---|
| 29 | //1/sqrt
|
|---|
| 30 | rsqrtps xmm0, xmm0
|
|---|
| 31 | mulps xmm0, xmm6 //apply
|
|---|
| 32 |
|
|---|
| 33 |
|
|---|
| 34 | //vH := vL + vV; (A)
|
|---|
| 35 | //xmm2 = vV
|
|---|
| 36 | addps xmm2, xmm0
|
|---|
| 37 |
|
|---|
| 38 |
|
|---|
| 39 |
|
|---|
| 40 | //vectproduct(vS1,vS2,vN);
|
|---|
| 41 | movups xmm6, vS1
|
|---|
| 42 | shufps xmm6, xmm6, Shift231
|
|---|
| 43 |
|
|---|
| 44 | movups xmm7, vS2
|
|---|
| 45 | shufps xmm7, xmm7, Shift312
|
|---|
| 46 |
|
|---|
| 47 | movaps xmm3,xmm6
|
|---|
| 48 | mulps xmm3,xmm7
|
|---|
| 49 |
|
|---|
| 50 | movups xmm6, vS1
|
|---|
| 51 | shufps xmm6, xmm6, Shift312
|
|---|
| 52 |
|
|---|
| 53 | movups xmm7, vS2
|
|---|
| 54 | shufps xmm7, xmm7, Shift231
|
|---|
| 55 |
|
|---|
| 56 | mulps xmm7,xmm6
|
|---|
| 57 | subps xmm3,xmm7
|
|---|
| 58 | //end of vectproduct
|
|---|
| 59 |
|
|---|
| 60 |
|
|---|
| 61 | //normalize(vH);
|
|---|
| 62 | movaps xmm6, xmm2 //save
|
|---|
| 63 | mulps xmm2, xmm2
|
|---|
| 64 |
|
|---|
| 65 | {$IFDEF PARAM_USESSE3}
|
|---|
| 66 | haddps xmm2,xmm2
|
|---|
| 67 | haddps xmm2,xmm2
|
|---|
| 68 | {$ELSE}
|
|---|
| 69 | //mix1
|
|---|
| 70 | movaps xmm7, xmm2
|
|---|
| 71 | shufps xmm7, xmm7, $4e
|
|---|
| 72 | addps xmm2, xmm7
|
|---|
| 73 | //mix2
|
|---|
| 74 | movaps xmm7, xmm2
|
|---|
| 75 | shufps xmm7, xmm7, $11
|
|---|
| 76 | addps xmm2, xmm7
|
|---|
| 77 | {$ENDIF}
|
|---|
| 78 |
|
|---|
| 79 | //1/sqrt
|
|---|
| 80 | rsqrtps xmm2, xmm2
|
|---|
| 81 | mulps xmm2, xmm6 //apply
|
|---|
| 82 |
|
|---|
| 83 | //normalize(vN);
|
|---|
| 84 | movaps xmm6, xmm3 //save
|
|---|
| 85 | mulps xmm3, xmm3
|
|---|
| 86 |
|
|---|
| 87 | {$IFDEF PARAM_USESSE3}
|
|---|
| 88 | haddps xmm3,xmm3
|
|---|
| 89 | haddps xmm3,xmm3
|
|---|
| 90 | {$ELSE}
|
|---|
| 91 | //mix1
|
|---|
| 92 | movaps xmm7, xmm3
|
|---|
| 93 | shufps xmm7, xmm7, $4e
|
|---|
| 94 | addps xmm3, xmm7
|
|---|
| 95 | //mix2
|
|---|
| 96 | movaps xmm7, xmm3
|
|---|
| 97 | shufps xmm7, xmm7, $11
|
|---|
| 98 | addps xmm3, xmm7
|
|---|
| 99 | {$ENDIF}
|
|---|
| 100 |
|
|---|
| 101 | //1/sqrt
|
|---|
| 102 | rsqrtps xmm3, xmm3
|
|---|
| 103 | mulps xmm3, xmm6 //apply
|
|---|
| 104 |
|
|---|
| 105 | //LdotN := vN * vL;
|
|---|
| 106 | movups xmm5, xmm3
|
|---|
| 107 | mulps xmm5, xmm0
|
|---|
| 108 | //mix1
|
|---|
| 109 | movaps xmm7, xmm5
|
|---|
| 110 | shufps xmm7, xmm7, $4e
|
|---|
| 111 | addps xmm5, xmm7
|
|---|
| 112 | //mix2
|
|---|
| 113 | movaps xmm7, xmm5
|
|---|
| 114 | shufps xmm7, xmm7, $11
|
|---|
| 115 | addps xmm5, xmm7
|
|---|
| 116 | //:=
|
|---|
| 117 | movss LdotN, xmm5
|
|---|
| 118 |
|
|---|
| 119 | //vD := vLS-vP;
|
|---|
| 120 | movaps xmm5, xmm1
|
|---|
| 121 | subps xmm5, xmm4
|
|---|
| 122 |
|
|---|
| 123 | //dist := sqrt(vD*vD);
|
|---|
| 124 | mulps xmm5, xmm5
|
|---|
| 125 |
|
|---|
| 126 | {$IFDEF PARAM_USESSE3}
|
|---|
| 127 | haddps xmm5,xmm5
|
|---|
| 128 | haddps xmm5,xmm5
|
|---|
| 129 | {$ELSE}
|
|---|
| 130 | //mix1
|
|---|
| 131 | movaps xmm7, xmm5
|
|---|
| 132 | shufps xmm7, xmm7, $4e
|
|---|
| 133 | addps xmm5, xmm7
|
|---|
| 134 | //mix2
|
|---|
| 135 | movaps xmm7, xmm5
|
|---|
| 136 | shufps xmm7, xmm7, $11
|
|---|
| 137 | addps xmm5, xmm7
|
|---|
| 138 | {$ENDIF}
|
|---|
| 139 |
|
|---|
| 140 | //:= sqrt
|
|---|
| 141 | sqrtss xmm5,xmm5
|
|---|
| 142 | movss dist, xmm5
|
|---|
| 143 |
|
|---|
| 144 | //NH := vH * vN;
|
|---|
| 145 | movups xmm6, xmm2
|
|---|
| 146 | mulps xmm6, xmm3
|
|---|
| 147 |
|
|---|
| 148 | {$IFDEF PARAM_USESSE3}
|
|---|
| 149 | haddps xmm6,xmm6
|
|---|
| 150 | haddps xmm6,xmm6
|
|---|
| 151 | {$ELSE}
|
|---|
| 152 | //mix1
|
|---|
| 153 | movaps xmm7, xmm6
|
|---|
| 154 | shufps xmm7, xmm7, $4e
|
|---|
| 155 | addps xmm6, xmm7
|
|---|
| 156 | //mix2
|
|---|
| 157 | movaps xmm7, xmm6
|
|---|
| 158 | shufps xmm7, xmm7, $11
|
|---|
| 159 | addss xmm6, xmm7
|
|---|
| 160 | {$ENDIF}
|
|---|
| 161 | //:=
|
|---|
| 162 | movss NH, xmm6
|
|---|
| 163 | end;
|
|---|
Note:
See
TracBrowser
for help on using the repository browser.