1848b8605Smrg 2848b8605Smrg/* 3848b8605Smrg * Mesa 3-D graphics library 4848b8605Smrg * 5848b8605Smrg * Copyright (C) 1999-2001 Brian Paul All Rights Reserved. 6848b8605Smrg * 7848b8605Smrg * Permission is hereby granted, free of charge, to any person obtaining a 8848b8605Smrg * copy of this software and associated documentation files (the "Software"), 9848b8605Smrg * to deal in the Software without restriction, including without limitation 10848b8605Smrg * the rights to use, copy, modify, merge, publish, distribute, sublicense, 11848b8605Smrg * and/or sell copies of the Software, and to permit persons to whom the 12848b8605Smrg * Software is furnished to do so, subject to the following conditions: 13848b8605Smrg * 14848b8605Smrg * The above copyright notice and this permission notice shall be included 15848b8605Smrg * in all copies or substantial portions of the Software. 16848b8605Smrg * 17848b8605Smrg * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 18848b8605Smrg * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 19848b8605Smrg * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20848b8605Smrg * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 21848b8605Smrg * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 22848b8605Smrg * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 23848b8605Smrg * OTHER DEALINGS IN THE SOFTWARE. 24848b8605Smrg */ 25848b8605Smrg 26848b8605Smrg/** TODO: 27848b8605Smrg * - insert PREFETCH instructions to avoid cache-misses ! 28848b8605Smrg * - some more optimizations are possible... 29848b8605Smrg * - for 40-50% more performance in the SSE-functions, the 30848b8605Smrg * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned ! 31848b8605Smrg */ 32848b8605Smrg 33848b8605Smrg#ifdef USE_SSE_ASM 34848b8605Smrg#include "assyntax.h" 35848b8605Smrg#include "matypes.h" 36848b8605Smrg#include "xform_args.h" 37848b8605Smrg 38848b8605Smrg SEG_TEXT 39848b8605Smrg 40848b8605Smrg#define S(i) REGOFF(i * 4, ESI) 41848b8605Smrg#define D(i) REGOFF(i * 4, EDI) 42848b8605Smrg#define M(i) REGOFF(i * 4, EDX) 43848b8605Smrg 44848b8605Smrg 45848b8605SmrgALIGNTEXT4 46848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points2_general) 47848b8605SmrgHIDDEN (_mesa_sse_transform_points2_general) 48848b8605SmrgGLNAME( _mesa_sse_transform_points2_general ): 49848b8605Smrg 50848b8605Smrg#define FRAME_OFFSET 8 51848b8605Smrg PUSH_L ( ESI ) 52848b8605Smrg PUSH_L ( EDI ) 53848b8605Smrg 54848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 55848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 56848b8605Smrg 57848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 58848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 59848b8605Smrg 60848b8605Smrg TEST_L( ECX, ECX ) 61848b8605Smrg JZ( LLBL(K_GTP2GR_finish) ) /* count was zero; go to finish */ 62848b8605Smrg 63848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 64848b8605Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 65848b8605Smrg 66848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 67848b8605Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 68848b8605Smrg 69848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 70848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 71848b8605Smrg 72848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 73848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 74848b8605Smrg 75848b8605SmrgALIGNTEXT32 76848b8605Smrg MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */ 77848b8605Smrg MOVAPS( M(4), XMM1 ) /* m7 | m6 | m5 | m4 */ 78848b8605Smrg MOVAPS( M(12), XMM2 ) /* m15 | m14 | m13 | m12 */ 79848b8605Smrg 80848b8605SmrgALIGNTEXT32 81848b8605SmrgLLBL(K_GTP2GR_top): 82848b8605Smrg MOVSS( S(0), XMM3 ) /* ox */ 83848b8605Smrg SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox | ox */ 84848b8605Smrg MULPS( XMM0, XMM3 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */ 85848b8605Smrg MOVSS( S(1), XMM4 ) /* oy */ 86848b8605Smrg SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy | oy */ 87848b8605Smrg MULPS( XMM1, XMM4 ) /* oy*m7 | oy*m6 | oy*m5 | oy*m4 */ 88848b8605Smrg 89848b8605Smrg ADDPS( XMM4, XMM3 ) 90848b8605Smrg ADDPS( XMM2, XMM3 ) 91848b8605Smrg MOVAPS( XMM3, D(0) ) 92848b8605Smrg 93848b8605SmrgLLBL(K_GTP2GR_skip): 94848b8605Smrg ADD_L ( CONST(16), EDI ) 95848b8605Smrg ADD_L ( EAX, ESI ) 96848b8605Smrg CMP_L ( ECX, EDI ) 97848b8605Smrg JNE ( LLBL(K_GTP2GR_top) ) 98848b8605Smrg 99848b8605SmrgLLBL(K_GTP2GR_finish): 100848b8605Smrg POP_L ( EDI ) 101848b8605Smrg POP_L ( ESI ) 102848b8605Smrg RET 103848b8605Smrg#undef FRAME_OFFSET 104848b8605Smrg 105848b8605Smrg 106848b8605SmrgALIGNTEXT4 107848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points2_identity) 108848b8605SmrgHIDDEN(_mesa_sse_transform_points2_identity) 109848b8605SmrgGLNAME( _mesa_sse_transform_points2_identity ): 110848b8605Smrg 111848b8605Smrg#define FRAME_OFFSET 8 112848b8605Smrg PUSH_L ( ESI ) 113848b8605Smrg PUSH_L ( EDI ) 114848b8605Smrg 115848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 116848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 117848b8605Smrg 118848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 119848b8605Smrg 120848b8605Smrg TEST_L( ECX, ECX) 121848b8605Smrg JZ( LLBL(K_GTP2IR_finish) ) /* count was zero; go to finish */ 122848b8605Smrg 123848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 124848b8605Smrg OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 125848b8605Smrg 126848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 127848b8605Smrg MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 128848b8605Smrg 129848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 130848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 131848b8605Smrg 132848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 133848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 134848b8605Smrg 135848b8605Smrg CMP_L( ESI, EDI ) 136848b8605Smrg JE( LLBL(K_GTP2IR_finish) ) 137848b8605Smrg 138848b8605Smrg 139848b8605SmrgALIGNTEXT32 140848b8605SmrgLLBL(K_GTP2IR_top): 141848b8605Smrg MOV_L ( S(0), EDX ) 142848b8605Smrg MOV_L ( EDX, D(0) ) 143848b8605Smrg MOV_L ( S(1), EDX ) 144848b8605Smrg MOV_L ( EDX, D(1) ) 145848b8605Smrg 146848b8605SmrgLLBL(K_GTP2IR_skip): 147848b8605Smrg ADD_L ( CONST(16), EDI ) 148848b8605Smrg ADD_L ( EAX, ESI ) 149848b8605Smrg CMP_L ( ECX, EDI ) 150848b8605Smrg JNE ( LLBL(K_GTP2IR_top) ) 151848b8605Smrg 152848b8605SmrgLLBL(K_GTP2IR_finish): 153848b8605Smrg POP_L ( EDI ) 154848b8605Smrg POP_L ( ESI ) 155848b8605Smrg RET 156848b8605Smrg#undef FRAME_OFFSET 157848b8605Smrg 158848b8605Smrg 159848b8605SmrgALIGNTEXT4 160848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot) 161848b8605SmrgHIDDEN(_mesa_sse_transform_points2_3d_no_rot) 162848b8605SmrgGLNAME(_mesa_sse_transform_points2_3d_no_rot): 163848b8605Smrg 164848b8605Smrg#define FRAME_OFFSET 8 165848b8605Smrg PUSH_L( ESI ) 166848b8605Smrg PUSH_L( EDI ) 167848b8605Smrg 168848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 169848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 170848b8605Smrg 171848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 172848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 173848b8605Smrg 174848b8605Smrg TEST_L( ECX, ECX) 175848b8605Smrg JZ( LLBL(K_GTP23DNRR_finish) ) /* count was zero; go to finish */ 176848b8605Smrg 177848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 178848b8605Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 179848b8605Smrg 180848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 181848b8605Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 182848b8605Smrg 183848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 184848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 185848b8605Smrg 186848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 187848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 188848b8605Smrg 189848b8605Smrg XORPS( XMM0, XMM0 ) /* clean the working register */ 190848b8605Smrg 191848b8605SmrgALIGNTEXT32 192848b8605Smrg MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 193848b8605Smrg MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 194848b8605Smrg UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 195848b8605Smrg MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */ 196848b8605Smrg MOVSS ( M(14), XMM3 ) /* - | - | - | m14 */ 197848b8605Smrg 198848b8605SmrgALIGNTEXT32 199848b8605SmrgLLBL(K_GTP23DNRR_top): 200848b8605Smrg MOVLPS ( S(0), XMM0 ) /* - | - | oy | ox */ 201848b8605Smrg MULPS ( XMM1, XMM0 ) /* - | - | oy*m5 | ox*m0 */ 202848b8605Smrg ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */ 203848b8605Smrg MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */ 204848b8605Smrg 205848b8605Smrg MOVSS ( XMM3, D(2) ) /* -> D(2) */ 206848b8605Smrg 207848b8605SmrgLLBL(K_GTP23DNRR_skip): 208848b8605Smrg ADD_L ( CONST(16), EDI ) 209848b8605Smrg ADD_L ( EAX, ESI ) 210848b8605Smrg CMP_L ( ECX, EDI ) 211848b8605Smrg JNE ( LLBL(K_GTP23DNRR_top) ) 212848b8605Smrg 213848b8605SmrgLLBL(K_GTP23DNRR_finish): 214848b8605Smrg POP_L ( EDI ) 215848b8605Smrg POP_L ( ESI ) 216848b8605Smrg RET 217848b8605Smrg#undef FRAME_OFFSET 218848b8605Smrg 219848b8605Smrg 220848b8605SmrgALIGNTEXT4 221848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points2_perspective) 222848b8605SmrgHIDDEN(_mesa_sse_transform_points2_perspective) 223848b8605SmrgGLNAME(_mesa_sse_transform_points2_perspective): 224848b8605Smrg 225848b8605Smrg#define FRAME_OFFSET 8 226848b8605Smrg PUSH_L ( ESI ) 227848b8605Smrg PUSH_L ( EDI ) 228848b8605Smrg 229848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 230848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 231848b8605Smrg 232848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 233848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 234848b8605Smrg 235848b8605Smrg TEST_L( ECX, ECX) 236848b8605Smrg JZ( LLBL(K_GTP23PR_finish) ) /* count was zero; go to finish */ 237848b8605Smrg 238848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 239848b8605Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 240848b8605Smrg 241848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 242848b8605Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 243848b8605Smrg 244848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 245848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 246848b8605Smrg 247848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 248848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 249848b8605Smrg 250848b8605SmrgALIGNTEXT32 251848b8605Smrg MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 252848b8605Smrg MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 253848b8605Smrg UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 254848b8605Smrg MOVSS ( M(14), XMM3 ) /* m14 */ 255848b8605Smrg XORPS ( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */ 256848b8605Smrg 257848b8605SmrgALIGNTEXT32 258848b8605SmrgLLBL(K_GTP23PR_top): 259848b8605Smrg MOVLPS( S(0), XMM4 ) /* oy | ox */ 260848b8605Smrg MULPS( XMM1, XMM4 ) /* oy*m5 | ox*m0 */ 261848b8605Smrg MOVLPS( XMM4, D(0) ) /* ->D(1) | ->D(0) */ 262848b8605Smrg MOVSS( XMM3, D(2) ) /* ->D(2) */ 263848b8605Smrg MOVSS( XMM0, D(3) ) /* ->D(3) */ 264848b8605Smrg 265848b8605SmrgLLBL(K_GTP23PR_skip): 266848b8605Smrg ADD_L( CONST(16), EDI ) 267848b8605Smrg ADD_L( EAX, ESI ) 268848b8605Smrg CMP_L( ECX, EDI ) 269848b8605Smrg JNE( LLBL(K_GTP23PR_top) ) 270848b8605Smrg 271848b8605SmrgLLBL(K_GTP23PR_finish): 272848b8605Smrg POP_L ( EDI ) 273848b8605Smrg POP_L ( ESI ) 274848b8605Smrg RET 275848b8605Smrg#undef FRAME_OFFSET 276848b8605Smrg 277848b8605Smrg 278848b8605Smrg 279848b8605SmrgALIGNTEXT4 280848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points2_2d) 281848b8605SmrgHIDDEN(_mesa_sse_transform_points2_2d) 282848b8605SmrgGLNAME(_mesa_sse_transform_points2_2d): 283848b8605Smrg 284848b8605Smrg#define FRAME_OFFSET 8 285848b8605Smrg PUSH_L( ESI ) 286848b8605Smrg PUSH_L( EDI ) 287848b8605Smrg 288848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 289848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 290848b8605Smrg 291848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 292848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 293848b8605Smrg 294848b8605Smrg TEST_L( ECX, ECX) 295848b8605Smrg JZ( LLBL(K_GTP23P2DR_finish) ) /* count was zero; go to finish */ 296848b8605Smrg 297848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 298848b8605Smrg OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 299848b8605Smrg 300848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 301848b8605Smrg MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 302848b8605Smrg 303848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 304848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 305848b8605Smrg 306848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 307848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 308848b8605Smrg 309848b8605SmrgALIGNTEXT32 310848b8605Smrg MOVLPS( M(0), XMM0 ) /* m1 | m0 */ 311848b8605Smrg MOVLPS( M(4), XMM1 ) /* m5 | m4 */ 312848b8605Smrg MOVLPS( M(12), XMM2 ) /* m13 | m12 */ 313848b8605Smrg 314848b8605SmrgALIGNTEXT32 315848b8605SmrgLLBL(K_GTP23P2DR_top): 316848b8605Smrg MOVSS( S(0), XMM3 ) /* ox */ 317848b8605Smrg SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox */ 318848b8605Smrg MULPS( XMM0, XMM3 ) /* ox*m1 | ox*m0 */ 319848b8605Smrg 320848b8605Smrg MOVSS( S(1), XMM4 ) /* oy */ 321848b8605Smrg SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy */ 322848b8605Smrg MULPS( XMM1, XMM4 ) /* oy*m5 | oy*m4 */ 323848b8605Smrg 324848b8605Smrg ADDPS( XMM4, XMM3 ) 325848b8605Smrg ADDPS( XMM2, XMM3 ) 326848b8605Smrg MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */ 327848b8605Smrg 328848b8605SmrgLLBL(K_GTP23P2DR_skip): 329848b8605Smrg ADD_L ( CONST(16), EDI ) 330848b8605Smrg ADD_L ( EAX, ESI ) 331848b8605Smrg CMP_L ( ECX, EDI ) 332848b8605Smrg JNE ( LLBL(K_GTP23P2DR_top) ) 333848b8605Smrg 334848b8605SmrgLLBL(K_GTP23P2DR_finish): 335848b8605Smrg POP_L ( EDI ) 336848b8605Smrg POP_L ( ESI ) 337848b8605Smrg RET 338848b8605Smrg#undef FRAME_OFFSET 339848b8605Smrg 340848b8605Smrg 341848b8605Smrg 342848b8605SmrgALIGNTEXT4 343848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot) 344848b8605SmrgHIDDEN(_mesa_sse_transform_points2_2d_no_rot) 345848b8605SmrgGLNAME(_mesa_sse_transform_points2_2d_no_rot): 346848b8605Smrg 347848b8605Smrg#define FRAME_OFFSET 8 348848b8605Smrg PUSH_L( ESI ) 349848b8605Smrg PUSH_L( EDI ) 350848b8605Smrg 351848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 352848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 353848b8605Smrg 354848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 355848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 356848b8605Smrg 357848b8605Smrg TEST_L( ECX, ECX) 358848b8605Smrg JZ( LLBL(K_GTP23P2DNRR_finish) ) /* count was zero; go to finish */ 359848b8605Smrg 360848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 361848b8605Smrg OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 362848b8605Smrg 363848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 364848b8605Smrg MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 365848b8605Smrg 366848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 367848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 368848b8605Smrg 369848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 370848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 371848b8605Smrg 372848b8605SmrgALIGNTEXT32 373848b8605Smrg MOVSS ( M(0), XMM1 ) /* m0 */ 374848b8605Smrg MOVSS ( M(5), XMM2 ) /* m5 */ 375848b8605Smrg UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */ 376848b8605Smrg MOVLPS ( M(12), XMM2 ) /* m13 | m12 */ 377848b8605Smrg 378848b8605SmrgALIGNTEXT32 379848b8605SmrgLLBL(K_GTP23P2DNRR_top): 380848b8605Smrg MOVLPS( S(0), XMM0 ) /* oy | ox */ 381848b8605Smrg MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 382848b8605Smrg ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */ 383848b8605Smrg MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 384848b8605Smrg 385848b8605SmrgLLBL(K_GTP23P2DNRR_skip): 386848b8605Smrg ADD_L( CONST(16), EDI ) 387848b8605Smrg ADD_L( EAX, ESI ) 388848b8605Smrg CMP_L( ECX, EDI ) 389848b8605Smrg JNE( LLBL(K_GTP23P2DNRR_top) ) 390848b8605Smrg 391848b8605SmrgLLBL(K_GTP23P2DNRR_finish): 392848b8605Smrg POP_L( EDI ) 393848b8605Smrg POP_L( ESI ) 394848b8605Smrg RET 395848b8605Smrg#undef FRAME_OFFSET 396848b8605Smrg 397848b8605Smrg 398848b8605Smrg 399848b8605SmrgALIGNTEXT4 400848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points2_3d) 401848b8605SmrgHIDDEN(_mesa_sse_transform_points2_3d) 402848b8605SmrgGLNAME(_mesa_sse_transform_points2_3d): 403848b8605Smrg 404848b8605Smrg#define FRAME_OFFSET 8 405848b8605Smrg PUSH_L( ESI ) 406848b8605Smrg PUSH_L( EDI ) 407848b8605Smrg 408848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 409848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 410848b8605Smrg 411848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 412848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 413848b8605Smrg 414848b8605Smrg TEST_L( ECX, ECX) 415848b8605Smrg JZ( LLBL(K_GTP23P3DR_finish) ) /* count was zero; go to finish */ 416848b8605Smrg 417848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 418848b8605Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 419848b8605Smrg 420848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 421848b8605Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 422848b8605Smrg 423848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 424848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 425848b8605Smrg 426848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 427848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 428848b8605Smrg 429848b8605SmrgALIGNTEXT32 430848b8605Smrg MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */ 431848b8605Smrg MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */ 432848b8605Smrg MOVAPS( M(12), XMM2 ) /* m14 | m13 | m12 */ 433848b8605Smrg 434848b8605SmrgALIGNTEXT32 435848b8605SmrgLLBL(K_GTP23P3DR_top): 436848b8605Smrg MOVSS( S(0), XMM3 ) /* ox */ 437848b8605Smrg SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox */ 438848b8605Smrg MULPS( XMM0, XMM3 ) /* ox*m2 | ox*m1 | ox*m0 */ 439848b8605Smrg 440848b8605Smrg MOVSS( S(1), XMM4 ) /* oy */ 441848b8605Smrg SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy */ 442848b8605Smrg MULPS( XMM1, XMM4 ) /* oy*m6 | oy*m5 | oy*m4 */ 443848b8605Smrg 444848b8605Smrg ADDPS( XMM4, XMM3 ) 445848b8605Smrg ADDPS( XMM2, XMM3 ) 446848b8605Smrg 447848b8605Smrg MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */ 448848b8605Smrg UNPCKHPS( XMM3, XMM3 ) 449848b8605Smrg MOVSS( XMM3, D(2) ) /* ->D(2) */ 450848b8605Smrg 451848b8605SmrgLLBL(K_GTP23P3DR_skip): 452848b8605Smrg ADD_L( CONST(16), EDI ) 453848b8605Smrg ADD_L( EAX, ESI ) 454848b8605Smrg CMP_L( ECX, EDI ) 455848b8605Smrg JNE( LLBL(K_GTP23P3DR_top) ) 456848b8605Smrg 457848b8605SmrgLLBL(K_GTP23P3DR_finish): 458848b8605Smrg POP_L( EDI ) 459848b8605Smrg POP_L( ESI ) 460848b8605Smrg RET 461848b8605Smrg#undef FRAME_OFFSET 462848b8605Smrg#endif 463848b8605Smrg 464848b8605Smrg#if defined (__ELF__) && defined (__linux__) 465848b8605Smrg .section .note.GNU-stack,"",%progbits 466848b8605Smrg#endif 467