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_points3_general) 47848b8605SmrgHIDDEN(_mesa_sse_transform_points3_general) 48848b8605SmrgGLNAME( _mesa_sse_transform_points3_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 CMP_L ( CONST(0), ECX ) /* count == 0 ? */ 61848b8605Smrg JE ( LLBL(K_GTPGR_finish) ) /* yes -> nothing to do. */ 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 75848b8605Smrg 76848b8605SmrgALIGNTEXT32 77848b8605Smrg MOVAPS ( REGOFF(0, EDX), XMM0 ) /* m0 | m1 | m2 | m3 */ 78848b8605Smrg MOVAPS ( REGOFF(16, EDX), XMM1 ) /* m4 | m5 | m6 | m7 */ 79848b8605Smrg MOVAPS ( REGOFF(32, EDX), XMM2 ) /* m8 | m9 | m10 | m11 */ 80848b8605Smrg MOVAPS ( REGOFF(48, EDX), XMM3 ) /* m12 | m13 | m14 | m15 */ 81848b8605Smrg 82848b8605Smrg 83848b8605SmrgALIGNTEXT32 84848b8605SmrgLLBL(K_GTPGR_top): 85848b8605Smrg MOVSS ( REGOFF(0, ESI), XMM4 ) /* | | | ox */ 86848b8605Smrg SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox | ox */ 87848b8605Smrg MOVSS ( REGOFF(4, ESI), XMM5 ) /* | | | oy */ 88848b8605Smrg SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy | oy */ 89848b8605Smrg MOVSS ( REGOFF(8, ESI), XMM6 ) /* | | | oz */ 90848b8605Smrg SHUFPS ( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz | oz */ 91848b8605Smrg 92848b8605Smrg MULPS ( XMM0, XMM4 ) /* m3*ox | m2*ox | m1*ox | m0*ox */ 93848b8605Smrg MULPS ( XMM1, XMM5 ) /* m7*oy | m6*oy | m5*oy | m4*oy */ 94848b8605Smrg MULPS ( XMM2, XMM6 ) /* m11*oz | m10*oz | m9*oz | m8*oz */ 95848b8605Smrg 96848b8605Smrg ADDPS ( XMM5, XMM4 ) 97848b8605Smrg ADDPS ( XMM6, XMM4 ) 98848b8605Smrg ADDPS ( XMM3, XMM4 ) 99848b8605Smrg 100848b8605Smrg MOVAPS ( XMM4, REGOFF(0, EDI) ) 101848b8605Smrg 102848b8605SmrgLLBL(K_GTPGR_skip): 103848b8605Smrg ADD_L ( CONST(16), EDI ) 104848b8605Smrg ADD_L ( EAX, ESI ) 105848b8605Smrg CMP_L ( ECX, EDI ) 106848b8605Smrg JNE ( LLBL(K_GTPGR_top) ) 107848b8605Smrg 108848b8605SmrgLLBL(K_GTPGR_finish): 109848b8605Smrg POP_L ( EDI ) 110848b8605Smrg POP_L ( ESI ) 111848b8605Smrg RET 112848b8605Smrg#undef FRAME_OFFSET 113848b8605Smrg 114848b8605Smrg 115848b8605SmrgALIGNTEXT4 116848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points3_identity) 117848b8605SmrgHIDDEN(_mesa_sse_transform_points3_identity) 118848b8605SmrgGLNAME( _mesa_sse_transform_points3_identity ): 119848b8605Smrg 120848b8605Smrg#define FRAME_OFFSET 8 121848b8605Smrg PUSH_L ( ESI ) 122848b8605Smrg PUSH_L ( EDI ) 123848b8605Smrg 124848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 125848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 126848b8605Smrg 127848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 128848b8605Smrg 129848b8605Smrg TEST_L( ECX, ECX) 130848b8605Smrg JZ( LLBL(K_GTPIR_finish) ) /* count was zero; go to finish */ 131848b8605Smrg 132848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 133848b8605Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 134848b8605Smrg 135848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 136848b8605Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 137848b8605Smrg 138848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 139848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 140848b8605Smrg 141848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 142848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 143848b8605Smrg 144848b8605Smrg CMP_L( ESI, EDI ) 145848b8605Smrg JE( LLBL(K_GTPIR_finish) ) 146848b8605Smrg 147848b8605Smrg 148848b8605SmrgALIGNTEXT32 149848b8605SmrgLLBL(K_GTPIR_top): 150848b8605Smrg MOVLPS ( S(0), XMM0 ) 151848b8605Smrg MOVLPS ( XMM0, D(0) ) 152848b8605Smrg MOVSS ( S(2), XMM0 ) 153848b8605Smrg MOVSS ( XMM0, D(2) ) 154848b8605Smrg 155848b8605SmrgLLBL(K_GTPIR_skip): 156848b8605Smrg ADD_L ( CONST(16), EDI ) 157848b8605Smrg ADD_L ( EAX, ESI ) 158848b8605Smrg CMP_L ( ECX, EDI ) 159848b8605Smrg JNE ( LLBL(K_GTPIR_top) ) 160848b8605Smrg 161848b8605SmrgLLBL(K_GTPIR_finish): 162848b8605Smrg POP_L ( EDI ) 163848b8605Smrg POP_L ( ESI ) 164848b8605Smrg RET 165848b8605Smrg#undef FRAME_OFFSET 166848b8605Smrg 167848b8605Smrg 168848b8605Smrg 169848b8605Smrg 170848b8605SmrgALIGNTEXT4 171848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points3_3d_no_rot) 172848b8605SmrgHIDDEN(_mesa_sse_transform_points3_3d_no_rot) 173848b8605SmrgGLNAME(_mesa_sse_transform_points3_3d_no_rot): 174848b8605Smrg 175848b8605Smrg#define FRAME_OFFSET 8 176848b8605Smrg PUSH_L( ESI ) 177848b8605Smrg PUSH_L( EDI ) 178848b8605Smrg 179848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 180848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 181848b8605Smrg 182848b8605Smrg 183848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 184848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 185848b8605Smrg 186848b8605Smrg TEST_L( ECX, ECX) 187848b8605Smrg JZ( LLBL(K_GTP3DNRR_finish) ) /* count was zero; go to finish */ 188848b8605Smrg 189848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 190848b8605Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 191848b8605Smrg 192848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 193848b8605Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 194848b8605Smrg 195848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 196848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 197848b8605Smrg 198848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 199848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 200848b8605Smrg 201848b8605Smrg XORPS( XMM0, XMM0 ) /* clean the working register */ 202848b8605Smrg 203848b8605SmrgALIGNTEXT32 204848b8605Smrg MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 205848b8605Smrg MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 206848b8605Smrg UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 207848b8605Smrg MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */ 208848b8605Smrg MOVSS ( M(10), XMM3 ) /* - | - | - | m10 */ 209848b8605Smrg MOVSS ( M(14), XMM4 ) /* - | - | - | m14 */ 210848b8605Smrg 211848b8605SmrgALIGNTEXT32 212848b8605SmrgLLBL(K_GTP3DNRR_top): 213848b8605Smrg 214848b8605Smrg MOVLPS ( S(0), XMM0 ) /* - | - | s1 | s0 */ 215848b8605Smrg MULPS ( XMM1, XMM0 ) /* - | - | s1*m5 | s0*m0 */ 216848b8605Smrg ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */ 217848b8605Smrg MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */ 218848b8605Smrg 219848b8605Smrg MOVSS ( S(2), XMM0 ) /* sz */ 220848b8605Smrg MULSS ( XMM3, XMM0 ) /* sz*m10 */ 221848b8605Smrg ADDSS ( XMM4, XMM0 ) /* +m14 */ 222848b8605Smrg MOVSS ( XMM0, D(2) ) /* -> D(2) */ 223848b8605Smrg 224848b8605SmrgLLBL(K_GTP3DNRR_skip): 225848b8605Smrg ADD_L ( CONST(16), EDI ) 226848b8605Smrg ADD_L ( EAX, ESI ) 227848b8605Smrg CMP_L ( ECX, EDI ) 228848b8605Smrg JNE ( LLBL(K_GTP3DNRR_top) ) 229848b8605Smrg 230848b8605SmrgLLBL(K_GTP3DNRR_finish): 231848b8605Smrg POP_L ( EDI ) 232848b8605Smrg POP_L ( ESI ) 233848b8605Smrg RET 234848b8605Smrg#undef FRAME_OFFSET 235848b8605Smrg 236848b8605Smrg 237848b8605Smrg 238848b8605SmrgALIGNTEXT4 239848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points3_perspective) 240848b8605SmrgHIDDEN(_mesa_sse_transform_points3_perspective) 241848b8605SmrgGLNAME(_mesa_sse_transform_points3_perspective): 242848b8605Smrg 243848b8605Smrg#define FRAME_OFFSET 8 244848b8605Smrg PUSH_L ( ESI ) 245848b8605Smrg PUSH_L ( EDI ) 246848b8605Smrg 247848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 248848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 249848b8605Smrg 250848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 251848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 252848b8605Smrg 253848b8605Smrg TEST_L( ECX, ECX) 254848b8605Smrg JZ( LLBL(K_GTP3PR_finish) ) /* count was zero; go to finish */ 255848b8605Smrg 256848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 257848b8605Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 258848b8605Smrg 259848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 260848b8605Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 261848b8605Smrg 262848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 263848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 264848b8605Smrg 265848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 266848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 267848b8605Smrg 268848b8605SmrgALIGNTEXT32 269848b8605Smrg MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 270848b8605Smrg MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 271848b8605Smrg UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 272848b8605Smrg MOVLPS ( M(8), XMM2 ) /* - | - | m9 | m8 */ 273848b8605Smrg MOVSS ( M(10), XMM3 ) /* m10 */ 274848b8605Smrg MOVSS ( M(14), XMM4 ) /* m14 */ 275848b8605Smrg XORPS ( XMM6, XMM6 ) /* 0 */ 276848b8605Smrg 277848b8605SmrgALIGNTEXT32 278848b8605SmrgLLBL(K_GTP3PR_top): 279848b8605Smrg MOVLPS ( S(0), XMM0 ) /* oy | ox */ 280848b8605Smrg MULPS ( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 281848b8605Smrg MOVSS ( S(2), XMM5 ) /* oz */ 282848b8605Smrg SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oz | oz */ 283848b8605Smrg MULPS ( XMM2, XMM5 ) /* oz*m9 | oz*m8 */ 284848b8605Smrg ADDPS ( XMM5, XMM0 ) /* +oy*m5 | +ox*m0 */ 285848b8605Smrg MOVLPS ( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 286848b8605Smrg 287848b8605Smrg MOVSS ( S(2), XMM0 ) /* oz */ 288848b8605Smrg MULSS ( XMM3, XMM0 ) /* oz*m10 */ 289848b8605Smrg ADDSS ( XMM4, XMM0 ) /* +m14 */ 290848b8605Smrg MOVSS ( XMM0, D(2) ) /* ->D(2) */ 291848b8605Smrg 292848b8605Smrg MOVSS ( S(2), XMM0 ) /* oz */ 293848b8605Smrg MOVSS ( XMM6, XMM5 ) /* 0 */ 294848b8605Smrg SUBPS ( XMM0, XMM5 ) /* -oz */ 295848b8605Smrg MOVSS ( XMM5, D(3) ) /* ->D(3) */ 296848b8605Smrg 297848b8605SmrgLLBL(K_GTP3PR_skip): 298848b8605Smrg ADD_L( CONST(16), EDI ) 299848b8605Smrg ADD_L( EAX, ESI ) 300848b8605Smrg CMP_L( ECX, EDI ) 301848b8605Smrg JNE( LLBL(K_GTP3PR_top) ) 302848b8605Smrg 303848b8605SmrgLLBL(K_GTP3PR_finish): 304848b8605Smrg POP_L ( EDI ) 305848b8605Smrg POP_L ( ESI ) 306848b8605Smrg RET 307848b8605Smrg#undef FRAME_OFFSET 308848b8605Smrg 309848b8605Smrg 310848b8605Smrg 311848b8605SmrgALIGNTEXT4 312848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points3_2d) 313848b8605SmrgHIDDEN(_mesa_sse_transform_points3_2d) 314848b8605SmrgGLNAME(_mesa_sse_transform_points3_2d): 315848b8605Smrg 316848b8605Smrg#define FRAME_OFFSET 8 317848b8605Smrg PUSH_L( ESI ) 318848b8605Smrg PUSH_L( EDI ) 319848b8605Smrg 320848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 321848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 322848b8605Smrg 323848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 324848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 325848b8605Smrg 326848b8605Smrg TEST_L( ECX, ECX) 327848b8605Smrg JZ( LLBL(K_GTP3P2DR_finish) ) /* count was zero; go to finish */ 328848b8605Smrg 329848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 330848b8605Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 331848b8605Smrg 332848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 333848b8605Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 334848b8605Smrg 335848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 336848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 337848b8605Smrg 338848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 339848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 340848b8605Smrg 341848b8605SmrgALIGNTEXT32 342848b8605Smrg MOVLPS( M(0), XMM0 ) /* m1 | m0 */ 343848b8605Smrg MOVLPS( M(4), XMM1 ) /* m5 | m4 */ 344848b8605Smrg MOVLPS( M(12), XMM2 ) /* m13 | m12 */ 345848b8605Smrg 346848b8605SmrgALIGNTEXT32 347848b8605SmrgLLBL(K_GTP3P2DR_top): 348848b8605Smrg MOVSS ( S(0), XMM3 ) /* ox */ 349848b8605Smrg SHUFPS ( CONST(0x0), XMM3, XMM3 ) /* ox | ox */ 350848b8605Smrg MULPS ( XMM0, XMM3 ) /* ox*m1 | ox*m0 */ 351848b8605Smrg MOVSS ( S(1), XMM4 ) /* oy */ 352848b8605Smrg SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* oy | oy */ 353848b8605Smrg MULPS ( XMM1, XMM4 ) /* oy*m5 | oy*m4 */ 354848b8605Smrg 355848b8605Smrg ADDPS ( XMM4, XMM3 ) 356848b8605Smrg ADDPS ( XMM2, XMM3 ) 357848b8605Smrg MOVLPS ( XMM3, D(0) ) 358848b8605Smrg 359848b8605Smrg MOVSS ( S(2), XMM3 ) 360848b8605Smrg MOVSS ( XMM3, D(2) ) 361848b8605Smrg 362848b8605SmrgLLBL(K_GTP3P2DR_skip): 363848b8605Smrg ADD_L ( CONST(16), EDI ) 364848b8605Smrg ADD_L ( EAX, ESI ) 365848b8605Smrg CMP_L ( ECX, EDI ) 366848b8605Smrg JNE ( LLBL(K_GTP3P2DR_top) ) 367848b8605Smrg 368848b8605SmrgLLBL(K_GTP3P2DR_finish): 369848b8605Smrg POP_L ( EDI ) 370848b8605Smrg POP_L ( ESI ) 371848b8605Smrg RET 372848b8605Smrg#undef FRAME_OFFSET 373848b8605Smrg 374848b8605Smrg 375848b8605Smrg 376848b8605SmrgALIGNTEXT4 377848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot) 378848b8605SmrgHIDDEN(_mesa_sse_transform_points3_2d_no_rot) 379848b8605SmrgGLNAME(_mesa_sse_transform_points3_2d_no_rot): 380848b8605Smrg 381848b8605Smrg#define FRAME_OFFSET 8 382848b8605Smrg PUSH_L( ESI ) 383848b8605Smrg PUSH_L( EDI ) 384848b8605Smrg 385848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 386848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 387848b8605Smrg 388848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 389848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 390848b8605Smrg 391848b8605Smrg TEST_L( ECX, ECX) 392848b8605Smrg JZ( LLBL(K_GTP3P2DNRR_finish) ) /* count was zero; go to finish */ 393848b8605Smrg 394848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 395848b8605Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 396848b8605Smrg 397848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 398848b8605Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 399848b8605Smrg 400848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 401848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 402848b8605Smrg 403848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 404848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 405848b8605Smrg 406848b8605SmrgALIGNTEXT32 407848b8605Smrg MOVSS ( M(0), XMM1 ) /* m0 */ 408848b8605Smrg MOVSS ( M(5), XMM2 ) /* m5 */ 409848b8605Smrg UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */ 410848b8605Smrg MOVLPS ( M(12), XMM2 ) /* m13 | m12 */ 411848b8605Smrg 412848b8605SmrgALIGNTEXT32 413848b8605SmrgLLBL(K_GTP3P2DNRR_top): 414848b8605Smrg MOVLPS( S(0), XMM0 ) /* oy | ox */ 415848b8605Smrg MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 416848b8605Smrg ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */ 417848b8605Smrg MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 418848b8605Smrg 419848b8605Smrg MOVSS( S(2), XMM0 ) 420848b8605Smrg MOVSS( XMM0, D(2) ) 421848b8605Smrg 422848b8605SmrgLLBL(K_GTP3P2DNRR_skip): 423848b8605Smrg ADD_L( CONST(16), EDI ) 424848b8605Smrg ADD_L( EAX, ESI ) 425848b8605Smrg CMP_L( ECX, EDI ) 426848b8605Smrg JNE( LLBL(K_GTP3P2DNRR_top) ) 427848b8605Smrg 428848b8605SmrgLLBL(K_GTP3P2DNRR_finish): 429848b8605Smrg POP_L( EDI ) 430848b8605Smrg POP_L( ESI ) 431848b8605Smrg RET 432848b8605Smrg#undef FRAME_OFFSET 433848b8605Smrg 434848b8605Smrg 435848b8605Smrg 436848b8605Smrg 437848b8605SmrgALIGNTEXT4 438848b8605SmrgGLOBL GLNAME(_mesa_sse_transform_points3_3d) 439848b8605SmrgHIDDEN(_mesa_sse_transform_points3_3d) 440848b8605SmrgGLNAME(_mesa_sse_transform_points3_3d): 441848b8605Smrg 442848b8605Smrg#define FRAME_OFFSET 8 443848b8605Smrg PUSH_L( ESI ) 444848b8605Smrg PUSH_L( EDI ) 445848b8605Smrg 446848b8605Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 447848b8605Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 448848b8605Smrg 449848b8605Smrg 450848b8605Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 451848b8605Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 452848b8605Smrg 453848b8605Smrg TEST_L( ECX, ECX) 454848b8605Smrg JZ( LLBL(K_GTP3P3DR_finish) ) /* count was zero; go to finish */ 455848b8605Smrg 456848b8605Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 457848b8605Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 458848b8605Smrg 459848b8605Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 460848b8605Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 461848b8605Smrg 462848b8605Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 463848b8605Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 464848b8605Smrg 465848b8605Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 466848b8605Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 467848b8605Smrg 468848b8605Smrg 469848b8605SmrgALIGNTEXT32 470848b8605Smrg MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */ 471848b8605Smrg MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */ 472848b8605Smrg MOVAPS( M(8), XMM2 ) /* m10 | m9 | m8 */ 473848b8605Smrg MOVAPS( M(12), XMM3 ) /* m14 | m13 | m12 */ 474848b8605Smrg 475848b8605SmrgALIGNTEXT32 476848b8605SmrgLLBL(K_GTP3P3DR_top): 477848b8605Smrg MOVSS( S(0), XMM4 ) 478848b8605Smrg SHUFPS( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox */ 479848b8605Smrg MULPS( XMM0, XMM4 ) /* ox*m2 | ox*m1 | ox*m0 */ 480848b8605Smrg 481848b8605Smrg MOVSS( S(1), XMM5 ) 482848b8605Smrg SHUFPS( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy */ 483848b8605Smrg MULPS( XMM1, XMM5 ) /* oy*m6 | oy*m5 | oy*m4 */ 484848b8605Smrg 485848b8605Smrg MOVSS( S(2), XMM6 ) 486848b8605Smrg SHUFPS( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz */ 487848b8605Smrg MULPS( XMM2, XMM6 ) /* oz*m10 | oz*m9 | oz*m8 */ 488848b8605Smrg 489848b8605Smrg ADDPS( XMM5, XMM4 ) /* + | + | + */ 490848b8605Smrg ADDPS( XMM6, XMM4 ) /* + | + | + */ 491848b8605Smrg ADDPS( XMM3, XMM4 ) /* + | + | + */ 492848b8605Smrg 493848b8605Smrg MOVLPS( XMM4, D(0) ) /* => D(1) | => D(0) */ 494848b8605Smrg UNPCKHPS( XMM4, XMM4 ) 495848b8605Smrg MOVSS( XMM4, D(2) ) 496848b8605Smrg 497848b8605SmrgLLBL(K_GTP3P3DR_skip): 498848b8605Smrg ADD_L( CONST(16), EDI ) 499848b8605Smrg ADD_L( EAX, ESI ) 500848b8605Smrg CMP_L( ECX, EDI ) 501848b8605Smrg JNE( LLBL(K_GTP3P3DR_top) ) 502848b8605Smrg 503848b8605SmrgLLBL(K_GTP3P3DR_finish): 504848b8605Smrg POP_L( EDI ) 505848b8605Smrg POP_L( ESI ) 506848b8605Smrg RET 507848b8605Smrg#undef FRAME_OFFSET 508848b8605Smrg#endif 509848b8605Smrg 510848b8605Smrg#if defined (__ELF__) && defined (__linux__) 511848b8605Smrg .section .note.GNU-stack,"",%progbits 512848b8605Smrg#endif 513