17117f1b4Smrg 27117f1b4Smrg/* 37117f1b4Smrg * Mesa 3-D graphics library 47117f1b4Smrg * 57117f1b4Smrg * Copyright (C) 1999-2001 Brian Paul All Rights Reserved. 67117f1b4Smrg * 77117f1b4Smrg * Permission is hereby granted, free of charge, to any person obtaining a 87117f1b4Smrg * copy of this software and associated documentation files (the "Software"), 97117f1b4Smrg * to deal in the Software without restriction, including without limitation 107117f1b4Smrg * the rights to use, copy, modify, merge, publish, distribute, sublicense, 117117f1b4Smrg * and/or sell copies of the Software, and to permit persons to whom the 127117f1b4Smrg * Software is furnished to do so, subject to the following conditions: 137117f1b4Smrg * 147117f1b4Smrg * The above copyright notice and this permission notice shall be included 157117f1b4Smrg * in all copies or substantial portions of the Software. 167117f1b4Smrg * 177117f1b4Smrg * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 187117f1b4Smrg * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 197117f1b4Smrg * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20af69d88dSmrg * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 21af69d88dSmrg * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 22af69d88dSmrg * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 23af69d88dSmrg * OTHER DEALINGS IN THE SOFTWARE. 247117f1b4Smrg */ 257117f1b4Smrg 267117f1b4Smrg/** TODO: 277117f1b4Smrg * - insert PREFETCH instructions to avoid cache-misses ! 287117f1b4Smrg * - some more optimizations are possible... 297117f1b4Smrg * - for 40-50% more performance in the SSE-functions, the 307117f1b4Smrg * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned ! 317117f1b4Smrg */ 327117f1b4Smrg 337117f1b4Smrg#ifdef USE_SSE_ASM 34c1f859d4Smrg#include "assyntax.h" 357ec681f3Smrg#define MATH_ASM_PTR_SIZE 4 367ec681f3Smrg#include "math/m_vector_asm.h" 377117f1b4Smrg#include "xform_args.h" 387117f1b4Smrg 397117f1b4Smrg SEG_TEXT 407117f1b4Smrg 417117f1b4Smrg#define S(i) REGOFF(i * 4, ESI) 427117f1b4Smrg#define D(i) REGOFF(i * 4, EDI) 437117f1b4Smrg#define M(i) REGOFF(i * 4, EDX) 447117f1b4Smrg 457117f1b4Smrg 467117f1b4SmrgALIGNTEXT4 477117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points3_general) 487117f1b4SmrgHIDDEN(_mesa_sse_transform_points3_general) 497117f1b4SmrgGLNAME( _mesa_sse_transform_points3_general ): 507ec681f3Smrg _CET_ENDBR 517117f1b4Smrg#define FRAME_OFFSET 8 527117f1b4Smrg PUSH_L ( ESI ) 537117f1b4Smrg PUSH_L ( EDI ) 547117f1b4Smrg 557117f1b4Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 567117f1b4Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 577117f1b4Smrg 587117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 597117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 607117f1b4Smrg 617117f1b4Smrg CMP_L ( CONST(0), ECX ) /* count == 0 ? */ 627117f1b4Smrg JE ( LLBL(K_GTPGR_finish) ) /* yes -> nothing to do. */ 637117f1b4Smrg 647117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 657117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 667117f1b4Smrg 677117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 687117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 697117f1b4Smrg 707117f1b4Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 717117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 727117f1b4Smrg 737117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 747117f1b4Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 757117f1b4Smrg 767117f1b4Smrg 777117f1b4SmrgALIGNTEXT32 787117f1b4Smrg MOVAPS ( REGOFF(0, EDX), XMM0 ) /* m0 | m1 | m2 | m3 */ 797117f1b4Smrg MOVAPS ( REGOFF(16, EDX), XMM1 ) /* m4 | m5 | m6 | m7 */ 807117f1b4Smrg MOVAPS ( REGOFF(32, EDX), XMM2 ) /* m8 | m9 | m10 | m11 */ 817117f1b4Smrg MOVAPS ( REGOFF(48, EDX), XMM3 ) /* m12 | m13 | m14 | m15 */ 827117f1b4Smrg 837117f1b4Smrg 847117f1b4SmrgALIGNTEXT32 857117f1b4SmrgLLBL(K_GTPGR_top): 867117f1b4Smrg MOVSS ( REGOFF(0, ESI), XMM4 ) /* | | | ox */ 877117f1b4Smrg SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox | ox */ 887117f1b4Smrg MOVSS ( REGOFF(4, ESI), XMM5 ) /* | | | oy */ 897117f1b4Smrg SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy | oy */ 907117f1b4Smrg MOVSS ( REGOFF(8, ESI), XMM6 ) /* | | | oz */ 917117f1b4Smrg SHUFPS ( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz | oz */ 927117f1b4Smrg 937117f1b4Smrg MULPS ( XMM0, XMM4 ) /* m3*ox | m2*ox | m1*ox | m0*ox */ 947117f1b4Smrg MULPS ( XMM1, XMM5 ) /* m7*oy | m6*oy | m5*oy | m4*oy */ 957117f1b4Smrg MULPS ( XMM2, XMM6 ) /* m11*oz | m10*oz | m9*oz | m8*oz */ 967117f1b4Smrg 977117f1b4Smrg ADDPS ( XMM5, XMM4 ) 987117f1b4Smrg ADDPS ( XMM6, XMM4 ) 997117f1b4Smrg ADDPS ( XMM3, XMM4 ) 1007117f1b4Smrg 1017117f1b4Smrg MOVAPS ( XMM4, REGOFF(0, EDI) ) 1027117f1b4Smrg 1037117f1b4SmrgLLBL(K_GTPGR_skip): 1047117f1b4Smrg ADD_L ( CONST(16), EDI ) 1057117f1b4Smrg ADD_L ( EAX, ESI ) 1067117f1b4Smrg CMP_L ( ECX, EDI ) 1077117f1b4Smrg JNE ( LLBL(K_GTPGR_top) ) 1087117f1b4Smrg 1097117f1b4SmrgLLBL(K_GTPGR_finish): 1107117f1b4Smrg POP_L ( EDI ) 1117117f1b4Smrg POP_L ( ESI ) 1127117f1b4Smrg RET 1137117f1b4Smrg#undef FRAME_OFFSET 1147117f1b4Smrg 1157117f1b4Smrg 1167117f1b4SmrgALIGNTEXT4 1177117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points3_identity) 1187117f1b4SmrgHIDDEN(_mesa_sse_transform_points3_identity) 1197117f1b4SmrgGLNAME( _mesa_sse_transform_points3_identity ): 1207ec681f3Smrg _CET_ENDBR 1217117f1b4Smrg#define FRAME_OFFSET 8 1227117f1b4Smrg PUSH_L ( ESI ) 1237117f1b4Smrg PUSH_L ( EDI ) 1247117f1b4Smrg 1257117f1b4Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 1267117f1b4Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 1277117f1b4Smrg 1287117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 1297117f1b4Smrg 1307117f1b4Smrg TEST_L( ECX, ECX) 1317117f1b4Smrg JZ( LLBL(K_GTPIR_finish) ) /* count was zero; go to finish */ 1327117f1b4Smrg 1337117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 1347117f1b4Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 1357117f1b4Smrg 1367117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 1377117f1b4Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 1387117f1b4Smrg 1397117f1b4Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 1407117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 1417117f1b4Smrg 1427117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 1437117f1b4Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 1447117f1b4Smrg 1457117f1b4Smrg CMP_L( ESI, EDI ) 1467117f1b4Smrg JE( LLBL(K_GTPIR_finish) ) 1477117f1b4Smrg 1487117f1b4Smrg 1497117f1b4SmrgALIGNTEXT32 1507117f1b4SmrgLLBL(K_GTPIR_top): 1517117f1b4Smrg MOVLPS ( S(0), XMM0 ) 1527117f1b4Smrg MOVLPS ( XMM0, D(0) ) 1537117f1b4Smrg MOVSS ( S(2), XMM0 ) 1547117f1b4Smrg MOVSS ( XMM0, D(2) ) 1557117f1b4Smrg 1567117f1b4SmrgLLBL(K_GTPIR_skip): 1577117f1b4Smrg ADD_L ( CONST(16), EDI ) 1587117f1b4Smrg ADD_L ( EAX, ESI ) 1597117f1b4Smrg CMP_L ( ECX, EDI ) 1607117f1b4Smrg JNE ( LLBL(K_GTPIR_top) ) 1617117f1b4Smrg 1627117f1b4SmrgLLBL(K_GTPIR_finish): 1637117f1b4Smrg POP_L ( EDI ) 1647117f1b4Smrg POP_L ( ESI ) 1657117f1b4Smrg RET 1667117f1b4Smrg#undef FRAME_OFFSET 1677117f1b4Smrg 1687117f1b4Smrg 1697117f1b4Smrg 1707117f1b4Smrg 1717117f1b4SmrgALIGNTEXT4 1727117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points3_3d_no_rot) 1737117f1b4SmrgHIDDEN(_mesa_sse_transform_points3_3d_no_rot) 1747117f1b4SmrgGLNAME(_mesa_sse_transform_points3_3d_no_rot): 1757ec681f3Smrg _CET_ENDBR 1767117f1b4Smrg#define FRAME_OFFSET 8 1777117f1b4Smrg PUSH_L( ESI ) 1787117f1b4Smrg PUSH_L( EDI ) 1797117f1b4Smrg 1807117f1b4Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 1817117f1b4Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 1827117f1b4Smrg 1837117f1b4Smrg 1847117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 1857117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 1867117f1b4Smrg 1877117f1b4Smrg TEST_L( ECX, ECX) 1887117f1b4Smrg JZ( LLBL(K_GTP3DNRR_finish) ) /* count was zero; go to finish */ 1897117f1b4Smrg 1907117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 1917117f1b4Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 1927117f1b4Smrg 1937117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 1947117f1b4Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 1957117f1b4Smrg 1967117f1b4Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 1977117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 1987117f1b4Smrg 1997117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 2007117f1b4Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 2017117f1b4Smrg 2024a49301eSmrg XORPS( XMM0, XMM0 ) /* clean the working register */ 2037117f1b4Smrg 2047117f1b4SmrgALIGNTEXT32 2057117f1b4Smrg MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 2067117f1b4Smrg MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 2077117f1b4Smrg UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 2087117f1b4Smrg MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */ 2097117f1b4Smrg MOVSS ( M(10), XMM3 ) /* - | - | - | m10 */ 2107117f1b4Smrg MOVSS ( M(14), XMM4 ) /* - | - | - | m14 */ 2117117f1b4Smrg 2127117f1b4SmrgALIGNTEXT32 2137117f1b4SmrgLLBL(K_GTP3DNRR_top): 2147117f1b4Smrg 2157117f1b4Smrg MOVLPS ( S(0), XMM0 ) /* - | - | s1 | s0 */ 2167117f1b4Smrg MULPS ( XMM1, XMM0 ) /* - | - | s1*m5 | s0*m0 */ 2177117f1b4Smrg ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */ 2187117f1b4Smrg MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */ 2197117f1b4Smrg 2207117f1b4Smrg MOVSS ( S(2), XMM0 ) /* sz */ 2217117f1b4Smrg MULSS ( XMM3, XMM0 ) /* sz*m10 */ 2227117f1b4Smrg ADDSS ( XMM4, XMM0 ) /* +m14 */ 2237117f1b4Smrg MOVSS ( XMM0, D(2) ) /* -> D(2) */ 2247117f1b4Smrg 2257117f1b4SmrgLLBL(K_GTP3DNRR_skip): 2267117f1b4Smrg ADD_L ( CONST(16), EDI ) 2277117f1b4Smrg ADD_L ( EAX, ESI ) 2287117f1b4Smrg CMP_L ( ECX, EDI ) 2297117f1b4Smrg JNE ( LLBL(K_GTP3DNRR_top) ) 2307117f1b4Smrg 2317117f1b4SmrgLLBL(K_GTP3DNRR_finish): 2327117f1b4Smrg POP_L ( EDI ) 2337117f1b4Smrg POP_L ( ESI ) 2347117f1b4Smrg RET 2357117f1b4Smrg#undef FRAME_OFFSET 2367117f1b4Smrg 2377117f1b4Smrg 2387117f1b4Smrg 2397117f1b4SmrgALIGNTEXT4 2407117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points3_perspective) 2417117f1b4SmrgHIDDEN(_mesa_sse_transform_points3_perspective) 2427117f1b4SmrgGLNAME(_mesa_sse_transform_points3_perspective): 2437ec681f3Smrg _CET_ENDBR 2447117f1b4Smrg#define FRAME_OFFSET 8 2457117f1b4Smrg PUSH_L ( ESI ) 2467117f1b4Smrg PUSH_L ( EDI ) 2477117f1b4Smrg 2487117f1b4Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 2497117f1b4Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 2507117f1b4Smrg 2517117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 2527117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 2537117f1b4Smrg 2547117f1b4Smrg TEST_L( ECX, ECX) 2557117f1b4Smrg JZ( LLBL(K_GTP3PR_finish) ) /* count was zero; go to finish */ 2567117f1b4Smrg 2577117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 2587117f1b4Smrg OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 2597117f1b4Smrg 2607117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 2617117f1b4Smrg MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 2627117f1b4Smrg 2637117f1b4Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 2647117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 2657117f1b4Smrg 2667117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 2677117f1b4Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 2687117f1b4Smrg 2697117f1b4SmrgALIGNTEXT32 2707117f1b4Smrg MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 2717117f1b4Smrg MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 2727117f1b4Smrg UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 2737117f1b4Smrg MOVLPS ( M(8), XMM2 ) /* - | - | m9 | m8 */ 2747117f1b4Smrg MOVSS ( M(10), XMM3 ) /* m10 */ 2757117f1b4Smrg MOVSS ( M(14), XMM4 ) /* m14 */ 2767117f1b4Smrg XORPS ( XMM6, XMM6 ) /* 0 */ 2777117f1b4Smrg 2787117f1b4SmrgALIGNTEXT32 2797117f1b4SmrgLLBL(K_GTP3PR_top): 2807117f1b4Smrg MOVLPS ( S(0), XMM0 ) /* oy | ox */ 2817117f1b4Smrg MULPS ( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 2827117f1b4Smrg MOVSS ( S(2), XMM5 ) /* oz */ 2837117f1b4Smrg SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oz | oz */ 2847117f1b4Smrg MULPS ( XMM2, XMM5 ) /* oz*m9 | oz*m8 */ 2857117f1b4Smrg ADDPS ( XMM5, XMM0 ) /* +oy*m5 | +ox*m0 */ 2867117f1b4Smrg MOVLPS ( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 2877117f1b4Smrg 2887117f1b4Smrg MOVSS ( S(2), XMM0 ) /* oz */ 2897117f1b4Smrg MULSS ( XMM3, XMM0 ) /* oz*m10 */ 2907117f1b4Smrg ADDSS ( XMM4, XMM0 ) /* +m14 */ 2917117f1b4Smrg MOVSS ( XMM0, D(2) ) /* ->D(2) */ 2927117f1b4Smrg 2937117f1b4Smrg MOVSS ( S(2), XMM0 ) /* oz */ 2947117f1b4Smrg MOVSS ( XMM6, XMM5 ) /* 0 */ 2957117f1b4Smrg SUBPS ( XMM0, XMM5 ) /* -oz */ 2967117f1b4Smrg MOVSS ( XMM5, D(3) ) /* ->D(3) */ 2977117f1b4Smrg 2987117f1b4SmrgLLBL(K_GTP3PR_skip): 2997117f1b4Smrg ADD_L( CONST(16), EDI ) 3007117f1b4Smrg ADD_L( EAX, ESI ) 3017117f1b4Smrg CMP_L( ECX, EDI ) 3027117f1b4Smrg JNE( LLBL(K_GTP3PR_top) ) 3037117f1b4Smrg 3047117f1b4SmrgLLBL(K_GTP3PR_finish): 3057117f1b4Smrg POP_L ( EDI ) 3067117f1b4Smrg POP_L ( ESI ) 3077117f1b4Smrg RET 3087117f1b4Smrg#undef FRAME_OFFSET 3097117f1b4Smrg 3107117f1b4Smrg 3117117f1b4Smrg 3127117f1b4SmrgALIGNTEXT4 3137117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points3_2d) 3147117f1b4SmrgHIDDEN(_mesa_sse_transform_points3_2d) 3157117f1b4SmrgGLNAME(_mesa_sse_transform_points3_2d): 3167ec681f3Smrg _CET_ENDBR 3177117f1b4Smrg#define FRAME_OFFSET 8 3187117f1b4Smrg PUSH_L( ESI ) 3197117f1b4Smrg PUSH_L( EDI ) 3207117f1b4Smrg 3217117f1b4Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 3227117f1b4Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 3237117f1b4Smrg 3247117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 3257117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 3267117f1b4Smrg 3277117f1b4Smrg TEST_L( ECX, ECX) 3287117f1b4Smrg JZ( LLBL(K_GTP3P2DR_finish) ) /* count was zero; go to finish */ 3297117f1b4Smrg 3307117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 3317117f1b4Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 3327117f1b4Smrg 3337117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 3347117f1b4Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 3357117f1b4Smrg 3367117f1b4Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 3377117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 3387117f1b4Smrg 3397117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 3407117f1b4Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 3417117f1b4Smrg 3427117f1b4SmrgALIGNTEXT32 3437117f1b4Smrg MOVLPS( M(0), XMM0 ) /* m1 | m0 */ 3447117f1b4Smrg MOVLPS( M(4), XMM1 ) /* m5 | m4 */ 3457117f1b4Smrg MOVLPS( M(12), XMM2 ) /* m13 | m12 */ 3467117f1b4Smrg 3477117f1b4SmrgALIGNTEXT32 3487117f1b4SmrgLLBL(K_GTP3P2DR_top): 3497117f1b4Smrg MOVSS ( S(0), XMM3 ) /* ox */ 3507117f1b4Smrg SHUFPS ( CONST(0x0), XMM3, XMM3 ) /* ox | ox */ 3517117f1b4Smrg MULPS ( XMM0, XMM3 ) /* ox*m1 | ox*m0 */ 3527117f1b4Smrg MOVSS ( S(1), XMM4 ) /* oy */ 3537117f1b4Smrg SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* oy | oy */ 3547117f1b4Smrg MULPS ( XMM1, XMM4 ) /* oy*m5 | oy*m4 */ 3557117f1b4Smrg 3567117f1b4Smrg ADDPS ( XMM4, XMM3 ) 3577117f1b4Smrg ADDPS ( XMM2, XMM3 ) 3587117f1b4Smrg MOVLPS ( XMM3, D(0) ) 3597117f1b4Smrg 3607117f1b4Smrg MOVSS ( S(2), XMM3 ) 3617117f1b4Smrg MOVSS ( XMM3, D(2) ) 3627117f1b4Smrg 3637117f1b4SmrgLLBL(K_GTP3P2DR_skip): 3647117f1b4Smrg ADD_L ( CONST(16), EDI ) 3657117f1b4Smrg ADD_L ( EAX, ESI ) 3667117f1b4Smrg CMP_L ( ECX, EDI ) 3677117f1b4Smrg JNE ( LLBL(K_GTP3P2DR_top) ) 3687117f1b4Smrg 3697117f1b4SmrgLLBL(K_GTP3P2DR_finish): 3707117f1b4Smrg POP_L ( EDI ) 3717117f1b4Smrg POP_L ( ESI ) 3727117f1b4Smrg RET 3737117f1b4Smrg#undef FRAME_OFFSET 3747117f1b4Smrg 3757117f1b4Smrg 3767117f1b4Smrg 3777117f1b4SmrgALIGNTEXT4 3787117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot) 3797117f1b4SmrgHIDDEN(_mesa_sse_transform_points3_2d_no_rot) 3807117f1b4SmrgGLNAME(_mesa_sse_transform_points3_2d_no_rot): 3817ec681f3Smrg _CET_ENDBR 3827117f1b4Smrg#define FRAME_OFFSET 8 3837117f1b4Smrg PUSH_L( ESI ) 3847117f1b4Smrg PUSH_L( EDI ) 3857117f1b4Smrg 3867117f1b4Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 3877117f1b4Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 3887117f1b4Smrg 3897117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 3907117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 3917117f1b4Smrg 3927117f1b4Smrg TEST_L( ECX, ECX) 3937117f1b4Smrg JZ( LLBL(K_GTP3P2DNRR_finish) ) /* count was zero; go to finish */ 3947117f1b4Smrg 3957117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 3967117f1b4Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 3977117f1b4Smrg 3987117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 3997117f1b4Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 4007117f1b4Smrg 4017117f1b4Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 4027117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 4037117f1b4Smrg 4047117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 4057117f1b4Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 4067117f1b4Smrg 4077117f1b4SmrgALIGNTEXT32 4087117f1b4Smrg MOVSS ( M(0), XMM1 ) /* m0 */ 4097117f1b4Smrg MOVSS ( M(5), XMM2 ) /* m5 */ 4107117f1b4Smrg UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */ 4117117f1b4Smrg MOVLPS ( M(12), XMM2 ) /* m13 | m12 */ 4127117f1b4Smrg 4137117f1b4SmrgALIGNTEXT32 4147117f1b4SmrgLLBL(K_GTP3P2DNRR_top): 4157117f1b4Smrg MOVLPS( S(0), XMM0 ) /* oy | ox */ 4167117f1b4Smrg MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 4177117f1b4Smrg ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */ 4187117f1b4Smrg MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 4197117f1b4Smrg 4207117f1b4Smrg MOVSS( S(2), XMM0 ) 4217117f1b4Smrg MOVSS( XMM0, D(2) ) 4227117f1b4Smrg 4237117f1b4SmrgLLBL(K_GTP3P2DNRR_skip): 4247117f1b4Smrg ADD_L( CONST(16), EDI ) 4257117f1b4Smrg ADD_L( EAX, ESI ) 4267117f1b4Smrg CMP_L( ECX, EDI ) 4277117f1b4Smrg JNE( LLBL(K_GTP3P2DNRR_top) ) 4287117f1b4Smrg 4297117f1b4SmrgLLBL(K_GTP3P2DNRR_finish): 4307117f1b4Smrg POP_L( EDI ) 4317117f1b4Smrg POP_L( ESI ) 4327117f1b4Smrg RET 4337117f1b4Smrg#undef FRAME_OFFSET 4347117f1b4Smrg 4357117f1b4Smrg 4367117f1b4Smrg 4377117f1b4Smrg 4387117f1b4SmrgALIGNTEXT4 4397117f1b4SmrgGLOBL GLNAME(_mesa_sse_transform_points3_3d) 4407117f1b4SmrgHIDDEN(_mesa_sse_transform_points3_3d) 4417117f1b4SmrgGLNAME(_mesa_sse_transform_points3_3d): 4427ec681f3Smrg _CET_ENDBR 4437117f1b4Smrg#define FRAME_OFFSET 8 4447117f1b4Smrg PUSH_L( ESI ) 4457117f1b4Smrg PUSH_L( EDI ) 4467117f1b4Smrg 4477117f1b4Smrg MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 4487117f1b4Smrg MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 4497117f1b4Smrg 4507117f1b4Smrg 4517117f1b4Smrg MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 4527117f1b4Smrg MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 4537117f1b4Smrg 4547117f1b4Smrg TEST_L( ECX, ECX) 4557117f1b4Smrg JZ( LLBL(K_GTP3P3DR_finish) ) /* count was zero; go to finish */ 4567117f1b4Smrg 4577117f1b4Smrg MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 4587117f1b4Smrg OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 4597117f1b4Smrg 4607117f1b4Smrg MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 4617117f1b4Smrg MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 4627117f1b4Smrg 4637117f1b4Smrg SHL_L( CONST(4), ECX ) /* count *= 16 */ 4647117f1b4Smrg MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 4657117f1b4Smrg 4667117f1b4Smrg MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 4677117f1b4Smrg ADD_L( EDI, ECX ) /* count += dest ptr */ 4687117f1b4Smrg 4697117f1b4Smrg 4707117f1b4SmrgALIGNTEXT32 4717117f1b4Smrg MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */ 4727117f1b4Smrg MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */ 4737117f1b4Smrg MOVAPS( M(8), XMM2 ) /* m10 | m9 | m8 */ 4747117f1b4Smrg MOVAPS( M(12), XMM3 ) /* m14 | m13 | m12 */ 4757117f1b4Smrg 4767117f1b4SmrgALIGNTEXT32 4777117f1b4SmrgLLBL(K_GTP3P3DR_top): 4787117f1b4Smrg MOVSS( S(0), XMM4 ) 4797117f1b4Smrg SHUFPS( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox */ 4807117f1b4Smrg MULPS( XMM0, XMM4 ) /* ox*m2 | ox*m1 | ox*m0 */ 4817117f1b4Smrg 4827117f1b4Smrg MOVSS( S(1), XMM5 ) 4837117f1b4Smrg SHUFPS( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy */ 4847117f1b4Smrg MULPS( XMM1, XMM5 ) /* oy*m6 | oy*m5 | oy*m4 */ 4857117f1b4Smrg 4867117f1b4Smrg MOVSS( S(2), XMM6 ) 4877117f1b4Smrg SHUFPS( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz */ 4887117f1b4Smrg MULPS( XMM2, XMM6 ) /* oz*m10 | oz*m9 | oz*m8 */ 4897117f1b4Smrg 4907117f1b4Smrg ADDPS( XMM5, XMM4 ) /* + | + | + */ 4917117f1b4Smrg ADDPS( XMM6, XMM4 ) /* + | + | + */ 4927117f1b4Smrg ADDPS( XMM3, XMM4 ) /* + | + | + */ 4937117f1b4Smrg 4947117f1b4Smrg MOVLPS( XMM4, D(0) ) /* => D(1) | => D(0) */ 4957117f1b4Smrg UNPCKHPS( XMM4, XMM4 ) 4967117f1b4Smrg MOVSS( XMM4, D(2) ) 4977117f1b4Smrg 4987117f1b4SmrgLLBL(K_GTP3P3DR_skip): 4997117f1b4Smrg ADD_L( CONST(16), EDI ) 5007117f1b4Smrg ADD_L( EAX, ESI ) 5017117f1b4Smrg CMP_L( ECX, EDI ) 5027117f1b4Smrg JNE( LLBL(K_GTP3P3DR_top) ) 5037117f1b4Smrg 5047117f1b4SmrgLLBL(K_GTP3P3DR_finish): 5057117f1b4Smrg POP_L( EDI ) 5067117f1b4Smrg POP_L( ESI ) 5077117f1b4Smrg RET 5087117f1b4Smrg#undef FRAME_OFFSET 5097117f1b4Smrg#endif 5107117f1b4Smrg 5117117f1b4Smrg#if defined (__ELF__) && defined (__linux__) 5127117f1b4Smrg .section .note.GNU-stack,"",%progbits 5137117f1b4Smrg#endif 514