1 /* $Id: sse_normal.S,v 1.6 2006/04/17 18:58:24 krh Exp $ */
4 * Mesa 3-D graphics library
7 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
9 * Permission is hereby granted, free of charge, to any person obtaining a
10 * copy of this software and associated documentation files (the "Software"),
11 * to deal in the Software without restriction, including without limitation
12 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
13 * and/or sell copies of the Software, and to permit persons to whom the
14 * Software is furnished to do so, subject to the following conditions:
16 * The above copyright notice and this permission notice shall be included
17 * in all copies or substantial portions of the Software.
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
23 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
24 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
28 * - insert PREFETCH instructions to avoid cache-misses !
29 * - some more optimizations are possible...
30 * - for 40-50% more performance in the SSE-functions, the
31 * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned !
36 #include "norm_args.h"
40 #define M(i) REGOFF(i * 4, EDX)
41 #define S(i) REGOFF(i * 4, ESI)
42 #define D(i) REGOFF(i * 4, EDI)
43 #define STRIDE REGOFF(12, ESI)
47 GLOBL GLNAME(_mesa_sse_transform_rescale_normals_no_rot)
48 HIDDEN(_mesa_sse_transform_rescale_normals_no_rot)
49 GLNAME(_mesa_sse_transform_rescale_normals_no_rot):
51 #define FRAME_OFFSET 8
55 MOV_L ( ARG_IN, ESI ) /* ptr to source GLvector3f */
56 MOV_L ( ARG_DEST, EDI ) /* ptr to dest GLvector3f */
58 MOV_L ( ARG_MAT, EDX ) /* ptr to matrix */
59 MOV_L ( REGOFF(MATRIX_INV, EDX), EDX) /* matrix->inv */
61 MOV_L ( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
64 JZ( LLBL(K_G3TRNNRR_finish) ) /* count was zero; go to finish */
66 MOV_L ( STRIDE, EAX ) /* stride */
67 MOV_L ( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest-count */
69 IMUL_L( CONST(16), ECX ) /* count *= 16 */
70 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
72 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
73 ADD_L( EDI, ECX ) /* count += dest ptr */
76 MOVSS ( M(0), XMM1 ) /* m0 */
77 MOVSS ( M(5), XMM2 ) /* m5 */
78 UNPCKLPS( XMM2, XMM1 ) /* m5 | m0 */
79 MOVSS ( ARG_SCALE, XMM0 ) /* scale */
80 SHUFPS ( CONST(0x0), XMM0, XMM0 ) /* scale | scale */
81 MULPS ( XMM0, XMM1 ) /* m5*scale | m0*scale */
82 MULSS ( M(10), XMM0 ) /* m10*scale */
86 MOVLPS ( S(0), XMM2 ) /* uy | ux */
87 MULPS ( XMM1, XMM2 ) /* uy*m5*scale | ux*m0*scale */
88 MOVLPS ( XMM2, D(0) ) /* ->D(1) | D(0) */
90 MOVSS ( S(2), XMM2 ) /* uz */
91 MULSS ( XMM0, XMM2 ) /* uz*m10*scale */
92 MOVSS ( XMM2, D(2) ) /* ->D(2) */
94 LLBL(K_G3TRNNRR_skip):
95 ADD_L ( CONST(16), EDI )
98 JNE ( LLBL(K_G3TRNNRR_top) )
100 LLBL(K_G3TRNNRR_finish):
109 GLOBL GLNAME(_mesa_sse_transform_rescale_normals)
110 HIDDEN(_mesa_sse_transform_rescale_normals)
111 GLNAME(_mesa_sse_transform_rescale_normals):
113 #define FRAME_OFFSET 8
117 MOV_L ( ARG_IN, ESI ) /* ptr to source GLvector3f */
118 MOV_L ( ARG_DEST, EDI ) /* ptr to dest GLvector3f */
120 MOV_L ( ARG_MAT, EDX ) /* ptr to matrix */
121 MOV_L ( REGOFF(MATRIX_INV, EDX), EDX) /* matrix->inv */
123 MOV_L ( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
126 JZ( LLBL(K_G3TRNR_finish) ) /* count was zero; go to finish */
128 MOV_L ( STRIDE, EAX ) /* stride */
129 MOV_L ( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest-count */
131 IMUL_L( CONST(16), ECX ) /* count *= 16 */
132 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
134 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
135 ADD_L( EDI, ECX ) /* count += dest ptr */
138 MOVSS ( M(0), XMM0 ) /* m0 */
139 MOVSS ( M(4), XMM1 ) /* m4 */
140 UNPCKLPS( XMM1, XMM0 ) /* m4 | m0 */
142 MOVSS ( ARG_SCALE, XMM4 ) /* scale */
143 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* scale | scale */
145 MULPS ( XMM4, XMM0 ) /* m4*scale | m0*scale */
146 MOVSS ( M(1), XMM1 ) /* m1 */
147 MOVSS ( M(5), XMM2 ) /* m5 */
148 UNPCKLPS( XMM2, XMM1 ) /* m5 | m1 */
149 MULPS ( XMM4, XMM1 ) /* m5*scale | m1*scale */
150 MOVSS ( M(2), XMM2 ) /* m2 */
151 MOVSS ( M(6), XMM3 ) /* m6 */
152 UNPCKLPS( XMM3, XMM2 ) /* m6 | m2 */
153 MULPS ( XMM4, XMM2 ) /* m6*scale | m2*scale */
155 MOVSS ( M(8), XMM6 ) /* m8 */
156 MULSS ( ARG_SCALE, XMM6 ) /* m8*scale */
157 MOVSS ( M(9), XMM7 ) /* m9 */
158 MULSS ( ARG_SCALE, XMM7 ) /* m9*scale */
162 MOVSS ( S(0), XMM3 ) /* ux */
163 SHUFPS ( CONST(0x0), XMM3, XMM3 ) /* ux | ux */
164 MULPS ( XMM0, XMM3 ) /* ux*m4 | ux*m0 */
165 MOVSS ( S(1), XMM4 ) /* uy */
166 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* uy | uy */
167 MULPS ( XMM1, XMM4 ) /* uy*m5 | uy*m1 */
168 MOVSS ( S(2), XMM5 ) /* uz */
169 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* uz | uz */
170 MULPS ( XMM2, XMM5 ) /* uz*m6 | uz*m2 */
174 MOVLPS ( XMM3, D(0) )
176 MOVSS ( M(10), XMM3 ) /* m10 */
177 MULSS ( ARG_SCALE, XMM3 ) /* m10*scale */
178 MULSS ( S(2), XMM3 ) /* m10*scale*uz */
179 MOVSS ( S(1), XMM4 ) /* uy */
180 MULSS ( XMM7, XMM4 ) /* uy*m9*scale */
181 MOVSS ( S(0), XMM5 ) /* ux */
182 MULSS ( XMM6, XMM5 ) /* ux*m8*scale */
189 ADD_L ( CONST(16), EDI )
192 JNE ( LLBL(K_G3TRNR_top) )
194 LLBL(K_G3TRNR_finish):
202 GLOBL GLNAME(_mesa_sse_transform_normals_no_rot)
203 HIDDEN(_mesa_sse_transform_normals_no_rot)
204 GLNAME(_mesa_sse_transform_normals_no_rot):
206 #define FRAME_OFFSET 8
210 MOV_L ( ARG_IN, ESI ) /* ptr to source GLvector3f */
211 MOV_L ( ARG_DEST, EDI ) /* ptr to dest GLvector3f */
213 MOV_L ( ARG_MAT, EDX ) /* ptr to matrix */
214 MOV_L ( REGOFF(MATRIX_INV, EDX), EDX) /* matrix->inv */
216 MOV_L ( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
219 JZ( LLBL(K_G3TNNRR_finish) ) /* count was zero; go to finish */
221 MOV_L ( STRIDE, EAX ) /* stride */
222 MOV_L ( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest-count */
224 IMUL_L( CONST(16), ECX ) /* count *= 16 */
225 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
227 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
228 ADD_L( EDI, ECX ) /* count += dest ptr */
231 MOVSS( M(0), XMM0 ) /* m0 */
232 MOVSS( M(5), XMM1 ) /* m5 */
233 UNPCKLPS( XMM1, XMM0 ) /* m5 | m0 */
234 MOVSS( M(10), XMM1 ) /* m10 */
238 MOVLPS( S(0), XMM2 ) /* uy | ux */
239 MULPS( XMM0, XMM2 ) /* uy*m5 | ux*m0 */
242 MOVSS( S(2), XMM2 ) /* uz */
243 MULSS( XMM1, XMM2 ) /* uz*m10 */
246 LLBL(K_G3TNNRR_skip):
247 ADD_L ( CONST(16), EDI )
250 JNE ( LLBL(K_G3TNNRR_top) )
252 LLBL(K_G3TNNRR_finish):
259 #if defined (__ELF__) && defined (__linux__)
260 .section .note.GNU-stack,"",%progbits