1 /**************************************************************************
3 * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
28 #include "pipe/p_config.h"
30 #if defined(PIPE_ARCH_X86)
32 #include "util/u_debug.h"
33 #include "pipe/p_shader_tokens.h"
34 #include "util/u_math.h"
35 #include "util/u_memory.h"
36 #if defined(PIPE_ARCH_SSE)
37 #include "util/u_sse.h"
39 #include "tgsi/tgsi_info.h"
40 #include "tgsi/tgsi_parse.h"
41 #include "tgsi/tgsi_util.h"
42 #include "tgsi/tgsi_dump.h"
43 #include "tgsi/tgsi_exec.h"
44 #include "tgsi/tgsi_sse2.h"
46 #include "rtasm/rtasm_x86sse.h"
50 * This costs about 100fps (close to 10%) in gears:
52 #define HIGH_PRECISION 1
57 #define FOR_EACH_CHANNEL( CHAN )\
58 for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++)
60 #define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
61 ((INST).Dst[0].DstRegister.WriteMask & (1 << (CHAN)))
63 #define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
64 if (IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
66 #define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
67 FOR_EACH_CHANNEL( CHAN )\
68 IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
75 #define TEMP_ONE_I TGSI_EXEC_TEMP_ONE_I
76 #define TEMP_ONE_C TGSI_EXEC_TEMP_ONE_C
78 #define TEMP_R0 TGSI_EXEC_TEMP_R0
79 #define TEMP_ADDR TGSI_EXEC_TEMP_ADDR
80 #define TEMP_EXEC_MASK_I TGSI_EXEC_MASK_I
81 #define TEMP_EXEC_MASK_C TGSI_EXEC_MASK_C
85 * X86 utility functions.
94 (enum x86_reg_name
) xmm
);
98 * X86 register mapping helpers.
101 static struct x86_reg
102 get_const_base( void )
109 static struct x86_reg
110 get_machine_base( void )
117 static struct x86_reg
118 get_input_base( void )
120 return x86_make_disp(
122 Offset(struct tgsi_exec_machine
, Inputs
) );
125 static struct x86_reg
126 get_output_base( void )
128 return x86_make_disp(
130 Offset(struct tgsi_exec_machine
, Outputs
) );
133 static struct x86_reg
134 get_temp_base( void )
136 return x86_make_disp(
138 Offset(struct tgsi_exec_machine
, Temps
) );
141 static struct x86_reg
142 get_coef_base( void )
149 static struct x86_reg
150 get_sampler_base( void )
157 static struct x86_reg
158 get_immediate_base( void )
167 * Data access helpers.
171 static struct x86_reg
176 return x86_make_disp(
177 get_immediate_base(),
178 (vec
* 4 + chan
) * 4 );
181 static struct x86_reg
186 return x86_make_disp(
188 (vec
* 4 + chan
) * 4 );
191 static struct x86_reg
195 return x86_make_disp(
197 unit
* sizeof( struct tgsi_sampler
* ) );
200 static struct x86_reg
205 return x86_make_disp(
207 (vec
* 4 + chan
) * 16 );
210 static struct x86_reg
215 return x86_make_disp(
217 (vec
* 4 + chan
) * 16 );
220 static struct x86_reg
225 return x86_make_disp(
227 (vec
* 4 + chan
) * 16 );
230 static struct x86_reg
236 return x86_make_disp(
238 ((vec
* 3 + member
) * 4 + chan
) * 4 );
244 struct x86_function
*func
)
251 * Data fetch helpers.
255 * Copy a shader constant to xmm register
256 * \param xmm the destination xmm register
257 * \param vec the src const buffer index
258 * \param chan src channel to fetch (X, Y, Z or W)
262 struct x86_function
*func
,
271 /* 'vec' is the offset from the address register's value.
272 * We're loading CONST[ADDR+vec] into an xmm register.
274 struct x86_reg r0
= get_immediate_base();
275 struct x86_reg r1
= get_coef_base();
278 assert( indirectFile
== TGSI_FILE_ADDRESS
);
279 assert( indirectIndex
== 0 );
280 assert( r0
.mod
== mod_REG
);
281 assert( r1
.mod
== mod_REG
);
283 x86_push( func
, r0
);
284 x86_push( func
, r1
);
287 * Loop over the four pixels or vertices in the quad.
288 * Get the value of the address (offset) register for pixel/vertex[i],
289 * add it to the src offset and index into the constant buffer.
290 * Note that we're working on SOA data.
291 * If any of the pixel/vertex execution channels are unused their
292 * values will be garbage. It's very important that we don't use
293 * those garbage values as indexes into the constant buffer since
294 * that'll cause segfaults.
295 * The solution is to bitwise-AND the offset with the execution mask
296 * register whose values are either 0 or ~0.
297 * The caller must setup the execution mask register to indicate
298 * which channels are valid/alive before running the shader.
299 * The execution mask will also figure into loops and conditionals
302 for (i
= 0; i
< QUAD_SIZE
; i
++) {
303 /* r1 = address register[i] */
304 x86_mov( func
, r1
, x86_make_disp( get_temp( TEMP_ADDR
, CHAN_X
), i
* 4 ) );
305 /* r0 = execution mask[i] */
306 x86_mov( func
, r0
, x86_make_disp( get_temp( TEMP_EXEC_MASK_I
, TEMP_EXEC_MASK_C
), i
* 4 ) );
308 x86_and( func
, r1
, r0
);
309 /* r0 = 'vec', the offset */
310 x86_lea( func
, r0
, get_const( vec
, chan
) );
312 /* Quick hack to multiply r1 by 16 -- need to add SHL to rtasm.
314 x86_add( func
, r1
, r1
);
315 x86_add( func
, r1
, r1
);
316 x86_add( func
, r1
, r1
);
317 x86_add( func
, r1
, r1
);
319 x86_add( func
, r0
, r1
); /* r0 = r0 + r1 */
320 x86_mov( func
, r1
, x86_deref( r0
) );
321 x86_mov( func
, x86_make_disp( get_temp( TEMP_R0
, CHAN_X
), i
* 4 ), r1
);
330 get_temp( TEMP_R0
, CHAN_X
) );
333 /* 'vec' is the index into the src register file, such as TEMP[vec] */
339 get_const( vec
, chan
) );
344 SHUF( 0, 0, 0, 0 ) );
350 struct x86_function
*func
,
358 get_immediate( vec
, chan
) );
363 SHUF( 0, 0, 0, 0 ) );
368 * Copy a shader input to xmm register
369 * \param xmm the destination xmm register
370 * \param vec the src input attrib
371 * \param chan src channel to fetch (X, Y, Z or W)
375 struct x86_function
*func
,
383 get_input( vec
, chan
) );
387 * Store an xmm register to a shader output
388 * \param xmm the source xmm register
389 * \param vec the dest output attrib
390 * \param chan src dest channel to store (X, Y, Z or W)
394 struct x86_function
*func
,
401 get_output( vec
, chan
),
406 * Copy a shader temporary to xmm register
407 * \param xmm the destination xmm register
408 * \param vec the src temp register
409 * \param chan src channel to fetch (X, Y, Z or W)
413 struct x86_function
*func
,
421 get_temp( vec
, chan
) );
425 * Load an xmm register with an input attrib coefficient (a0, dadx or dady)
426 * \param xmm the destination xmm register
427 * \param vec the src input/attribute coefficient index
428 * \param chan src channel to fetch (X, Y, Z or W)
429 * \param member 0=a0, 1=dadx, 2=dady
433 struct x86_function
*func
,
442 get_coef( vec
, chan
, member
) );
447 SHUF( 0, 0, 0, 0 ) );
451 * Data store helpers.
456 struct x86_function
*func
,
463 get_input( vec
, chan
),
469 struct x86_function
*func
,
476 get_temp( vec
, chan
),
482 struct x86_function
*func
,
492 vec
+ TGSI_EXEC_TEMP_ADDR
,
497 * Coefficent fetch helpers.
502 struct x86_function
*func
,
517 struct x86_function
*func
,
532 struct x86_function
*func
,
546 * Function call helpers.
550 * NOTE: In gcc, if the destination uses the SSE intrinsics, then it must be
551 * defined with __attribute__((force_align_arg_pointer)), as we do not guarantee
552 * that the stack pointer is 16 byte aligned, as expected.
556 struct x86_function
*func
,
557 unsigned xmm_save_mask
,
558 const struct x86_reg
*arg
,
560 void (PIPE_CDECL
*code
)() )
562 struct x86_reg ecx
= x86_make_reg( file_REG32
, reg_CX
);
567 x86_make_reg( file_REG32
, reg_AX
) );
570 x86_make_reg( file_REG32
, reg_CX
) );
573 x86_make_reg( file_REG32
, reg_DX
) );
575 /* Store XMM regs to the stack
577 for(i
= 0, n
= 0; i
< 8; ++i
)
578 if(xmm_save_mask
& (1 << i
))
583 x86_make_reg( file_REG32
, reg_SP
),
586 for(i
= 0, n
= 0; i
< 8; ++i
)
587 if(xmm_save_mask
& (1 << i
)) {
590 x86_make_disp( x86_make_reg( file_REG32
, reg_SP
), n
*16 ),
595 for (i
= 0; i
< nr_args
; i
++) {
596 /* Load the address of the buffer we use for passing arguments and
604 /* Push actual function arguments (currently just the pointer to
605 * the buffer above), and call the function:
607 x86_push( func
, ecx
);
610 x86_mov_reg_imm( func
, ecx
, (unsigned long) code
);
611 x86_call( func
, ecx
);
613 /* Pop the arguments (or just add an immediate to esp)
615 for (i
= 0; i
< nr_args
; i
++) {
619 /* Pop the saved XMM regs:
621 for(i
= 0, n
= 0; i
< 8; ++i
)
622 if(xmm_save_mask
& (1 << i
)) {
626 x86_make_disp( x86_make_reg( file_REG32
, reg_SP
), n
*16 ) );
632 x86_make_reg( file_REG32
, reg_SP
),
635 /* Restore GP registers in a reverse order.
639 x86_make_reg( file_REG32
, reg_DX
) );
642 x86_make_reg( file_REG32
, reg_CX
) );
645 x86_make_reg( file_REG32
, reg_AX
) );
649 emit_func_call_dst_src1(
650 struct x86_function
*func
,
654 void (PIPE_CDECL
*code
)() )
656 struct x86_reg store
= get_temp( TEMP_R0
, 0 );
657 unsigned xmm_mask
= ((1 << xmm_save
) - 1) & ~(1 << xmm_dst
);
659 /* Store our input parameters (in xmm regs) to the buffer we use
660 * for passing arguments. We will pass a pointer to this buffer as
661 * the actual function argument.
666 make_xmm( xmm_src0
) );
668 emit_func_call( func
,
682 emit_func_call_dst_src2(
683 struct x86_function
*func
,
688 void (PIPE_CDECL
*code
)() )
690 struct x86_reg store
= get_temp( TEMP_R0
, 0 );
691 unsigned xmm_mask
= ((1 << xmm_save
) - 1) & ~(1 << xmm_dst
);
693 /* Store two inputs to parameter buffer.
698 make_xmm( xmm_src0
) );
702 x86_make_disp( store
, 4 * sizeof(float) ),
703 make_xmm( xmm_src1
) );
708 emit_func_call( func
,
714 /* Retrieve the results:
726 #if defined(PIPE_ARCH_SSE)
729 * Fast SSE2 implementation of special math functions.
732 #define POLY0(x, c0) _mm_set1_ps(c0)
733 #define POLY1(x, c0, c1) _mm_add_ps(_mm_mul_ps(POLY0(x, c1), x), _mm_set1_ps(c0))
734 #define POLY2(x, c0, c1, c2) _mm_add_ps(_mm_mul_ps(POLY1(x, c1, c2), x), _mm_set1_ps(c0))
735 #define POLY3(x, c0, c1, c2, c3) _mm_add_ps(_mm_mul_ps(POLY2(x, c1, c2, c3), x), _mm_set1_ps(c0))
736 #define POLY4(x, c0, c1, c2, c3, c4) _mm_add_ps(_mm_mul_ps(POLY3(x, c1, c2, c3, c4), x), _mm_set1_ps(c0))
737 #define POLY5(x, c0, c1, c2, c3, c4, c5) _mm_add_ps(_mm_mul_ps(POLY4(x, c1, c2, c3, c4, c5), x), _mm_set1_ps(c0))
739 #define EXP_POLY_DEGREE 3
740 #define LOG_POLY_DEGREE 5
743 * See http://www.devmaster.net/forums/showthread.php?p=43580
749 __m128 fpart
, expipart
, expfpart
;
751 x
= _mm_min_ps(x
, _mm_set1_ps( 129.00000f
));
752 x
= _mm_max_ps(x
, _mm_set1_ps(-126.99999f
));
754 /* ipart = int(x - 0.5) */
755 ipart
= _mm_cvtps_epi32(_mm_sub_ps(x
, _mm_set1_ps(0.5f
)));
757 /* fpart = x - ipart */
758 fpart
= _mm_sub_ps(x
, _mm_cvtepi32_ps(ipart
));
760 /* expipart = (float) (1 << ipart) */
761 expipart
= _mm_castsi128_ps(_mm_slli_epi32(_mm_add_epi32(ipart
, _mm_set1_epi32(127)), 23));
763 /* minimax polynomial fit of 2**x, in range [-0.5, 0.5[ */
764 #if EXP_POLY_DEGREE == 5
765 expfpart
= POLY5(fpart
, 9.9999994e-1f
, 6.9315308e-1f
, 2.4015361e-1f
, 5.5826318e-2f
, 8.9893397e-3f
, 1.8775767e-3f
);
766 #elif EXP_POLY_DEGREE == 4
767 expfpart
= POLY4(fpart
, 1.0000026f
, 6.9300383e-1f
, 2.4144275e-1f
, 5.2011464e-2f
, 1.3534167e-2f
);
768 #elif EXP_POLY_DEGREE == 3
769 expfpart
= POLY3(fpart
, 9.9992520e-1f
, 6.9583356e-1f
, 2.2606716e-1f
, 7.8024521e-2f
);
770 #elif EXP_POLY_DEGREE == 2
771 expfpart
= POLY2(fpart
, 1.0017247f
, 6.5763628e-1f
, 3.3718944e-1f
);
776 return _mm_mul_ps(expipart
, expfpart
);
781 * See http://www.devmaster.net/forums/showthread.php?p=43580
786 __m128i expmask
= _mm_set1_epi32(0x7f800000);
787 __m128i mantmask
= _mm_set1_epi32(0x007fffff);
788 __m128 one
= _mm_set1_ps(1.0f
);
790 __m128i i
= _mm_castps_si128(x
);
792 /* exp = (float) exponent(x) */
793 __m128 exp
= _mm_cvtepi32_ps(_mm_sub_epi32(_mm_srli_epi32(_mm_and_si128(i
, expmask
), 23), _mm_set1_epi32(127)));
795 /* mant = (float) mantissa(x) */
796 __m128 mant
= _mm_or_ps(_mm_castsi128_ps(_mm_and_si128(i
, mantmask
)), one
);
800 /* Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[
801 * These coefficients can be generate with
802 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
804 #if LOG_POLY_DEGREE == 6
805 logmant
= POLY5(mant
, 3.11578814719469302614f
, -3.32419399085241980044f
, 2.59883907202499966007f
, -1.23152682416275988241f
, 0.318212422185251071475f
, -0.0344359067839062357313f
);
806 #elif LOG_POLY_DEGREE == 5
807 logmant
= POLY4(mant
, 2.8882704548164776201f
, -2.52074962577807006663f
, 1.48116647521213171641f
, -0.465725644288844778798f
, 0.0596515482674574969533f
);
808 #elif LOG_POLY_DEGREE == 4
809 logmant
= POLY3(mant
, 2.61761038894603480148f
, -1.75647175389045657003f
, 0.688243882994381274313f
, -0.107254423828329604454f
);
810 #elif LOG_POLY_DEGREE == 3
811 logmant
= POLY2(mant
, 2.28330284476918490682f
, -1.04913055217340124191f
, 0.204446009836232697516f
);
816 /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
817 logmant
= _mm_mul_ps(logmant
, _mm_sub_ps(mant
, one
));
819 return _mm_add_ps(logmant
, exp
);
824 powf4(__m128 x
, __m128 y
)
826 return exp2f4(_mm_mul_ps(log2f4(x
), y
));
829 #endif /* PIPE_ARCH_SSE */
834 * Low-level instruction translators.
839 struct x86_function
*func
,
846 TGSI_EXEC_TEMP_7FFFFFFF_I
,
847 TGSI_EXEC_TEMP_7FFFFFFF_C
) );
852 struct x86_function
*func
,
859 make_xmm( xmm_src
) );
862 static void PIPE_CDECL
866 store
[0] = cosf( store
[0] );
867 store
[1] = cosf( store
[1] );
868 store
[2] = cosf( store
[2] );
869 store
[3] = cosf( store
[3] );
874 struct x86_function
*func
,
878 emit_func_call_dst_src1(
886 static void PIPE_CDECL
887 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_SSE)
888 __attribute__((force_align_arg_pointer
))
893 #if defined(PIPE_ARCH_SSE)
894 _mm_store_ps(&store
[0], exp2f4( _mm_load_ps(&store
[0]) ));
896 store
[0] = util_fast_exp2( store
[0] );
897 store
[1] = util_fast_exp2( store
[1] );
898 store
[2] = util_fast_exp2( store
[2] );
899 store
[3] = util_fast_exp2( store
[3] );
905 struct x86_function
*func
,
909 emit_func_call_dst_src1(
919 struct x86_function
*func
,
930 struct x86_function
*func
,
939 static void PIPE_CDECL
943 store
[0] = floorf( store
[0] );
944 store
[1] = floorf( store
[1] );
945 store
[2] = floorf( store
[2] );
946 store
[3] = floorf( store
[3] );
951 struct x86_function
*func
,
955 emit_func_call_dst_src1(
963 static void PIPE_CDECL
967 store
[0] -= floorf( store
[0] );
968 store
[1] -= floorf( store
[1] );
969 store
[2] -= floorf( store
[2] );
970 store
[3] -= floorf( store
[3] );
975 struct x86_function
*func
,
979 emit_func_call_dst_src1(
987 static void PIPE_CDECL
988 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_SSE)
989 __attribute__((force_align_arg_pointer
))
994 #if defined(PIPE_ARCH_SSE)
995 _mm_store_ps(&store
[0], log2f4( _mm_load_ps(&store
[0]) ));
997 store
[0] = util_fast_log2( store
[0] );
998 store
[1] = util_fast_log2( store
[1] );
999 store
[2] = util_fast_log2( store
[2] );
1000 store
[3] = util_fast_log2( store
[3] );
1006 struct x86_function
*func
,
1010 emit_func_call_dst_src1(
1020 struct x86_function
*func
,
1026 make_xmm( xmm_dst
),
1027 make_xmm( xmm_src
) );
1031 emit_mul (struct x86_function
*func
,
1037 make_xmm( xmm_dst
),
1038 make_xmm( xmm_src
) );
1043 struct x86_function
*func
,
1050 TGSI_EXEC_TEMP_80000000_I
,
1051 TGSI_EXEC_TEMP_80000000_C
) );
1054 static void PIPE_CDECL
1055 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_SSE)
1056 __attribute__((force_align_arg_pointer
))
1061 #if defined(PIPE_ARCH_SSE)
1062 _mm_store_ps(&store
[0], powf4( _mm_load_ps(&store
[0]), _mm_load_ps(&store
[4]) ));
1064 store
[0] = util_fast_pow( store
[0], store
[4] );
1065 store
[1] = util_fast_pow( store
[1], store
[5] );
1066 store
[2] = util_fast_pow( store
[2], store
[6] );
1067 store
[3] = util_fast_pow( store
[3], store
[7] );
1073 struct x86_function
*func
,
1079 emit_func_call_dst_src2(
1090 struct x86_function
*func
,
1094 /* On Intel CPUs at least, this is only accurate to 12 bits -- not
1095 * good enough. Need to either emit a proper divide or use the
1096 * iterative technique described below in emit_rsqrt().
1100 make_xmm( xmm_dst
),
1101 make_xmm( xmm_src
) );
1104 static void PIPE_CDECL
1108 store
[0] = floorf( store
[0] + 0.5f
);
1109 store
[1] = floorf( store
[1] + 0.5f
);
1110 store
[2] = floorf( store
[2] + 0.5f
);
1111 store
[3] = floorf( store
[3] + 0.5f
);
1116 struct x86_function
*func
,
1120 emit_func_call_dst_src1(
1130 struct x86_function
*func
,
1135 /* Although rsqrtps() and rcpps() are low precision on some/all SSE
1136 * implementations, it is possible to improve its precision at
1137 * fairly low cost, using a newton/raphson step, as below:
1139 * x1 = 2 * rcpps(a) - a * rcpps(a) * rcpps(a)
1140 * x1 = 0.5 * rsqrtps(a) * [3.0 - (a * rsqrtps(a))* rsqrtps(a)]
1142 * See: http://softwarecommunity.intel.com/articles/eng/1818.htm
1145 struct x86_reg dst
= make_xmm( xmm_dst
);
1146 struct x86_reg src
= make_xmm( xmm_src
);
1147 struct x86_reg tmp0
= make_xmm( 2 );
1148 struct x86_reg tmp1
= make_xmm( 3 );
1150 assert( xmm_dst
!= xmm_src
);
1151 assert( xmm_dst
!= 2 && xmm_dst
!= 3 );
1152 assert( xmm_src
!= 2 && xmm_src
!= 3 );
1154 sse_movaps( func
, dst
, get_temp( TGSI_EXEC_TEMP_HALF_I
, TGSI_EXEC_TEMP_HALF_C
) );
1155 sse_movaps( func
, tmp0
, get_temp( TGSI_EXEC_TEMP_THREE_I
, TGSI_EXEC_TEMP_THREE_C
) );
1156 sse_rsqrtps( func
, tmp1
, src
);
1157 sse_mulps( func
, src
, tmp1
);
1158 sse_mulps( func
, dst
, tmp1
);
1159 sse_mulps( func
, src
, tmp1
);
1160 sse_subps( func
, tmp0
, src
);
1161 sse_mulps( func
, dst
, tmp0
);
1164 /* On Intel CPUs at least, this is only accurate to 12 bits -- not
1169 make_xmm( xmm_dst
),
1170 make_xmm( xmm_src
) );
1176 struct x86_function
*func
,
1183 TGSI_EXEC_TEMP_80000000_I
,
1184 TGSI_EXEC_TEMP_80000000_C
) );
1187 static void PIPE_CDECL
1191 store
[0] = store
[0] < 0.0f
? -1.0f
: store
[0] > 0.0f
? 1.0f
: 0.0f
;
1192 store
[1] = store
[1] < 0.0f
? -1.0f
: store
[1] > 0.0f
? 1.0f
: 0.0f
;
1193 store
[2] = store
[2] < 0.0f
? -1.0f
: store
[2] > 0.0f
? 1.0f
: 0.0f
;
1194 store
[3] = store
[3] < 0.0f
? -1.0f
: store
[3] > 0.0f
? 1.0f
: 0.0f
;
1199 struct x86_function
*func
,
1203 emit_func_call_dst_src1(
1211 static void PIPE_CDECL
1215 store
[0] = sinf( store
[0] );
1216 store
[1] = sinf( store
[1] );
1217 store
[2] = sinf( store
[2] );
1218 store
[3] = sinf( store
[3] );
1222 emit_sin (struct x86_function
*func
,
1226 emit_func_call_dst_src1(
1236 struct x86_function
*func
,
1242 make_xmm( xmm_dst
),
1243 make_xmm( xmm_src
) );
1258 struct x86_function
*func
,
1260 const struct tgsi_full_src_register
*reg
,
1261 const unsigned chan_index
)
1263 unsigned swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
1266 case TGSI_SWIZZLE_X
:
1267 case TGSI_SWIZZLE_Y
:
1268 case TGSI_SWIZZLE_Z
:
1269 case TGSI_SWIZZLE_W
:
1270 switch (reg
->SrcRegister
.File
) {
1271 case TGSI_FILE_CONSTANT
:
1275 reg
->SrcRegister
.Index
,
1277 reg
->SrcRegister
.Indirect
,
1278 reg
->SrcRegisterInd
.File
,
1279 reg
->SrcRegisterInd
.Index
);
1282 case TGSI_FILE_IMMEDIATE
:
1286 reg
->SrcRegister
.Index
,
1290 case TGSI_FILE_INPUT
:
1294 reg
->SrcRegister
.Index
,
1298 case TGSI_FILE_TEMPORARY
:
1302 reg
->SrcRegister
.Index
,
1315 switch( tgsi_util_get_full_src_register_sign_mode( reg
, chan_index
) ) {
1316 case TGSI_UTIL_SIGN_CLEAR
:
1317 emit_abs( func
, xmm
);
1320 case TGSI_UTIL_SIGN_SET
:
1321 emit_setsign( func
, xmm
);
1324 case TGSI_UTIL_SIGN_TOGGLE
:
1325 emit_neg( func
, xmm
);
1328 case TGSI_UTIL_SIGN_KEEP
:
1333 #define FETCH( FUNC, INST, XMM, INDEX, CHAN )\
1334 emit_fetch( FUNC, XMM, &(INST).Src[INDEX], CHAN )
1342 struct x86_function
*func
,
1344 const struct tgsi_full_dst_register
*reg
,
1345 const struct tgsi_full_instruction
*inst
,
1346 unsigned chan_index
)
1348 switch( inst
->Instruction
.Saturate
) {
1352 case TGSI_SAT_ZERO_ONE
:
1357 TGSI_EXEC_TEMP_00000000_I
,
1358 TGSI_EXEC_TEMP_00000000_C
) );
1364 TGSI_EXEC_TEMP_ONE_I
,
1365 TGSI_EXEC_TEMP_ONE_C
) );
1368 case TGSI_SAT_MINUS_PLUS_ONE
:
1374 switch( reg
->DstRegister
.File
) {
1375 case TGSI_FILE_OUTPUT
:
1379 reg
->DstRegister
.Index
,
1383 case TGSI_FILE_TEMPORARY
:
1387 reg
->DstRegister
.Index
,
1391 case TGSI_FILE_ADDRESS
:
1395 reg
->DstRegister
.Index
,
1404 #define STORE( FUNC, INST, XMM, INDEX, CHAN )\
1405 emit_store( FUNC, XMM, &(INST).Dst[INDEX], &(INST), CHAN )
1408 static void PIPE_CDECL
1409 fetch_texel( struct tgsi_sampler
**sampler
,
1415 debug_printf("%s sampler: %p (%p) store: %p\n",
1420 debug_printf("lodbias %f\n", store
[12]);
1422 for (j
= 0; j
< 4; j
++)
1423 debug_printf("sample %d texcoord %f %f\n",
1430 float rgba
[NUM_CHANNELS
][QUAD_SIZE
];
1431 (*sampler
)->get_samples(*sampler
,
1435 store
[12], /* lodbias */
1436 rgba
); /* results */
1438 memcpy( store
, rgba
, 16 * sizeof(float));
1442 for (j
= 0; j
< 4; j
++)
1443 debug_printf("sample %d result %f %f %f %f\n",
1453 * High-level instruction translators.
1457 emit_tex( struct x86_function
*func
,
1458 const struct tgsi_full_instruction
*inst
,
1462 const uint unit
= inst
->Src
[1].SrcRegister
.Index
;
1463 struct x86_reg args
[2];
1467 assert(inst
->Instruction
.Texture
);
1468 switch (inst
->Texture
.Texture
) {
1469 case TGSI_TEXTURE_1D
:
1472 case TGSI_TEXTURE_2D
:
1473 case TGSI_TEXTURE_RECT
:
1476 case TGSI_TEXTURE_SHADOW1D
:
1477 case TGSI_TEXTURE_SHADOW2D
:
1478 case TGSI_TEXTURE_SHADOWRECT
:
1479 case TGSI_TEXTURE_3D
:
1480 case TGSI_TEXTURE_CUBE
:
1489 FETCH( func
, *inst
, 3, 0, 3 );
1495 TGSI_EXEC_TEMP_00000000_I
,
1496 TGSI_EXEC_TEMP_00000000_C
);
1500 /* store lodbias whether enabled or not -- fetch_texel currently
1501 * respects it always.
1504 get_temp( TEMP_R0
, 3 ),
1509 FETCH( func
, *inst
, 3, 0, 3 );
1511 emit_rcp( func
, 3, 3 );
1514 for (i
= 0; i
< count
; i
++) {
1515 FETCH( func
, *inst
, i
, 0, i
);
1524 /* Store in the argument buffer:
1528 get_temp( TEMP_R0
, i
),
1532 args
[0] = get_temp( TEMP_R0
, 0 );
1533 args
[1] = get_sampler_ptr( unit
);
1536 emit_func_call( func
,
1542 /* If all four channels are enabled, could use a pointer to
1543 * dst[0].x instead of TEMP_R0 for store?
1545 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, i
) {
1550 get_temp( TEMP_R0
, i
) );
1552 STORE( func
, *inst
, 0, 0, i
);
1559 struct x86_function
*func
,
1560 const struct tgsi_full_src_register
*reg
)
1562 unsigned uniquemask
;
1563 unsigned unique_count
= 0;
1564 unsigned chan_index
;
1567 /* This mask stores component bits that were already tested. Note that
1568 * we test if the value is less than zero, so 1.0 and 0.0 need not to be
1572 FOR_EACH_CHANNEL( chan_index
) {
1575 /* unswizzle channel */
1576 swizzle
= tgsi_util_get_full_src_register_swizzle(
1580 /* check if the component has not been already tested */
1581 if( !(uniquemask
& (1 << swizzle
)) ) {
1582 uniquemask
|= 1 << swizzle
;
1584 /* allocate register */
1595 x86_make_reg( file_REG32
, reg_AX
) );
1598 x86_make_reg( file_REG32
, reg_DX
) );
1600 for (i
= 0 ; i
< unique_count
; i
++ ) {
1601 struct x86_reg dataXMM
= make_xmm(i
);
1607 TGSI_EXEC_TEMP_00000000_I
,
1608 TGSI_EXEC_TEMP_00000000_C
),
1614 x86_make_reg( file_REG32
, reg_AX
),
1620 x86_make_reg( file_REG32
, reg_DX
),
1624 x86_make_reg( file_REG32
, reg_AX
),
1625 x86_make_reg( file_REG32
, reg_DX
) );
1632 TGSI_EXEC_TEMP_KILMASK_I
,
1633 TGSI_EXEC_TEMP_KILMASK_C
),
1634 x86_make_reg( file_REG32
, reg_AX
) );
1638 x86_make_reg( file_REG32
, reg_DX
) );
1641 x86_make_reg( file_REG32
, reg_AX
) );
1647 struct x86_function
*func
)
1649 /* XXX todo / fix me */
1655 struct x86_function
*func
,
1656 struct tgsi_full_instruction
*inst
,
1659 unsigned chan_index
;
1661 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1662 FETCH( func
, *inst
, 0, 0, chan_index
);
1663 FETCH( func
, *inst
, 1, 1, chan_index
);
1675 STORE( func
, *inst
, 0, 0, chan_index
);
1681 struct x86_function
*func
,
1682 struct tgsi_full_instruction
*inst
)
1684 unsigned chan_index
;
1686 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1687 FETCH( func
, *inst
, 0, 0, chan_index
);
1688 FETCH( func
, *inst
, 1, 1, chan_index
);
1689 FETCH( func
, *inst
, 2, 2, chan_index
);
1694 TGSI_EXEC_TEMP_00000000_I
,
1695 TGSI_EXEC_TEMP_00000000_C
),
1709 STORE( func
, *inst
, 0, 0, chan_index
);
1715 * Check if inst src/dest regs use indirect addressing into temporary
1719 indirect_temp_reference(const struct tgsi_full_instruction
*inst
)
1722 for (i
= 0; i
< inst
->Instruction
.NumSrcRegs
; i
++) {
1723 const struct tgsi_full_src_register
*reg
= &inst
->Src
[i
];
1724 if (reg
->SrcRegister
.File
== TGSI_FILE_TEMPORARY
&&
1725 reg
->SrcRegister
.Indirect
)
1728 for (i
= 0; i
< inst
->Instruction
.NumDstRegs
; i
++) {
1729 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[i
];
1730 if (reg
->DstRegister
.File
== TGSI_FILE_TEMPORARY
&&
1731 reg
->DstRegister
.Indirect
)
1740 struct x86_function
*func
,
1741 struct tgsi_full_instruction
*inst
)
1743 unsigned chan_index
;
1745 /* we can't handle indirect addressing into temp register file yet */
1746 if (indirect_temp_reference(inst
))
1749 switch (inst
->Instruction
.Opcode
) {
1750 case TGSI_OPCODE_ARL
:
1751 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1752 FETCH( func
, *inst
, 0, 0, chan_index
);
1753 emit_flr(func
, 0, 0);
1754 emit_f2it( func
, 0 );
1755 STORE( func
, *inst
, 0, 0, chan_index
);
1759 case TGSI_OPCODE_MOV
:
1760 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1761 FETCH( func
, *inst
, 4 + chan_index
, 0, chan_index
);
1763 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1764 STORE( func
, *inst
, 4 + chan_index
, 0, chan_index
);
1768 case TGSI_OPCODE_LIT
:
1769 if( IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) ||
1770 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_W
) ) {
1776 if( IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) ) {
1777 STORE( func
, *inst
, 0, 0, CHAN_X
);
1779 if( IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_W
) ) {
1780 STORE( func
, *inst
, 0, 0, CHAN_W
);
1783 if( IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
) ||
1784 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
) ) {
1785 if( IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
) ) {
1786 FETCH( func
, *inst
, 0, 0, CHAN_X
);
1791 TGSI_EXEC_TEMP_00000000_I
,
1792 TGSI_EXEC_TEMP_00000000_C
) );
1793 STORE( func
, *inst
, 0, 0, CHAN_Y
);
1795 if( IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
) ) {
1796 /* XMM[1] = SrcReg[0].yyyy */
1797 FETCH( func
, *inst
, 1, 0, CHAN_Y
);
1798 /* XMM[1] = max(XMM[1], 0) */
1803 TGSI_EXEC_TEMP_00000000_I
,
1804 TGSI_EXEC_TEMP_00000000_C
) );
1805 /* XMM[2] = SrcReg[0].wwww */
1806 FETCH( func
, *inst
, 2, 0, CHAN_W
);
1807 /* XMM[2] = min(XMM[2], 128.0) */
1812 TGSI_EXEC_TEMP_128_I
,
1813 TGSI_EXEC_TEMP_128_C
) );
1814 /* XMM[2] = max(XMM[2], -128.0) */
1819 TGSI_EXEC_TEMP_MINUS_128_I
,
1820 TGSI_EXEC_TEMP_MINUS_128_C
) );
1821 emit_pow( func
, 3, 1, 1, 2 );
1822 FETCH( func
, *inst
, 0, 0, CHAN_X
);
1836 STORE( func
, *inst
, 2, 0, CHAN_Z
);
1841 case TGSI_OPCODE_RCP
:
1842 FETCH( func
, *inst
, 0, 0, CHAN_X
);
1843 emit_rcp( func
, 0, 0 );
1844 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1845 STORE( func
, *inst
, 0, 0, chan_index
);
1849 case TGSI_OPCODE_RSQ
:
1850 FETCH( func
, *inst
, 0, 0, CHAN_X
);
1851 emit_abs( func
, 0 );
1852 emit_rsqrt( func
, 1, 0 );
1853 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1854 STORE( func
, *inst
, 1, 0, chan_index
);
1858 case TGSI_OPCODE_EXP
:
1859 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) ||
1860 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
) ||
1861 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
)) {
1862 FETCH( func
, *inst
, 0, 0, CHAN_X
);
1863 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) ||
1864 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
)) {
1865 emit_MOV( func
, 1, 0 );
1866 emit_flr( func
, 2, 1 );
1867 /* dst.x = ex2(floor(src.x)) */
1868 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
)) {
1869 emit_MOV( func
, 2, 1 );
1870 emit_ex2( func
, 3, 2 );
1871 STORE( func
, *inst
, 2, 0, CHAN_X
);
1873 /* dst.y = src.x - floor(src.x) */
1874 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
)) {
1875 emit_MOV( func
, 2, 0 );
1876 emit_sub( func
, 2, 1 );
1877 STORE( func
, *inst
, 2, 0, CHAN_Y
);
1880 /* dst.z = ex2(src.x) */
1881 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
)) {
1882 emit_ex2( func
, 3, 0 );
1883 STORE( func
, *inst
, 0, 0, CHAN_Z
);
1887 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_W
)) {
1888 emit_tempf( func
, 0, TEMP_ONE_I
, TEMP_ONE_C
);
1889 STORE( func
, *inst
, 0, 0, CHAN_W
);
1893 case TGSI_OPCODE_LOG
:
1894 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) ||
1895 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
) ||
1896 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
)) {
1897 FETCH( func
, *inst
, 0, 0, CHAN_X
);
1898 emit_abs( func
, 0 );
1899 emit_MOV( func
, 1, 0 );
1900 emit_lg2( func
, 2, 1 );
1901 /* dst.z = lg2(abs(src.x)) */
1902 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
)) {
1903 STORE( func
, *inst
, 1, 0, CHAN_Z
);
1905 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) ||
1906 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
)) {
1907 emit_flr( func
, 2, 1 );
1908 /* dst.x = floor(lg2(abs(src.x))) */
1909 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
)) {
1910 STORE( func
, *inst
, 1, 0, CHAN_X
);
1912 /* dst.x = abs(src)/ex2(floor(lg2(abs(src.x)))) */
1913 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
)) {
1914 emit_ex2( func
, 2, 1 );
1915 emit_rcp( func
, 1, 1 );
1916 emit_mul( func
, 0, 1 );
1917 STORE( func
, *inst
, 0, 0, CHAN_Y
);
1922 if (IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_W
)) {
1923 emit_tempf( func
, 0, TEMP_ONE_I
, TEMP_ONE_C
);
1924 STORE( func
, *inst
, 0, 0, CHAN_W
);
1928 case TGSI_OPCODE_MUL
:
1929 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1930 FETCH( func
, *inst
, 0, 0, chan_index
);
1931 FETCH( func
, *inst
, 1, 1, chan_index
);
1932 emit_mul( func
, 0, 1 );
1933 STORE( func
, *inst
, 0, 0, chan_index
);
1937 case TGSI_OPCODE_ADD
:
1938 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1939 FETCH( func
, *inst
, 0, 0, chan_index
);
1940 FETCH( func
, *inst
, 1, 1, chan_index
);
1941 emit_add( func
, 0, 1 );
1942 STORE( func
, *inst
, 0, 0, chan_index
);
1946 case TGSI_OPCODE_DP3
:
1947 FETCH( func
, *inst
, 0, 0, CHAN_X
);
1948 FETCH( func
, *inst
, 1, 1, CHAN_X
);
1949 emit_mul( func
, 0, 1 );
1950 FETCH( func
, *inst
, 1, 0, CHAN_Y
);
1951 FETCH( func
, *inst
, 2, 1, CHAN_Y
);
1952 emit_mul( func
, 1, 2 );
1953 emit_add( func
, 0, 1 );
1954 FETCH( func
, *inst
, 1, 0, CHAN_Z
);
1955 FETCH( func
, *inst
, 2, 1, CHAN_Z
);
1956 emit_mul( func
, 1, 2 );
1957 emit_add( func
, 0, 1 );
1958 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1959 STORE( func
, *inst
, 0, 0, chan_index
);
1963 case TGSI_OPCODE_DP4
:
1964 FETCH( func
, *inst
, 0, 0, CHAN_X
);
1965 FETCH( func
, *inst
, 1, 1, CHAN_X
);
1966 emit_mul( func
, 0, 1 );
1967 FETCH( func
, *inst
, 1, 0, CHAN_Y
);
1968 FETCH( func
, *inst
, 2, 1, CHAN_Y
);
1969 emit_mul( func
, 1, 2 );
1970 emit_add( func
, 0, 1 );
1971 FETCH( func
, *inst
, 1, 0, CHAN_Z
);
1972 FETCH( func
, *inst
, 2, 1, CHAN_Z
);
1973 emit_mul(func
, 1, 2 );
1974 emit_add(func
, 0, 1 );
1975 FETCH( func
, *inst
, 1, 0, CHAN_W
);
1976 FETCH( func
, *inst
, 2, 1, CHAN_W
);
1977 emit_mul( func
, 1, 2 );
1978 emit_add( func
, 0, 1 );
1979 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
1980 STORE( func
, *inst
, 0, 0, chan_index
);
1984 case TGSI_OPCODE_DST
:
1985 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) {
1991 STORE( func
, *inst
, 0, 0, CHAN_X
);
1993 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
) {
1994 FETCH( func
, *inst
, 0, 0, CHAN_Y
);
1995 FETCH( func
, *inst
, 1, 1, CHAN_Y
);
1996 emit_mul( func
, 0, 1 );
1997 STORE( func
, *inst
, 0, 0, CHAN_Y
);
1999 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
) {
2000 FETCH( func
, *inst
, 0, 0, CHAN_Z
);
2001 STORE( func
, *inst
, 0, 0, CHAN_Z
);
2003 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_W
) {
2004 FETCH( func
, *inst
, 0, 1, CHAN_W
);
2005 STORE( func
, *inst
, 0, 0, CHAN_W
);
2009 case TGSI_OPCODE_MIN
:
2010 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2011 FETCH( func
, *inst
, 0, 0, chan_index
);
2012 FETCH( func
, *inst
, 1, 1, chan_index
);
2017 STORE( func
, *inst
, 0, 0, chan_index
);
2021 case TGSI_OPCODE_MAX
:
2022 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2023 FETCH( func
, *inst
, 0, 0, chan_index
);
2024 FETCH( func
, *inst
, 1, 1, chan_index
);
2029 STORE( func
, *inst
, 0, 0, chan_index
);
2033 case TGSI_OPCODE_SLT
:
2034 emit_setcc( func
, inst
, cc_LessThan
);
2037 case TGSI_OPCODE_SGE
:
2038 emit_setcc( func
, inst
, cc_NotLessThan
);
2041 case TGSI_OPCODE_MAD
:
2042 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2043 FETCH( func
, *inst
, 0, 0, chan_index
);
2044 FETCH( func
, *inst
, 1, 1, chan_index
);
2045 FETCH( func
, *inst
, 2, 2, chan_index
);
2046 emit_mul( func
, 0, 1 );
2047 emit_add( func
, 0, 2 );
2048 STORE( func
, *inst
, 0, 0, chan_index
);
2052 case TGSI_OPCODE_SUB
:
2053 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2054 FETCH( func
, *inst
, 0, 0, chan_index
);
2055 FETCH( func
, *inst
, 1, 1, chan_index
);
2056 emit_sub( func
, 0, 1 );
2057 STORE( func
, *inst
, 0, 0, chan_index
);
2061 case TGSI_OPCODE_LRP
:
2062 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2063 FETCH( func
, *inst
, 0, 0, chan_index
);
2064 FETCH( func
, *inst
, 1, 1, chan_index
);
2065 FETCH( func
, *inst
, 2, 2, chan_index
);
2066 emit_sub( func
, 1, 2 );
2067 emit_mul( func
, 0, 1 );
2068 emit_add( func
, 0, 2 );
2069 STORE( func
, *inst
, 0, 0, chan_index
);
2073 case TGSI_OPCODE_CND
:
2077 case TGSI_OPCODE_DP2A
:
2078 FETCH( func
, *inst
, 0, 0, CHAN_X
); /* xmm0 = src[0].x */
2079 FETCH( func
, *inst
, 1, 1, CHAN_X
); /* xmm1 = src[1].x */
2080 emit_mul( func
, 0, 1 ); /* xmm0 = xmm0 * xmm1 */
2081 FETCH( func
, *inst
, 1, 0, CHAN_Y
); /* xmm1 = src[0].y */
2082 FETCH( func
, *inst
, 2, 1, CHAN_Y
); /* xmm2 = src[1].y */
2083 emit_mul( func
, 1, 2 ); /* xmm1 = xmm1 * xmm2 */
2084 emit_add( func
, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2085 FETCH( func
, *inst
, 1, 2, CHAN_X
); /* xmm1 = src[2].x */
2086 emit_add( func
, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2087 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2088 STORE( func
, *inst
, 0, 0, chan_index
); /* dest[ch] = xmm0 */
2092 case TGSI_OPCODE_FRC
:
2093 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2094 FETCH( func
, *inst
, 0, 0, chan_index
);
2095 emit_frc( func
, 0, 0 );
2096 STORE( func
, *inst
, 0, 0, chan_index
);
2100 case TGSI_OPCODE_CLAMP
:
2104 case TGSI_OPCODE_FLR
:
2105 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2106 FETCH( func
, *inst
, 0, 0, chan_index
);
2107 emit_flr( func
, 0, 0 );
2108 STORE( func
, *inst
, 0, 0, chan_index
);
2112 case TGSI_OPCODE_ROUND
:
2113 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2114 FETCH( func
, *inst
, 0, 0, chan_index
);
2115 emit_rnd( func
, 0, 0 );
2116 STORE( func
, *inst
, 0, 0, chan_index
);
2120 case TGSI_OPCODE_EX2
:
2121 FETCH( func
, *inst
, 0, 0, CHAN_X
);
2122 emit_ex2( func
, 0, 0 );
2123 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2124 STORE( func
, *inst
, 0, 0, chan_index
);
2128 case TGSI_OPCODE_LG2
:
2129 FETCH( func
, *inst
, 0, 0, CHAN_X
);
2130 emit_lg2( func
, 0, 0 );
2131 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2132 STORE( func
, *inst
, 0, 0, chan_index
);
2136 case TGSI_OPCODE_POW
:
2137 FETCH( func
, *inst
, 0, 0, CHAN_X
);
2138 FETCH( func
, *inst
, 1, 1, CHAN_X
);
2139 emit_pow( func
, 0, 0, 0, 1 );
2140 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2141 STORE( func
, *inst
, 0, 0, chan_index
);
2145 case TGSI_OPCODE_XPD
:
2146 if( IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) ||
2147 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
) ) {
2148 FETCH( func
, *inst
, 1, 1, CHAN_Z
);
2149 FETCH( func
, *inst
, 3, 0, CHAN_Z
);
2151 if( IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) ||
2152 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
) ) {
2153 FETCH( func
, *inst
, 0, 0, CHAN_Y
);
2154 FETCH( func
, *inst
, 4, 1, CHAN_Y
);
2156 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) {
2157 emit_MOV( func
, 2, 0 );
2158 emit_mul( func
, 2, 1 );
2159 emit_MOV( func
, 5, 3 );
2160 emit_mul( func
, 5, 4 );
2161 emit_sub( func
, 2, 5 );
2162 STORE( func
, *inst
, 2, 0, CHAN_X
);
2164 if( IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
) ||
2165 IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
) ) {
2166 FETCH( func
, *inst
, 2, 1, CHAN_X
);
2167 FETCH( func
, *inst
, 5, 0, CHAN_X
);
2169 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
) {
2170 emit_mul( func
, 3, 2 );
2171 emit_mul( func
, 1, 5 );
2172 emit_sub( func
, 3, 1 );
2173 STORE( func
, *inst
, 3, 0, CHAN_Y
);
2175 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
) {
2176 emit_mul( func
, 5, 4 );
2177 emit_mul( func
, 0, 2 );
2178 emit_sub( func
, 5, 0 );
2179 STORE( func
, *inst
, 5, 0, CHAN_Z
);
2181 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_W
) {
2187 STORE( func
, *inst
, 0, 0, CHAN_W
);
2191 case TGSI_OPCODE_ABS
:
2192 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2193 FETCH( func
, *inst
, 0, 0, chan_index
);
2194 emit_abs( func
, 0) ;
2196 STORE( func
, *inst
, 0, 0, chan_index
);
2200 case TGSI_OPCODE_RCC
:
2204 case TGSI_OPCODE_DPH
:
2205 FETCH( func
, *inst
, 0, 0, CHAN_X
);
2206 FETCH( func
, *inst
, 1, 1, CHAN_X
);
2207 emit_mul( func
, 0, 1 );
2208 FETCH( func
, *inst
, 1, 0, CHAN_Y
);
2209 FETCH( func
, *inst
, 2, 1, CHAN_Y
);
2210 emit_mul( func
, 1, 2 );
2211 emit_add( func
, 0, 1 );
2212 FETCH( func
, *inst
, 1, 0, CHAN_Z
);
2213 FETCH( func
, *inst
, 2, 1, CHAN_Z
);
2214 emit_mul( func
, 1, 2 );
2215 emit_add( func
, 0, 1 );
2216 FETCH( func
, *inst
, 1, 1, CHAN_W
);
2217 emit_add( func
, 0, 1 );
2218 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2219 STORE( func
, *inst
, 0, 0, chan_index
);
2223 case TGSI_OPCODE_COS
:
2224 FETCH( func
, *inst
, 0, 0, CHAN_X
);
2225 emit_cos( func
, 0, 0 );
2226 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2227 STORE( func
, *inst
, 0, 0, chan_index
);
2231 case TGSI_OPCODE_DDX
:
2235 case TGSI_OPCODE_DDY
:
2239 case TGSI_OPCODE_KILP
:
2240 /* predicated kill */
2242 return 0; /* XXX fix me */
2245 case TGSI_OPCODE_KIL
:
2246 /* conditional kill */
2247 emit_kil( func
, &inst
->Src
[0] );
2250 case TGSI_OPCODE_PK2H
:
2254 case TGSI_OPCODE_PK2US
:
2258 case TGSI_OPCODE_PK4B
:
2262 case TGSI_OPCODE_PK4UB
:
2266 case TGSI_OPCODE_RFL
:
2270 case TGSI_OPCODE_SEQ
:
2271 emit_setcc( func
, inst
, cc_Equal
);
2274 case TGSI_OPCODE_SFL
:
2278 case TGSI_OPCODE_SGT
:
2279 emit_setcc( func
, inst
, cc_NotLessThanEqual
);
2282 case TGSI_OPCODE_SIN
:
2283 FETCH( func
, *inst
, 0, 0, CHAN_X
);
2284 emit_sin( func
, 0, 0 );
2285 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2286 STORE( func
, *inst
, 0, 0, chan_index
);
2290 case TGSI_OPCODE_SLE
:
2291 emit_setcc( func
, inst
, cc_LessThanEqual
);
2294 case TGSI_OPCODE_SNE
:
2295 emit_setcc( func
, inst
, cc_NotEqual
);
2298 case TGSI_OPCODE_STR
:
2302 case TGSI_OPCODE_TEX
:
2303 emit_tex( func
, inst
, FALSE
, FALSE
);
2306 case TGSI_OPCODE_TXD
:
2310 case TGSI_OPCODE_UP2H
:
2314 case TGSI_OPCODE_UP2US
:
2318 case TGSI_OPCODE_UP4B
:
2322 case TGSI_OPCODE_UP4UB
:
2326 case TGSI_OPCODE_X2D
:
2330 case TGSI_OPCODE_ARA
:
2334 case TGSI_OPCODE_ARR
:
2335 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2336 FETCH( func
, *inst
, 0, 0, chan_index
);
2337 emit_rnd( func
, 0, 0 );
2338 emit_f2it( func
, 0 );
2339 STORE( func
, *inst
, 0, 0, chan_index
);
2343 case TGSI_OPCODE_BRA
:
2347 case TGSI_OPCODE_CAL
:
2351 case TGSI_OPCODE_RET
:
2355 case TGSI_OPCODE_END
:
2358 case TGSI_OPCODE_SSG
:
2359 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2360 FETCH( func
, *inst
, 0, 0, chan_index
);
2361 emit_sgn( func
, 0, 0 );
2362 STORE( func
, *inst
, 0, 0, chan_index
);
2366 case TGSI_OPCODE_CMP
:
2367 emit_cmp (func
, inst
);
2370 case TGSI_OPCODE_SCS
:
2371 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_X
) {
2372 FETCH( func
, *inst
, 0, 0, CHAN_X
);
2373 emit_cos( func
, 0, 0 );
2374 STORE( func
, *inst
, 0, 0, CHAN_X
);
2376 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Y
) {
2377 FETCH( func
, *inst
, 0, 0, CHAN_X
);
2378 emit_sin( func
, 0, 0 );
2379 STORE( func
, *inst
, 0, 0, CHAN_Y
);
2381 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_Z
) {
2385 TGSI_EXEC_TEMP_00000000_I
,
2386 TGSI_EXEC_TEMP_00000000_C
);
2387 STORE( func
, *inst
, 0, 0, CHAN_Z
);
2389 IF_IS_DST0_CHANNEL_ENABLED( *inst
, CHAN_W
) {
2395 STORE( func
, *inst
, 0, 0, CHAN_W
);
2399 case TGSI_OPCODE_TXB
:
2400 emit_tex( func
, inst
, TRUE
, FALSE
);
2403 case TGSI_OPCODE_NRM
:
2405 case TGSI_OPCODE_NRM4
:
2406 /* 3 or 4-component normalization */
2408 uint dims
= (inst
->Instruction
.Opcode
== TGSI_OPCODE_NRM
) ? 3 : 4;
2410 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_X
) ||
2411 IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_Y
) ||
2412 IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_Z
) ||
2413 (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_W
) && dims
== 4)) {
2415 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
2418 /* xmm0 = src.x * src.x */
2419 FETCH(func
, *inst
, 0, 0, CHAN_X
);
2420 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_X
)) {
2421 emit_MOV(func
, 4, 0);
2423 emit_mul(func
, 0, 0);
2426 /* xmm0 = xmm0 + src.y * src.y */
2427 FETCH(func
, *inst
, 1, 0, CHAN_Y
);
2428 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_Y
)) {
2429 emit_MOV(func
, 5, 1);
2431 emit_mul(func
, 1, 1);
2432 emit_add(func
, 0, 1);
2435 /* xmm0 = xmm0 + src.z * src.z */
2436 FETCH(func
, *inst
, 1, 0, CHAN_Z
);
2437 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_Z
)) {
2438 emit_MOV(func
, 6, 1);
2440 emit_mul(func
, 1, 1);
2441 emit_add(func
, 0, 1);
2445 /* xmm0 = xmm0 + src.w * src.w */
2446 FETCH(func
, *inst
, 1, 0, CHAN_W
);
2447 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_W
)) {
2448 emit_MOV(func
, 7, 1);
2450 emit_mul(func
, 1, 1);
2451 emit_add(func
, 0, 1);
2454 /* xmm1 = 1 / sqrt(xmm0) */
2455 emit_rsqrt(func
, 1, 0);
2457 /* dst.x = xmm1 * src.x */
2458 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_X
)) {
2459 emit_mul(func
, 4, 1);
2460 STORE(func
, *inst
, 4, 0, CHAN_X
);
2463 /* dst.y = xmm1 * src.y */
2464 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_Y
)) {
2465 emit_mul(func
, 5, 1);
2466 STORE(func
, *inst
, 5, 0, CHAN_Y
);
2469 /* dst.z = xmm1 * src.z */
2470 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_Z
)) {
2471 emit_mul(func
, 6, 1);
2472 STORE(func
, *inst
, 6, 0, CHAN_Z
);
2475 /* dst.w = xmm1 * src.w */
2476 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_X
) && dims
== 4) {
2477 emit_mul(func
, 7, 1);
2478 STORE(func
, *inst
, 7, 0, CHAN_W
);
2483 if (IS_DST0_CHANNEL_ENABLED(*inst
, CHAN_W
) && dims
== 3) {
2484 emit_tempf(func
, 0, TEMP_ONE_I
, TEMP_ONE_C
);
2485 STORE(func
, *inst
, 0, 0, CHAN_W
);
2490 case TGSI_OPCODE_DIV
:
2494 case TGSI_OPCODE_DP2
:
2495 FETCH( func
, *inst
, 0, 0, CHAN_X
); /* xmm0 = src[0].x */
2496 FETCH( func
, *inst
, 1, 1, CHAN_X
); /* xmm1 = src[1].x */
2497 emit_mul( func
, 0, 1 ); /* xmm0 = xmm0 * xmm1 */
2498 FETCH( func
, *inst
, 1, 0, CHAN_Y
); /* xmm1 = src[0].y */
2499 FETCH( func
, *inst
, 2, 1, CHAN_Y
); /* xmm2 = src[1].y */
2500 emit_mul( func
, 1, 2 ); /* xmm1 = xmm1 * xmm2 */
2501 emit_add( func
, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2502 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2503 STORE( func
, *inst
, 0, 0, chan_index
); /* dest[ch] = xmm0 */
2507 case TGSI_OPCODE_TXL
:
2508 emit_tex( func
, inst
, TRUE
, FALSE
);
2511 case TGSI_OPCODE_TXP
:
2512 emit_tex( func
, inst
, FALSE
, TRUE
);
2515 case TGSI_OPCODE_BRK
:
2519 case TGSI_OPCODE_IF
:
2523 case TGSI_OPCODE_BGNFOR
:
2527 case TGSI_OPCODE_REP
:
2531 case TGSI_OPCODE_ELSE
:
2535 case TGSI_OPCODE_ENDIF
:
2539 case TGSI_OPCODE_ENDFOR
:
2543 case TGSI_OPCODE_ENDREP
:
2547 case TGSI_OPCODE_PUSHA
:
2551 case TGSI_OPCODE_POPA
:
2555 case TGSI_OPCODE_CEIL
:
2559 case TGSI_OPCODE_I2F
:
2563 case TGSI_OPCODE_NOT
:
2567 case TGSI_OPCODE_TRUNC
:
2568 FOR_EACH_DST0_ENABLED_CHANNEL( *inst
, chan_index
) {
2569 FETCH( func
, *inst
, 0, 0, chan_index
);
2570 emit_f2it( func
, 0 );
2571 emit_i2f( func
, 0 );
2572 STORE( func
, *inst
, 0, 0, chan_index
);
2576 case TGSI_OPCODE_SHL
:
2580 case TGSI_OPCODE_SHR
:
2584 case TGSI_OPCODE_AND
:
2588 case TGSI_OPCODE_OR
:
2592 case TGSI_OPCODE_MOD
:
2596 case TGSI_OPCODE_XOR
:
2600 case TGSI_OPCODE_SAD
:
2604 case TGSI_OPCODE_TXF
:
2608 case TGSI_OPCODE_TXQ
:
2612 case TGSI_OPCODE_CONT
:
2616 case TGSI_OPCODE_EMIT
:
2620 case TGSI_OPCODE_ENDPRIM
:
2633 struct x86_function
*func
,
2634 struct tgsi_full_declaration
*decl
)
2636 if( decl
->Declaration
.File
== TGSI_FILE_INPUT
) {
2637 unsigned first
, last
, mask
;
2640 first
= decl
->Range
.First
;
2641 last
= decl
->Range
.Last
;
2642 mask
= decl
->Declaration
.UsageMask
;
2644 for( i
= first
; i
<= last
; i
++ ) {
2645 for( j
= 0; j
< NUM_CHANNELS
; j
++ ) {
2646 if( mask
& (1 << j
) ) {
2647 switch( decl
->Declaration
.Interpolate
) {
2648 case TGSI_INTERPOLATE_CONSTANT
:
2649 emit_coef_a0( func
, 0, i
, j
);
2650 emit_inputs( func
, 0, i
, j
);
2653 case TGSI_INTERPOLATE_LINEAR
:
2654 emit_tempf( func
, 0, 0, TGSI_SWIZZLE_X
);
2655 emit_coef_dadx( func
, 1, i
, j
);
2656 emit_tempf( func
, 2, 0, TGSI_SWIZZLE_Y
);
2657 emit_coef_dady( func
, 3, i
, j
);
2658 emit_mul( func
, 0, 1 ); /* x * dadx */
2659 emit_coef_a0( func
, 4, i
, j
);
2660 emit_mul( func
, 2, 3 ); /* y * dady */
2661 emit_add( func
, 0, 4 ); /* x * dadx + a0 */
2662 emit_add( func
, 0, 2 ); /* x * dadx + y * dady + a0 */
2663 emit_inputs( func
, 0, i
, j
);
2666 case TGSI_INTERPOLATE_PERSPECTIVE
:
2667 emit_tempf( func
, 0, 0, TGSI_SWIZZLE_X
);
2668 emit_coef_dadx( func
, 1, i
, j
);
2669 emit_tempf( func
, 2, 0, TGSI_SWIZZLE_Y
);
2670 emit_coef_dady( func
, 3, i
, j
);
2671 emit_mul( func
, 0, 1 ); /* x * dadx */
2672 emit_tempf( func
, 4, 0, TGSI_SWIZZLE_W
);
2673 emit_coef_a0( func
, 5, i
, j
);
2674 emit_rcp( func
, 4, 4 ); /* 1.0 / w */
2675 emit_mul( func
, 2, 3 ); /* y * dady */
2676 emit_add( func
, 0, 5 ); /* x * dadx + a0 */
2677 emit_add( func
, 0, 2 ); /* x * dadx + y * dady + a0 */
2678 emit_mul( func
, 0, 4 ); /* (x * dadx + y * dady + a0) / w */
2679 emit_inputs( func
, 0, i
, j
);
2692 static void aos_to_soa( struct x86_function
*func
,
2698 struct x86_reg soa_input
= x86_make_reg( file_REG32
, reg_AX
);
2699 struct x86_reg aos_input
= x86_make_reg( file_REG32
, reg_BX
);
2700 struct x86_reg num_inputs
= x86_make_reg( file_REG32
, reg_CX
);
2701 struct x86_reg stride
= x86_make_reg( file_REG32
, reg_DX
);
2706 x86_push( func
, x86_make_reg( file_REG32
, reg_BX
) );
2708 x86_mov( func
, aos_input
, x86_fn_arg( func
, arg_aos
) );
2709 x86_mov( func
, soa_input
, x86_fn_arg( func
, arg_machine
) );
2710 x86_lea( func
, soa_input
,
2711 x86_make_disp( soa_input
,
2712 Offset(struct tgsi_exec_machine
, Inputs
) ) );
2713 x86_mov( func
, num_inputs
, x86_fn_arg( func
, arg_num
) );
2714 x86_mov( func
, stride
, x86_fn_arg( func
, arg_stride
) );
2717 inner_loop
= x86_get_label( func
);
2719 x86_push( func
, aos_input
);
2720 sse_movlps( func
, make_xmm( 0 ), x86_make_disp( aos_input
, 0 ) );
2721 sse_movlps( func
, make_xmm( 3 ), x86_make_disp( aos_input
, 8 ) );
2722 x86_add( func
, aos_input
, stride
);
2723 sse_movhps( func
, make_xmm( 0 ), x86_make_disp( aos_input
, 0 ) );
2724 sse_movhps( func
, make_xmm( 3 ), x86_make_disp( aos_input
, 8 ) );
2725 x86_add( func
, aos_input
, stride
);
2726 sse_movlps( func
, make_xmm( 1 ), x86_make_disp( aos_input
, 0 ) );
2727 sse_movlps( func
, make_xmm( 4 ), x86_make_disp( aos_input
, 8 ) );
2728 x86_add( func
, aos_input
, stride
);
2729 sse_movhps( func
, make_xmm( 1 ), x86_make_disp( aos_input
, 0 ) );
2730 sse_movhps( func
, make_xmm( 4 ), x86_make_disp( aos_input
, 8 ) );
2731 x86_pop( func
, aos_input
);
2733 sse_movaps( func
, make_xmm( 2 ), make_xmm( 0 ) );
2734 sse_movaps( func
, make_xmm( 5 ), make_xmm( 3 ) );
2735 sse_shufps( func
, make_xmm( 0 ), make_xmm( 1 ), 0x88 );
2736 sse_shufps( func
, make_xmm( 2 ), make_xmm( 1 ), 0xdd );
2737 sse_shufps( func
, make_xmm( 3 ), make_xmm( 4 ), 0x88 );
2738 sse_shufps( func
, make_xmm( 5 ), make_xmm( 4 ), 0xdd );
2740 sse_movups( func
, x86_make_disp( soa_input
, 0 ), make_xmm( 0 ) );
2741 sse_movups( func
, x86_make_disp( soa_input
, 16 ), make_xmm( 2 ) );
2742 sse_movups( func
, x86_make_disp( soa_input
, 32 ), make_xmm( 3 ) );
2743 sse_movups( func
, x86_make_disp( soa_input
, 48 ), make_xmm( 5 ) );
2745 /* Advance to next input */
2746 x86_lea( func
, aos_input
, x86_make_disp(aos_input
, 16) );
2747 x86_lea( func
, soa_input
, x86_make_disp(soa_input
, 64) );
2749 /* while --num_inputs */
2750 x86_dec( func
, num_inputs
);
2751 x86_jcc( func
, cc_NE
, inner_loop
);
2754 x86_pop( func
, x86_make_reg( file_REG32
, reg_BX
) );
2757 static void soa_to_aos( struct x86_function
*func
,
2763 struct x86_reg soa_output
= x86_make_reg( file_REG32
, reg_AX
);
2764 struct x86_reg aos_output
= x86_make_reg( file_REG32
, reg_BX
);
2765 struct x86_reg num_outputs
= x86_make_reg( file_REG32
, reg_CX
);
2766 struct x86_reg temp
= x86_make_reg( file_REG32
, reg_DX
);
2770 x86_push( func
, x86_make_reg( file_REG32
, reg_BX
) );
2772 x86_mov( func
, aos_output
, x86_fn_arg( func
, arg_aos
) );
2773 x86_mov( func
, soa_output
, x86_fn_arg( func
, arg_machine
) );
2774 x86_lea( func
, soa_output
,
2775 x86_make_disp( soa_output
,
2776 Offset(struct tgsi_exec_machine
, Outputs
) ) );
2777 x86_mov( func
, num_outputs
, x86_fn_arg( func
, arg_num
) );
2780 inner_loop
= x86_get_label( func
);
2782 sse_movups( func
, make_xmm( 0 ), x86_make_disp( soa_output
, 0 ) );
2783 sse_movups( func
, make_xmm( 1 ), x86_make_disp( soa_output
, 16 ) );
2784 sse_movups( func
, make_xmm( 3 ), x86_make_disp( soa_output
, 32 ) );
2785 sse_movups( func
, make_xmm( 4 ), x86_make_disp( soa_output
, 48 ) );
2787 sse_movaps( func
, make_xmm( 2 ), make_xmm( 0 ) );
2788 sse_movaps( func
, make_xmm( 5 ), make_xmm( 3 ) );
2789 sse_unpcklps( func
, make_xmm( 0 ), make_xmm( 1 ) );
2790 sse_unpckhps( func
, make_xmm( 2 ), make_xmm( 1 ) );
2791 sse_unpcklps( func
, make_xmm( 3 ), make_xmm( 4 ) );
2792 sse_unpckhps( func
, make_xmm( 5 ), make_xmm( 4 ) );
2794 x86_mov( func
, temp
, x86_fn_arg( func
, arg_stride
) );
2795 x86_push( func
, aos_output
);
2796 sse_movlps( func
, x86_make_disp( aos_output
, 0 ), make_xmm( 0 ) );
2797 sse_movlps( func
, x86_make_disp( aos_output
, 8 ), make_xmm( 3 ) );
2798 x86_add( func
, aos_output
, temp
);
2799 sse_movhps( func
, x86_make_disp( aos_output
, 0 ), make_xmm( 0 ) );
2800 sse_movhps( func
, x86_make_disp( aos_output
, 8 ), make_xmm( 3 ) );
2801 x86_add( func
, aos_output
, temp
);
2802 sse_movlps( func
, x86_make_disp( aos_output
, 0 ), make_xmm( 2 ) );
2803 sse_movlps( func
, x86_make_disp( aos_output
, 8 ), make_xmm( 5 ) );
2804 x86_add( func
, aos_output
, temp
);
2805 sse_movhps( func
, x86_make_disp( aos_output
, 0 ), make_xmm( 2 ) );
2806 sse_movhps( func
, x86_make_disp( aos_output
, 8 ), make_xmm( 5 ) );
2807 x86_pop( func
, aos_output
);
2809 /* Advance to next output */
2810 x86_lea( func
, aos_output
, x86_make_disp(aos_output
, 16) );
2811 x86_lea( func
, soa_output
, x86_make_disp(soa_output
, 64) );
2813 /* while --num_outputs */
2814 x86_dec( func
, num_outputs
);
2815 x86_jcc( func
, cc_NE
, inner_loop
);
2818 x86_pop( func
, x86_make_reg( file_REG32
, reg_BX
) );
2822 * Translate a TGSI vertex/fragment shader to SSE2 code.
2823 * Slightly different things are done for vertex vs. fragment shaders.
2825 * \param tokens the TGSI input shader
2826 * \param func the output SSE code/function
2827 * \param immediates buffer to place immediates, later passed to SSE func
2828 * \param return 1 for success, 0 if translation failed
2832 const struct tgsi_token
*tokens
,
2833 struct x86_function
*func
,
2834 float (*immediates
)[4],
2835 boolean do_swizzles
)
2837 struct tgsi_parse_context parse
;
2839 uint num_immediates
= 0;
2843 func
->csr
= func
->store
;
2845 tgsi_parse_init( &parse
, tokens
);
2847 /* Can't just use EDI, EBX without save/restoring them:
2849 x86_push( func
, x86_make_reg( file_REG32
, reg_BX
) );
2850 x86_push( func
, x86_make_reg( file_REG32
, reg_DI
) );
2853 * Different function args for vertex/fragment shaders:
2855 if (parse
.FullHeader
.Processor
.Processor
== TGSI_PROCESSOR_VERTEX
) {
2861 6 ); /* input_stride */
2867 x86_fn_arg( func
, 1 ) );
2871 x86_fn_arg( func
, 2 ) );
2874 get_immediate_base(),
2875 x86_fn_arg( func
, 3 ) );
2877 if (parse
.FullHeader
.Processor
.Processor
== TGSI_PROCESSOR_FRAGMENT
) {
2881 x86_fn_arg( func
, 4 ) );
2887 x86_make_disp( get_machine_base(),
2888 Offset( struct tgsi_exec_machine
, Samplers
) ) );
2891 while( !tgsi_parse_end_of_tokens( &parse
) && ok
) {
2892 tgsi_parse_token( &parse
);
2894 switch( parse
.FullToken
.Token
.Type
) {
2895 case TGSI_TOKEN_TYPE_DECLARATION
:
2896 if (parse
.FullHeader
.Processor
.Processor
== TGSI_PROCESSOR_FRAGMENT
) {
2899 &parse
.FullToken
.FullDeclaration
);
2903 case TGSI_TOKEN_TYPE_INSTRUCTION
:
2904 ok
= emit_instruction(
2906 &parse
.FullToken
.FullInstruction
);
2909 uint opcode
= parse
.FullToken
.FullInstruction
.Instruction
.Opcode
;
2910 debug_printf("failed to translate tgsi opcode %d (%s) to SSE (%s)\n",
2912 tgsi_get_opcode_name(opcode
),
2913 parse
.FullHeader
.Processor
.Processor
== TGSI_PROCESSOR_VERTEX
?
2914 "vertex shader" : "fragment shader");
2917 if (tgsi_check_soa_dependencies(&parse
.FullToken
.FullInstruction
)) {
2918 uint opcode
= parse
.FullToken
.FullInstruction
.Instruction
.Opcode
;
2920 /* XXX: we only handle src/dst aliasing in a few opcodes
2921 * currently. Need to use an additional temporay to hold
2922 * the result in the cases where the code is too opaque to
2925 if (opcode
!= TGSI_OPCODE_MOV
) {
2926 debug_printf("Warning: src/dst aliasing in instruction"
2927 " is not handled:\n");
2928 tgsi_dump_instruction(&parse
.FullToken
.FullInstruction
, 1);
2933 case TGSI_TOKEN_TYPE_IMMEDIATE
:
2934 /* simply copy the immediate values into the next immediates[] slot */
2936 const uint size
= parse
.FullToken
.FullImmediate
.Immediate
.NrTokens
- 1;
2939 assert(num_immediates
< TGSI_EXEC_NUM_IMMEDIATES
);
2940 for( i
= 0; i
< size
; i
++ ) {
2941 immediates
[num_immediates
][i
] =
2942 parse
.FullToken
.FullImmediate
.u
[i
].Float
;
2945 debug_printf("SSE FS immediate[%d] = %f %f %f %f\n",
2947 immediates
[num_immediates
][0],
2948 immediates
[num_immediates
][1],
2949 immediates
[num_immediates
][2],
2950 immediates
[num_immediates
][3]);
2962 if (parse
.FullHeader
.Processor
.Processor
== TGSI_PROCESSOR_VERTEX
) {
2967 8, /* num_outputs */
2968 9 ); /* output_stride */
2971 /* Can't just use EBX, EDI without save/restoring them:
2973 x86_pop( func
, x86_make_reg( file_REG32
, reg_DI
) );
2974 x86_pop( func
, x86_make_reg( file_REG32
, reg_BX
) );
2978 tgsi_parse_free( &parse
);
2983 #endif /* PIPE_ARCH_X86 */