1 /**************************************************************************
3 * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
5 * Copyright 2009-2010 VMware, Inc. All rights Reserved.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
29 #include "pipe/p_config.h"
31 #include "tgsi/tgsi_sse2.h"
33 #if defined(PIPE_ARCH_X86) && 0 /* See FIXME notes below */
35 #include "util/u_debug.h"
36 #include "pipe/p_shader_tokens.h"
37 #include "util/u_math.h"
38 #include "util/u_memory.h"
39 #if defined(PIPE_ARCH_SSE)
40 #include "util/u_sse.h"
42 #include "tgsi/tgsi_info.h"
43 #include "tgsi/tgsi_parse.h"
44 #include "tgsi/tgsi_util.h"
45 #include "tgsi/tgsi_dump.h"
46 #include "tgsi/tgsi_exec.h"
48 #include "rtasm/rtasm_x86sse.h"
52 * This costs about 100fps (close to 10%) in gears:
54 #define HIGH_PRECISION 1
59 #define FOR_EACH_CHANNEL( CHAN )\
60 for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++)
62 #define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
63 ((INST).Dst[0].Register.WriteMask & (1 << (CHAN)))
65 #define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
66 if (IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
68 #define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
69 FOR_EACH_CHANNEL( CHAN )\
70 IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
77 #define TEMP_ONE_I TGSI_EXEC_TEMP_ONE_I
78 #define TEMP_ONE_C TGSI_EXEC_TEMP_ONE_C
80 #define TEMP_R0 TGSI_EXEC_TEMP_R0
81 #define TEMP_ADDR TGSI_EXEC_TEMP_ADDR
82 #define TEMP_EXEC_MASK_I TGSI_EXEC_MASK_I
83 #define TEMP_EXEC_MASK_C TGSI_EXEC_MASK_C
87 * X86 utility functions.
96 (enum x86_reg_name) xmm );
100 * X86 register mapping helpers.
103 static struct x86_reg
104 get_const_base( void )
111 static struct x86_reg
112 get_machine_base( void )
119 static struct x86_reg
120 get_input_base( void )
122 /* FIXME: tgsi_exec_machine::Inputs is a pointer now! */
123 return x86_make_disp(
125 Offset(struct tgsi_exec_machine, Inputs) );
128 static struct x86_reg
129 get_output_base( void )
131 /* FIXME: tgsi_exec_machine::Ouputs is a pointer now! */
132 return x86_make_disp(
134 Offset(struct tgsi_exec_machine, Outputs) );
137 static struct x86_reg
138 get_temp_base( void )
140 return x86_make_disp(
142 Offset(struct tgsi_exec_machine, Temps) );
145 static struct x86_reg
146 get_coef_base( void )
153 static struct x86_reg
154 get_sampler_base( void )
161 static struct x86_reg
162 get_immediate_base( void )
169 static struct x86_reg
170 get_system_value_base( void )
172 return x86_make_disp(
174 Offset(struct tgsi_exec_machine, SystemValue) );
179 * Data access helpers.
183 static struct x86_reg
188 return x86_make_disp(
189 get_immediate_base(),
190 (vec * 4 + chan) * 4 );
193 static struct x86_reg
198 return x86_make_disp(
200 (vec * 4 + chan) * 4 );
203 static struct x86_reg
207 return x86_make_disp(
209 unit * sizeof( struct tgsi_sampler * ) );
212 static struct x86_reg
217 return x86_make_disp(
219 (vec * 4 + chan) * 16 );
222 static struct x86_reg
227 return x86_make_disp(
229 (vec * 4 + chan) * 16 );
232 static struct x86_reg
237 return x86_make_disp(
239 (vec * 4 + chan) * 16 );
242 static struct x86_reg
247 return x86_make_disp(
248 get_system_value_base(), /* base */
249 (vec * 4 + chan) * 4 ); /* byte offset from base */
252 static struct x86_reg
258 return x86_make_disp(
260 ((vec * 3 + member) * 4 + chan) * 4 );
266 struct x86_function *func )
273 * Data fetch helpers.
277 * Copy a shader constant to xmm register
278 * \param xmm the destination xmm register
279 * \param vec the src const buffer index
280 * \param chan src channel to fetch (X, Y, Z or W)
284 struct x86_function *func,
293 /* 'vec' is the offset from the address register's value.
294 * We're loading CONST[ADDR+vec] into an xmm register.
296 struct x86_reg r0 = get_immediate_base();
297 struct x86_reg r1 = get_coef_base();
300 assert( indirectFile == TGSI_FILE_ADDRESS );
301 assert( indirectIndex == 0 );
302 assert( r0.mod == mod_REG );
303 assert( r1.mod == mod_REG );
305 x86_push( func, r0 );
306 x86_push( func, r1 );
309 * Loop over the four pixels or vertices in the quad.
310 * Get the value of the address (offset) register for pixel/vertex[i],
311 * add it to the src offset and index into the constant buffer.
312 * Note that we're working on SOA data.
313 * If any of the pixel/vertex execution channels are unused their
314 * values will be garbage. It's very important that we don't use
315 * those garbage values as indexes into the constant buffer since
316 * that'll cause segfaults.
317 * The solution is to bitwise-AND the offset with the execution mask
318 * register whose values are either 0 or ~0.
319 * The caller must setup the execution mask register to indicate
320 * which channels are valid/alive before running the shader.
321 * The execution mask will also figure into loops and conditionals
324 for (i = 0; i < QUAD_SIZE; i++) {
325 /* r1 = address register[i] */
326 x86_mov( func, r1, x86_make_disp( get_temp( TEMP_ADDR, CHAN_X ), i * 4 ) );
327 /* r0 = execution mask[i] */
328 x86_mov( func, r0, x86_make_disp( get_temp( TEMP_EXEC_MASK_I, TEMP_EXEC_MASK_C ), i * 4 ) );
330 x86_and( func, r1, r0 );
331 /* r0 = 'vec', the offset */
332 x86_lea( func, r0, get_const( vec, chan ) );
334 /* Quick hack to multiply r1 by 16 -- need to add SHL to rtasm.
336 x86_add( func, r1, r1 );
337 x86_add( func, r1, r1 );
338 x86_add( func, r1, r1 );
339 x86_add( func, r1, r1 );
341 x86_add( func, r0, r1 ); /* r0 = r0 + r1 */
342 x86_mov( func, r1, x86_deref( r0 ) );
343 x86_mov( func, x86_make_disp( get_temp( TEMP_R0, CHAN_X ), i * 4 ), r1 );
352 get_temp( TEMP_R0, CHAN_X ) );
355 /* 'vec' is the index into the src register file, such as TEMP[vec] */
361 get_const( vec, chan ) );
366 SHUF( 0, 0, 0, 0 ) );
372 struct x86_function *func,
380 get_immediate( vec, chan ) );
385 SHUF( 0, 0, 0, 0 ) );
390 * Copy a shader input to xmm register
391 * \param xmm the destination xmm register
392 * \param vec the src input attrib
393 * \param chan src channel to fetch (X, Y, Z or W)
397 struct x86_function *func,
405 get_input( vec, chan ) );
409 * Store an xmm register to a shader output
410 * \param xmm the source xmm register
411 * \param vec the dest output attrib
412 * \param chan src dest channel to store (X, Y, Z or W)
416 struct x86_function *func,
423 get_output( vec, chan ),
428 * Copy a shader temporary to xmm register
429 * \param xmm the destination xmm register
430 * \param vec the src temp register
431 * \param chan src channel to fetch (X, Y, Z or W)
435 struct x86_function *func,
443 get_temp( vec, chan ) );
447 * Copy a system value to xmm register
448 * \param xmm the destination xmm register
449 * \param vec the source system value register
450 * \param chan src channel to fetch (X, Y, Z or W)
454 struct x86_function *func,
462 get_system_value( vec, chan ) );
467 SHUF( 0, 0, 0, 0 ) );
471 * Load an xmm register with an input attrib coefficient (a0, dadx or dady)
472 * \param xmm the destination xmm register
473 * \param vec the src input/attribute coefficient index
474 * \param chan src channel to fetch (X, Y, Z or W)
475 * \param member 0=a0, 1=dadx, 2=dady
479 struct x86_function *func,
488 get_coef( vec, chan, member ) );
493 SHUF( 0, 0, 0, 0 ) );
497 * Data store helpers.
502 struct x86_function *func,
509 get_input( vec, chan ),
515 struct x86_function *func,
522 get_temp( vec, chan ),
528 struct x86_function *func,
538 vec + TGSI_EXEC_TEMP_ADDR,
543 * Coefficent fetch helpers.
548 struct x86_function *func,
563 struct x86_function *func,
578 struct x86_function *func,
592 * Function call helpers.
596 * NOTE: In gcc, if the destination uses the SSE intrinsics, then it must be
597 * defined with __attribute__((force_align_arg_pointer)), as we do not guarantee
598 * that the stack pointer is 16 byte aligned, as expected.
602 struct x86_function *func,
603 unsigned xmm_save_mask,
604 const struct x86_reg *arg,
606 void (PIPE_CDECL *code)() )
608 struct x86_reg ecx = x86_make_reg( file_REG32, reg_CX );
613 x86_make_reg( file_REG32, reg_AX) );
616 x86_make_reg( file_REG32, reg_CX) );
619 x86_make_reg( file_REG32, reg_DX) );
621 /* Store XMM regs to the stack
623 for(i = 0, n = 0; i < 8; ++i)
624 if(xmm_save_mask & (1 << i))
629 x86_make_reg( file_REG32, reg_SP ),
632 for(i = 0, n = 0; i < 8; ++i)
633 if(xmm_save_mask & (1 << i)) {
636 x86_make_disp( x86_make_reg( file_REG32, reg_SP ), n*16 ),
641 for (i = 0; i < nr_args; i++) {
642 /* Load the address of the buffer we use for passing arguments and
650 /* Push actual function arguments (currently just the pointer to
651 * the buffer above), and call the function:
653 x86_push( func, ecx );
656 x86_mov_reg_imm( func, ecx, (unsigned long) code );
657 x86_call( func, ecx );
659 /* Pop the arguments (or just add an immediate to esp)
661 for (i = 0; i < nr_args; i++) {
665 /* Pop the saved XMM regs:
667 for(i = 0, n = 0; i < 8; ++i)
668 if(xmm_save_mask & (1 << i)) {
672 x86_make_disp( x86_make_reg( file_REG32, reg_SP ), n*16 ) );
678 x86_make_reg( file_REG32, reg_SP ),
681 /* Restore GP registers in a reverse order.
685 x86_make_reg( file_REG32, reg_DX) );
688 x86_make_reg( file_REG32, reg_CX) );
691 x86_make_reg( file_REG32, reg_AX) );
695 emit_func_call_dst_src1(
696 struct x86_function *func,
700 void (PIPE_CDECL *code)() )
702 struct x86_reg store = get_temp( TEMP_R0, 0 );
703 unsigned xmm_mask = ((1 << xmm_save) - 1) & ~(1 << xmm_dst);
705 /* Store our input parameters (in xmm regs) to the buffer we use
706 * for passing arguments. We will pass a pointer to this buffer as
707 * the actual function argument.
712 make_xmm( xmm_src0 ) );
714 emit_func_call( func,
728 emit_func_call_dst_src2(
729 struct x86_function *func,
734 void (PIPE_CDECL *code)() )
736 struct x86_reg store = get_temp( TEMP_R0, 0 );
737 unsigned xmm_mask = ((1 << xmm_save) - 1) & ~(1 << xmm_dst);
739 /* Store two inputs to parameter buffer.
744 make_xmm( xmm_src0 ) );
748 x86_make_disp( store, 4 * sizeof(float) ),
749 make_xmm( xmm_src1 ) );
754 emit_func_call( func,
760 /* Retrieve the results:
772 #if defined(PIPE_ARCH_SSE)
775 * Fast SSE2 implementation of special math functions.
778 #define POLY0(x, c0) _mm_set1_ps(c0)
779 #define POLY1(x, c0, c1) _mm_add_ps(_mm_mul_ps(POLY0(x, c1), x), _mm_set1_ps(c0))
780 #define POLY2(x, c0, c1, c2) _mm_add_ps(_mm_mul_ps(POLY1(x, c1, c2), x), _mm_set1_ps(c0))
781 #define POLY3(x, c0, c1, c2, c3) _mm_add_ps(_mm_mul_ps(POLY2(x, c1, c2, c3), x), _mm_set1_ps(c0))
782 #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))
783 #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))
785 #define EXP_POLY_DEGREE 3
786 #define LOG_POLY_DEGREE 5
789 * See http://www.devmaster.net/forums/showthread.php?p=43580
795 __m128 fpart, expipart, expfpart;
797 x = _mm_min_ps(x, _mm_set1_ps( 129.00000f));
798 x = _mm_max_ps(x, _mm_set1_ps(-126.99999f));
800 /* ipart = int(x - 0.5) */
801 ipart = _mm_cvtps_epi32(_mm_sub_ps(x, _mm_set1_ps(0.5f)));
803 /* fpart = x - ipart */
804 fpart = _mm_sub_ps(x, _mm_cvtepi32_ps(ipart));
806 /* expipart = (float) (1 << ipart) */
807 expipart = _mm_castsi128_ps(_mm_slli_epi32(_mm_add_epi32(ipart, _mm_set1_epi32(127)), 23));
809 /* minimax polynomial fit of 2**x, in range [-0.5, 0.5[ */
810 #if EXP_POLY_DEGREE == 5
811 expfpart = POLY5(fpart, 9.9999994e-1f, 6.9315308e-1f, 2.4015361e-1f, 5.5826318e-2f, 8.9893397e-3f, 1.8775767e-3f);
812 #elif EXP_POLY_DEGREE == 4
813 expfpart = POLY4(fpart, 1.0000026f, 6.9300383e-1f, 2.4144275e-1f, 5.2011464e-2f, 1.3534167e-2f);
814 #elif EXP_POLY_DEGREE == 3
815 expfpart = POLY3(fpart, 9.9992520e-1f, 6.9583356e-1f, 2.2606716e-1f, 7.8024521e-2f);
816 #elif EXP_POLY_DEGREE == 2
817 expfpart = POLY2(fpart, 1.0017247f, 6.5763628e-1f, 3.3718944e-1f);
822 return _mm_mul_ps(expipart, expfpart);
827 * See http://www.devmaster.net/forums/showthread.php?p=43580
832 __m128i expmask = _mm_set1_epi32(0x7f800000);
833 __m128i mantmask = _mm_set1_epi32(0x007fffff);
834 __m128 one = _mm_set1_ps(1.0f);
836 __m128i i = _mm_castps_si128(x);
838 /* exp = (float) exponent(x) */
839 __m128 exp = _mm_cvtepi32_ps(_mm_sub_epi32(_mm_srli_epi32(_mm_and_si128(i, expmask), 23), _mm_set1_epi32(127)));
841 /* mant = (float) mantissa(x) */
842 __m128 mant = _mm_or_ps(_mm_castsi128_ps(_mm_and_si128(i, mantmask)), one);
846 /* Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[
847 * These coefficients can be generate with
848 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
850 #if LOG_POLY_DEGREE == 6
851 logmant = POLY5(mant, 3.11578814719469302614f, -3.32419399085241980044f, 2.59883907202499966007f, -1.23152682416275988241f, 0.318212422185251071475f, -0.0344359067839062357313f);
852 #elif LOG_POLY_DEGREE == 5
853 logmant = POLY4(mant, 2.8882704548164776201f, -2.52074962577807006663f, 1.48116647521213171641f, -0.465725644288844778798f, 0.0596515482674574969533f);
854 #elif LOG_POLY_DEGREE == 4
855 logmant = POLY3(mant, 2.61761038894603480148f, -1.75647175389045657003f, 0.688243882994381274313f, -0.107254423828329604454f);
856 #elif LOG_POLY_DEGREE == 3
857 logmant = POLY2(mant, 2.28330284476918490682f, -1.04913055217340124191f, 0.204446009836232697516f);
862 /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
863 logmant = _mm_mul_ps(logmant, _mm_sub_ps(mant, one));
865 return _mm_add_ps(logmant, exp);
870 powf4(__m128 x, __m128 y)
872 return exp2f4(_mm_mul_ps(log2f4(x), y));
875 #endif /* PIPE_ARCH_SSE */
880 * Low-level instruction translators.
885 struct x86_function *func,
892 TGSI_EXEC_TEMP_7FFFFFFF_I,
893 TGSI_EXEC_TEMP_7FFFFFFF_C ) );
898 struct x86_function *func,
905 make_xmm( xmm_src ) );
908 static void PIPE_CDECL
912 store[0] = cosf( store[0] );
913 store[1] = cosf( store[1] );
914 store[2] = cosf( store[2] );
915 store[3] = cosf( store[3] );
920 struct x86_function *func,
924 emit_func_call_dst_src1(
932 static void PIPE_CDECL
933 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_SSE)
934 __attribute__((force_align_arg_pointer))
939 #if defined(PIPE_ARCH_SSE)
940 _mm_store_ps(&store[0], exp2f4( _mm_load_ps(&store[0]) ));
942 store[0] = util_fast_exp2( store[0] );
943 store[1] = util_fast_exp2( store[1] );
944 store[2] = util_fast_exp2( store[2] );
945 store[3] = util_fast_exp2( store[3] );
951 struct x86_function *func,
955 emit_func_call_dst_src1(
965 struct x86_function *func,
976 struct x86_function *func,
985 static void PIPE_CDECL
989 store[0] = floorf( store[0] );
990 store[1] = floorf( store[1] );
991 store[2] = floorf( store[2] );
992 store[3] = floorf( store[3] );
997 struct x86_function *func,
1001 emit_func_call_dst_src1(
1009 static void PIPE_CDECL
1013 store[0] -= floorf( store[0] );
1014 store[1] -= floorf( store[1] );
1015 store[2] -= floorf( store[2] );
1016 store[3] -= floorf( store[3] );
1021 struct x86_function *func,
1025 emit_func_call_dst_src1(
1033 static void PIPE_CDECL
1034 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_SSE)
1035 __attribute__((force_align_arg_pointer))
1040 #if defined(PIPE_ARCH_SSE)
1041 _mm_store_ps(&store[0], log2f4( _mm_load_ps(&store[0]) ));
1043 store[0] = util_fast_log2( store[0] );
1044 store[1] = util_fast_log2( store[1] );
1045 store[2] = util_fast_log2( store[2] );
1046 store[3] = util_fast_log2( store[3] );
1052 struct x86_function *func,
1056 emit_func_call_dst_src1(
1066 struct x86_function *func,
1072 make_xmm( xmm_dst ),
1073 make_xmm( xmm_src ) );
1077 emit_mul (struct x86_function *func,
1083 make_xmm( xmm_dst ),
1084 make_xmm( xmm_src ) );
1089 struct x86_function *func,
1096 TGSI_EXEC_TEMP_80000000_I,
1097 TGSI_EXEC_TEMP_80000000_C ) );
1100 static void PIPE_CDECL
1101 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_SSE)
1102 __attribute__((force_align_arg_pointer))
1107 #if defined(PIPE_ARCH_SSE)
1108 _mm_store_ps(&store[0], powf4( _mm_load_ps(&store[0]), _mm_load_ps(&store[4]) ));
1110 store[0] = util_fast_pow( store[0], store[4] );
1111 store[1] = util_fast_pow( store[1], store[5] );
1112 store[2] = util_fast_pow( store[2], store[6] );
1113 store[3] = util_fast_pow( store[3], store[7] );
1119 struct x86_function *func,
1125 emit_func_call_dst_src2(
1136 struct x86_function *func,
1140 /* On Intel CPUs at least, this is only accurate to 12 bits -- not
1141 * good enough. Need to either emit a proper divide or use the
1142 * iterative technique described below in emit_rsqrt().
1146 make_xmm( xmm_dst ),
1147 make_xmm( xmm_src ) );
1150 static void PIPE_CDECL
1154 store[0] = floorf( store[0] + 0.5f );
1155 store[1] = floorf( store[1] + 0.5f );
1156 store[2] = floorf( store[2] + 0.5f );
1157 store[3] = floorf( store[3] + 0.5f );
1162 struct x86_function *func,
1166 emit_func_call_dst_src1(
1176 struct x86_function *func,
1181 /* Although rsqrtps() and rcpps() are low precision on some/all SSE
1182 * implementations, it is possible to improve its precision at
1183 * fairly low cost, using a newton/raphson step, as below:
1185 * x1 = 2 * rcpps(a) - a * rcpps(a) * rcpps(a)
1186 * x1 = 0.5 * rsqrtps(a) * [3.0 - (a * rsqrtps(a))* rsqrtps(a)]
1188 * See: http://softwarecommunity.intel.com/articles/eng/1818.htm
1191 struct x86_reg dst = make_xmm( xmm_dst );
1192 struct x86_reg src = make_xmm( xmm_src );
1193 struct x86_reg tmp0 = make_xmm( 2 );
1194 struct x86_reg tmp1 = make_xmm( 3 );
1196 assert( xmm_dst != xmm_src );
1197 assert( xmm_dst != 2 && xmm_dst != 3 );
1198 assert( xmm_src != 2 && xmm_src != 3 );
1200 sse_movaps( func, dst, get_temp( TGSI_EXEC_TEMP_HALF_I, TGSI_EXEC_TEMP_HALF_C ) );
1201 sse_movaps( func, tmp0, get_temp( TGSI_EXEC_TEMP_THREE_I, TGSI_EXEC_TEMP_THREE_C ) );
1202 sse_rsqrtps( func, tmp1, src );
1203 sse_mulps( func, src, tmp1 );
1204 sse_mulps( func, dst, tmp1 );
1205 sse_mulps( func, src, tmp1 );
1206 sse_subps( func, tmp0, src );
1207 sse_mulps( func, dst, tmp0 );
1210 /* On Intel CPUs at least, this is only accurate to 12 bits -- not
1215 make_xmm( xmm_dst ),
1216 make_xmm( xmm_src ) );
1222 struct x86_function *func,
1229 TGSI_EXEC_TEMP_80000000_I,
1230 TGSI_EXEC_TEMP_80000000_C ) );
1233 static void PIPE_CDECL
1237 store[0] = store[0] < 0.0f ? -1.0f : store[0] > 0.0f ? 1.0f : 0.0f;
1238 store[1] = store[1] < 0.0f ? -1.0f : store[1] > 0.0f ? 1.0f : 0.0f;
1239 store[2] = store[2] < 0.0f ? -1.0f : store[2] > 0.0f ? 1.0f : 0.0f;
1240 store[3] = store[3] < 0.0f ? -1.0f : store[3] > 0.0f ? 1.0f : 0.0f;
1245 struct x86_function *func,
1249 emit_func_call_dst_src1(
1257 static void PIPE_CDECL
1261 store[0] = sinf( store[0] );
1262 store[1] = sinf( store[1] );
1263 store[2] = sinf( store[2] );
1264 store[3] = sinf( store[3] );
1268 emit_sin (struct x86_function *func,
1272 emit_func_call_dst_src1(
1282 struct x86_function *func,
1288 make_xmm( xmm_dst ),
1289 make_xmm( xmm_src ) );
1297 struct x86_function *func,
1299 const struct tgsi_full_src_register *reg,
1300 const unsigned chan_index )
1302 unsigned swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1305 case TGSI_SWIZZLE_X:
1306 case TGSI_SWIZZLE_Y:
1307 case TGSI_SWIZZLE_Z:
1308 case TGSI_SWIZZLE_W:
1309 switch (reg->Register.File) {
1310 case TGSI_FILE_CONSTANT:
1314 reg->Register.Index,
1316 reg->Register.Indirect,
1318 reg->Indirect.Index );
1321 case TGSI_FILE_IMMEDIATE:
1325 reg->Register.Index,
1329 case TGSI_FILE_SYSTEM_VALUE:
1333 reg->Register.Index,
1337 case TGSI_FILE_INPUT:
1341 reg->Register.Index,
1345 case TGSI_FILE_TEMPORARY:
1349 reg->Register.Index,
1362 switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) {
1363 case TGSI_UTIL_SIGN_CLEAR:
1364 emit_abs( func, xmm );
1367 case TGSI_UTIL_SIGN_SET:
1368 emit_setsign( func, xmm );
1371 case TGSI_UTIL_SIGN_TOGGLE:
1372 emit_neg( func, xmm );
1375 case TGSI_UTIL_SIGN_KEEP:
1380 #define FETCH( FUNC, INST, XMM, INDEX, CHAN )\
1381 emit_fetch( FUNC, XMM, &(INST).Src[INDEX], CHAN )
1388 struct x86_function *func,
1390 const struct tgsi_full_dst_register *reg,
1391 const struct tgsi_full_instruction *inst,
1392 unsigned chan_index )
1394 switch( inst->Instruction.Saturate ) {
1398 case TGSI_SAT_ZERO_ONE:
1403 TGSI_EXEC_TEMP_00000000_I,
1404 TGSI_EXEC_TEMP_00000000_C ) );
1410 TGSI_EXEC_TEMP_ONE_I,
1411 TGSI_EXEC_TEMP_ONE_C ) );
1414 case TGSI_SAT_MINUS_PLUS_ONE:
1420 switch( reg->Register.File ) {
1421 case TGSI_FILE_OUTPUT:
1425 reg->Register.Index,
1429 case TGSI_FILE_TEMPORARY:
1433 reg->Register.Index,
1437 case TGSI_FILE_ADDRESS:
1441 reg->Register.Index,
1450 #define STORE( FUNC, INST, XMM, INDEX, CHAN )\
1451 emit_store( FUNC, XMM, &(INST).Dst[INDEX], &(INST), CHAN )
1454 static void PIPE_CDECL
1455 fetch_texel( struct tgsi_sampler **sampler,
1461 debug_printf("%s sampler: %p (%p) store: %p\n",
1466 for (j = 0; j < 4; j++)
1467 debug_printf("sample %d texcoord %f %f %f lodbias %f\n",
1476 float rgba[NUM_CHANNELS][QUAD_SIZE];
1477 (*sampler)->get_samples(*sampler,
1481 &store[12], /* lodbias */
1482 tgsi_sampler_lod_bias,
1483 rgba); /* results */
1485 memcpy( store, rgba, 16 * sizeof(float));
1489 for (j = 0; j < 4; j++)
1490 debug_printf("sample %d result %f %f %f %f\n",
1500 * High-level instruction translators.
1503 emit_tex( struct x86_function *func,
1504 const struct tgsi_full_instruction *inst,
1508 const uint unit = inst->Src[1].Register.Index;
1509 struct x86_reg args[2];
1513 assert(inst->Instruction.Texture);
1514 switch (inst->Texture.Texture) {
1515 case TGSI_TEXTURE_1D:
1518 case TGSI_TEXTURE_2D:
1519 case TGSI_TEXTURE_RECT:
1520 case TGSI_TEXTURE_1D_ARRAY:
1523 case TGSI_TEXTURE_SHADOW1D:
1524 case TGSI_TEXTURE_SHADOW2D:
1525 case TGSI_TEXTURE_SHADOWRECT:
1526 case TGSI_TEXTURE_3D:
1527 case TGSI_TEXTURE_CUBE:
1528 case TGSI_TEXTURE_2D_ARRAY:
1537 FETCH( func, *inst, 3, 0, 3 );
1543 TGSI_EXEC_TEMP_00000000_I,
1544 TGSI_EXEC_TEMP_00000000_C );
1548 /* store lodbias whether enabled or not -- fetch_texel currently
1549 * respects it always.
1552 get_temp( TEMP_R0, 3 ),
1556 FETCH( func, *inst, 3, 0, 3 );
1558 emit_rcp( func, 3, 3 );
1561 for (i = 0; i < count; i++) {
1562 FETCH( func, *inst, i, 0, i );
1571 /* Store in the argument buffer:
1575 get_temp( TEMP_R0, i ),
1579 args[0] = get_temp( TEMP_R0, 0 );
1580 args[1] = get_sampler_ptr( unit );
1582 emit_func_call( func,
1588 /* If all four channels are enabled, could use a pointer to
1589 * dst[0].x instead of TEMP_R0 for store?
1591 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, i ) {
1596 get_temp( TEMP_R0, i ) );
1598 STORE( func, *inst, 0, 0, i );
1605 struct x86_function *func,
1606 const struct tgsi_full_src_register *reg )
1608 unsigned uniquemask;
1609 unsigned unique_count = 0;
1610 unsigned chan_index;
1613 /* This mask stores component bits that were already tested. Note that
1614 * we test if the value is less than zero, so 1.0 and 0.0 need not to be
1619 FOR_EACH_CHANNEL( chan_index ) {
1622 /* unswizzle channel */
1623 swizzle = tgsi_util_get_full_src_register_swizzle(
1627 /* check if the component has not been already tested */
1628 if( !(uniquemask & (1 << swizzle)) ) {
1629 uniquemask |= 1 << swizzle;
1631 /* allocate register */
1642 x86_make_reg( file_REG32, reg_AX ) );
1645 x86_make_reg( file_REG32, reg_DX ) );
1647 for (i = 0 ; i < unique_count; i++ ) {
1648 struct x86_reg dataXMM = make_xmm(i);
1654 TGSI_EXEC_TEMP_00000000_I,
1655 TGSI_EXEC_TEMP_00000000_C ),
1661 x86_make_reg( file_REG32, reg_AX ),
1667 x86_make_reg( file_REG32, reg_DX ),
1671 x86_make_reg( file_REG32, reg_AX ),
1672 x86_make_reg( file_REG32, reg_DX ) );
1679 TGSI_EXEC_TEMP_KILMASK_I,
1680 TGSI_EXEC_TEMP_KILMASK_C ),
1681 x86_make_reg( file_REG32, reg_AX ) );
1685 x86_make_reg( file_REG32, reg_DX ) );
1688 x86_make_reg( file_REG32, reg_AX ) );
1694 struct x86_function *func )
1696 /* XXX todo / fix me */
1702 struct x86_function *func,
1703 struct tgsi_full_instruction *inst,
1706 unsigned chan_index;
1708 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1709 FETCH( func, *inst, 0, 0, chan_index );
1710 FETCH( func, *inst, 1, 1, chan_index );
1722 STORE( func, *inst, 0, 0, chan_index );
1728 struct x86_function *func,
1729 struct tgsi_full_instruction *inst )
1731 unsigned chan_index;
1733 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1734 FETCH( func, *inst, 0, 0, chan_index );
1735 FETCH( func, *inst, 1, 1, chan_index );
1736 FETCH( func, *inst, 2, 2, chan_index );
1741 TGSI_EXEC_TEMP_00000000_I,
1742 TGSI_EXEC_TEMP_00000000_C ),
1756 STORE( func, *inst, 0, 0, chan_index );
1762 * Check if inst src/dest regs use indirect addressing into temporary,
1763 * input or output register files.
1766 indirect_reg_reference(const struct tgsi_full_instruction *inst)
1769 for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
1770 const struct tgsi_full_src_register *reg = &inst->Src[i];
1771 if ((reg->Register.File == TGSI_FILE_TEMPORARY ||
1772 reg->Register.File == TGSI_FILE_INPUT ||
1773 reg->Register.File == TGSI_FILE_OUTPUT) &&
1774 reg->Register.Indirect)
1777 for (i = 0; i < inst->Instruction.NumDstRegs; i++) {
1778 const struct tgsi_full_dst_register *reg = &inst->Dst[i];
1779 if ((reg->Register.File == TGSI_FILE_TEMPORARY ||
1780 reg->Register.File == TGSI_FILE_INPUT ||
1781 reg->Register.File == TGSI_FILE_OUTPUT) &&
1782 reg->Register.Indirect)
1791 struct x86_function *func,
1792 struct tgsi_full_instruction *inst )
1794 unsigned chan_index;
1796 /* we can't handle indirect addressing into temp register file yet */
1797 if (indirect_reg_reference(inst))
1800 switch (inst->Instruction.Opcode) {
1801 case TGSI_OPCODE_ARL:
1802 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1803 FETCH( func, *inst, 0, 0, chan_index );
1804 emit_flr(func, 0, 0);
1805 emit_f2it( func, 0 );
1806 STORE( func, *inst, 0, 0, chan_index );
1810 case TGSI_OPCODE_MOV:
1811 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1812 FETCH( func, *inst, 4 + chan_index, 0, chan_index );
1814 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1815 STORE( func, *inst, 4 + chan_index, 0, chan_index );
1819 case TGSI_OPCODE_LIT:
1820 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1821 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
1827 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ) {
1828 STORE( func, *inst, 0, 0, CHAN_X );
1830 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
1831 STORE( func, *inst, 0, 0, CHAN_W );
1834 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
1835 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
1836 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
1837 FETCH( func, *inst, 0, 0, CHAN_X );
1842 TGSI_EXEC_TEMP_00000000_I,
1843 TGSI_EXEC_TEMP_00000000_C ) );
1844 STORE( func, *inst, 0, 0, CHAN_Y );
1846 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
1847 /* XMM[1] = SrcReg[0].yyyy */
1848 FETCH( func, *inst, 1, 0, CHAN_Y );
1849 /* XMM[1] = max(XMM[1], 0) */
1854 TGSI_EXEC_TEMP_00000000_I,
1855 TGSI_EXEC_TEMP_00000000_C ) );
1856 /* XMM[2] = SrcReg[0].wwww */
1857 FETCH( func, *inst, 2, 0, CHAN_W );
1858 /* XMM[2] = min(XMM[2], 128.0) */
1863 TGSI_EXEC_TEMP_128_I,
1864 TGSI_EXEC_TEMP_128_C ) );
1865 /* XMM[2] = max(XMM[2], -128.0) */
1870 TGSI_EXEC_TEMP_MINUS_128_I,
1871 TGSI_EXEC_TEMP_MINUS_128_C ) );
1872 emit_pow( func, 3, 1, 1, 2 );
1873 FETCH( func, *inst, 0, 0, CHAN_X );
1887 STORE( func, *inst, 2, 0, CHAN_Z );
1892 case TGSI_OPCODE_RCP:
1893 FETCH( func, *inst, 0, 0, CHAN_X );
1894 emit_rcp( func, 0, 0 );
1895 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1896 STORE( func, *inst, 0, 0, chan_index );
1900 case TGSI_OPCODE_RSQ:
1901 FETCH( func, *inst, 0, 0, CHAN_X );
1902 emit_abs( func, 0 );
1903 emit_rsqrt( func, 1, 0 );
1904 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1905 STORE( func, *inst, 1, 0, chan_index );
1909 case TGSI_OPCODE_EXP:
1910 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1911 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
1912 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1913 FETCH( func, *inst, 0, 0, CHAN_X );
1914 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1915 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1916 emit_MOV( func, 1, 0 );
1917 emit_flr( func, 2, 1 );
1918 /* dst.x = ex2(floor(src.x)) */
1919 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X )) {
1920 emit_MOV( func, 2, 1 );
1921 emit_ex2( func, 3, 2 );
1922 STORE( func, *inst, 2, 0, CHAN_X );
1924 /* dst.y = src.x - floor(src.x) */
1925 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1926 emit_MOV( func, 2, 0 );
1927 emit_sub( func, 2, 1 );
1928 STORE( func, *inst, 2, 0, CHAN_Y );
1931 /* dst.z = ex2(src.x) */
1932 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1933 emit_ex2( func, 3, 0 );
1934 STORE( func, *inst, 0, 0, CHAN_Z );
1938 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W )) {
1939 emit_tempf( func, 0, TEMP_ONE_I, TEMP_ONE_C );
1940 STORE( func, *inst, 0, 0, CHAN_W );
1944 case TGSI_OPCODE_LOG:
1945 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1946 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
1947 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1948 FETCH( func, *inst, 0, 0, CHAN_X );
1949 emit_abs( func, 0 );
1950 emit_MOV( func, 1, 0 );
1951 emit_lg2( func, 2, 1 );
1952 /* dst.z = lg2(abs(src.x)) */
1953 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1954 STORE( func, *inst, 1, 0, CHAN_Z );
1956 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1957 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1958 emit_flr( func, 2, 1 );
1959 /* dst.x = floor(lg2(abs(src.x))) */
1960 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X )) {
1961 STORE( func, *inst, 1, 0, CHAN_X );
1963 /* dst.x = abs(src)/ex2(floor(lg2(abs(src.x)))) */
1964 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1965 emit_ex2( func, 2, 1 );
1966 emit_rcp( func, 1, 1 );
1967 emit_mul( func, 0, 1 );
1968 STORE( func, *inst, 0, 0, CHAN_Y );
1973 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W )) {
1974 emit_tempf( func, 0, TEMP_ONE_I, TEMP_ONE_C );
1975 STORE( func, *inst, 0, 0, CHAN_W );
1979 case TGSI_OPCODE_MUL:
1980 /* do all fetches and adds, storing results in temp regs */
1981 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1982 int r = chan_index + 1;
1983 FETCH( func, *inst, 0, 0, chan_index ); /* load xmm[0] */
1984 FETCH( func, *inst, r, 1, chan_index ); /* load xmm[r] */
1985 emit_mul( func, r, 0 ); /* xmm[r] = xmm[r] * xmm[0] */
1987 /* do all stores of the temp regs */
1988 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1989 int r = chan_index + 1;
1990 STORE( func, *inst, r, 0, chan_index ); /* store xmm[r] */
1994 case TGSI_OPCODE_ADD:
1995 /* do all fetches and adds, storing results in temp regs */
1996 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1997 int r = chan_index + 1;
1998 FETCH( func, *inst, 0, 0, chan_index ); /* load xmm[0] */
1999 FETCH( func, *inst, r, 1, chan_index ); /* load xmm[r] */
2000 emit_add( func, r, 0 ); /* xmm[r] = xmm[r] + xmm[0] */
2002 /* do all stores of the temp regs */
2003 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2004 int r = chan_index + 1;
2005 STORE( func, *inst, r, 0, chan_index ); /* store xmm[r] */
2009 case TGSI_OPCODE_DP3:
2010 FETCH( func, *inst, 0, 0, CHAN_X );
2011 FETCH( func, *inst, 1, 1, CHAN_X );
2012 emit_mul( func, 0, 1 );
2013 FETCH( func, *inst, 1, 0, CHAN_Y );
2014 FETCH( func, *inst, 2, 1, CHAN_Y );
2015 emit_mul( func, 1, 2 );
2016 emit_add( func, 0, 1 );
2017 FETCH( func, *inst, 1, 0, CHAN_Z );
2018 FETCH( func, *inst, 2, 1, CHAN_Z );
2019 emit_mul( func, 1, 2 );
2020 emit_add( func, 0, 1 );
2021 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2022 STORE( func, *inst, 0, 0, chan_index );
2026 case TGSI_OPCODE_DP4:
2027 FETCH( func, *inst, 0, 0, CHAN_X );
2028 FETCH( func, *inst, 1, 1, CHAN_X );
2029 emit_mul( func, 0, 1 );
2030 FETCH( func, *inst, 1, 0, CHAN_Y );
2031 FETCH( func, *inst, 2, 1, CHAN_Y );
2032 emit_mul( func, 1, 2 );
2033 emit_add( func, 0, 1 );
2034 FETCH( func, *inst, 1, 0, CHAN_Z );
2035 FETCH( func, *inst, 2, 1, CHAN_Z );
2036 emit_mul(func, 1, 2 );
2037 emit_add(func, 0, 1 );
2038 FETCH( func, *inst, 1, 0, CHAN_W );
2039 FETCH( func, *inst, 2, 1, CHAN_W );
2040 emit_mul( func, 1, 2 );
2041 emit_add( func, 0, 1 );
2042 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2043 STORE( func, *inst, 0, 0, chan_index );
2047 case TGSI_OPCODE_DST:
2048 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
2054 STORE( func, *inst, 0, 0, CHAN_X );
2056 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
2057 FETCH( func, *inst, 0, 0, CHAN_Y );
2058 FETCH( func, *inst, 1, 1, CHAN_Y );
2059 emit_mul( func, 0, 1 );
2060 STORE( func, *inst, 0, 0, CHAN_Y );
2062 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
2063 FETCH( func, *inst, 0, 0, CHAN_Z );
2064 STORE( func, *inst, 0, 0, CHAN_Z );
2066 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
2067 FETCH( func, *inst, 0, 1, CHAN_W );
2068 STORE( func, *inst, 0, 0, CHAN_W );
2072 case TGSI_OPCODE_MIN:
2073 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2074 FETCH( func, *inst, 0, 0, chan_index );
2075 FETCH( func, *inst, 1, 1, chan_index );
2080 STORE( func, *inst, 0, 0, chan_index );
2084 case TGSI_OPCODE_MAX:
2085 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2086 FETCH( func, *inst, 0, 0, chan_index );
2087 FETCH( func, *inst, 1, 1, chan_index );
2092 STORE( func, *inst, 0, 0, chan_index );
2096 case TGSI_OPCODE_SLT:
2097 emit_setcc( func, inst, cc_LessThan );
2100 case TGSI_OPCODE_SGE:
2101 emit_setcc( func, inst, cc_NotLessThan );
2104 case TGSI_OPCODE_MAD:
2105 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2106 FETCH( func, *inst, 0, 0, chan_index );
2107 FETCH( func, *inst, 1, 1, chan_index );
2108 FETCH( func, *inst, 2, 2, chan_index );
2109 emit_mul( func, 0, 1 );
2110 emit_add( func, 0, 2 );
2111 STORE( func, *inst, 0, 0, chan_index );
2115 case TGSI_OPCODE_SUB:
2116 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2117 FETCH( func, *inst, 0, 0, chan_index );
2118 FETCH( func, *inst, 1, 1, chan_index );
2119 emit_sub( func, 0, 1 );
2120 STORE( func, *inst, 0, 0, chan_index );
2124 case TGSI_OPCODE_LRP:
2125 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2126 FETCH( func, *inst, 0, 0, chan_index );
2127 FETCH( func, *inst, 1, 1, chan_index );
2128 FETCH( func, *inst, 2, 2, chan_index );
2129 emit_sub( func, 1, 2 );
2130 emit_mul( func, 0, 1 );
2131 emit_add( func, 0, 2 );
2132 STORE( func, *inst, 0, 0, chan_index );
2136 case TGSI_OPCODE_CND:
2140 case TGSI_OPCODE_DP2A:
2141 FETCH( func, *inst, 0, 0, CHAN_X ); /* xmm0 = src[0].x */
2142 FETCH( func, *inst, 1, 1, CHAN_X ); /* xmm1 = src[1].x */
2143 emit_mul( func, 0, 1 ); /* xmm0 = xmm0 * xmm1 */
2144 FETCH( func, *inst, 1, 0, CHAN_Y ); /* xmm1 = src[0].y */
2145 FETCH( func, *inst, 2, 1, CHAN_Y ); /* xmm2 = src[1].y */
2146 emit_mul( func, 1, 2 ); /* xmm1 = xmm1 * xmm2 */
2147 emit_add( func, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2148 FETCH( func, *inst, 1, 2, CHAN_X ); /* xmm1 = src[2].x */
2149 emit_add( func, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2150 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2151 STORE( func, *inst, 0, 0, chan_index ); /* dest[ch] = xmm0 */
2155 case TGSI_OPCODE_FRC:
2156 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2157 FETCH( func, *inst, 0, 0, chan_index );
2158 emit_frc( func, 0, 0 );
2159 STORE( func, *inst, 0, 0, chan_index );
2163 case TGSI_OPCODE_CLAMP:
2167 case TGSI_OPCODE_FLR:
2168 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2169 FETCH( func, *inst, 0, 0, chan_index );
2170 emit_flr( func, 0, 0 );
2171 STORE( func, *inst, 0, 0, chan_index );
2175 case TGSI_OPCODE_ROUND:
2176 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2177 FETCH( func, *inst, 0, 0, chan_index );
2178 emit_rnd( func, 0, 0 );
2179 STORE( func, *inst, 0, 0, chan_index );
2183 case TGSI_OPCODE_EX2:
2184 FETCH( func, *inst, 0, 0, CHAN_X );
2185 emit_ex2( func, 0, 0 );
2186 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2187 STORE( func, *inst, 0, 0, chan_index );
2191 case TGSI_OPCODE_LG2:
2192 FETCH( func, *inst, 0, 0, CHAN_X );
2193 emit_lg2( func, 0, 0 );
2194 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2195 STORE( func, *inst, 0, 0, chan_index );
2199 case TGSI_OPCODE_POW:
2200 FETCH( func, *inst, 0, 0, CHAN_X );
2201 FETCH( func, *inst, 1, 1, CHAN_X );
2202 emit_pow( func, 0, 0, 0, 1 );
2203 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2204 STORE( func, *inst, 0, 0, chan_index );
2208 case TGSI_OPCODE_XPD:
2209 /* Note: we do all stores after all operands have been fetched
2210 * to avoid src/dst register aliasing issues for an instruction
2211 * such as: XPD TEMP[2].xyz, TEMP[0], TEMP[2];
2213 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
2214 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
2215 FETCH( func, *inst, 1, 1, CHAN_Z ); /* xmm[1] = src[1].z */
2216 FETCH( func, *inst, 3, 0, CHAN_Z ); /* xmm[3] = src[0].z */
2218 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
2219 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
2220 FETCH( func, *inst, 0, 0, CHAN_Y ); /* xmm[0] = src[0].y */
2221 FETCH( func, *inst, 4, 1, CHAN_Y ); /* xmm[4] = src[1].y */
2223 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
2224 emit_MOV( func, 7, 0 ); /* xmm[7] = xmm[0] */
2225 emit_mul( func, 7, 1 ); /* xmm[7] = xmm[2] * xmm[1] */
2226 emit_MOV( func, 5, 3 ); /* xmm[5] = xmm[3] */
2227 emit_mul( func, 5, 4 ); /* xmm[5] = xmm[5] * xmm[4] */
2228 emit_sub( func, 7, 5 ); /* xmm[7] = xmm[2] - xmm[5] */
2229 /* store xmm[7] in dst.x below */
2231 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
2232 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
2233 FETCH( func, *inst, 2, 1, CHAN_X ); /* xmm[2] = src[1].x */
2234 FETCH( func, *inst, 5, 0, CHAN_X ); /* xmm[5] = src[0].x */
2236 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
2237 emit_mul( func, 3, 2 ); /* xmm[3] = xmm[3] * xmm[2] */
2238 emit_mul( func, 1, 5 ); /* xmm[1] = xmm[1] * xmm[5] */
2239 emit_sub( func, 3, 1 ); /* xmm[3] = xmm[3] - xmm[1] */
2240 /* store xmm[3] in dst.y below */
2242 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
2243 emit_mul( func, 5, 4 ); /* xmm[5] = xmm[5] * xmm[4] */
2244 emit_mul( func, 0, 2 ); /* xmm[0] = xmm[0] * xmm[2] */
2245 emit_sub( func, 5, 0 ); /* xmm[5] = xmm[5] - xmm[0] */
2246 STORE( func, *inst, 5, 0, CHAN_Z ); /* dst.z = xmm[5] */
2248 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
2249 STORE( func, *inst, 7, 0, CHAN_X ); /* dst.x = xmm[7] */
2251 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
2252 STORE( func, *inst, 3, 0, CHAN_Y ); /* dst.y = xmm[3] */
2254 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
2260 STORE( func, *inst, 0, 0, CHAN_W );
2264 case TGSI_OPCODE_ABS:
2265 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2266 FETCH( func, *inst, 0, 0, chan_index );
2267 emit_abs( func, 0) ;
2269 STORE( func, *inst, 0, 0, chan_index );
2273 case TGSI_OPCODE_RCC:
2277 case TGSI_OPCODE_DPH:
2278 FETCH( func, *inst, 0, 0, CHAN_X );
2279 FETCH( func, *inst, 1, 1, CHAN_X );
2280 emit_mul( func, 0, 1 );
2281 FETCH( func, *inst, 1, 0, CHAN_Y );
2282 FETCH( func, *inst, 2, 1, CHAN_Y );
2283 emit_mul( func, 1, 2 );
2284 emit_add( func, 0, 1 );
2285 FETCH( func, *inst, 1, 0, CHAN_Z );
2286 FETCH( func, *inst, 2, 1, CHAN_Z );
2287 emit_mul( func, 1, 2 );
2288 emit_add( func, 0, 1 );
2289 FETCH( func, *inst, 1, 1, CHAN_W );
2290 emit_add( func, 0, 1 );
2291 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2292 STORE( func, *inst, 0, 0, chan_index );
2296 case TGSI_OPCODE_COS:
2297 FETCH( func, *inst, 0, 0, CHAN_X );
2298 emit_cos( func, 0, 0 );
2299 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2300 STORE( func, *inst, 0, 0, chan_index );
2304 case TGSI_OPCODE_DDX:
2308 case TGSI_OPCODE_DDY:
2312 case TGSI_OPCODE_KILP:
2313 /* predicated kill */
2315 return 0; /* XXX fix me */
2318 case TGSI_OPCODE_KIL:
2319 /* conditional kill */
2320 emit_kil( func, &inst->Src[0] );
2323 case TGSI_OPCODE_PK2H:
2327 case TGSI_OPCODE_PK2US:
2331 case TGSI_OPCODE_PK4B:
2335 case TGSI_OPCODE_PK4UB:
2339 case TGSI_OPCODE_RFL:
2343 case TGSI_OPCODE_SEQ:
2344 emit_setcc( func, inst, cc_Equal );
2347 case TGSI_OPCODE_SFL:
2351 case TGSI_OPCODE_SGT:
2352 emit_setcc( func, inst, cc_NotLessThanEqual );
2355 case TGSI_OPCODE_SIN:
2356 FETCH( func, *inst, 0, 0, CHAN_X );
2357 emit_sin( func, 0, 0 );
2358 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2359 STORE( func, *inst, 0, 0, chan_index );
2363 case TGSI_OPCODE_SLE:
2364 emit_setcc( func, inst, cc_LessThanEqual );
2367 case TGSI_OPCODE_SNE:
2368 emit_setcc( func, inst, cc_NotEqual );
2371 case TGSI_OPCODE_STR:
2375 case TGSI_OPCODE_TEX:
2376 emit_tex( func, inst, FALSE, FALSE );
2379 case TGSI_OPCODE_TXD:
2383 case TGSI_OPCODE_UP2H:
2387 case TGSI_OPCODE_UP2US:
2391 case TGSI_OPCODE_UP4B:
2395 case TGSI_OPCODE_UP4UB:
2399 case TGSI_OPCODE_X2D:
2403 case TGSI_OPCODE_ARA:
2407 case TGSI_OPCODE_ARR:
2408 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2409 FETCH( func, *inst, 0, 0, chan_index );
2410 emit_rnd( func, 0, 0 );
2411 emit_f2it( func, 0 );
2412 STORE( func, *inst, 0, 0, chan_index );
2416 case TGSI_OPCODE_BRA:
2420 case TGSI_OPCODE_CAL:
2424 case TGSI_OPCODE_RET:
2428 case TGSI_OPCODE_END:
2431 case TGSI_OPCODE_SSG:
2432 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2433 FETCH( func, *inst, 0, 0, chan_index );
2434 emit_sgn( func, 0, 0 );
2435 STORE( func, *inst, 0, 0, chan_index );
2439 case TGSI_OPCODE_CMP:
2440 emit_cmp (func, inst);
2443 case TGSI_OPCODE_SCS:
2444 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
2445 FETCH( func, *inst, 0, 0, CHAN_X );
2446 emit_cos( func, 0, 0 );
2447 STORE( func, *inst, 0, 0, CHAN_X );
2449 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
2450 FETCH( func, *inst, 0, 0, CHAN_X );
2451 emit_sin( func, 0, 0 );
2452 STORE( func, *inst, 0, 0, CHAN_Y );
2454 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
2458 TGSI_EXEC_TEMP_00000000_I,
2459 TGSI_EXEC_TEMP_00000000_C );
2460 STORE( func, *inst, 0, 0, CHAN_Z );
2462 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
2468 STORE( func, *inst, 0, 0, CHAN_W );
2472 case TGSI_OPCODE_TXB:
2473 emit_tex( func, inst, TRUE, FALSE );
2476 case TGSI_OPCODE_NRM:
2478 case TGSI_OPCODE_NRM4:
2479 /* 3 or 4-component normalization */
2481 uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
2483 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) ||
2484 IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y) ||
2485 IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z) ||
2486 (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W) && dims == 4)) {
2488 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
2491 /* xmm0 = src.x * src.x */
2492 FETCH(func, *inst, 0, 0, CHAN_X);
2493 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) {
2494 emit_MOV(func, 4, 0);
2496 emit_mul(func, 0, 0);
2499 /* xmm0 = xmm0 + src.y * src.y */
2500 FETCH(func, *inst, 1, 0, CHAN_Y);
2501 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) {
2502 emit_MOV(func, 5, 1);
2504 emit_mul(func, 1, 1);
2505 emit_add(func, 0, 1);
2508 /* xmm0 = xmm0 + src.z * src.z */
2509 FETCH(func, *inst, 1, 0, CHAN_Z);
2510 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
2511 emit_MOV(func, 6, 1);
2513 emit_mul(func, 1, 1);
2514 emit_add(func, 0, 1);
2518 /* xmm0 = xmm0 + src.w * src.w */
2519 FETCH(func, *inst, 1, 0, CHAN_W);
2520 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W)) {
2521 emit_MOV(func, 7, 1);
2523 emit_mul(func, 1, 1);
2524 emit_add(func, 0, 1);
2527 /* xmm1 = 1 / sqrt(xmm0) */
2528 emit_rsqrt(func, 1, 0);
2530 /* dst.x = xmm1 * src.x */
2531 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) {
2532 emit_mul(func, 4, 1);
2533 STORE(func, *inst, 4, 0, CHAN_X);
2536 /* dst.y = xmm1 * src.y */
2537 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) {
2538 emit_mul(func, 5, 1);
2539 STORE(func, *inst, 5, 0, CHAN_Y);
2542 /* dst.z = xmm1 * src.z */
2543 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
2544 emit_mul(func, 6, 1);
2545 STORE(func, *inst, 6, 0, CHAN_Z);
2548 /* dst.w = xmm1 * src.w */
2549 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) && dims == 4) {
2550 emit_mul(func, 7, 1);
2551 STORE(func, *inst, 7, 0, CHAN_W);
2556 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W) && dims == 3) {
2557 emit_tempf(func, 0, TEMP_ONE_I, TEMP_ONE_C);
2558 STORE(func, *inst, 0, 0, CHAN_W);
2563 case TGSI_OPCODE_DIV:
2567 case TGSI_OPCODE_DP2:
2568 FETCH( func, *inst, 0, 0, CHAN_X ); /* xmm0 = src[0].x */
2569 FETCH( func, *inst, 1, 1, CHAN_X ); /* xmm1 = src[1].x */
2570 emit_mul( func, 0, 1 ); /* xmm0 = xmm0 * xmm1 */
2571 FETCH( func, *inst, 1, 0, CHAN_Y ); /* xmm1 = src[0].y */
2572 FETCH( func, *inst, 2, 1, CHAN_Y ); /* xmm2 = src[1].y */
2573 emit_mul( func, 1, 2 ); /* xmm1 = xmm1 * xmm2 */
2574 emit_add( func, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2575 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2576 STORE( func, *inst, 0, 0, chan_index ); /* dest[ch] = xmm0 */
2580 case TGSI_OPCODE_TXL:
2584 case TGSI_OPCODE_TXP:
2585 emit_tex( func, inst, FALSE, TRUE );
2588 case TGSI_OPCODE_BRK:
2592 case TGSI_OPCODE_IF:
2596 case TGSI_OPCODE_ELSE:
2600 case TGSI_OPCODE_ENDIF:
2604 case TGSI_OPCODE_PUSHA:
2608 case TGSI_OPCODE_POPA:
2612 case TGSI_OPCODE_CEIL:
2616 case TGSI_OPCODE_I2F:
2620 case TGSI_OPCODE_NOT:
2624 case TGSI_OPCODE_TRUNC:
2625 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2626 FETCH( func, *inst, 0, 0, chan_index );
2627 emit_f2it( func, 0 );
2628 emit_i2f( func, 0 );
2629 STORE( func, *inst, 0, 0, chan_index );
2633 case TGSI_OPCODE_SHL:
2637 case TGSI_OPCODE_ISHR:
2641 case TGSI_OPCODE_AND:
2645 case TGSI_OPCODE_OR:
2649 case TGSI_OPCODE_MOD:
2653 case TGSI_OPCODE_XOR:
2657 case TGSI_OPCODE_SAD:
2661 case TGSI_OPCODE_TXF:
2665 case TGSI_OPCODE_TXQ:
2669 case TGSI_OPCODE_CONT:
2673 case TGSI_OPCODE_EMIT:
2677 case TGSI_OPCODE_ENDPRIM:
2690 struct x86_function *func,
2691 struct tgsi_full_declaration *decl )
2693 if( decl->Declaration.File == TGSI_FILE_INPUT ) {
2694 unsigned first, last, mask;
2697 first = decl->Range.First;
2698 last = decl->Range.Last;
2699 mask = decl->Declaration.UsageMask;
2701 for( i = first; i <= last; i++ ) {
2702 for( j = 0; j < NUM_CHANNELS; j++ ) {
2703 if( mask & (1 << j) ) {
2704 switch( decl->Declaration.Interpolate ) {
2705 case TGSI_INTERPOLATE_CONSTANT:
2706 emit_coef_a0( func, 0, i, j );
2707 emit_inputs( func, 0, i, j );
2710 case TGSI_INTERPOLATE_LINEAR:
2711 emit_tempf( func, 0, 0, TGSI_SWIZZLE_X );
2712 emit_coef_dadx( func, 1, i, j );
2713 emit_tempf( func, 2, 0, TGSI_SWIZZLE_Y );
2714 emit_coef_dady( func, 3, i, j );
2715 emit_mul( func, 0, 1 ); /* x * dadx */
2716 emit_coef_a0( func, 4, i, j );
2717 emit_mul( func, 2, 3 ); /* y * dady */
2718 emit_add( func, 0, 4 ); /* x * dadx + a0 */
2719 emit_add( func, 0, 2 ); /* x * dadx + y * dady + a0 */
2720 emit_inputs( func, 0, i, j );
2723 case TGSI_INTERPOLATE_PERSPECTIVE:
2724 emit_tempf( func, 0, 0, TGSI_SWIZZLE_X );
2725 emit_coef_dadx( func, 1, i, j );
2726 emit_tempf( func, 2, 0, TGSI_SWIZZLE_Y );
2727 emit_coef_dady( func, 3, i, j );
2728 emit_mul( func, 0, 1 ); /* x * dadx */
2729 emit_tempf( func, 4, 0, TGSI_SWIZZLE_W );
2730 emit_coef_a0( func, 5, i, j );
2731 emit_rcp( func, 4, 4 ); /* 1.0 / w */
2732 emit_mul( func, 2, 3 ); /* y * dady */
2733 emit_add( func, 0, 5 ); /* x * dadx + a0 */
2734 emit_add( func, 0, 2 ); /* x * dadx + y * dady + a0 */
2735 emit_mul( func, 0, 4 ); /* (x * dadx + y * dady + a0) / w */
2736 emit_inputs( func, 0, i, j );
2749 static void aos_to_soa( struct x86_function *func,
2755 struct x86_reg soa_input = x86_make_reg( file_REG32, reg_AX );
2756 struct x86_reg aos_input = x86_make_reg( file_REG32, reg_BX );
2757 struct x86_reg num_inputs = x86_make_reg( file_REG32, reg_CX );
2758 struct x86_reg stride = x86_make_reg( file_REG32, reg_DX );
2759 int loop_top, loop_exit_fixup;
2762 x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
2764 x86_mov( func, aos_input, x86_fn_arg( func, arg_aos ) );
2765 x86_mov( func, soa_input, x86_fn_arg( func, arg_machine ) );
2766 /* FIXME: tgsi_exec_machine::Inputs is a pointer now! */
2767 x86_lea( func, soa_input,
2768 x86_make_disp( soa_input,
2769 Offset(struct tgsi_exec_machine, Inputs) ) );
2770 x86_mov( func, num_inputs, x86_fn_arg( func, arg_num ) );
2771 x86_mov( func, stride, x86_fn_arg( func, arg_stride ) );
2773 /* while (num_inputs != 0) */
2774 loop_top = x86_get_label( func );
2775 x86_cmp_imm( func, num_inputs, 0 );
2776 loop_exit_fixup = x86_jcc_forward( func, cc_E );
2779 x86_push( func, aos_input );
2780 sse_movlps( func, make_xmm( 0 ), x86_make_disp( aos_input, 0 ) );
2781 sse_movlps( func, make_xmm( 3 ), x86_make_disp( aos_input, 8 ) );
2782 x86_add( func, aos_input, stride );
2783 sse_movhps( func, make_xmm( 0 ), x86_make_disp( aos_input, 0 ) );
2784 sse_movhps( func, make_xmm( 3 ), x86_make_disp( aos_input, 8 ) );
2785 x86_add( func, aos_input, stride );
2786 sse_movlps( func, make_xmm( 1 ), x86_make_disp( aos_input, 0 ) );
2787 sse_movlps( func, make_xmm( 4 ), x86_make_disp( aos_input, 8 ) );
2788 x86_add( func, aos_input, stride );
2789 sse_movhps( func, make_xmm( 1 ), x86_make_disp( aos_input, 0 ) );
2790 sse_movhps( func, make_xmm( 4 ), x86_make_disp( aos_input, 8 ) );
2791 x86_pop( func, aos_input );
2793 sse_movaps( func, make_xmm( 2 ), make_xmm( 0 ) );
2794 sse_movaps( func, make_xmm( 5 ), make_xmm( 3 ) );
2795 sse_shufps( func, make_xmm( 0 ), make_xmm( 1 ), 0x88 );
2796 sse_shufps( func, make_xmm( 2 ), make_xmm( 1 ), 0xdd );
2797 sse_shufps( func, make_xmm( 3 ), make_xmm( 4 ), 0x88 );
2798 sse_shufps( func, make_xmm( 5 ), make_xmm( 4 ), 0xdd );
2800 sse_movups( func, x86_make_disp( soa_input, 0 ), make_xmm( 0 ) );
2801 sse_movups( func, x86_make_disp( soa_input, 16 ), make_xmm( 2 ) );
2802 sse_movups( func, x86_make_disp( soa_input, 32 ), make_xmm( 3 ) );
2803 sse_movups( func, x86_make_disp( soa_input, 48 ), make_xmm( 5 ) );
2805 /* Advance to next input */
2806 x86_lea( func, aos_input, x86_make_disp(aos_input, 16) );
2807 x86_lea( func, soa_input, x86_make_disp(soa_input, 64) );
2810 x86_dec( func, num_inputs );
2811 x86_jmp( func, loop_top );
2812 x86_fixup_fwd_jump( func, loop_exit_fixup );
2815 x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
2818 static void soa_to_aos( struct x86_function *func,
2824 struct x86_reg soa_output = x86_make_reg( file_REG32, reg_AX );
2825 struct x86_reg aos_output = x86_make_reg( file_REG32, reg_BX );
2826 struct x86_reg num_outputs = x86_make_reg( file_REG32, reg_CX );
2827 struct x86_reg temp = x86_make_reg( file_REG32, reg_DX );
2831 x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
2833 x86_mov( func, aos_output, x86_fn_arg( func, arg_aos ) );
2834 x86_mov( func, soa_output, x86_fn_arg( func, arg_machine ) );
2835 /* FIXME: tgsi_exec_machine::Ouputs is a pointer now! */
2836 x86_lea( func, soa_output,
2837 x86_make_disp( soa_output,
2838 Offset(struct tgsi_exec_machine, Outputs) ) );
2839 x86_mov( func, num_outputs, x86_fn_arg( func, arg_num ) );
2842 inner_loop = x86_get_label( func );
2844 sse_movups( func, make_xmm( 0 ), x86_make_disp( soa_output, 0 ) );
2845 sse_movups( func, make_xmm( 1 ), x86_make_disp( soa_output, 16 ) );
2846 sse_movups( func, make_xmm( 3 ), x86_make_disp( soa_output, 32 ) );
2847 sse_movups( func, make_xmm( 4 ), x86_make_disp( soa_output, 48 ) );
2849 sse_movaps( func, make_xmm( 2 ), make_xmm( 0 ) );
2850 sse_movaps( func, make_xmm( 5 ), make_xmm( 3 ) );
2851 sse_unpcklps( func, make_xmm( 0 ), make_xmm( 1 ) );
2852 sse_unpckhps( func, make_xmm( 2 ), make_xmm( 1 ) );
2853 sse_unpcklps( func, make_xmm( 3 ), make_xmm( 4 ) );
2854 sse_unpckhps( func, make_xmm( 5 ), make_xmm( 4 ) );
2856 x86_mov( func, temp, x86_fn_arg( func, arg_stride ) );
2857 x86_push( func, aos_output );
2858 sse_movlps( func, x86_make_disp( aos_output, 0 ), make_xmm( 0 ) );
2859 sse_movlps( func, x86_make_disp( aos_output, 8 ), make_xmm( 3 ) );
2860 x86_add( func, aos_output, temp );
2861 sse_movhps( func, x86_make_disp( aos_output, 0 ), make_xmm( 0 ) );
2862 sse_movhps( func, x86_make_disp( aos_output, 8 ), make_xmm( 3 ) );
2863 x86_add( func, aos_output, temp );
2864 sse_movlps( func, x86_make_disp( aos_output, 0 ), make_xmm( 2 ) );
2865 sse_movlps( func, x86_make_disp( aos_output, 8 ), make_xmm( 5 ) );
2866 x86_add( func, aos_output, temp );
2867 sse_movhps( func, x86_make_disp( aos_output, 0 ), make_xmm( 2 ) );
2868 sse_movhps( func, x86_make_disp( aos_output, 8 ), make_xmm( 5 ) );
2869 x86_pop( func, aos_output );
2871 /* Advance to next output */
2872 x86_lea( func, aos_output, x86_make_disp(aos_output, 16) );
2873 x86_lea( func, soa_output, x86_make_disp(soa_output, 64) );
2875 /* while --num_outputs */
2876 x86_dec( func, num_outputs );
2877 x86_jcc( func, cc_NE, inner_loop );
2880 x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
2885 * Check if the instructions dst register is the same as any src
2886 * register and warn if there's a posible SOA dependency.
2889 check_soa_dependencies(const struct tgsi_full_instruction *inst)
2891 uint opcode = inst->Instruction.Opcode;
2893 /* XXX: we only handle src/dst aliasing in a few opcodes currently.
2894 * Need to use an additional temporay to hold the result in the
2895 * cases where the code is too opaque to fix.
2899 case TGSI_OPCODE_ADD:
2900 case TGSI_OPCODE_MOV:
2901 case TGSI_OPCODE_MUL:
2902 case TGSI_OPCODE_RCP:
2903 case TGSI_OPCODE_RSQ:
2904 case TGSI_OPCODE_EXP:
2905 case TGSI_OPCODE_LOG:
2906 case TGSI_OPCODE_DP3:
2907 case TGSI_OPCODE_DP4:
2908 case TGSI_OPCODE_DP2A:
2909 case TGSI_OPCODE_EX2:
2910 case TGSI_OPCODE_LG2:
2911 case TGSI_OPCODE_POW:
2912 case TGSI_OPCODE_XPD:
2913 case TGSI_OPCODE_DPH:
2914 case TGSI_OPCODE_COS:
2915 case TGSI_OPCODE_SIN:
2916 case TGSI_OPCODE_TEX:
2917 case TGSI_OPCODE_TXB:
2918 case TGSI_OPCODE_TXP:
2919 case TGSI_OPCODE_NRM:
2920 case TGSI_OPCODE_NRM4:
2921 case TGSI_OPCODE_DP2:
2922 /* OK - these opcodes correctly handle SOA dependencies */
2925 if (!tgsi_check_soa_dependencies(inst))
2928 debug_printf("Warning: src/dst aliasing in instruction"
2929 " is not handled:\n");
2930 debug_printf("Warning: ");
2931 tgsi_dump_instruction(inst, 1);
2939 * Translate a TGSI vertex/fragment shader to SSE2 code.
2940 * Slightly different things are done for vertex vs. fragment shaders.
2942 * \param tokens the TGSI input shader
2943 * \param func the output SSE code/function
2944 * \param immediates buffer to place immediates, later passed to SSE func
2945 * \param return 1 for success, 0 if translation failed
2949 const struct tgsi_token *tokens,
2950 struct x86_function *func,
2951 float (*immediates)[4],
2952 boolean do_swizzles )
2954 struct tgsi_parse_context parse;
2956 uint num_immediates = 0;
2960 func->csr = func->store;
2962 tgsi_parse_init( &parse, tokens );
2964 /* Can't just use EDI, EBX without save/restoring them:
2966 x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
2967 x86_push( func, x86_make_reg( file_REG32, reg_DI ) );
2970 * Different function args for vertex/fragment shaders:
2972 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX) {
2978 6 ); /* input_stride */
2984 x86_fn_arg( func, 1 ) );
2988 x86_fn_arg( func, 2 ) );
2991 get_immediate_base(),
2992 x86_fn_arg( func, 3 ) );
2994 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
2998 x86_fn_arg( func, 4 ) );
3004 x86_make_disp( get_machine_base(),
3005 Offset( struct tgsi_exec_machine, Samplers ) ) );
3007 while( !tgsi_parse_end_of_tokens( &parse ) && ok ) {
3008 tgsi_parse_token( &parse );
3010 switch( parse.FullToken.Token.Type ) {
3011 case TGSI_TOKEN_TYPE_DECLARATION:
3012 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
3015 &parse.FullToken.FullDeclaration );
3019 case TGSI_TOKEN_TYPE_INSTRUCTION:
3020 ok = emit_instruction(
3022 &parse.FullToken.FullInstruction );
3025 uint opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
3026 uint proc = parse.FullHeader.Processor.Processor;
3027 debug_printf("failed to translate tgsi opcode %d (%s) to SSE (%s)\n",
3029 tgsi_get_opcode_name(opcode),
3030 tgsi_get_processor_name(proc));
3034 ok = check_soa_dependencies(&parse.FullToken.FullInstruction);
3037 case TGSI_TOKEN_TYPE_IMMEDIATE:
3038 /* simply copy the immediate values into the next immediates[] slot */
3040 const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
3043 assert(num_immediates < TGSI_EXEC_NUM_IMMEDIATES);
3044 for( i = 0; i < size; i++ ) {
3045 immediates[num_immediates][i] =
3046 parse.FullToken.FullImmediate.u[i].Float;
3049 debug_printf("SSE FS immediate[%d] = %f %f %f %f\n",
3051 immediates[num_immediates][0],
3052 immediates[num_immediates][1],
3053 immediates[num_immediates][2],
3054 immediates[num_immediates][3]);
3059 case TGSI_TOKEN_TYPE_PROPERTY:
3060 /* we just ignore them for now */
3069 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX) {
3074 8, /* num_outputs */
3075 9 ); /* output_stride */
3078 /* Can't just use EBX, EDI without save/restoring them:
3080 x86_pop( func, x86_make_reg( file_REG32, reg_DI ) );
3081 x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
3085 tgsi_parse_free( &parse );
3090 #else /* !PIPE_ARCH_X86 */
3094 const struct tgsi_token *tokens,
3095 struct x86_function *func,
3096 float (*immediates)[4],
3097 boolean do_swizzles )
3102 #endif /* !PIPE_ARCH_X86 */