2 * Copyright 2003 Tungsten Graphics, inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * TUNGSTEN GRAPHICS AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 * Keith Whitwell <keithw@tungstengraphics.com>
29 #include "pipe/p_config.h"
30 #include "pipe/p_compiler.h"
31 #include "util/u_memory.h"
32 #include "util/u_math.h"
34 #include "translate.h"
37 #if defined(PIPE_ARCH_X86)
39 #include "rtasm/rtasm_cpu.h"
40 #include "rtasm/rtasm_x86sse.h"
49 typedef void (PIPE_CDECL
*run_func
)( struct translate
*translate
,
55 typedef void (PIPE_CDECL
*run_elts_func
)( struct translate
*translate
,
61 struct translate_buffer
{
66 struct translate_buffer_varient
{
67 unsigned buffer_index
;
68 unsigned instance_divisor
;
69 void *ptr
; /* updated either per vertex or per instance */
73 #define ELEMENT_BUFFER_INSTANCE_ID 1001
76 struct translate_sse
{
77 struct translate translate
;
79 struct x86_function linear_func
;
80 struct x86_function elt_func
;
81 struct x86_function
*func
;
83 boolean loaded_identity
;
85 boolean loaded_inv_255
;
91 struct translate_buffer buffer
[PIPE_MAX_ATTRIBS
];
94 /* Multiple buffer varients can map to a single buffer. */
95 struct translate_buffer_varient buffer_varient
[PIPE_MAX_ATTRIBS
];
96 unsigned nr_buffer_varients
;
98 /* Multiple elements can map to a single buffer varient. */
99 unsigned element_to_buffer_varient
[PIPE_MAX_ATTRIBS
];
101 boolean use_instancing
;
102 unsigned instance_id
;
105 run_elts_func gen_run_elts
;
107 /* these are actually known values, but putting them in a struct
108 * like this is helpful to keep them in sync across the file.
110 struct x86_reg tmp_EAX
;
111 struct x86_reg idx_EBX
; /* either start+i or &elt[i] */
112 struct x86_reg outbuf_ECX
;
113 struct x86_reg machine_EDX
;
114 struct x86_reg count_ESI
; /* decrements to zero */
117 static int get_offset( const void *a
, const void *b
)
119 return (const char *)b
- (const char *)a
;
124 static struct x86_reg
get_identity( struct translate_sse
*p
)
126 struct x86_reg reg
= x86_make_reg(file_XMM
, 6);
128 if (!p
->loaded_identity
) {
129 p
->loaded_identity
= TRUE
;
135 sse_movups(p
->func
, reg
,
136 x86_make_disp(p
->machine_EDX
,
137 get_offset(p
, &p
->identity
[0])));
143 static struct x86_reg
get_255( struct translate_sse
*p
)
145 struct x86_reg reg
= x86_make_reg(file_XMM
, 7);
147 if (!p
->loaded_255
) {
148 p
->loaded_255
= TRUE
;
152 p
->float_255
[3] = 255.0f
;
154 sse_movups(p
->func
, reg
,
155 x86_make_disp(p
->machine_EDX
,
156 get_offset(p
, &p
->float_255
[0])));
162 static struct x86_reg
get_inv_255( struct translate_sse
*p
)
164 struct x86_reg reg
= x86_make_reg(file_XMM
, 5);
166 if (!p
->loaded_inv_255
) {
167 p
->loaded_inv_255
= TRUE
;
171 p
->inv_255
[3] = 1.0f
/ 255.0f
;
173 sse_movups(p
->func
, reg
,
174 x86_make_disp(p
->machine_EDX
,
175 get_offset(p
, &p
->inv_255
[0])));
182 static void emit_load_R32G32B32A32( struct translate_sse
*p
,
184 struct x86_reg arg0
)
186 sse_movups(p
->func
, data
, arg0
);
189 static void emit_load_R32G32B32( struct translate_sse
*p
,
191 struct x86_reg arg0
)
193 /* Have to jump through some hoops:
200 sse_movss(p
->func
, data
, x86_make_disp(arg0
, 8));
201 sse_shufps(p
->func
, data
, get_identity(p
), SHUF(X
,Y
,Z
,W
) );
202 sse_shufps(p
->func
, data
, data
, SHUF(Y
,Z
,X
,W
) );
203 sse_movlps(p
->func
, data
, arg0
);
206 static void emit_load_R32G32( struct translate_sse
*p
,
208 struct x86_reg arg0
)
213 sse_movups(p
->func
, data
, get_identity(p
) );
214 sse_movlps(p
->func
, data
, arg0
);
218 static void emit_load_R32( struct translate_sse
*p
,
220 struct x86_reg arg0
)
225 sse_movss(p
->func
, data
, arg0
);
226 sse_orps(p
->func
, data
, get_identity(p
) );
230 static void emit_load_R8G8B8A8_UNORM( struct translate_sse
*p
,
235 /* Load and unpack twice:
237 sse_movss(p
->func
, data
, src
);
238 sse2_punpcklbw(p
->func
, data
, get_identity(p
));
239 sse2_punpcklbw(p
->func
, data
, get_identity(p
));
243 sse2_cvtdq2ps(p
->func
, data
, data
);
248 sse_mulps(p
->func
, data
, get_inv_255(p
));
254 static void emit_store_R32G32B32A32( struct translate_sse
*p
,
256 struct x86_reg dataXMM
)
258 sse_movups(p
->func
, dest
, dataXMM
);
261 static void emit_store_R32G32B32( struct translate_sse
*p
,
263 struct x86_reg dataXMM
)
265 /* Emit two, shuffle, emit one.
267 sse_movlps(p
->func
, dest
, dataXMM
);
268 sse_shufps(p
->func
, dataXMM
, dataXMM
, SHUF(Z
,Z
,Z
,Z
) ); /* NOTE! destructive */
269 sse_movss(p
->func
, x86_make_disp(dest
,8), dataXMM
);
272 static void emit_store_R32G32( struct translate_sse
*p
,
274 struct x86_reg dataXMM
)
276 sse_movlps(p
->func
, dest
, dataXMM
);
279 static void emit_store_R32( struct translate_sse
*p
,
281 struct x86_reg dataXMM
)
283 sse_movss(p
->func
, dest
, dataXMM
);
288 static void emit_store_R8G8B8A8_UNORM( struct translate_sse
*p
,
290 struct x86_reg dataXMM
)
294 sse_mulps(p
->func
, dataXMM
, get_255(p
));
298 sse2_cvtps2dq(p
->func
, dataXMM
, dataXMM
);
299 sse2_packssdw(p
->func
, dataXMM
, dataXMM
);
300 sse2_packuswb(p
->func
, dataXMM
, dataXMM
);
301 sse_movss(p
->func
, dest
, dataXMM
);
308 /* Extended swizzles? Maybe later.
310 static void emit_swizzle( struct translate_sse
*p
,
313 unsigned char shuffle
)
315 sse_shufps(p
->func
, dest
, src
, shuffle
);
319 static boolean
translate_attr( struct translate_sse
*p
,
320 const struct translate_element
*a
,
321 struct x86_reg srcECX
,
322 struct x86_reg dstEAX
)
324 struct x86_reg dataXMM
= x86_make_reg(file_XMM
, 0);
326 switch (a
->input_format
) {
327 case PIPE_FORMAT_R32_FLOAT
:
328 emit_load_R32(p
, dataXMM
, srcECX
);
330 case PIPE_FORMAT_R32G32_FLOAT
:
331 emit_load_R32G32(p
, dataXMM
, srcECX
);
333 case PIPE_FORMAT_R32G32B32_FLOAT
:
334 emit_load_R32G32B32(p
, dataXMM
, srcECX
);
336 case PIPE_FORMAT_R32G32B32A32_FLOAT
:
337 emit_load_R32G32B32A32(p
, dataXMM
, srcECX
);
339 case PIPE_FORMAT_B8G8R8A8_UNORM
:
340 emit_load_R8G8B8A8_UNORM(p
, dataXMM
, srcECX
);
341 emit_swizzle(p
, dataXMM
, dataXMM
, SHUF(Z
,Y
,X
,W
));
343 case PIPE_FORMAT_R8G8B8A8_UNORM
:
344 emit_load_R8G8B8A8_UNORM(p
, dataXMM
, srcECX
);
350 switch (a
->output_format
) {
351 case PIPE_FORMAT_R32_FLOAT
:
352 emit_store_R32(p
, dstEAX
, dataXMM
);
354 case PIPE_FORMAT_R32G32_FLOAT
:
355 emit_store_R32G32(p
, dstEAX
, dataXMM
);
357 case PIPE_FORMAT_R32G32B32_FLOAT
:
358 emit_store_R32G32B32(p
, dstEAX
, dataXMM
);
360 case PIPE_FORMAT_R32G32B32A32_FLOAT
:
361 emit_store_R32G32B32A32(p
, dstEAX
, dataXMM
);
363 case PIPE_FORMAT_B8G8R8A8_UNORM
:
364 emit_swizzle(p
, dataXMM
, dataXMM
, SHUF(Z
,Y
,X
,W
));
365 emit_store_R8G8B8A8_UNORM(p
, dstEAX
, dataXMM
);
367 case PIPE_FORMAT_R8G8B8A8_UNORM
:
368 emit_store_R8G8B8A8_UNORM(p
, dstEAX
, dataXMM
);
378 static boolean
init_inputs( struct translate_sse
*p
,
382 struct x86_reg instance_id
= x86_make_disp(p
->machine_EDX
,
383 get_offset(p
, &p
->instance_id
));
385 for (i
= 0; i
< p
->nr_buffer_varients
; i
++) {
386 struct translate_buffer_varient
*varient
= &p
->buffer_varient
[i
];
387 struct translate_buffer
*buffer
= &p
->buffer
[varient
->buffer_index
];
389 if (linear
|| varient
->instance_divisor
) {
390 struct x86_reg buf_stride
= x86_make_disp(p
->machine_EDX
,
391 get_offset(p
, &buffer
->stride
));
392 struct x86_reg buf_ptr
= x86_make_disp(p
->machine_EDX
,
393 get_offset(p
, &varient
->ptr
));
394 struct x86_reg buf_base_ptr
= x86_make_disp(p
->machine_EDX
,
395 get_offset(p
, &buffer
->base_ptr
));
396 struct x86_reg elt
= p
->idx_EBX
;
397 struct x86_reg tmp_EAX
= p
->tmp_EAX
;
399 /* Calculate pointer to first attrib:
400 * base_ptr + stride * index, where index depends on instance divisor
402 if (varient
->instance_divisor
) {
403 /* Our index is instance ID divided by instance divisor.
405 x86_mov(p
->func
, tmp_EAX
, instance_id
);
407 if (varient
->instance_divisor
!= 1) {
408 struct x86_reg tmp_EDX
= p
->machine_EDX
;
409 struct x86_reg tmp_ECX
= p
->outbuf_ECX
;
411 /* TODO: Add x86_shr() to rtasm and use it whenever
412 * instance divisor is power of two.
415 x86_push(p
->func
, tmp_EDX
);
416 x86_push(p
->func
, tmp_ECX
);
417 x86_xor(p
->func
, tmp_EDX
, tmp_EDX
);
418 x86_mov_reg_imm(p
->func
, tmp_ECX
, varient
->instance_divisor
);
419 x86_div(p
->func
, tmp_ECX
); /* EAX = EDX:EAX / ECX */
420 x86_pop(p
->func
, tmp_ECX
);
421 x86_pop(p
->func
, tmp_EDX
);
424 x86_mov(p
->func
, tmp_EAX
, elt
);
426 x86_imul(p
->func
, tmp_EAX
, buf_stride
);
427 x86_add(p
->func
, tmp_EAX
, buf_base_ptr
);
430 /* In the linear case, keep the buffer pointer instead of the
433 if (linear
&& p
->nr_buffer_varients
== 1)
434 x86_mov(p
->func
, elt
, tmp_EAX
);
436 x86_mov(p
->func
, buf_ptr
, tmp_EAX
);
444 static struct x86_reg
get_buffer_ptr( struct translate_sse
*p
,
449 if (var_idx
== ELEMENT_BUFFER_INSTANCE_ID
) {
450 return x86_make_disp(p
->machine_EDX
,
451 get_offset(p
, &p
->instance_id
));
453 if (linear
&& p
->nr_buffer_varients
== 1) {
456 else if (linear
|| p
->buffer_varient
[var_idx
].instance_divisor
) {
457 struct x86_reg ptr
= p
->tmp_EAX
;
458 struct x86_reg buf_ptr
=
459 x86_make_disp(p
->machine_EDX
,
460 get_offset(p
, &p
->buffer_varient
[var_idx
].ptr
));
462 x86_mov(p
->func
, ptr
, buf_ptr
);
466 struct x86_reg ptr
= p
->tmp_EAX
;
467 const struct translate_buffer_varient
*varient
= &p
->buffer_varient
[var_idx
];
469 struct x86_reg buf_stride
=
470 x86_make_disp(p
->machine_EDX
,
471 get_offset(p
, &p
->buffer
[varient
->buffer_index
].stride
));
473 struct x86_reg buf_base_ptr
=
474 x86_make_disp(p
->machine_EDX
,
475 get_offset(p
, &p
->buffer
[varient
->buffer_index
].base_ptr
));
479 /* Calculate pointer to current attrib:
481 x86_mov(p
->func
, ptr
, buf_stride
);
482 x86_imul(p
->func
, ptr
, elt
);
483 x86_add(p
->func
, ptr
, buf_base_ptr
);
490 static boolean
incr_inputs( struct translate_sse
*p
,
493 if (linear
&& p
->nr_buffer_varients
== 1) {
494 struct x86_reg stride
= x86_make_disp(p
->machine_EDX
,
495 get_offset(p
, &p
->buffer
[0].stride
));
497 if (p
->buffer_varient
[0].instance_divisor
== 0) {
498 x86_add(p
->func
, p
->idx_EBX
, stride
);
499 sse_prefetchnta(p
->func
, x86_make_disp(p
->idx_EBX
, 192));
505 /* Is this worthwhile??
507 for (i
= 0; i
< p
->nr_buffer_varients
; i
++) {
508 struct translate_buffer_varient
*varient
= &p
->buffer_varient
[i
];
509 struct x86_reg buf_ptr
= x86_make_disp(p
->machine_EDX
,
510 get_offset(p
, &varient
->ptr
));
511 struct x86_reg buf_stride
= x86_make_disp(p
->machine_EDX
,
512 get_offset(p
, &p
->buffer
[varient
->buffer_index
].stride
));
514 if (varient
->instance_divisor
== 0) {
515 x86_mov(p
->func
, p
->tmp_EAX
, buf_ptr
);
516 x86_add(p
->func
, p
->tmp_EAX
, buf_stride
);
517 if (i
== 0) sse_prefetchnta(p
->func
, x86_make_disp(p
->tmp_EAX
, 192));
518 x86_mov(p
->func
, buf_ptr
, p
->tmp_EAX
);
523 x86_lea(p
->func
, p
->idx_EBX
, x86_make_disp(p
->idx_EBX
, 4));
530 /* Build run( struct translate *machine,
533 * void *output_buffer )
535 * run_elts( struct translate *machine,
538 * void *output_buffer )
542 * EAX -- pointer to current output vertex
543 * ECX -- pointer to current attribute
546 static boolean
build_vertex_emit( struct translate_sse
*p
,
547 struct x86_function
*func
,
553 p
->tmp_EAX
= x86_make_reg(file_REG32
, reg_AX
);
554 p
->idx_EBX
= x86_make_reg(file_REG32
, reg_BX
);
555 p
->outbuf_ECX
= x86_make_reg(file_REG32
, reg_CX
);
556 p
->machine_EDX
= x86_make_reg(file_REG32
, reg_DX
);
557 p
->count_ESI
= x86_make_reg(file_REG32
, reg_SI
);
560 p
->loaded_inv_255
= FALSE
;
561 p
->loaded_255
= FALSE
;
562 p
->loaded_identity
= FALSE
;
564 x86_init_func(p
->func
);
568 x86_push(p
->func
, p
->idx_EBX
);
569 x86_push(p
->func
, p
->count_ESI
);
571 /* Load arguments into regs:
573 x86_mov(p
->func
, p
->machine_EDX
, x86_fn_arg(p
->func
, 1));
574 x86_mov(p
->func
, p
->idx_EBX
, x86_fn_arg(p
->func
, 2));
575 x86_mov(p
->func
, p
->count_ESI
, x86_fn_arg(p
->func
, 3));
576 x86_mov(p
->func
, p
->outbuf_ECX
, x86_fn_arg(p
->func
, 5));
580 if (p
->use_instancing
) {
583 x86_fn_arg(p
->func
, 4));
585 x86_make_disp(p
->machine_EDX
, get_offset(p
, &p
->instance_id
)),
589 /* Get vertex count, compare to zero
591 x86_xor(p
->func
, p
->tmp_EAX
, p
->tmp_EAX
);
592 x86_cmp(p
->func
, p
->count_ESI
, p
->tmp_EAX
);
593 fixup
= x86_jcc_forward(p
->func
, cc_E
);
595 /* always load, needed or not:
597 init_inputs(p
, linear
);
599 /* Note address for loop jump
601 label
= x86_get_label(p
->func
);
603 struct x86_reg elt
= linear
? p
->idx_EBX
: x86_deref(p
->idx_EBX
);
604 int last_varient
= -1;
607 for (j
= 0; j
< p
->translate
.key
.nr_elements
; j
++) {
608 const struct translate_element
*a
= &p
->translate
.key
.element
[j
];
609 unsigned varient
= p
->element_to_buffer_varient
[j
];
611 /* Figure out source pointer address:
613 if (varient
!= last_varient
) {
614 last_varient
= varient
;
615 vb
= get_buffer_ptr(p
, linear
, varient
, elt
);
618 if (!translate_attr( p
, a
,
619 x86_make_disp(vb
, a
->input_offset
),
620 x86_make_disp(p
->outbuf_ECX
, a
->output_offset
)))
624 /* Next output vertex:
628 x86_make_disp(p
->outbuf_ECX
,
629 p
->translate
.key
.output_stride
));
633 incr_inputs( p
, linear
);
636 /* decr count, loop if not zero
638 x86_dec(p
->func
, p
->count_ESI
);
639 x86_jcc(p
->func
, cc_NZ
, label
);
643 if (p
->func
->need_emms
)
646 /* Land forward jump here:
648 x86_fixup_fwd_jump(p
->func
, fixup
);
650 /* Pop regs and return
653 x86_pop(p
->func
, p
->count_ESI
);
654 x86_pop(p
->func
, p
->idx_EBX
);
666 static void translate_sse_set_buffer( struct translate
*translate
,
671 struct translate_sse
*p
= (struct translate_sse
*)translate
;
673 if (buf
< p
->nr_buffers
) {
674 p
->buffer
[buf
].base_ptr
= (char *)ptr
;
675 p
->buffer
[buf
].stride
= stride
;
678 if (0) debug_printf("%s %d/%d: %p %d\n",
685 static void translate_sse_release( struct translate
*translate
)
687 struct translate_sse
*p
= (struct translate_sse
*)translate
;
689 x86_release_func( &p
->linear_func
);
690 x86_release_func( &p
->elt_func
);
695 static void PIPE_CDECL
translate_sse_run_elts( struct translate
*translate
,
696 const unsigned *elts
,
698 unsigned instance_id
,
699 void *output_buffer
)
701 struct translate_sse
*p
= (struct translate_sse
*)translate
;
703 p
->gen_run_elts( translate
,
710 static void PIPE_CDECL
translate_sse_run( struct translate
*translate
,
713 unsigned instance_id
,
714 void *output_buffer
)
716 struct translate_sse
*p
= (struct translate_sse
*)translate
;
718 p
->gen_run( translate
,
726 struct translate
*translate_sse2_create( const struct translate_key
*key
)
728 struct translate_sse
*p
= NULL
;
731 if (!rtasm_cpu_has_sse() || !rtasm_cpu_has_sse2())
734 p
= CALLOC_STRUCT( translate_sse
);
738 p
->translate
.key
= *key
;
739 p
->translate
.release
= translate_sse_release
;
740 p
->translate
.set_buffer
= translate_sse_set_buffer
;
741 p
->translate
.run_elts
= translate_sse_run_elts
;
742 p
->translate
.run
= translate_sse_run
;
744 for (i
= 0; i
< key
->nr_elements
; i
++) {
745 if (key
->element
[i
].type
== TRANSLATE_ELEMENT_NORMAL
) {
748 p
->nr_buffers
= MAX2(p
->nr_buffers
, key
->element
[i
].input_buffer
+ 1);
750 if (key
->element
[i
].instance_divisor
) {
751 p
->use_instancing
= TRUE
;
755 * Map vertex element to vertex buffer varient.
757 for (j
= 0; j
< p
->nr_buffer_varients
; j
++) {
758 if (p
->buffer_varient
[j
].buffer_index
== key
->element
[i
].input_buffer
&&
759 p
->buffer_varient
[j
].instance_divisor
== key
->element
[i
].instance_divisor
) {
763 if (j
== p
->nr_buffer_varients
) {
764 p
->buffer_varient
[j
].buffer_index
= key
->element
[i
].input_buffer
;
765 p
->buffer_varient
[j
].instance_divisor
= key
->element
[i
].instance_divisor
;
766 p
->nr_buffer_varients
++;
768 p
->element_to_buffer_varient
[i
] = j
;
770 assert(key
->element
[i
].type
== TRANSLATE_ELEMENT_INSTANCE_ID
);
772 p
->element_to_buffer_varient
[i
] = ELEMENT_BUFFER_INSTANCE_ID
;
776 if (0) debug_printf("nr_buffers: %d\n", p
->nr_buffers
);
778 if (!build_vertex_emit(p
, &p
->linear_func
, TRUE
))
781 if (!build_vertex_emit(p
, &p
->elt_func
, FALSE
))
784 p
->gen_run
= (run_func
)x86_get_func(&p
->linear_func
);
785 if (p
->gen_run
== NULL
)
788 p
->gen_run_elts
= (run_elts_func
)x86_get_func(&p
->elt_func
);
789 if (p
->gen_run_elts
== NULL
)
792 return &p
->translate
;
796 translate_sse_release( &p
->translate
);
805 struct translate
*translate_sse2_create( const struct translate_key
*key
)