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
{
67 struct translate_buffer_varient
{
68 unsigned buffer_index
;
69 unsigned instance_divisor
;
70 void *ptr
; /* updated either per vertex or per instance */
74 #define ELEMENT_BUFFER_INSTANCE_ID 1001
77 struct translate_sse
{
78 struct translate translate
;
80 struct x86_function linear_func
;
81 struct x86_function elt_func
;
82 struct x86_function
*func
;
84 boolean loaded_identity
;
86 boolean loaded_inv_255
;
92 struct translate_buffer buffer
[PIPE_MAX_ATTRIBS
];
95 /* Multiple buffer varients can map to a single buffer. */
96 struct translate_buffer_varient buffer_varient
[PIPE_MAX_ATTRIBS
];
97 unsigned nr_buffer_varients
;
99 /* Multiple elements can map to a single buffer varient. */
100 unsigned element_to_buffer_varient
[PIPE_MAX_ATTRIBS
];
102 boolean use_instancing
;
103 unsigned instance_id
;
106 run_elts_func gen_run_elts
;
108 /* these are actually known values, but putting them in a struct
109 * like this is helpful to keep them in sync across the file.
111 struct x86_reg tmp_EAX
;
112 struct x86_reg idx_EBX
; /* either start+i or &elt[i] */
113 struct x86_reg outbuf_ECX
;
114 struct x86_reg machine_EDX
;
115 struct x86_reg count_ESI
; /* decrements to zero */
118 static int get_offset( const void *a
, const void *b
)
120 return (const char *)b
- (const char *)a
;
125 static struct x86_reg
get_identity( struct translate_sse
*p
)
127 struct x86_reg reg
= x86_make_reg(file_XMM
, 6);
129 if (!p
->loaded_identity
) {
130 p
->loaded_identity
= TRUE
;
136 sse_movups(p
->func
, reg
,
137 x86_make_disp(p
->machine_EDX
,
138 get_offset(p
, &p
->identity
[0])));
144 static struct x86_reg
get_255( struct translate_sse
*p
)
146 struct x86_reg reg
= x86_make_reg(file_XMM
, 7);
148 if (!p
->loaded_255
) {
149 p
->loaded_255
= TRUE
;
153 p
->float_255
[3] = 255.0f
;
155 sse_movups(p
->func
, reg
,
156 x86_make_disp(p
->machine_EDX
,
157 get_offset(p
, &p
->float_255
[0])));
163 static struct x86_reg
get_inv_255( struct translate_sse
*p
)
165 struct x86_reg reg
= x86_make_reg(file_XMM
, 5);
167 if (!p
->loaded_inv_255
) {
168 p
->loaded_inv_255
= TRUE
;
172 p
->inv_255
[3] = 1.0f
/ 255.0f
;
174 sse_movups(p
->func
, reg
,
175 x86_make_disp(p
->machine_EDX
,
176 get_offset(p
, &p
->inv_255
[0])));
183 static void emit_load_R32G32B32A32( struct translate_sse
*p
,
185 struct x86_reg arg0
)
187 sse_movups(p
->func
, data
, arg0
);
190 static void emit_load_R32G32B32( struct translate_sse
*p
,
192 struct x86_reg arg0
)
194 /* Have to jump through some hoops:
201 sse_movss(p
->func
, data
, x86_make_disp(arg0
, 8));
202 sse_shufps(p
->func
, data
, get_identity(p
), SHUF(X
,Y
,Z
,W
) );
203 sse_shufps(p
->func
, data
, data
, SHUF(Y
,Z
,X
,W
) );
204 sse_movlps(p
->func
, data
, arg0
);
207 static void emit_load_R32G32( struct translate_sse
*p
,
209 struct x86_reg arg0
)
214 sse_movups(p
->func
, data
, get_identity(p
) );
215 sse_movlps(p
->func
, data
, arg0
);
219 static void emit_load_R32( struct translate_sse
*p
,
221 struct x86_reg arg0
)
226 sse_movss(p
->func
, data
, arg0
);
227 sse_orps(p
->func
, data
, get_identity(p
) );
231 static void emit_load_R8G8B8A8_UNORM( struct translate_sse
*p
,
236 /* Load and unpack twice:
238 sse_movss(p
->func
, data
, src
);
239 sse2_punpcklbw(p
->func
, data
, get_identity(p
));
240 sse2_punpcklbw(p
->func
, data
, get_identity(p
));
244 sse2_cvtdq2ps(p
->func
, data
, data
);
249 sse_mulps(p
->func
, data
, get_inv_255(p
));
255 static void emit_store_R32G32B32A32( struct translate_sse
*p
,
257 struct x86_reg dataXMM
)
259 sse_movups(p
->func
, dest
, dataXMM
);
262 static void emit_store_R32G32B32( struct translate_sse
*p
,
264 struct x86_reg dataXMM
)
266 /* Emit two, shuffle, emit one.
268 sse_movlps(p
->func
, dest
, dataXMM
);
269 sse_shufps(p
->func
, dataXMM
, dataXMM
, SHUF(Z
,Z
,Z
,Z
) ); /* NOTE! destructive */
270 sse_movss(p
->func
, x86_make_disp(dest
,8), dataXMM
);
273 static void emit_store_R32G32( struct translate_sse
*p
,
275 struct x86_reg dataXMM
)
277 sse_movlps(p
->func
, dest
, dataXMM
);
280 static void emit_store_R32( struct translate_sse
*p
,
282 struct x86_reg dataXMM
)
284 sse_movss(p
->func
, dest
, dataXMM
);
289 static void emit_store_R8G8B8A8_UNORM( struct translate_sse
*p
,
291 struct x86_reg dataXMM
)
295 sse_mulps(p
->func
, dataXMM
, get_255(p
));
299 sse2_cvtps2dq(p
->func
, dataXMM
, dataXMM
);
300 sse2_packssdw(p
->func
, dataXMM
, dataXMM
);
301 sse2_packuswb(p
->func
, dataXMM
, dataXMM
);
302 sse_movss(p
->func
, dest
, dataXMM
);
309 /* Extended swizzles? Maybe later.
311 static void emit_swizzle( struct translate_sse
*p
,
314 unsigned char shuffle
)
316 sse_shufps(p
->func
, dest
, src
, shuffle
);
320 static boolean
translate_attr( struct translate_sse
*p
,
321 const struct translate_element
*a
,
322 struct x86_reg srcECX
,
323 struct x86_reg dstEAX
)
325 struct x86_reg dataXMM
= x86_make_reg(file_XMM
, 0);
327 switch (a
->input_format
) {
328 case PIPE_FORMAT_R32_FLOAT
:
329 emit_load_R32(p
, dataXMM
, srcECX
);
331 case PIPE_FORMAT_R32G32_FLOAT
:
332 emit_load_R32G32(p
, dataXMM
, srcECX
);
334 case PIPE_FORMAT_R32G32B32_FLOAT
:
335 emit_load_R32G32B32(p
, dataXMM
, srcECX
);
337 case PIPE_FORMAT_R32G32B32A32_FLOAT
:
338 emit_load_R32G32B32A32(p
, dataXMM
, srcECX
);
340 case PIPE_FORMAT_B8G8R8A8_UNORM
:
341 emit_load_R8G8B8A8_UNORM(p
, dataXMM
, srcECX
);
342 emit_swizzle(p
, dataXMM
, dataXMM
, SHUF(Z
,Y
,X
,W
));
344 case PIPE_FORMAT_R8G8B8A8_UNORM
:
345 emit_load_R8G8B8A8_UNORM(p
, dataXMM
, srcECX
);
351 switch (a
->output_format
) {
352 case PIPE_FORMAT_R32_FLOAT
:
353 emit_store_R32(p
, dstEAX
, dataXMM
);
355 case PIPE_FORMAT_R32G32_FLOAT
:
356 emit_store_R32G32(p
, dstEAX
, dataXMM
);
358 case PIPE_FORMAT_R32G32B32_FLOAT
:
359 emit_store_R32G32B32(p
, dstEAX
, dataXMM
);
361 case PIPE_FORMAT_R32G32B32A32_FLOAT
:
362 emit_store_R32G32B32A32(p
, dstEAX
, dataXMM
);
364 case PIPE_FORMAT_B8G8R8A8_UNORM
:
365 emit_swizzle(p
, dataXMM
, dataXMM
, SHUF(Z
,Y
,X
,W
));
366 emit_store_R8G8B8A8_UNORM(p
, dstEAX
, dataXMM
);
368 case PIPE_FORMAT_R8G8B8A8_UNORM
:
369 emit_store_R8G8B8A8_UNORM(p
, dstEAX
, dataXMM
);
379 static boolean
init_inputs( struct translate_sse
*p
,
383 struct x86_reg instance_id
= x86_make_disp(p
->machine_EDX
,
384 get_offset(p
, &p
->instance_id
));
386 for (i
= 0; i
< p
->nr_buffer_varients
; i
++) {
387 struct translate_buffer_varient
*varient
= &p
->buffer_varient
[i
];
388 struct translate_buffer
*buffer
= &p
->buffer
[varient
->buffer_index
];
390 if (linear
|| varient
->instance_divisor
) {
391 struct x86_reg buf_stride
= x86_make_disp(p
->machine_EDX
,
392 get_offset(p
, &buffer
->stride
));
393 struct x86_reg buf_ptr
= x86_make_disp(p
->machine_EDX
,
394 get_offset(p
, &varient
->ptr
));
395 struct x86_reg buf_base_ptr
= x86_make_disp(p
->machine_EDX
,
396 get_offset(p
, &buffer
->base_ptr
));
397 struct x86_reg elt
= p
->idx_EBX
;
398 struct x86_reg tmp_EAX
= p
->tmp_EAX
;
400 /* Calculate pointer to first attrib:
401 * base_ptr + stride * index, where index depends on instance divisor
403 if (varient
->instance_divisor
) {
404 /* Our index is instance ID divided by instance divisor.
406 x86_mov(p
->func
, tmp_EAX
, instance_id
);
408 if (varient
->instance_divisor
!= 1) {
409 struct x86_reg tmp_EDX
= p
->machine_EDX
;
410 struct x86_reg tmp_ECX
= p
->outbuf_ECX
;
412 /* TODO: Add x86_shr() to rtasm and use it whenever
413 * instance divisor is power of two.
416 x86_push(p
->func
, tmp_EDX
);
417 x86_push(p
->func
, tmp_ECX
);
418 x86_xor(p
->func
, tmp_EDX
, tmp_EDX
);
419 x86_mov_reg_imm(p
->func
, tmp_ECX
, varient
->instance_divisor
);
420 x86_div(p
->func
, tmp_ECX
); /* EAX = EDX:EAX / ECX */
421 x86_pop(p
->func
, tmp_ECX
);
422 x86_pop(p
->func
, tmp_EDX
);
425 x86_mov(p
->func
, tmp_EAX
, elt
);
429 * TODO: Respect translate_buffer::max_index.
432 x86_imul(p
->func
, tmp_EAX
, buf_stride
);
433 x86_add(p
->func
, tmp_EAX
, buf_base_ptr
);
436 /* In the linear case, keep the buffer pointer instead of the
439 if (linear
&& p
->nr_buffer_varients
== 1)
440 x86_mov(p
->func
, elt
, tmp_EAX
);
442 x86_mov(p
->func
, buf_ptr
, tmp_EAX
);
450 static struct x86_reg
get_buffer_ptr( struct translate_sse
*p
,
455 if (var_idx
== ELEMENT_BUFFER_INSTANCE_ID
) {
456 return x86_make_disp(p
->machine_EDX
,
457 get_offset(p
, &p
->instance_id
));
459 if (linear
&& p
->nr_buffer_varients
== 1) {
462 else if (linear
|| p
->buffer_varient
[var_idx
].instance_divisor
) {
463 struct x86_reg ptr
= p
->tmp_EAX
;
464 struct x86_reg buf_ptr
=
465 x86_make_disp(p
->machine_EDX
,
466 get_offset(p
, &p
->buffer_varient
[var_idx
].ptr
));
468 x86_mov(p
->func
, ptr
, buf_ptr
);
472 struct x86_reg ptr
= p
->tmp_EAX
;
473 const struct translate_buffer_varient
*varient
= &p
->buffer_varient
[var_idx
];
475 struct x86_reg buf_stride
=
476 x86_make_disp(p
->machine_EDX
,
477 get_offset(p
, &p
->buffer
[varient
->buffer_index
].stride
));
479 struct x86_reg buf_base_ptr
=
480 x86_make_disp(p
->machine_EDX
,
481 get_offset(p
, &p
->buffer
[varient
->buffer_index
].base_ptr
));
485 /* Calculate pointer to current attrib:
487 x86_mov(p
->func
, ptr
, buf_stride
);
488 x86_imul(p
->func
, ptr
, elt
);
489 x86_add(p
->func
, ptr
, buf_base_ptr
);
496 static boolean
incr_inputs( struct translate_sse
*p
,
499 if (linear
&& p
->nr_buffer_varients
== 1) {
500 struct x86_reg stride
= x86_make_disp(p
->machine_EDX
,
501 get_offset(p
, &p
->buffer
[0].stride
));
503 if (p
->buffer_varient
[0].instance_divisor
== 0) {
504 x86_add(p
->func
, p
->idx_EBX
, stride
);
505 sse_prefetchnta(p
->func
, x86_make_disp(p
->idx_EBX
, 192));
511 /* Is this worthwhile??
513 for (i
= 0; i
< p
->nr_buffer_varients
; i
++) {
514 struct translate_buffer_varient
*varient
= &p
->buffer_varient
[i
];
515 struct x86_reg buf_ptr
= x86_make_disp(p
->machine_EDX
,
516 get_offset(p
, &varient
->ptr
));
517 struct x86_reg buf_stride
= x86_make_disp(p
->machine_EDX
,
518 get_offset(p
, &p
->buffer
[varient
->buffer_index
].stride
));
520 if (varient
->instance_divisor
== 0) {
521 x86_mov(p
->func
, p
->tmp_EAX
, buf_ptr
);
522 x86_add(p
->func
, p
->tmp_EAX
, buf_stride
);
523 if (i
== 0) sse_prefetchnta(p
->func
, x86_make_disp(p
->tmp_EAX
, 192));
524 x86_mov(p
->func
, buf_ptr
, p
->tmp_EAX
);
529 x86_lea(p
->func
, p
->idx_EBX
, x86_make_disp(p
->idx_EBX
, 4));
536 /* Build run( struct translate *machine,
539 * void *output_buffer )
541 * run_elts( struct translate *machine,
544 * void *output_buffer )
548 * EAX -- pointer to current output vertex
549 * ECX -- pointer to current attribute
552 static boolean
build_vertex_emit( struct translate_sse
*p
,
553 struct x86_function
*func
,
559 p
->tmp_EAX
= x86_make_reg(file_REG32
, reg_AX
);
560 p
->idx_EBX
= x86_make_reg(file_REG32
, reg_BX
);
561 p
->outbuf_ECX
= x86_make_reg(file_REG32
, reg_CX
);
562 p
->machine_EDX
= x86_make_reg(file_REG32
, reg_DX
);
563 p
->count_ESI
= x86_make_reg(file_REG32
, reg_SI
);
566 p
->loaded_inv_255
= FALSE
;
567 p
->loaded_255
= FALSE
;
568 p
->loaded_identity
= FALSE
;
570 x86_init_func(p
->func
);
574 x86_push(p
->func
, p
->idx_EBX
);
575 x86_push(p
->func
, p
->count_ESI
);
577 /* Load arguments into regs:
579 x86_mov(p
->func
, p
->machine_EDX
, x86_fn_arg(p
->func
, 1));
580 x86_mov(p
->func
, p
->idx_EBX
, x86_fn_arg(p
->func
, 2));
581 x86_mov(p
->func
, p
->count_ESI
, x86_fn_arg(p
->func
, 3));
582 x86_mov(p
->func
, p
->outbuf_ECX
, x86_fn_arg(p
->func
, 5));
586 if (p
->use_instancing
) {
589 x86_fn_arg(p
->func
, 4));
591 x86_make_disp(p
->machine_EDX
, get_offset(p
, &p
->instance_id
)),
595 /* Get vertex count, compare to zero
597 x86_xor(p
->func
, p
->tmp_EAX
, p
->tmp_EAX
);
598 x86_cmp(p
->func
, p
->count_ESI
, p
->tmp_EAX
);
599 fixup
= x86_jcc_forward(p
->func
, cc_E
);
601 /* always load, needed or not:
603 init_inputs(p
, linear
);
605 /* Note address for loop jump
607 label
= x86_get_label(p
->func
);
609 struct x86_reg elt
= linear
? p
->idx_EBX
: x86_deref(p
->idx_EBX
);
610 int last_varient
= -1;
613 for (j
= 0; j
< p
->translate
.key
.nr_elements
; j
++) {
614 const struct translate_element
*a
= &p
->translate
.key
.element
[j
];
615 unsigned varient
= p
->element_to_buffer_varient
[j
];
617 /* Figure out source pointer address:
619 if (varient
!= last_varient
) {
620 last_varient
= varient
;
621 vb
= get_buffer_ptr(p
, linear
, varient
, elt
);
624 if (!translate_attr( p
, a
,
625 x86_make_disp(vb
, a
->input_offset
),
626 x86_make_disp(p
->outbuf_ECX
, a
->output_offset
)))
630 /* Next output vertex:
634 x86_make_disp(p
->outbuf_ECX
,
635 p
->translate
.key
.output_stride
));
639 incr_inputs( p
, linear
);
642 /* decr count, loop if not zero
644 x86_dec(p
->func
, p
->count_ESI
);
645 x86_jcc(p
->func
, cc_NZ
, label
);
649 if (p
->func
->need_emms
)
652 /* Land forward jump here:
654 x86_fixup_fwd_jump(p
->func
, fixup
);
656 /* Pop regs and return
659 x86_pop(p
->func
, p
->count_ESI
);
660 x86_pop(p
->func
, p
->idx_EBX
);
672 static void translate_sse_set_buffer( struct translate
*translate
,
678 struct translate_sse
*p
= (struct translate_sse
*)translate
;
680 if (buf
< p
->nr_buffers
) {
681 p
->buffer
[buf
].base_ptr
= (char *)ptr
;
682 p
->buffer
[buf
].stride
= stride
;
683 p
->buffer
[buf
].max_index
= max_index
;
686 if (0) debug_printf("%s %d/%d: %p %d\n",
693 static void translate_sse_release( struct translate
*translate
)
695 struct translate_sse
*p
= (struct translate_sse
*)translate
;
697 x86_release_func( &p
->linear_func
);
698 x86_release_func( &p
->elt_func
);
703 static void PIPE_CDECL
translate_sse_run_elts( struct translate
*translate
,
704 const unsigned *elts
,
706 unsigned instance_id
,
707 void *output_buffer
)
709 struct translate_sse
*p
= (struct translate_sse
*)translate
;
711 p
->gen_run_elts( translate
,
718 static void PIPE_CDECL
translate_sse_run( struct translate
*translate
,
721 unsigned instance_id
,
722 void *output_buffer
)
724 struct translate_sse
*p
= (struct translate_sse
*)translate
;
726 p
->gen_run( translate
,
734 struct translate
*translate_sse2_create( const struct translate_key
*key
)
736 struct translate_sse
*p
= NULL
;
739 if (!rtasm_cpu_has_sse() || !rtasm_cpu_has_sse2())
742 p
= CALLOC_STRUCT( translate_sse
);
746 p
->translate
.key
= *key
;
747 p
->translate
.release
= translate_sse_release
;
748 p
->translate
.set_buffer
= translate_sse_set_buffer
;
749 p
->translate
.run_elts
= translate_sse_run_elts
;
750 p
->translate
.run
= translate_sse_run
;
752 for (i
= 0; i
< key
->nr_elements
; i
++) {
753 if (key
->element
[i
].type
== TRANSLATE_ELEMENT_NORMAL
) {
756 p
->nr_buffers
= MAX2(p
->nr_buffers
, key
->element
[i
].input_buffer
+ 1);
758 if (key
->element
[i
].instance_divisor
) {
759 p
->use_instancing
= TRUE
;
763 * Map vertex element to vertex buffer varient.
765 for (j
= 0; j
< p
->nr_buffer_varients
; j
++) {
766 if (p
->buffer_varient
[j
].buffer_index
== key
->element
[i
].input_buffer
&&
767 p
->buffer_varient
[j
].instance_divisor
== key
->element
[i
].instance_divisor
) {
771 if (j
== p
->nr_buffer_varients
) {
772 p
->buffer_varient
[j
].buffer_index
= key
->element
[i
].input_buffer
;
773 p
->buffer_varient
[j
].instance_divisor
= key
->element
[i
].instance_divisor
;
774 p
->nr_buffer_varients
++;
776 p
->element_to_buffer_varient
[i
] = j
;
778 assert(key
->element
[i
].type
== TRANSLATE_ELEMENT_INSTANCE_ID
);
780 p
->element_to_buffer_varient
[i
] = ELEMENT_BUFFER_INSTANCE_ID
;
784 if (0) debug_printf("nr_buffers: %d\n", p
->nr_buffers
);
786 if (!build_vertex_emit(p
, &p
->linear_func
, TRUE
))
789 if (!build_vertex_emit(p
, &p
->elt_func
, FALSE
))
792 p
->gen_run
= (run_func
)x86_get_func(&p
->linear_func
);
793 if (p
->gen_run
== NULL
)
796 p
->gen_run_elts
= (run_elts_func
)x86_get_func(&p
->elt_func
);
797 if (p
->gen_run_elts
== NULL
)
800 return &p
->translate
;
804 translate_sse_release( &p
->translate
);
813 struct translate
*translate_sse2_create( const struct translate_key
*key
)