#ifndef DRAW_VS_AOS_H
#define DRAW_VS_AOS_H
+#include "pipe/p_config.h"
+#include "tgsi/tgsi_exec.h"
+#include "draw_vs.h"
+
+#ifdef PIPE_ARCH_X86
struct tgsi_token;
struct x86_function;
#define W 3
#define MAX_INPUTS PIPE_MAX_ATTRIBS
-#define MAX_OUTPUTS PIPE_MAX_ATTRIBS
-#define MAX_TEMPS PIPE_MAX_ATTRIBS /* say */
-#define MAX_CONSTANTS PIPE_MAX_ATTRIBS /* say */
-#define MAX_IMMEDIATES PIPE_MAX_ATTRIBS /* say */
-#define MAX_INTERNALS 4
+#define MAX_OUTPUTS PIPE_MAX_SHADER_OUTPUTS
+#define MAX_TEMPS TGSI_EXEC_NUM_TEMPS
+#define MAX_CONSTANTS 1024 /** only used for sanity checking */
+#define MAX_IMMEDIATES 1024 /** only used for sanity checking */
+#define MAX_INTERNALS 8 /** see IMM_x values below */
#define AOS_FILE_INTERNAL TGSI_FILE_COUNT
-/* This is the temporary storage used by all the aos_sse vs varients.
+#define FPU_RND_NEG 1
+#define FPU_RND_NEAREST 2
+
+struct aos_machine;
+typedef void (PIPE_CDECL *lit_func)( struct aos_machine *,
+ float *result,
+ const float *in,
+ unsigned count );
+
+void PIPE_CDECL aos_do_lit( struct aos_machine *machine,
+ float *result,
+ const float *in,
+ unsigned count );
+
+struct shine_tab {
+ float exponent;
+ float values[258];
+ unsigned last_used;
+};
+
+struct lit_info {
+ lit_func func;
+ struct shine_tab *shine_tab;
+};
+
+#define MAX_SHINE_TAB 4
+#define MAX_LIT_INFO 16
+
+struct aos_buffer {
+ const void *base_ptr;
+ unsigned stride;
+ void *ptr; /* updated per vertex */
+};
+
+
+
+
+/* This is the temporary storage used by all the aos_sse vs variants.
* Create one per context and reuse by passing a pointer in at
- * vs_varient creation??
+ * vs_variant creation??
*/
struct aos_machine {
float input [MAX_INPUTS ][4];
float output [MAX_OUTPUTS ][4];
float temp [MAX_TEMPS ][4];
- float constant [MAX_CONSTANTS ][4]; /* fixme -- should just be a pointer */
- float immediate[MAX_IMMEDIATES][4]; /* fixme -- should just be a pointer */
float internal [MAX_INTERNALS ][4];
float scale[4]; /* viewport */
float translate[4]; /* viewport */
- ushort fpu_round_nearest;
- ushort fpu_round_neg_inf;
+ float tmp[2][4]; /* scratch space for LIT */
+
+ struct shine_tab shine_tab[MAX_SHINE_TAB];
+ struct lit_info lit_info[MAX_LIT_INFO];
+ unsigned now;
+
+
+ ushort fpu_rnd_nearest;
+ ushort fpu_rnd_neg_inf;
ushort fpu_restore;
ushort fpucntl; /* one of FPU_* above */
- struct {
- const void *input_ptr;
- unsigned input_stride;
+ const float (*immediates)[4]; /* points to shader data */
+ const void *constants[PIPE_MAX_CONSTANT_BUFFERS]; /* points to draw data */
- unsigned output_offset;
- } attrib[PIPE_MAX_ATTRIBS];
+ const struct aos_buffer *buffer; /* points to ? */
};
struct aos_compilation {
struct x86_function *func;
- struct draw_vs_varient_aos_sse *vaos;
+ struct draw_vs_variant_aos_sse *vaos;
unsigned insn_counter;
unsigned num_immediates;
+ unsigned count;
+ unsigned lit_count;
struct {
unsigned idx:16;
unsigned last_used;
} xmm[8];
+ unsigned x86_reg[2]; /* one of X86_* */
boolean input_fetched[PIPE_MAX_ATTRIBS];
unsigned output_last_write[PIPE_MAX_ATTRIBS];
struct x86_reg outbuf_ECX;
struct x86_reg machine_EDX;
struct x86_reg count_ESI; /* decrements to zero */
+ struct x86_reg temp_EBP;
+ struct x86_reg stack_ESP;
};
struct x86_reg aos_get_xmm_reg( struct aos_compilation *cp );
unsigned idx,
unsigned dirty );
+void aos_spill_all( struct aos_compilation *cp );
+
struct x86_reg aos_get_shader_reg( struct aos_compilation *cp,
unsigned file,
unsigned idx );
-boolean aos_fetch_inputs( struct aos_compilation *cp,
- boolean linear );
+boolean aos_init_inputs( struct aos_compilation *cp, boolean linear );
+boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear );
+boolean aos_incr_inputs( struct aos_compilation *cp, boolean linear );
boolean aos_emit_outputs( struct aos_compilation *cp );
#define IMM_ONES 0 /* 1, 1,1,1 */
-#define IMM_NEGS 1 /* 1,-1,0,0 */
+#define IMM_SWZ 1 /* 1,-1,0, 0xffffffff */
#define IMM_IDENTITY 2 /* 0, 0,0,1 */
#define IMM_INV_255 3 /* 1/255, 1/255, 1/255, 1/255 */
#define IMM_255 4 /* 255, 255, 255, 255 */
+#define IMM_NEGS 5 /* -1,-1,-1,-1 */
+#define IMM_RSQ 6 /* -.5,1.5,_,_ */
+#define IMM_PSIZE 7 /* not really an immediate - updated each run */
struct x86_reg aos_get_internal( struct aos_compilation *cp,
unsigned imm );
+struct x86_reg aos_get_internal_xmm( struct aos_compilation *cp,
+ unsigned imm );
-#define ERROR(cp, msg) \
+#define AOS_ERROR(cp, msg) \
do { \
- debug_printf("%s: x86 translation failed: %s\n", __FUNCTION__, msg); \
+ if (0) debug_printf("%s: x86 translation failed: %s\n", __FUNCTION__, msg); \
cp->error = 1; \
- assert(0); \
} while (0)
+#define X86_NULL 0
+#define X86_IMMEDIATES 1
+#define X86_CONSTANTS 2
+#define X86_BUFFERS 3
+struct x86_reg aos_get_x86( struct aos_compilation *cp,
+ unsigned which_reg,
+ unsigned value );
+typedef void (PIPE_CDECL *vaos_run_elts_func)( struct aos_machine *,
+ const unsigned *elts,
+ unsigned count,
+ void *output_buffer);
-struct draw_vs_varient_aos_sse {
- struct draw_vs_varient base;
- struct draw_context *draw;
+typedef void (PIPE_CDECL *vaos_run_linear_func)( struct aos_machine *,
+ unsigned start,
+ unsigned count,
+ void *output_buffer);
-#if 0
- struct {
- const void *ptr;
- unsigned stride;
- } attrib[PIPE_MAX_ATTRIBS];
-#endif
- struct aos_machine *machine; /* XXX: temporarily unshared */
+struct draw_vs_variant_aos_sse {
+ struct draw_vs_variant base;
+ struct draw_context *draw;
+
+ struct aos_buffer *buffer;
+ unsigned nr_vb;
- vsv_run_linear_func gen_run_linear;
- vsv_run_elts_func gen_run_elts;
+ vaos_run_linear_func gen_run_linear;
+ vaos_run_elts_func gen_run_elts;
struct x86_function func[2];
};
+#endif
#endif
projects / mesa.git / blobdiff