AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

nv50: add support for address regs

Browse Source 
Allow indirect uniform access and increase the
limit on parameters from 128 to 512.
This commit is contained in:
Christoph Bumiller
2009-10-19 18:17:45 +02:00
parent eb7ea97e7f
commit ec5c23551c
2 changed files with 175 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/gallium/drivers/nv50/nv50_program.c
+170 -8
View File
@@ -32,6 +32,7 @@
#include "nv50_context.h"
#define NV50_SU_MAX_TEMP 64
#define NV50_SU_MAX_ADDR 7
//#define NV50_PROGRAM_DUMP
/* ARL - gallium craps itself on progs/vp/arl.txt
@@ -79,7 +80,8 @@ struct nv50_reg {
P_ATTR,
P_RESULT,
P_CONST,
P_IMMD
P_IMMD,
P_ADDR
} type;
int index;
@@ -99,6 +101,7 @@ struct nv50_pc {
/* hw resources */
struct nv50_reg *r_temp[NV50_SU_MAX_TEMP];
struct nv50_reg r_addr[NV50_SU_MAX_ADDR];
/* tgsi resources */
struct nv50_reg *temp;
@@ -112,6 +115,8 @@ struct nv50_pc {
struct nv50_reg *immd;
float *immd_buf;
int immd_nr;
struct nv50_reg **addr;
int addr_nr;
struct nv50_reg *temp_temp[16];
unsigned temp_temp_nr;
@@ -158,6 +163,17 @@ popcnt4(uint32_t val)
return cnt[val & 0xf];
}
static void
terminate_mbb(struct nv50_pc *pc)
{
int i;
/* remove records of temporary address register values */
for (i = 0; i < NV50_SU_MAX_ADDR; ++i)
if (pc->r_addr[i].index < 0)
pc->r_addr[i].rhw = -1;
}
static void
alloc_reg(struct nv50_pc *pc, struct nv50_reg *reg)
{
@@ -454,9 +470,68 @@ set_immd(struct nv50_pc *pc, struct nv50_reg *imm, struct nv50_program_exec *e)
e->inst[1] |= (val >> 6) << 2;
}
static void
emit_set_addr(struct nv50_pc *pc, struct nv50_reg *dst, unsigned val)
{
struct nv50_program_exec *e = exec(pc);
assert(val <= 0xffff);
e->inst[0] = 0xd0000000 | ((val & 0xffff) << 9);
e->inst[1] = 0x20000000;
e->inst[0] |= dst->hw << 2;
set_long(pc, e);
emit(pc, e);
}
static struct nv50_reg *
alloc_addr(struct nv50_pc *pc, struct nv50_reg *ref)
{
int i;
struct nv50_reg *a = NULL;
if (!ref) {
for (i = 0; i < NV50_SU_MAX_ADDR; ++i) {
if (pc->r_addr[i].index >= 0)
continue;
if (pc->r_addr[i].rhw >= 0 &&
pc->r_addr[i].acc == pc->insn_cur)
continue;
pc->r_addr[i].rhw = -1;
pc->r_addr[i].index = i;
return &pc->r_addr[i];
}
assert(0);
return NULL;
}
for (i = NV50_SU_MAX_ADDR - 1; i >= 0; --i) {
if (pc->r_addr[i].index >= 0) /* occupied for TGSI */
continue;
if (pc->r_addr[i].rhw < 0) { /* unused */
a = &pc->r_addr[i];
continue;
}
if (!a && pc->r_addr[i].acc != pc->insn_cur)
a = &pc->r_addr[i];
if (ref->hw - pc->r_addr[i].rhw < 128) {
/* alloc'd & suitable */
pc->r_addr[i].acc = pc->insn_cur;
return &pc->r_addr[i];
}
}
assert(a);
emit_set_addr(pc, a, ref->hw * 4);
a->rhw = ref->hw % 128;
a->acc = pc->insn_cur;
return a;
}
#define INTERP_LINEAR 0
#define INTERP_FLAT 1
#define INTERP_FLAT 1
#define INTERP_PERSPECTIVE 2
#define INTERP_CENTROID 4
@@ -488,6 +563,16 @@ emit_interp(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *iv,
emit(pc, e);
}
static INLINE void
set_addr(struct nv50_program_exec *e, struct nv50_reg *a)
{
assert(!(e->inst[0] & 0x0c000000));
assert(!(e->inst[1] & 0x00000004));
e->inst[0] |= (a->hw & 3) << 26;
e->inst[1] |= (a->hw >> 2) << 2;
}
static void
set_data(struct nv50_pc *pc, struct nv50_reg *src, unsigned m, unsigned s,
struct nv50_program_exec *e)
@@ -498,6 +583,14 @@ set_data(struct nv50_pc *pc, struct nv50_reg *src, unsigned m, unsigned s,
e->param.shift = s;
e->param.mask = m << (s % 32);
if (src->hw > 127)
set_addr(e, alloc_addr(pc, src));
else
if (src->acc < 0) {
assert(src->type == P_CONST);
set_addr(e, pc->addr[src->index]);
}
e->inst[1] |= (((src->type == P_IMMD) ? 0 : 1) << 22);
}
@@ -632,7 +725,7 @@ set_src_1(struct nv50_pc *pc, struct nv50_reg *src, struct nv50_program_exec *e)
}
alloc_reg(pc, src);
e->inst[0] |= (src->hw << 16);
e->inst[0] |= ((src->hw & 127) << 16);
}
static void
@@ -660,7 +753,7 @@ set_src_2(struct nv50_pc *pc, struct nv50_reg *src, struct nv50_program_exec *e)
}
alloc_reg(pc, src);
e->inst[1] |= (src->hw << 14);
e->inst[1] |= ((src->hw & 127) << 14);
}
static void
@@ -722,6 +815,22 @@ emit_add(struct nv50_pc *pc, struct nv50_reg *dst,
emit(pc, e);
}
static void
emit_arl(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src,
uint8_t s)
{
struct nv50_program_exec *e = exec(pc);
set_long(pc, e);
e->inst[1] |= 0xc0000000;
e->inst[0] |= dst->hw << 2;
e->inst[0] |= s << 16; /* shift left */
set_src_0_restricted(pc, src, e);
emit(pc, e);
}
static void
emit_minmax(struct nv50_pc *pc, unsigned sub, struct nv50_reg *dst,
struct nv50_reg *src0, struct nv50_reg *src1)
@@ -1403,6 +1512,16 @@ tgsi_dst(struct nv50_pc *pc, int c, const struct tgsi_full_dst_register *dst)
return &pc->temp[dst->DstRegister.Index * 4 + c];
case TGSI_FILE_OUTPUT:
return &pc->result[dst->DstRegister.Index * 4 + c];
case TGSI_FILE_ADDRESS:
{
struct nv50_reg *r = pc->addr[dst->DstRegister.Index * 4 + c];
if (!r) {
r = alloc_addr(pc, NULL);
pc->addr[dst->DstRegister.Index * 4 + c] = r;
}
assert(r);
return r;
}
case TGSI_FILE_NULL:
return NULL;
default:
@@ -1418,7 +1537,10 @@ tgsi_src(struct nv50_pc *pc, int chan, const struct tgsi_full_src_register *src,
{
struct nv50_reg *r = NULL;
struct nv50_reg *temp;
unsigned sgn, c;
unsigned sgn, c, swz;
if (src->SrcRegister.File != TGSI_FILE_CONSTANT)
assert(!src->SrcRegister.Indirect);
sgn = tgsi_util_get_full_src_register_sign_mode(src, chan);
@@ -1436,13 +1558,29 @@ tgsi_src(struct nv50_pc *pc, int chan, const struct tgsi_full_src_register *src,
r = &pc->temp[src->SrcRegister.Index * 4 + c];
break;
case TGSI_FILE_CONSTANT:
r = &pc->param[src->SrcRegister.Index * 4 + c];
if (!src->SrcRegister.Indirect) {
r = &pc->param[src->SrcRegister.Index * 4 + c];
break;
}
/* Indicate indirection by setting r->acc < 0 and
* use the index field to select the address reg.
*/
r = MALLOC_STRUCT(nv50_reg);
swz = tgsi_util_get_src_register_swizzle(
&src->SrcRegisterInd, 0);
ctor_reg(r, P_CONST,
src->SrcRegisterInd.Index * 4 + swz, c);
r->acc = -1;
break;
case TGSI_FILE_IMMEDIATE:
r = &pc->immd[src->SrcRegister.Index * 4 + c];
break;
case TGSI_FILE_SAMPLER:
break;
case TGSI_FILE_ADDRESS:
r = pc->addr[src->SrcRegister.Index * 4 + c];
assert(r);
break;
default:
assert(0);
break;
@@ -1678,8 +1816,15 @@ nv50_program_tx_insn(struct nv50_pc *pc,
emit_add(pc, dst[c], src[0][c], src[1][c]);
}
break;
case TGSI_OPCODE_ARL:
assert(src[0][0]);
temp = temp_temp(pc);
emit_cvt(pc, temp, src[0][0], -1, CVTOP_FLOOR, CVT_S32_F32);
emit_arl(pc, dst[0], temp, 4);
break;
case TGSI_OPCODE_BGNLOOP:
pc->loop_pos[pc->loop_lvl++] = pc->p->exec_size;
terminate_mbb(pc);
break;
case TGSI_OPCODE_BRK:
emit_branch(pc, -1, 0, NULL);
@@ -1763,6 +1908,7 @@ nv50_program_tx_insn(struct nv50_pc *pc,
emit_branch(pc, -1, 0, NULL);
pc->if_insn[--pc->if_lvl]->param.index = pc->p->exec_size;
pc->if_insn[pc->if_lvl++] = pc->p->exec_tail;
terminate_mbb(pc);
break;
case TGSI_OPCODE_ENDIF:
pc->if_insn[--pc->if_lvl]->param.index = pc->p->exec_size;
@@ -1775,6 +1921,7 @@ nv50_program_tx_insn(struct nv50_pc *pc,
pc->br_join[pc->if_lvl]->param.index = pc->p->exec_size;
pc->br_join[pc->if_lvl] = NULL;
}
terminate_mbb(pc);
/* emit a NOP as join point, we could set it on the next
* one, but would have to make sure it is long and !immd
*/
@@ -1785,6 +1932,7 @@ nv50_program_tx_insn(struct nv50_pc *pc,
emit_branch(pc, -1, 0, NULL);
pc->p->exec_tail->param.index = pc->loop_pos[--pc->loop_lvl];
pc->br_loop[pc->loop_lvl]->param.index = pc->p->exec_size;
terminate_mbb(pc);
break;
case TGSI_OPCODE_EX2:
emit_preex2(pc, temp, src[0][0]);
@@ -1812,6 +1960,7 @@ nv50_program_tx_insn(struct nv50_pc *pc,
set_pred_wr(pc, 1, 0, pc->if_cond);
emit_branch(pc, 0, 2, &pc->br_join[pc->if_lvl]);
pc->if_insn[pc->if_lvl++] = pc->p->exec_tail;
terminate_mbb(pc);
break;
case TGSI_OPCODE_KIL:
emit_kil(pc, src[0][0]);
@@ -1989,6 +2138,9 @@ nv50_program_tx_insn(struct nv50_pc *pc,
src[i][c]->neg = 0;
if (src[i][c]->index == -1 && src[i][c]->type == P_IMMD)
FREE(src[i][c]);
else
if (src[i][c]->acc < 0 && src[i][c]->type == P_CONST)
FREE(src[i][c]); /* indirect constant */
}
}
@@ -2332,8 +2484,8 @@ nv50_program_tx_prep(struct nv50_pc *pc)
pc->interp_mode[i] = mode;
}
break;
case TGSI_FILE_ADDRESS:
case TGSI_FILE_CONSTANT:
break;
case TGSI_FILE_SAMPLER:
break;
default:
@@ -2527,6 +2679,8 @@ ctor_nv50_pc(struct nv50_pc *pc, struct nv50_program *p)
pc->attr_nr = p->info.file_max[TGSI_FILE_INPUT] + 1;
pc->result_nr = p->info.file_max[TGSI_FILE_OUTPUT] + 1;
pc->param_nr = p->info.file_max[TGSI_FILE_CONSTANT] + 1;
pc->addr_nr = p->info.file_max[TGSI_FILE_ADDRESS] + 1;
assert(pc->addr_nr <= 2);
p->cfg.high_temp = 4;
@@ -2595,6 +2749,14 @@ ctor_nv50_pc(struct nv50_pc *pc, struct nv50_program *p)
ctor_reg(&pc->param[rid], P_CONST, i, rid);
}
if (pc->addr_nr) {
pc->addr = CALLOC(pc->addr_nr * 4, sizeof(struct nv50_reg *));
if (!pc->addr)
return FALSE;
}
for (i = 0; i < NV50_SU_MAX_ADDR; ++i)
ctor_reg(&pc->r_addr[i], P_ADDR, -1, i + 1);
return TRUE;
}
@@ -2774,7 +2936,7 @@ nv50_program_validate_data(struct nv50_context *nv50, struct nv50_program *p)
p->immd_nr, NV50_CB_PMISC);
}
assert(p->param_nr <= 128);
assert(p->param_nr <= 512);
if (p->param_nr) {
unsigned cb;
src/gallium/drivers/nv50/nv50_screen.c
+5 -5
View File
@@ -301,7 +301,7 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
so_data (so, 8);
/* constant buffers for immediates and VP/FP parameters */
ret = nouveau_bo_new(dev, NOUVEAU_BO_VRAM, 0, 128*4*4,
ret = nouveau_bo_new(dev, NOUVEAU_BO_VRAM, 0, (32 * 4) * 4,
&screen->constbuf_misc[0]);
if (ret) {
nv50_screen_destroy(pscreen);
@@ -309,7 +309,7 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
}
for (i = 0; i < 2; i++) {
ret = nouveau_bo_new(dev, NOUVEAU_BO_VRAM, 0, 128*4*4,
ret = nouveau_bo_new(dev, NOUVEAU_BO_VRAM, 0, (128 * 4) * 4,
&screen->constbuf_parm[i]);
if (ret) {
nv50_screen_destroy(pscreen);
@@ -318,8 +318,8 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
}
if (nouveau_resource_init(&screen->immd_heap[0], 0, 128) ||
nouveau_resource_init(&screen->parm_heap[0], 0, 128) ||
nouveau_resource_init(&screen->parm_heap[1], 0, 128))
nouveau_resource_init(&screen->parm_heap[0], 0, 512) ||
nouveau_resource_init(&screen->parm_heap[1], 0, 512))
{
NOUVEAU_ERR("Error initialising constant buffers.\n");
nv50_screen_destroy(pscreen);
@@ -340,7 +340,7 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
so_reloc (so, screen->constbuf_misc[0], 0, NOUVEAU_BO_VRAM |
NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
so_data (so, (NV50_CB_PMISC << 16) | 0x00000800);
so_data (so, (NV50_CB_PMISC << 16) | 0x00000200);
so_method(so, screen->tesla, NV50TCL_SET_PROGRAM_CB, 1);
so_data (so, 0x00000001 | (NV50_CB_PMISC << 12));
so_method(so, screen->tesla, NV50TCL_SET_PROGRAM_CB, 1);