int rhw; /* result hw for FP outputs, or interpolant index */
int acc; /* instruction where this reg is last read (first insn == 1) */
+
+ int vtx; /* vertex index, for GP inputs (TGSI Dimension.Index) */
+ int indirect[2]; /* index into pc->addr, or -1 */
};
#define NV50_MOD_NEG 1
#define NV50_MOD_ABS 2
+#define NV50_MOD_NEG_ABS (NV50_MOD_NEG | NV50_MOD_ABS)
#define NV50_MOD_SAT 4
+#define NV50_MOD_I32 8
+
+/* NV50_MOD_I32 is used to indicate integer mode for neg/abs */
/* STACK: Conditionals and loops have to use the (per warp) stack.
* Stack entries consist of an entry type (divergent path, join at),
int immd_nr;
struct nv50_reg **addr;
int addr_nr;
- uint8_t addr_alloc; /* set bit indicates used for TGSI_FILE_ADDRESS */
struct nv50_reg *temp_temp[16];
+ struct nv50_program_exec *temp_temp_exec[16];
unsigned temp_temp_nr;
/* broadcast and destination replacement regs */
uint8_t edgeflag_out;
};
+static struct nv50_reg *get_address_reg(struct nv50_pc *, struct nv50_reg *);
+
static INLINE void
ctor_reg(struct nv50_reg *reg, unsigned type, int index, int hw)
{
reg->hw = hw;
reg->mod = 0;
reg->rhw = -1;
+ reg->vtx = -1;
reg->acc = 0;
+ reg->indirect[0] = reg->indirect[1] = -1;
}
static INLINE unsigned
/* remove records of temporary address register values */
for (i = 0; i < NV50_SU_MAX_ADDR; ++i)
- pc->r_addr[i].rhw = -1;
+ if (pc->r_addr[i].index < 0)
+ pc->r_addr[i].acc = 0;
}
static void
}
}
- assert(0);
+ NOUVEAU_ERR("out of registers\n");
+ abort();
}
static INLINE struct nv50_reg *
if (reg) {
alloc_reg(pc, reg);
*ri = *reg;
+ reg->indirect[0] = reg->indirect[1] = -1;
reg->mod = 0;
}
return ri;
}
}
- assert(0);
+ NOUVEAU_ERR("out of registers\n");
+ abort();
return NULL;
}
}
static struct nv50_reg *
-temp_temp(struct nv50_pc *pc)
+temp_temp(struct nv50_pc *pc, struct nv50_program_exec *e)
{
if (pc->temp_temp_nr >= 16)
assert(0);
pc->temp_temp[pc->temp_temp_nr] = alloc_temp(pc, NULL);
+ pc->temp_temp_exec[pc->temp_temp_nr] = e;
return pc->temp_temp[pc->temp_temp_nr++];
}
+/* This *must* be called for all nv50_program_exec that have been
+ * given as argument to temp_temp, or the temps will be leaked !
+ */
static void
-kill_temp_temp(struct nv50_pc *pc)
+kill_temp_temp(struct nv50_pc *pc, struct nv50_program_exec *e)
{
int i;
for (i = 0; i < pc->temp_temp_nr; i++)
- free_temp(pc, pc->temp_temp[i]);
- pc->temp_temp_nr = 0;
+ if (pc->temp_temp_exec[i] == e)
+ free_temp(pc, pc->temp_temp[i]);
+ if (!e)
+ pc->temp_temp_nr = 0;
}
static int
p->exec_head = e;
p->exec_tail = e;
p->exec_size += (e->inst[0] & 1) ? 2 : 1;
+
+ kill_temp_temp(pc, e);
}
static INLINE void set_long(struct nv50_pc *, struct nv50_program_exec *);
static INLINE void
set_addr(struct nv50_program_exec *e, struct nv50_reg *a)
{
+ assert(a->type == P_ADDR);
+
assert(!(e->inst[0] & 0x0c000000));
assert(!(e->inst[1] & 0x00000004));
e->inst[0] |= (a->hw & 3) << 26;
- e->inst[1] |= (a->hw >> 2) << 2;
+ e->inst[1] |= a->hw & 4;
+}
+
+static void
+emit_arl(struct nv50_pc *, struct nv50_reg *, struct nv50_reg *, uint8_t);
+
+static void
+emit_shl_imm(struct nv50_pc *, struct nv50_reg *, struct nv50_reg *, int);
+
+static void
+emit_mov_from_addr(struct nv50_pc *pc, struct nv50_reg *dst,
+ struct nv50_reg *src)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[1] = 0x40000000;
+ set_long(pc, e);
+ set_dst(pc, dst, e);
+ set_addr(e, src);
+
+ emit(pc, e);
}
static void
emit(pc, e);
}
-static struct nv50_reg *
-alloc_addr(struct nv50_pc *pc, struct nv50_reg *ref)
-{
- struct nv50_reg *a_tgsi = NULL, *a = NULL;
- int i;
- uint8_t avail = ~pc->addr_alloc;
-
- if (!ref) {
- /* allocate for TGSI_FILE_ADDRESS */
- while (avail) {
- i = ffs(avail) - 1;
-
- if (pc->r_addr[i].rhw < 0 ||
- pc->r_addr[i].acc != pc->insn_cur) {
- pc->addr_alloc |= (1 << i);
-
- pc->r_addr[i].rhw = -1;
- pc->r_addr[i].index = i;
- return &pc->r_addr[i];
- }
- avail &= ~(1 << i);
- }
- assert(0);
- return NULL;
- }
-
- /* Allocate and set an address reg so we can access 'ref'.
- *
- * If and r_addr->index will be -1 or the hw index the value
- * value in rhw is relative to. If rhw < 0, the reg has not
- * been initialized or is in use for TGSI_FILE_ADDRESS.
- */
- while (avail) { /* only consider regs that are not TGSI */
- i = ffs(avail) - 1;
- avail &= ~(1 << i);
-
- if ((!a || a->rhw >= 0) && pc->r_addr[i].rhw < 0) {
- /* prefer an usused reg with low hw index */
- 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)
- continue;
-
- if ((ref->acc >= 0 && pc->r_addr[i].index < 0) ||
- (ref->acc < 0 && pc->r_addr[i].index == ref->index)) {
- pc->r_addr[i].acc = pc->insn_cur;
- return &pc->r_addr[i];
- }
- }
- assert(a);
-
- if (ref->acc < 0)
- a_tgsi = pc->addr[ref->index];
-
- emit_add_addr_imm(pc, a, a_tgsi, (ref->hw & ~0x7f) * 4);
-
- a->rhw = ref->hw & ~0x7f;
- a->acc = pc->insn_cur;
- a->index = a_tgsi ? ref->index : -1;
- return a;
-}
-
#define INTERP_LINEAR 0
#define INTERP_FLAT 1
#define INTERP_PERSPECTIVE 2
e->param.shift = s;
e->param.mask = m << (s % 32);
- if (src->hw > 127)
- set_addr(e, alloc_addr(pc, src));
+ if (src->hw < 0 || src->hw > 127) /* need (additional) address reg */
+ set_addr(e, get_address_reg(pc, src));
else
if (src->acc < 0) {
assert(src->type == P_CONST);
- set_addr(e, pc->addr[src->index]);
+ set_addr(e, pc->addr[src->indirect[0]]);
}
e->inst[1] |= (((src->type == P_IMMD) ? 0 : 1) << 22);
if (src->type == P_ATTR) {
set_long(pc, e);
e->inst[1] |= 0x00200000;
+
+ if (src->vtx >= 0) {
+ /* indirect (vertex base + c) load from p[] */
+ e->inst[0] |= 0x01800000;
+ set_addr(e, get_address_reg(pc, src));
+ }
}
alloc_reg(pc, src);
struct nv50_reg *temp;
if (src->type != P_TEMP) {
- temp = temp_temp(pc);
+ temp = temp_temp(pc, e);
emit_mov(pc, temp, src);
src = temp;
}
if (src->type == P_ATTR) {
set_long(pc, e);
e->inst[1] |= 0x00200000;
+
+ if (src->vtx >= 0) {
+ e->inst[0] |= 0x01800000; /* src from p[] */
+ set_addr(e, get_address_reg(pc, src));
+ }
} else
if (src->type == P_CONST || src->type == P_IMMD) {
- struct nv50_reg *temp = temp_temp(pc);
+ struct nv50_reg *temp = temp_temp(pc, e);
emit_mov(pc, temp, src);
src = temp;
set_src_1(struct nv50_pc *pc, struct nv50_reg *src, struct nv50_program_exec *e)
{
if (src->type == P_ATTR) {
- struct nv50_reg *temp = temp_temp(pc);
+ struct nv50_reg *temp = temp_temp(pc, e);
emit_mov(pc, temp, src);
src = temp;
} else
if (src->type == P_CONST || src->type == P_IMMD) {
- assert(!(e->inst[0] & 0x00800000));
- if (e->inst[0] & 0x01000000) {
- struct nv50_reg *temp = temp_temp(pc);
+ if (e->inst[0] & 0x01800000) {
+ struct nv50_reg *temp = temp_temp(pc, e);
emit_mov(pc, temp, src);
src = temp;
} else {
+ assert(!(e->inst[0] & 0x00800000));
set_data(pc, src, 0x7f, 16, e);
e->inst[0] |= 0x00800000;
}
set_long(pc, e);
if (src->type == P_ATTR) {
- struct nv50_reg *temp = temp_temp(pc);
+ struct nv50_reg *temp = temp_temp(pc, e);
emit_mov(pc, temp, src);
src = temp;
} else
if (src->type == P_CONST || src->type == P_IMMD) {
- assert(!(e->inst[0] & 0x01000000));
- if (e->inst[0] & 0x00800000) {
- struct nv50_reg *temp = temp_temp(pc);
+ if (e->inst[0] & 0x01800000) {
+ struct nv50_reg *temp = temp_temp(pc, e);
emit_mov(pc, temp, src);
src = temp;
} else {
+ assert(!(e->inst[0] & 0x01000000));
set_data(pc, src, 0x7f, 32+14, e);
e->inst[0] |= 0x01000000;
}
e->inst[1] |= ((src->hw & 127) << 14);
}
+static void
+set_half_src(struct nv50_pc *pc, struct nv50_reg *src, int lh,
+ struct nv50_program_exec *e, int pos)
+{
+ struct nv50_reg *r = src;
+
+ alloc_reg(pc, r);
+ if (r->type != P_TEMP) {
+ r = temp_temp(pc, e);
+ emit_mov(pc, r, src);
+ }
+
+ if (r->hw > (NV50_SU_MAX_TEMP / 2)) {
+ NOUVEAU_ERR("out of low GPRs\n");
+ abort();
+ }
+
+ e->inst[pos / 32] |= ((src->hw * 2) + lh) << (pos % 32);
+}
+
static void
emit_mov_from_pred(struct nv50_pc *pc, struct nv50_reg *dst, int pred)
{
e->inst[0] |= dst->hw << 2;
e->inst[0] |= s << 16; /* shift left */
- set_src_0_restricted(pc, src, e);
+ set_src_0(pc, src, e);
emit(pc, e);
}
+static boolean
+address_reg_suitable(struct nv50_reg *a, struct nv50_reg *r)
+{
+ if (!r)
+ return FALSE;
+
+ if (r->vtx != a->vtx)
+ return FALSE;
+ if (r->vtx >= 0)
+ return (r->indirect[1] == a->indirect[1]);
+
+ if (r->hw < a->rhw || (r->hw - a->rhw) >= 128)
+ return FALSE;
+
+ if (a->index >= 0)
+ return (a->index == r->indirect[0]);
+ return (a->indirect[0] == r->indirect[0]);
+}
+
+static void
+load_vertex_base(struct nv50_pc *pc, struct nv50_reg *dst,
+ struct nv50_reg *a, int shift)
+{
+ struct nv50_reg mem, *temp;
+
+ ctor_reg(&mem, P_ATTR, -1, dst->vtx);
+
+ assert(dst->type == P_ADDR);
+ if (!a) {
+ emit_arl(pc, dst, &mem, 0);
+ return;
+ }
+ temp = alloc_temp(pc, NULL);
+
+ if (shift) {
+ emit_mov_from_addr(pc, temp, a);
+ if (shift < 0)
+ emit_shl_imm(pc, temp, temp, shift);
+ emit_arl(pc, dst, temp, MAX2(shift, 0));
+ }
+ emit_mov(pc, temp, &mem);
+ set_addr(pc->p->exec_tail, dst);
+
+ emit_arl(pc, dst, temp, 0);
+ free_temp(pc, temp);
+}
+
+/* case (ref == NULL): allocate address register for TGSI_FILE_ADDRESS
+ * case (vtx >= 0, acc >= 0): load vertex base from a[vtx * 4] to $aX
+ * case (vtx >= 0, acc < 0): load vertex base from s[$aY + vtx * 4] to $aX
+ * case (vtx < 0, acc >= 0): memory address too high to encode
+ * case (vtx < 0, acc < 0): get source register for TGSI_FILE_ADDRESS
+ */
+static struct nv50_reg *
+get_address_reg(struct nv50_pc *pc, struct nv50_reg *ref)
+{
+ int i;
+ struct nv50_reg *a_ref, *a = NULL;
+
+ for (i = 0; i < NV50_SU_MAX_ADDR; ++i) {
+ if (pc->r_addr[i].acc == 0)
+ a = &pc->r_addr[i]; /* an unused address reg */
+ else
+ if (address_reg_suitable(&pc->r_addr[i], ref)) {
+ pc->r_addr[i].acc = pc->insn_cur;
+ return &pc->r_addr[i];
+ } else
+ if (!a && pc->r_addr[i].index < 0 &&
+ pc->r_addr[i].acc < pc->insn_cur)
+ a = &pc->r_addr[i];
+ }
+ if (!a) {
+ /* We'll be able to spill address regs when this
+ * mess is replaced with a proper compiler ...
+ */
+ NOUVEAU_ERR("out of address regs\n");
+ abort();
+ return NULL;
+ }
+
+ /* initialize and reserve for this TGSI instruction */
+ a->rhw = 0;
+ a->index = a->indirect[0] = a->indirect[1] = -1;
+ a->acc = pc->insn_cur;
+
+ if (!ref) {
+ a->vtx = -1;
+ return a;
+ }
+ a->vtx = ref->vtx;
+
+ /* now put in the correct value ... */
+
+ if (ref->vtx >= 0) {
+ a->indirect[1] = ref->indirect[1];
+
+ /* For an indirect vertex index, we need to shift address right
+ * by 2, the address register will contain vtx * 16, we need to
+ * load from a[vtx * 4].
+ */
+ load_vertex_base(pc, a, (ref->acc < 0) ?
+ pc->addr[ref->indirect[1]] : NULL, -2);
+ } else {
+ assert(ref->acc < 0 || ref->indirect[0] < 0);
+
+ a->rhw = ref->hw & ~0x7f;
+ a->indirect[0] = ref->indirect[0];
+ a_ref = (ref->acc < 0) ? pc->addr[ref->indirect[0]] : NULL;
+
+ emit_add_addr_imm(pc, a, a_ref, a->rhw * 4);
+ }
+ return a;
+}
+
+#define NV50_MAX_F32 0x880
+#define NV50_MAX_S32 0x08c
+#define NV50_MAX_U32 0x084
+#define NV50_MIN_F32 0x8a0
+#define NV50_MIN_S32 0x0ac
+#define NV50_MIN_U32 0x0a4
+
static void
emit_minmax(struct nv50_pc *pc, unsigned sub, struct nv50_reg *dst,
struct nv50_reg *src0, struct nv50_reg *src1)
struct nv50_program_exec *e = exec(pc);
set_long(pc, e);
- e->inst[0] |= 0xb0000000;
- e->inst[1] |= (sub << 29);
+ e->inst[0] |= 0x30000000 | ((sub & 0x800) << 20);
+ e->inst[1] |= (sub << 24);
check_swap_src_0_1(pc, &src0, &src1);
set_dst(pc, dst, e);
emit(pc, e);
}
+static void
+emit_not(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0xd0000000;
+ e->inst[1] = 0x0402c000;
+ set_long(pc, e);
+ set_dst(pc, dst, e);
+ set_src_1(pc, src, e);
+
+ emit(pc, e);
+}
+
+static void
+emit_shift(struct nv50_pc *pc, struct nv50_reg *dst,
+ struct nv50_reg *src0, struct nv50_reg *src1, unsigned dir)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0x30000000;
+ e->inst[1] = 0xc4000000;
+
+ set_long(pc, e);
+ set_dst(pc, dst, e);
+ set_src_0(pc, src0, e);
+
+ if (src1->type == P_IMMD) {
+ e->inst[1] |= (1 << 20);
+ e->inst[0] |= (pc->immd_buf[src1->hw] & 0x7f) << 16;
+ } else
+ set_src_1(pc, src1, e);
+
+ if (dir != TGSI_OPCODE_SHL)
+ e->inst[1] |= (1 << 29);
+
+ if (dir == TGSI_OPCODE_ISHR)
+ e->inst[1] |= (1 << 27);
+
+ emit(pc, e);
+}
+
+static void
+emit_shl_imm(struct nv50_pc *pc, struct nv50_reg *dst,
+ struct nv50_reg *src, int s)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0x30000000;
+ e->inst[1] = 0xc4100000;
+ if (s < 0) {
+ e->inst[1] |= 1 << 29;
+ s = -s;
+ }
+ e->inst[1] |= ((s & 0x7f) << 16);
+
+ set_long(pc, e);
+ set_dst(pc, dst, e);
+ set_src_0(pc, src, e);
+
+ emit(pc, e);
+}
+
static void
emit_mad(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src0,
struct nv50_reg *src1, struct nv50_reg *src2)
emit(pc, e);
}
-#define CVTOP_RN 0x01
-#define CVTOP_FLOOR 0x03
-#define CVTOP_CEIL 0x05
-#define CVTOP_TRUNC 0x07
-#define CVTOP_SAT 0x08
-#define CVTOP_ABS 0x10
-
-/* 0x04 == 32 bit dst */
-/* 0x40 == dst is float */
-/* 0x80 == src is float */
-#define CVT_F32_F32 0xc4
-#define CVT_F32_S32 0x44
-#define CVT_S32_F32 0x8c
-#define CVT_S32_S32 0x0c
-#define CVT_NEG 0x20
-#define CVT_RI 0x08
+#define CVT_RN (0x00 << 16)
+#define CVT_FLOOR (0x02 << 16)
+#define CVT_CEIL (0x04 << 16)
+#define CVT_TRUNC (0x06 << 16)
+#define CVT_SAT (0x08 << 16)
+#define CVT_ABS (0x10 << 16)
+
+#define CVT_X32_X32 0x04004000
+#define CVT_X32_S32 0x04014000
+#define CVT_F32_F32 ((0xc0 << 24) | CVT_X32_X32)
+#define CVT_S32_F32 ((0x88 << 24) | CVT_X32_X32)
+#define CVT_U32_F32 ((0x80 << 24) | CVT_X32_X32)
+#define CVT_F32_S32 ((0x40 << 24) | CVT_X32_S32)
+#define CVT_F32_U32 ((0x40 << 24) | CVT_X32_X32)
+#define CVT_S32_S32 ((0x08 << 24) | CVT_X32_S32)
+#define CVT_S32_U32 ((0x08 << 24) | CVT_X32_X32)
+#define CVT_U32_S32 ((0x00 << 24) | CVT_X32_S32)
+
+#define CVT_NEG 0x20000000
+#define CVT_RI 0x08000000
static void
emit_cvt(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src,
- int wp, unsigned cvn, unsigned fmt)
+ int wp, uint32_t cvn)
{
struct nv50_program_exec *e;
e = exec(pc);
- set_long(pc, e);
- e->inst[0] |= 0xa0000000;
- e->inst[1] |= 0x00004000; /* 32 bit src */
- e->inst[1] |= (cvn << 16);
- e->inst[1] |= (fmt << 24);
+ if (src->mod & NV50_MOD_NEG) cvn |= CVT_NEG;
+ if (src->mod & NV50_MOD_ABS) cvn |= CVT_ABS;
+
+ e->inst[0] = 0xa0000000;
+ e->inst[1] = cvn;
+ set_long(pc, e);
set_src_0(pc, src, e);
if (wp >= 0)
* 0x6 = GE
* 0x7 = set condition code ? (used before bra.lt/le/gt/ge)
* 0x8 = unordered bit (allows NaN)
+ *
+ * mode = 0x04 (u32), 0x0c (s32), 0x80 (f32)
*/
static void
emit_set(struct nv50_pc *pc, unsigned ccode, struct nv50_reg *dst, int wp,
- struct nv50_reg *src0, struct nv50_reg *src1)
+ struct nv50_reg *src0, struct nv50_reg *src1, uint8_t mode)
{
static const unsigned cc_swapped[8] = { 0, 4, 2, 6, 1, 5, 3, 7 };
if (dst && dst->type != P_TEMP)
dst = alloc_temp(pc, NULL);
- /* set.u32 */
set_long(pc, e);
- e->inst[0] |= 0xb0000000;
+ e->inst[0] |= 0x30000000 | (mode << 24);
e->inst[1] |= 0x60000000 | (ccode << 14);
- /* XXX: decuda will disasm as .u16 and use .lo/.hi regs, but
- * that doesn't seem to match what the hw actually does
- e->inst[1] |= 0x04000000; << breaks things, u32 by default ?
- */
-
if (wp >= 0)
set_pred_wr(pc, 1, wp, e);
if (dst)
emit(pc, e);
- /* cvt.f32.u32/s32 (?) if we didn't only write the predicate */
- if (rdst)
- emit_cvt(pc, rdst, dst, -1, CVTOP_ABS | CVTOP_RN, CVT_F32_S32);
+ if (rdst && mode == 0x80) /* convert to float ? */
+ emit_cvt(pc, rdst, dst, -1, CVT_ABS | CVT_F32_S32);
if (rdst && rdst != dst)
free_temp(pc, dst);
}
-static INLINE unsigned
-map_tgsi_setop_cc(unsigned op)
+static INLINE void
+map_tgsi_setop_hw(unsigned op, uint8_t *cc, uint8_t *ty)
{
switch (op) {
- case TGSI_OPCODE_SLT: return 0x1;
- case TGSI_OPCODE_SGE: return 0x6;
- case TGSI_OPCODE_SEQ: return 0x2;
- case TGSI_OPCODE_SGT: return 0x4;
- case TGSI_OPCODE_SLE: return 0x3;
- case TGSI_OPCODE_SNE: return 0xd;
+ case TGSI_OPCODE_SLT: *cc = 0x1; *ty = 0x80; break;
+ case TGSI_OPCODE_SGE: *cc = 0x6; *ty = 0x80; break;
+ case TGSI_OPCODE_SEQ: *cc = 0x2; *ty = 0x80; break;
+ case TGSI_OPCODE_SGT: *cc = 0x4; *ty = 0x80; break;
+ case TGSI_OPCODE_SLE: *cc = 0x3; *ty = 0x80; break;
+ case TGSI_OPCODE_SNE: *cc = 0xd; *ty = 0x80; break;
+
+ case TGSI_OPCODE_ISLT: *cc = 0x1; *ty = 0x0c; break;
+ case TGSI_OPCODE_ISGE: *cc = 0x6; *ty = 0x0c; break;
+ case TGSI_OPCODE_USEQ: *cc = 0x2; *ty = 0x04; break;
+ case TGSI_OPCODE_USGE: *cc = 0x6; *ty = 0x04; break;
+ case TGSI_OPCODE_USLT: *cc = 0x1; *ty = 0x04; break;
+ case TGSI_OPCODE_USNE: *cc = 0x5; *ty = 0x04; break;
default:
assert(0);
- return 0;
+ return;
}
}
+static void
+emit_add_b32(struct nv50_pc *pc, struct nv50_reg *dst,
+ struct nv50_reg *src0, struct nv50_reg *rsrc1)
+{
+ struct nv50_program_exec *e = exec(pc);
+ struct nv50_reg *src1;
+
+ e->inst[0] = 0x20000000;
+
+ alloc_reg(pc, rsrc1);
+ check_swap_src_0_1(pc, &src0, &rsrc1);
+
+ src1 = rsrc1;
+ if (src0->mod & rsrc1->mod & NV50_MOD_NEG) {
+ src1 = temp_temp(pc, e);
+ emit_cvt(pc, src1, rsrc1, -1, CVT_S32_S32);
+ }
+
+ if (!pc->allow32 || src1->hw > 63 ||
+ (src1->type != P_TEMP && src1->type != P_IMMD))
+ set_long(pc, e);
+
+ set_dst(pc, dst, e);
+ set_src_0(pc, src0, e);
+
+ if (is_long(e)) {
+ e->inst[1] |= 1 << 26;
+ set_src_2(pc, src1, e);
+ } else {
+ e->inst[0] |= 0x8000;
+ if (src1->type == P_IMMD)
+ set_immd(pc, src1, e);
+ else
+ set_src_1(pc, src1, e);
+ }
+
+ if (src0->mod & NV50_MOD_NEG)
+ e->inst[0] |= 1 << 28;
+ else
+ if (src1->mod & NV50_MOD_NEG)
+ e->inst[0] |= 1 << 22;
+
+ emit(pc, e);
+}
+
+static void
+emit_mad_u16(struct nv50_pc *pc, struct nv50_reg *dst,
+ struct nv50_reg *src0, int lh_0, struct nv50_reg *src1, int lh_1,
+ struct nv50_reg *src2)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0x60000000;
+ if (!pc->allow32)
+ set_long(pc, e);
+ set_dst(pc, dst, e);
+
+ set_half_src(pc, src0, lh_0, e, 9);
+ set_half_src(pc, src1, lh_1, e, 16);
+ alloc_reg(pc, src2);
+ if (is_long(e) || (src2->type != P_TEMP) || (src2->hw != dst->hw))
+ set_src_2(pc, src2, e);
+
+ emit(pc, e);
+}
+
+static void
+emit_mul_u16(struct nv50_pc *pc, struct nv50_reg *dst,
+ struct nv50_reg *src0, int lh_0, struct nv50_reg *src1, int lh_1)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0x40000000;
+ set_long(pc, e);
+ set_dst(pc, dst, e);
+
+ set_half_src(pc, src0, lh_0, e, 9);
+ set_half_src(pc, src1, lh_1, e, 16);
+
+ emit(pc, e);
+}
+
+static void
+emit_sad(struct nv50_pc *pc, struct nv50_reg *dst,
+ struct nv50_reg *src0, struct nv50_reg *src1, struct nv50_reg *src2)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0x50000000;
+ if (!pc->allow32)
+ set_long(pc, e);
+ check_swap_src_0_1(pc, &src0, &src1);
+ set_dst(pc, dst, e);
+ set_src_0(pc, src0, e);
+ set_src_1(pc, src1, e);
+ alloc_reg(pc, src2);
+ if (is_long(e) || (src2->type != dst->type) || (src2->hw != dst->hw))
+ set_src_2(pc, src2, e);
+
+ if (is_long(e))
+ e->inst[1] |= 0x0c << 24;
+ else
+ e->inst[0] |= 0x81 << 8;
+
+ emit(pc, e);
+}
+
static INLINE void
emit_flr(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
{
- emit_cvt(pc, dst, src, -1, CVTOP_FLOOR, CVT_F32_F32 | CVT_RI);
+ emit_cvt(pc, dst, src, -1, CVT_FLOOR | CVT_F32_F32 | CVT_RI);
}
static void
free_temp(pc, temp);
}
-static INLINE void
-emit_abs(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
-{
- emit_cvt(pc, dst, src, -1, CVTOP_ABS, CVT_F32_F32);
-}
-
static INLINE void
emit_sat(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
{
- emit_cvt(pc, dst, src, -1, CVTOP_SAT, CVT_F32_F32);
+ emit_cvt(pc, dst, src, -1, CVT_SAT | CVT_F32_F32);
}
static void
if (mask & (3 << 1)) {
tmp[0] = alloc_temp(pc, NULL);
- emit_minmax(pc, 4, tmp[0], src[0], zero);
+ emit_minmax(pc, NV50_MAX_F32, tmp[0], src[0], zero);
}
if (mask & (1 << 2)) {
set_pred_wr(pc, 1, 0, pc->p->exec_tail);
- tmp[1] = temp_temp(pc);
- emit_minmax(pc, 4, tmp[1], src[1], zero);
+ tmp[1] = temp_temp(pc, NULL);
+ emit_minmax(pc, NV50_MAX_F32, tmp[1], src[1], zero);
- tmp[3] = temp_temp(pc);
- emit_minmax(pc, 4, tmp[3], src[3], neg128);
- emit_minmax(pc, 5, tmp[3], tmp[3], pos128);
+ tmp[3] = temp_temp(pc, NULL);
+ emit_minmax(pc, NV50_MAX_F32, tmp[3], src[3], neg128);
+ emit_minmax(pc, NV50_MIN_F32, tmp[3], tmp[3], pos128);
emit_pow(pc, dst[2], tmp[1], tmp[3]);
emit_mov(pc, dst[2], zero);
FREE(one);
}
-static INLINE void
-emit_neg(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
-{
- emit_cvt(pc, dst, src, -1, CVTOP_RN, CVT_F32_F32 | CVT_NEG);
-}
-
static void
emit_kil(struct nv50_pc *pc, struct nv50_reg *src)
{
set_long(pc, e); /* sets cond code to ALWAYS */
if (src) {
- unsigned cvn = CVT_F32_F32;
-
set_pred(pc, 0x1 /* cc = LT */, r_pred, e);
-
- if (src->mod & NV50_MOD_NEG)
- cvn |= CVT_NEG;
- /* write predicate reg */
- emit_cvt(pc, NULL, src, r_pred, CVTOP_RN, cvn);
+ /* write to predicate reg */
+ emit_cvt(pc, NULL, src, r_pred, CVT_F32_F32);
}
emit(pc, e);
emit_control_flow(pc, 0x3, pred, cc);
}
+static void
+emit_prim_cmd(struct nv50_pc *pc, unsigned cmd)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0xf0000000 | (cmd << 9);
+ e->inst[1] = 0xc0000000;
+ set_long(pc, e);
+
+ emit(pc, e);
+}
+
#define QOP_ADD 0
#define QOP_SUBR 1
#define QOP_SUB 2
src[1]->mod |= NV50_MOD_ABS;
src[2]->mod |= NV50_MOD_ABS;
- emit_minmax(pc, 4, t[2], src[0], src[1]);
- emit_minmax(pc, 4, t[2], src[2], t[2]);
+ emit_minmax(pc, NV50_MAX_F32, t[2], src[0], src[1]);
+ emit_minmax(pc, NV50_MAX_F32, t[2], src[2], t[2]);
src[0]->mod = mod[0];
src[1]->mod = mod[1];
q = 0x0403c000;
m = 0xffff7fff;
break;
+ case 0x2:
+ case 0x3:
+ /* ADD, SUB, SUBR b32 */
+ m = ~(0x8000 | (127 << 16));
+ q = ((e->inst[0] & (~m)) >> 2) | (1 << 26);
+ break;
+ case 0x5:
+ /* SAD */
+ m = ~(0x81 << 8);
+ q = (0x0c << 24) | ((e->inst[0] & (0x7f << 2)) << 12);
+ break;
+ case 0x6:
+ /* MAD u16 */
+ q = (e->inst[0] & (0x7f << 2)) << 12;
+ break;
case 0x8:
/* INTERP (move centroid, perspective and flat bits) */
m = ~0x03000100;
}
/* Some operations support an optional negation flag. */
-static boolean
-negate_supported(const struct tgsi_full_instruction *insn, int i)
+static int
+get_supported_mods(const struct tgsi_full_instruction *insn, int i)
{
switch (insn->Instruction.Opcode) {
case TGSI_OPCODE_ADD:
case TGSI_OPCODE_SCS:
case TGSI_OPCODE_SIN:
case TGSI_OPCODE_SUB:
- return TRUE;
+ return NV50_MOD_NEG;
+ case TGSI_OPCODE_MAX:
+ case TGSI_OPCODE_MIN:
+ case TGSI_OPCODE_INEG: /* tgsi src sign toggle/set would be stupid */
+ return NV50_MOD_ABS;
+ case TGSI_OPCODE_CEIL:
+ case TGSI_OPCODE_FLR:
+ case TGSI_OPCODE_TRUNC:
+ return NV50_MOD_NEG | NV50_MOD_ABS;
+ case TGSI_OPCODE_F2I:
+ case TGSI_OPCODE_F2U:
+ case TGSI_OPCODE_I2F:
+ case TGSI_OPCODE_U2F:
+ return NV50_MOD_NEG | NV50_MOD_ABS | NV50_MOD_I32;
+ case TGSI_OPCODE_UADD:
+ return NV50_MOD_NEG | NV50_MOD_I32;
+ case TGSI_OPCODE_SAD:
+ case TGSI_OPCODE_SHL:
+ case TGSI_OPCODE_IMAX:
+ case TGSI_OPCODE_IMIN:
+ case TGSI_OPCODE_ISHR:
+ case TGSI_OPCODE_NOT:
+ case TGSI_OPCODE_UMAD:
+ case TGSI_OPCODE_UMAX:
+ case TGSI_OPCODE_UMIN:
+ case TGSI_OPCODE_UMUL:
+ case TGSI_OPCODE_USHR:
+ return NV50_MOD_I32;
default:
- return FALSE;
+ return 0;
}
}
{
struct nv50_reg *r = pc->addr[dst->Register.Index * 4 + c];
if (!r) {
- r = alloc_addr(pc, NULL);
- pc->addr[dst->Register.Index * 4 + c] = r;
+ r = get_address_reg(pc, NULL);
+ r->index = dst->Register.Index * 4 + c;
+ pc->addr[r->index] = r;
}
assert(r);
return r;
static struct nv50_reg *
tgsi_src(struct nv50_pc *pc, int chan, const struct tgsi_full_src_register *src,
- boolean neg)
+ int mod)
{
struct nv50_reg *r = NULL;
- struct nv50_reg *temp;
- unsigned sgn, c, swz;
+ struct nv50_reg *temp = NULL;
+ unsigned sgn, c, swz, cvn;
if (src->Register.File != TGSI_FILE_CONSTANT)
assert(!src->Register.Indirect);
switch (src->Register.File) {
case TGSI_FILE_INPUT:
r = &pc->attr[src->Register.Index * 4 + c];
+
+ if (!src->Dimension.Dimension)
+ break;
+ r = reg_instance(pc, r);
+ r->vtx = src->Dimension.Index;
+
+ if (!src->Dimension.Indirect)
+ break;
+ swz = tgsi_util_get_src_register_swizzle(
+ &src->DimIndirect, 0);
+ r->acc = -1;
+ r->indirect[1] = src->DimIndirect.Index * 4 + swz;
break;
case TGSI_FILE_TEMPORARY:
r = &pc->temp[src->Register.Index * 4 + c];
* use the index field to select the address reg.
*/
r = reg_instance(pc, NULL);
+ ctor_reg(r, P_CONST, -1, src->Register.Index * 4 + c);
+
swz = tgsi_util_get_src_register_swizzle(
- &src->Indirect, 0);
- ctor_reg(r, P_CONST,
- src->Indirect.Index * 4 + swz,
- src->Register.Index * 4 + c);
+ &src->Indirect, 0);
r->acc = -1;
+ r->indirect[0] = src->Indirect.Index * 4 + swz;
break;
case TGSI_FILE_IMMEDIATE:
r = &pc->immd[src->Register.Index * 4 + c];
break;
case TGSI_FILE_SAMPLER:
- break;
+ return NULL;
case TGSI_FILE_ADDRESS:
r = pc->addr[src->Register.Index * 4 + c];
assert(r);
break;
}
+ cvn = (mod & NV50_MOD_I32) ? CVT_S32_S32 : CVT_F32_F32;
+
switch (sgn) {
- case TGSI_UTIL_SIGN_KEEP:
- break;
case TGSI_UTIL_SIGN_CLEAR:
- temp = temp_temp(pc);
- emit_abs(pc, temp, r);
- r = temp;
- break;
- case TGSI_UTIL_SIGN_TOGGLE:
- if (neg)
- r->mod = NV50_MOD_NEG;
- else {
- temp = temp_temp(pc);
- emit_neg(pc, temp, r);
- r = temp;
- }
+ r->mod = NV50_MOD_ABS;
break;
case TGSI_UTIL_SIGN_SET:
- temp = temp_temp(pc);
- emit_cvt(pc, temp, r, -1, CVTOP_ABS, CVT_F32_F32 | CVT_NEG);
- r = temp;
+ r->mod = NV50_MOD_NEG_ABS;
+ break;
+ case TGSI_UTIL_SIGN_TOGGLE:
+ r->mod = NV50_MOD_NEG;
break;
default:
- assert(0);
+ assert(!r->mod && sgn == TGSI_UTIL_SIGN_KEEP);
break;
}
- if (r && r->acc >= 0 && r != temp)
- return reg_instance(pc, r);
+ if ((r->mod & mod) != r->mod) {
+ temp = temp_temp(pc, NULL);
+ emit_cvt(pc, temp, r, -1, cvn);
+ r->mod = 0;
+ r = temp;
+ } else
+ r->mod |= mod & NV50_MOD_I32;
+
+ assert(r);
+ if (r->acc >= 0 && r->vtx < 0 && r != temp)
+ return reg_instance(pc, r); /* will clear r->mod */
return r;
}
for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
const struct tgsi_full_src_register *fs = &inst->Src[i];
unsigned src_mask;
- boolean neg_supp;
+ int mod_supp;
src_mask = nv50_tgsi_src_mask(inst, i);
- neg_supp = negate_supported(inst, i);
+ mod_supp = get_supported_mods(inst, i);
if (fs->Register.File == TGSI_FILE_SAMPLER)
unit = fs->Register.Index;
for (c = 0; c < 4; c++)
if (src_mask & (1 << c))
- src[i][c] = tgsi_src(pc, c, fs, neg_supp);
+ src[i][c] = tgsi_src(pc, c, fs, mod_supp);
}
brdc = temp = pc->r_brdc;
if (brdc && brdc->type != P_TEMP) {
- temp = temp_temp(pc);
+ temp = temp_temp(pc, NULL);
if (sat)
brdc = temp;
} else
if (!(mask & (1 << c)) || dst[c]->type == P_TEMP)
continue;
/* rdst[c] = dst[c]; */ /* done above */
- dst[c] = temp_temp(pc);
+ dst[c] = temp_temp(pc, NULL);
}
}
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
- emit_abs(pc, dst[c], src[0][c]);
+ emit_cvt(pc, dst[c], src[0][c], -1,
+ CVT_ABS | CVT_F32_F32);
}
break;
case TGSI_OPCODE_ADD:
}
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);
+ temp = temp_temp(pc, NULL);
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_cvt(pc, temp, src[0][c], -1,
+ CVT_FLOOR | CVT_S32_F32);
+ emit_arl(pc, dst[c], temp, 4);
+ }
break;
case TGSI_OPCODE_BGNLOOP:
pc->loop_brka[pc->loop_lvl] = emit_breakaddr(pc);
if (!(mask & (1 << c)))
continue;
emit_cvt(pc, dst[c], src[0][c], -1,
- CVTOP_CEIL, CVT_F32_F32 | CVT_RI);
+ CVT_CEIL | CVT_F32_F32 | CVT_RI);
}
break;
case TGSI_OPCODE_CMP:
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
- emit_cvt(pc, NULL, src[0][c], 1, CVTOP_RN, CVT_F32_F32);
+ emit_cvt(pc, NULL, src[0][c], 1, CVT_F32_F32);
emit_mov(pc, dst[c], src[1][c]);
set_pred(pc, 0x1, 1, pc->p->exec_tail); /* @SF */
emit_mov(pc, dst[c], src[2][c]);
if (!(mask &= 7))
break;
if (temp == dst[3])
- temp = brdc = temp_temp(pc);
+ temp = brdc = temp_temp(pc, NULL);
}
emit_precossin(pc, temp, src[0][0]);
emit_flop(pc, NV50_FLOP_COS, brdc, temp);
pc->if_insn[pc->if_lvl++] = pc->p->exec_tail;
terminate_mbb(pc);
break;
+ case TGSI_OPCODE_EMIT:
+ emit_prim_cmd(pc, 1);
+ break;
case TGSI_OPCODE_ENDIF:
pc->if_insn[--pc->if_lvl]->param.index = pc->p->exec_size;
pc->loop_brka[pc->loop_lvl]->param.index = pc->p->exec_size;
terminate_mbb(pc);
break;
+ case TGSI_OPCODE_ENDPRIM:
+ emit_prim_cmd(pc, 2);
+ break;
case TGSI_OPCODE_ENDSUB:
assert(pc->in_subroutine);
+ terminate_mbb(pc);
pc->in_subroutine = FALSE;
break;
case TGSI_OPCODE_EX2:
struct nv50_reg *t[2];
assert(!temp);
- t[0] = temp_temp(pc);
- t[1] = temp_temp(pc);
+ t[0] = temp_temp(pc, NULL);
+ t[1] = temp_temp(pc, NULL);
if (mask & 0x6)
emit_mov(pc, t[0], src[0][0]);
emit_mov_immdval(pc, dst[3], 1.0f);
}
break;
+ case TGSI_OPCODE_F2I:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_cvt(pc, dst[c], src[0][c], -1,
+ CVT_TRUNC | CVT_S32_F32);
+ }
+ break;
+ case TGSI_OPCODE_F2U:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_cvt(pc, dst[c], src[0][c], -1,
+ CVT_TRUNC | CVT_U32_F32);
+ }
+ break;
case TGSI_OPCODE_FLR:
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
}
break;
case TGSI_OPCODE_FRC:
- temp = temp_temp(pc);
+ temp = temp_temp(pc, NULL);
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
emit_sub(pc, dst[c], src[0][c], temp);
}
break;
+ case TGSI_OPCODE_I2F:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_cvt(pc, dst[c], src[0][c], -1, CVT_F32_S32);
+ }
+ break;
case TGSI_OPCODE_IF:
assert(pc->if_lvl < NV50_MAX_COND_NESTING);
- emit_cvt(pc, NULL, src[0][0], 0, CVTOP_ABS | CVTOP_RN,
- CVT_F32_F32);
+ emit_cvt(pc, NULL, src[0][0], 0, CVT_ABS | CVT_F32_F32);
pc->if_join[pc->if_lvl] = emit_joinat(pc);
pc->if_insn[pc->if_lvl++] = emit_branch(pc, 0, 2);;
terminate_mbb(pc);
break;
+ case TGSI_OPCODE_IMAX:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_minmax(pc, 0x08c, dst[c], src[0][c], src[1][c]);
+ }
+ break;
+ case TGSI_OPCODE_IMIN:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_minmax(pc, 0x0ac, dst[c], src[0][c], src[1][c]);
+ }
+ break;
+ case TGSI_OPCODE_INEG:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_cvt(pc, dst[c], src[0][c], -1,
+ CVT_S32_S32 | CVT_NEG);
+ }
+ break;
case TGSI_OPCODE_KIL:
assert(src[0][0] && src[0][1] && src[0][2] && src[0][3]);
emit_kil(pc, src[0][0]);
{
struct nv50_reg *t[2];
- t[0] = temp_temp(pc);
+ t[0] = temp_temp(pc, NULL);
if (mask & (1 << 1))
- t[1] = temp_temp(pc);
+ t[1] = temp_temp(pc, NULL);
else
t[1] = t[0];
- emit_abs(pc, t[0], src[0][0]);
+ emit_cvt(pc, t[0], src[0][0], -1, CVT_ABS | CVT_F32_F32);
emit_flop(pc, NV50_FLOP_LG2, t[1], t[0]);
if (mask & (1 << 2))
emit_mov(pc, dst[2], t[1]);
}
break;
case TGSI_OPCODE_LRP:
- temp = temp_temp(pc);
+ temp = temp_temp(pc, NULL);
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
- emit_minmax(pc, 4, dst[c], src[0][c], src[1][c]);
+ emit_minmax(pc, 0x880, dst[c], src[0][c], src[1][c]);
}
break;
case TGSI_OPCODE_MIN:
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
- emit_minmax(pc, 5, dst[c], src[0][c], src[1][c]);
+ emit_minmax(pc, 0x8a0, dst[c], src[0][c], src[1][c]);
}
break;
case TGSI_OPCODE_MOV:
emit_mul(pc, dst[c], src[0][c], src[1][c]);
}
break;
+ case TGSI_OPCODE_NOT:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_not(pc, dst[c], src[0][c]);
+ }
+ break;
case TGSI_OPCODE_POW:
emit_pow(pc, brdc, src[0][0], src[1][0]);
break;
case TGSI_OPCODE_RCP:
+ if (!sat && popcnt4(mask) == 1)
+ brdc = dst[ffs(mask) - 1];
emit_flop(pc, NV50_FLOP_RCP, brdc, src[0][0]);
break;
case TGSI_OPCODE_RET:
emit_ret(pc, -1, 0);
break;
case TGSI_OPCODE_RSQ:
+ if (!sat && popcnt4(mask) == 1)
+ brdc = dst[ffs(mask) - 1];
src[0][0]->mod |= NV50_MOD_ABS;
emit_flop(pc, NV50_FLOP_RSQ, brdc, src[0][0]);
break;
+ case TGSI_OPCODE_SAD:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_sad(pc, dst[c], src[0][c], src[1][c], src[2][c]);
+ }
+ break;
case TGSI_OPCODE_SCS:
- temp = temp_temp(pc);
+ temp = temp_temp(pc, NULL);
if (mask & 3)
emit_precossin(pc, temp, src[0][0]);
if (mask & (1 << 0))
if (mask & (1 << 3))
emit_mov_immdval(pc, dst[3], 1.0);
break;
+ case TGSI_OPCODE_SHL:
+ case TGSI_OPCODE_ISHR:
+ case TGSI_OPCODE_USHR:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_shift(pc, dst[c], src[0][c], src[1][c],
+ inst->Instruction.Opcode);
+ }
+ break;
case TGSI_OPCODE_SIN:
if (mask & 8) {
emit_precossin(pc, temp, src[0][3]);
if (!(mask &= 7))
break;
if (temp == dst[3])
- temp = brdc = temp_temp(pc);
+ temp = brdc = temp_temp(pc, NULL);
}
emit_precossin(pc, temp, src[0][0]);
emit_flop(pc, NV50_FLOP_SIN, brdc, temp);
case TGSI_OPCODE_SGT:
case TGSI_OPCODE_SLE:
case TGSI_OPCODE_SNE:
- i = map_tgsi_setop_cc(inst->Instruction.Opcode);
+ case TGSI_OPCODE_ISLT:
+ case TGSI_OPCODE_ISGE:
+ case TGSI_OPCODE_USEQ:
+ case TGSI_OPCODE_USGE:
+ case TGSI_OPCODE_USLT:
+ case TGSI_OPCODE_USNE:
+ {
+ uint8_t cc, ty;
+
+ map_tgsi_setop_hw(inst->Instruction.Opcode, &cc, &ty);
+
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
- emit_set(pc, i, dst[c], -1, src[0][c], src[1][c]);
+ emit_set(pc, cc, dst[c], -1, src[0][c], src[1][c], ty);
}
+ }
break;
case TGSI_OPCODE_SUB:
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
emit_cvt(pc, dst[c], src[0][c], -1,
- CVTOP_TRUNC, CVT_F32_F32 | CVT_RI);
+ CVT_TRUNC | CVT_F32_F32 | CVT_RI);
+ }
+ break;
+ case TGSI_OPCODE_U2F:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_cvt(pc, dst[c], src[0][c], -1, CVT_F32_U32);
}
break;
+ case TGSI_OPCODE_UADD:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_add_b32(pc, dst[c], src[0][c], src[1][c]);
+ }
+ break;
+ case TGSI_OPCODE_UMAX:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_minmax(pc, 0x084, dst[c], src[0][c], src[1][c]);
+ }
+ break;
+ case TGSI_OPCODE_UMIN:
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_minmax(pc, 0x0a4, dst[c], src[0][c], src[1][c]);
+ }
+ break;
+ case TGSI_OPCODE_UMAD:
+ {
+ assert(!temp);
+ temp = temp_temp(pc, NULL);
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_mul_u16(pc, temp, src[0][c], 0, src[1][c], 1);
+ emit_mad_u16(pc, temp, src[0][c], 1, src[1][c], 0,
+ temp);
+ emit_shl_imm(pc, temp, temp, 16);
+ emit_mad_u16(pc, temp, src[0][c], 0, src[1][c], 0,
+ temp);
+ emit_add_b32(pc, dst[c], temp, src[2][c]);
+ }
+ }
+ break;
+ case TGSI_OPCODE_UMUL:
+ {
+ assert(!temp);
+ temp = temp_temp(pc, NULL);
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ emit_mul_u16(pc, temp, src[0][c], 0, src[1][c], 1);
+ emit_mad_u16(pc, temp, src[0][c], 1, src[1][c], 0,
+ temp);
+ emit_shl_imm(pc, temp, temp, 16);
+ emit_mad_u16(pc, dst[c], src[0][c], 0, src[1][c], 0,
+ temp);
+ }
+ }
+ break;
case TGSI_OPCODE_XPD:
- temp = temp_temp(pc);
+ temp = temp_temp(pc, NULL);
if (mask & (1 << 0)) {
emit_mul(pc, temp, src[0][2], src[1][1]);
emit_msb(pc, dst[0], src[0][1], src[1][2], temp);
emit_nop(pc);
pc->p->exec_tail->inst[1] |= 1; /* set exit bit */
+
+ terminate_mbb(pc);
break;
default:
NOUVEAU_ERR("invalid opcode %d\n", inst->Instruction.Opcode);
}
}
- kill_temp_temp(pc);
+ kill_temp_temp(pc, NULL);
pc->reg_instance_nr = 0;
return TRUE;
static void
prep_inspect_insn(struct nv50_pc *pc, const struct tgsi_full_instruction *insn)
{
- struct nv50_reg *reg = NULL;
+ struct nv50_reg *r, *reg = NULL;
const struct tgsi_full_src_register *src;
const struct tgsi_dst_register *dst;
unsigned i, c, k, mask;
continue;
k = tgsi_util_get_full_src_register_swizzle(src, c);
- reg[src->Register.Index * 4 + k].acc = pc->insn_nr;
+ r = ®[src->Register.Index * 4 + k];
+
+ /* If used before written, pre-allocate the reg,
+ * lest we overwrite results from a subroutine.
+ */
+ if (!r->acc && r->type == P_TEMP)
+ alloc_reg(pc, r);
+
+ r->acc = pc->insn_nr;
}
}
}
for (i = 0; i < insn->Instruction.NumSrcRegs; i++) {
unsigned chn, mask = nv50_tgsi_src_mask(insn, i);
- boolean neg_supp = negate_supported(insn, i);
+ int ms = get_supported_mods(insn, i);
fs = &insn->Src[i];
if (fs->Register.File != fd->Register.File ||
if (!(fd->Register.WriteMask & (1 << c)))
continue;
- /* no danger if src is copied to TEMP first */
- if ((s != TGSI_UTIL_SIGN_KEEP) &&
- (s != TGSI_UTIL_SIGN_TOGGLE || !neg_supp))
- continue;
+ if (s == TGSI_UTIL_SIGN_TOGGLE && !(ms & NV50_MOD_NEG))
+ continue;
+ if (s == TGSI_UTIL_SIGN_CLEAR && !(ms & NV50_MOD_ABS))
+ continue;
+ if ((s == TGSI_UTIL_SIGN_SET) && ((ms & 3) != 3))
+ continue;
rdep[c] |= nv50_tgsi_dst_revdep(
insn->Instruction.Opcode, i, chn);
if (is_scalar_op(insn.Instruction.Opcode)) {
pc->r_brdc = tgsi_broadcast_dst(pc, fd, deqs);
if (!pc->r_brdc)
- pc->r_brdc = temp_temp(pc);
+ pc->r_brdc = temp_temp(pc, NULL);
return nv50_program_tx_insn(pc, &insn);
}
pc->r_brdc = NULL;
FREE(one);
}
+static void
+copy_semantic_info(struct nv50_program *p)
+{
+ unsigned i, id;
+
+ for (i = 0; i < p->cfg.in_nr; ++i) {
+ id = p->cfg.in[i].id;
+ p->cfg.in[i].sn = p->info.input_semantic_name[id];
+ p->cfg.in[i].si = p->info.input_semantic_index[id];
+ }
+
+ for (i = 0; i < p->cfg.out_nr; ++i) {
+ id = p->cfg.out[i].id;
+ p->cfg.out[i].sn = p->info.output_semantic_name[id];
+ p->cfg.out[i].si = p->info.output_semantic_index[id];
+ }
+}
+
static boolean
nv50_program_tx_prep(struct nv50_pc *pc)
{
switch (d->Semantic.Name) {
case TGSI_SEMANTIC_BCOLOR:
p->cfg.two_side[si].hw = first;
- if (p->cfg.io_nr > first)
- p->cfg.io_nr = first;
+ if (p->cfg.out_nr > first)
+ p->cfg.out_nr = first;
break;
case TGSI_SEMANTIC_PSIZE:
p->cfg.psiz = first;
- if (p->cfg.io_nr > first)
- p->cfg.io_nr = first;
+ if (p->cfg.out_nr > first)
+ p->cfg.out_nr = first;
break;
case TGSI_SEMANTIC_EDGEFLAG:
pc->edgeflag_out = first;
}
}
- if (p->type == PIPE_SHADER_VERTEX) {
+ if (p->type == PIPE_SHADER_VERTEX || p->type == PIPE_SHADER_GEOMETRY) {
int rid = 0;
- for (i = 0; i < pc->attr_nr * 4; ++i) {
- if (pc->attr[i].acc) {
- pc->attr[i].hw = rid++;
- p->cfg.attr[i / 32] |= 1 << (i % 32);
+ if (p->type == PIPE_SHADER_GEOMETRY) {
+ for (i = 0; i < pc->attr_nr; ++i) {
+ p->cfg.in[i].hw = rid;
+ p->cfg.in[i].id = i;
+
+ for (c = 0; c < 4; ++c) {
+ int n = i * 4 + c;
+ if (!pc->attr[n].acc)
+ continue;
+ pc->attr[n].hw = rid++;
+ p->cfg.in[i].mask |= 1 << c;
+ }
+ }
+ } else {
+ for (i = 0; i < pc->attr_nr * 4; ++i) {
+ if (pc->attr[i].acc) {
+ pc->attr[i].hw = rid++;
+ p->cfg.attr[i / 32] |= 1 << (i % 32);
+ }
}
}
for (i = 0, rid = 0; i < pc->result_nr; ++i) {
- p->cfg.io[i].hw = rid;
- p->cfg.io[i].id = i;
+ p->cfg.out[i].hw = rid;
+ p->cfg.out[i].id = i;
for (c = 0; c < 4; ++c) {
int n = i * 4 + c;
if (!pc->result[n].acc)
continue;
pc->result[n].hw = rid++;
- p->cfg.io[i].mask |= 1 << c;
+ p->cfg.out[i].mask |= 1 << c;
}
}
for (c = 0; c < 2; ++c)
if (p->cfg.two_side[c].hw < 0x40)
- p->cfg.two_side[c] = p->cfg.io[
+ p->cfg.two_side[c] = p->cfg.out[
p->cfg.two_side[c].hw];
if (p->cfg.psiz < 0x40)
- p->cfg.psiz = p->cfg.io[p->cfg.psiz].hw;
+ p->cfg.psiz = p->cfg.out[p->cfg.psiz].hw;
+
+ copy_semantic_info(p);
} else
if (p->type == PIPE_SHADER_FRAGMENT) {
- int rid, aid;
+ int rid, aid, base;
unsigned n = 0, m = pc->attr_nr - flat_nr;
pc->allow32 = TRUE;
- int base = (TGSI_SEMANTIC_POSITION ==
- p->info.input_semantic_name[0]) ? 0 : 1;
+ base = (TGSI_SEMANTIC_POSITION ==
+ p->info.input_semantic_name[0]) ? 0 : 1;
/* non-flat interpolants have to be mapped to
* the lower hardware IDs, so sort them:
*/
for (i = 0; i < pc->attr_nr; i++) {
if (pc->interp_mode[i] == INTERP_FLAT)
- p->cfg.io[m++].id = i;
+ p->cfg.in[m++].id = i;
else {
if (!(pc->interp_mode[i] & INTERP_PERSPECTIVE))
- p->cfg.io[n].linear = TRUE;
- p->cfg.io[n++].id = i;
+ p->cfg.in[n].linear = TRUE;
+ p->cfg.in[n++].id = i;
}
}
+ copy_semantic_info(p);
if (!base) /* set w-coordinate mask from perspective interp */
- p->cfg.io[0].mask |= p->cfg.regs[1] >> 24;
+ p->cfg.in[0].mask |= p->cfg.regs[1] >> 24;
aid = popcnt4( /* if fcrd isn't contained in cfg.io */
- base ? (p->cfg.regs[1] >> 24) : p->cfg.io[0].mask);
+ base ? (p->cfg.regs[1] >> 24) : p->cfg.in[0].mask);
for (n = 0; n < pc->attr_nr; ++n) {
- p->cfg.io[n].hw = rid = aid;
- i = p->cfg.io[n].id;
+ p->cfg.in[n].hw = rid = aid;
+ i = p->cfg.in[n].id;
if (p->info.input_semantic_name[n] ==
TGSI_SEMANTIC_FACE) {
if (!pc->attr[i * 4 + c].acc)
continue;
pc->attr[i * 4 + c].rhw = rid++;
- p->cfg.io[n].mask |= 1 << c;
+ p->cfg.in[n].mask |= 1 << c;
load_interpolant(pc, &pc->attr[i * 4 + c]);
}
- aid += popcnt4(p->cfg.io[n].mask);
+ aid += popcnt4(p->cfg.in[n].mask);
}
if (!base)
- p->cfg.regs[1] |= p->cfg.io[0].mask << 24;
+ p->cfg.regs[1] |= p->cfg.in[0].mask << 24;
m = popcnt4(p->cfg.regs[1] >> 24);
p->cfg.regs[1] |= aid - m;
if (flat_nr) {
- i = p->cfg.io[pc->attr_nr - flat_nr].hw;
+ i = p->cfg.in[pc->attr_nr - flat_nr].hw;
p->cfg.regs[1] |= (i - m) << 16;
} else
p->cfg.regs[1] |= p->cfg.regs[1] << 16;
/* mark color semantic for light-twoside */
- n = 0x40;
- for (i = 0; i < pc->attr_nr; i++) {
- ubyte si, sn;
-
- sn = p->info.input_semantic_name[p->cfg.io[i].id];
- si = p->info.input_semantic_index[p->cfg.io[i].id];
-
- if (sn == TGSI_SEMANTIC_COLOR) {
- p->cfg.two_side[si] = p->cfg.io[i];
-
- /* increase colour count */
- p->cfg.regs[0] += popcnt4(
- p->cfg.two_side[si].mask) << 16;
-
- n = MIN2(n, p->cfg.io[i].hw - m);
+ n = 0x80;
+ for (i = 0; i < p->cfg.in_nr; i++) {
+ if (p->cfg.in[i].sn == TGSI_SEMANTIC_COLOR) {
+ n = MIN2(n, p->cfg.in[i].hw - m);
+ p->cfg.two_side[p->cfg.in[i].si] = p->cfg.in[i];
+
+ p->cfg.regs[0] += /* increase colour count */
+ popcnt4(p->cfg.in[i].mask) << 16;
}
}
- if (n < 0x40)
+ if (n < 0x80)
p->cfg.regs[0] += n;
/* Initialize FP results:
FREE(pc->attr);
if (pc->temp)
FREE(pc->temp);
+ if (pc->insn_pos)
+ FREE(pc->insn_pos);
FREE(pc);
}
+static INLINE uint32_t
+nv50_map_gs_output_prim(unsigned pprim)
+{
+ switch (pprim) {
+ case PIPE_PRIM_POINTS:
+ return NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE_POINTS;
+ case PIPE_PRIM_LINE_STRIP:
+ return NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE_LINE_STRIP;
+ case PIPE_PRIM_TRIANGLE_STRIP:
+ return NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE_TRIANGLE_STRIP;
+ default:
+ NOUVEAU_ERR("invalid GS_OUTPUT_PRIMITIVE: %u\n", pprim);
+ abort();
+ return 0;
+ }
+}
+
static boolean
ctor_nv50_pc(struct nv50_pc *pc, struct nv50_program *p)
{
p->cfg.edgeflag_in = pc->edgeflag_out = 0xff;
+ for (i = 0; i < p->info.num_properties; ++i) {
+ unsigned *data = &p->info.properties[i].data[0];
+
+ switch (p->info.properties[i].name) {
+ case TGSI_PROPERTY_GS_OUTPUT_PRIM:
+ p->cfg.prim_type = nv50_map_gs_output_prim(data[0]);
+ break;
+ case TGSI_PROPERTY_GS_MAX_VERTICES:
+ p->cfg.vert_count = data[0];
+ break;
+ default:
+ break;
+ }
+ }
+
switch (p->type) {
case PIPE_SHADER_VERTEX:
p->cfg.psiz = 0x40;
p->cfg.clpd = 0x40;
- p->cfg.io_nr = pc->result_nr;
+ p->cfg.out_nr = pc->result_nr;
+ break;
+ case PIPE_SHADER_GEOMETRY:
+ assert(p->cfg.prim_type);
+ assert(p->cfg.vert_count);
+
+ p->cfg.psiz = 0x80;
+ p->cfg.clpd = 0x80;
+ p->cfg.out_nr = pc->result_nr;
+ p->cfg.in_nr = pc->attr_nr;
+
+ p->cfg.two_side[0].hw = 0x80;
+ p->cfg.two_side[1].hw = 0x80;
break;
case PIPE_SHADER_FRAGMENT:
rtype[0] = rtype[1] = P_TEMP;
p->cfg.regs[0] = 0x01000004;
- p->cfg.io_nr = pc->attr_nr;
+ p->cfg.in_nr = pc->attr_nr;
if (p->info.writes_z) {
p->cfg.regs[2] |= 0x00000100;
return FALSE;
}
for (i = 0; i < NV50_SU_MAX_ADDR; ++i)
- ctor_reg(&pc->r_addr[i], P_ADDR, -256, i + 1);
+ ctor_reg(&pc->r_addr[i], P_ADDR, -1, i + 1);
return TRUE;
}
if (p->param_nr) {
unsigned cb;
- uint32_t *map = pipe_buffer_map(pscreen, nv50->constbuf[p->type],
+ uint32_t *map = pipe_buffer_map(pscreen,
+ nv50->constbuf[p->type],
PIPE_BUFFER_USAGE_CPU_READ);
-
- if (p->type == PIPE_SHADER_VERTEX)
+ switch (p->type) {
+ case PIPE_SHADER_GEOMETRY: cb = NV50_CB_PGP; break;
+ case PIPE_SHADER_FRAGMENT: cb = NV50_CB_PFP; break;
+ default:
cb = NV50_CB_PVP;
- else
- cb = NV50_CB_PFP;
+ assert(p->type == PIPE_SHADER_VERTEX);
+ break;
+ }
nv50_program_upload_data(nv50, map, 0, p->param_nr, cb);
pipe_buffer_unmap(pscreen, nv50->constbuf[p->type]);
nv50_program_validate_data(nv50, p);
nv50_program_validate_code(nv50, p);
- so = so_new(13, 2);
+ so = so_new(5, 7, 2);
so_method(so, tesla, NV50TCL_VP_ADDRESS_HIGH, 2);
so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
- NOUVEAU_BO_HIGH, 0, 0);
+ NOUVEAU_BO_HIGH, 0, 0);
so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
- NOUVEAU_BO_LOW, 0, 0);
+ NOUVEAU_BO_LOW, 0, 0);
so_method(so, tesla, NV50TCL_VP_ATTR_EN_0, 2);
so_data (so, p->cfg.attr[0]);
so_data (so, p->cfg.attr[1]);
so_method(so, tesla, NV50TCL_VP_REG_ALLOC_RESULT, 1);
so_data (so, p->cfg.high_result);
- so_method(so, tesla, NV50TCL_VP_RESULT_MAP_SIZE, 2);
- so_data (so, p->cfg.high_result); //8);
+ so_method(so, tesla, NV50TCL_VP_REG_ALLOC_TEMP, 1);
so_data (so, p->cfg.high_temp);
so_method(so, tesla, NV50TCL_VP_START_ID, 1);
so_data (so, 0); /* program start offset */
nv50_program_validate_data(nv50, p);
nv50_program_validate_code(nv50, p);
- so = so_new(64, 2);
+ so = so_new(6, 7, 2);
so_method(so, tesla, NV50TCL_FP_ADDRESS_HIGH, 2);
so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
NOUVEAU_BO_HIGH, 0, 0);
so_ref(NULL, &so);
}
-static void
+void
+nv50_geomprog_validate(struct nv50_context *nv50)
+{
+ struct nouveau_grobj *tesla = nv50->screen->tesla;
+ struct nv50_program *p = nv50->geomprog;
+ struct nouveau_stateobj *so;
+
+ if (!p->translated) {
+ nv50_program_validate(nv50, p);
+ if (!p->translated)
+ assert(0);
+ }
+
+ nv50_program_validate_data(nv50, p);
+ nv50_program_validate_code(nv50, p);
+
+ so = so_new(6, 7, 2);
+ so_method(so, tesla, NV50TCL_GP_ADDRESS_HIGH, 2);
+ so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
+ NOUVEAU_BO_HIGH, 0, 0);
+ so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
+ NOUVEAU_BO_LOW, 0, 0);
+ so_method(so, tesla, NV50TCL_GP_REG_ALLOC_TEMP, 1);
+ so_data (so, p->cfg.high_temp);
+ so_method(so, tesla, NV50TCL_GP_REG_ALLOC_RESULT, 1);
+ so_data (so, p->cfg.high_result);
+ so_method(so, tesla, NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE, 1);
+ so_data (so, p->cfg.prim_type);
+ so_method(so, tesla, NV50TCL_GP_VERTEX_OUTPUT_COUNT, 1);
+ so_data (so, p->cfg.vert_count);
+ so_method(so, tesla, NV50TCL_GP_START_ID, 1);
+ so_data (so, 0);
+ so_ref(so, &nv50->state.geomprog);
+ so_ref(NULL, &so);
+}
+
+static uint32_t
nv50_pntc_replace(struct nv50_context *nv50, uint32_t pntc[8], unsigned base)
{
+ struct nv50_program *vp;
struct nv50_program *fp = nv50->fragprog;
- struct nv50_program *vp = nv50->vertprog;
unsigned i, c, m = base;
+ uint32_t origin = 0x00000010;
+
+ vp = nv50->geomprog ? nv50->geomprog : nv50->vertprog;
/* XXX: this might not work correctly in all cases yet - we'll
* just assume that an FP generic input that is not written in
*/
memset(pntc, 0, 8 * sizeof(uint32_t));
- for (i = 0; i < fp->cfg.io_nr; i++) {
- uint8_t sn, si;
- uint8_t j, k = fp->cfg.io[i].id;
- unsigned n = popcnt4(fp->cfg.io[i].mask);
+ for (i = 0; i < fp->cfg.in_nr; i++) {
+ unsigned j, n = popcnt4(fp->cfg.in[i].mask);
- if (fp->info.input_semantic_name[k] != TGSI_SEMANTIC_GENERIC) {
+ if (fp->cfg.in[i].sn != TGSI_SEMANTIC_GENERIC) {
m += n;
continue;
}
- for (j = 0; j < vp->info.num_outputs; ++j) {
- sn = vp->info.output_semantic_name[j];
- si = vp->info.output_semantic_index[j];
-
- if (sn == fp->info.input_semantic_name[k] &&
- si == fp->info.input_semantic_index[k])
+ for (j = 0; j < vp->cfg.out_nr; ++j)
+ if (vp->cfg.out[j].sn == fp->cfg.in[i].sn &&
+ vp->cfg.out[j].si == fp->cfg.in[i].si)
break;
- }
- if (j < vp->info.num_outputs) {
- ubyte mode =
- nv50->rasterizer->pipe.sprite_coord_mode[si];
+ if (j < vp->cfg.out_nr) {
+ ubyte mode = nv50->rasterizer->pipe.sprite_coord_mode[
+ vp->cfg.out[j].si];
if (mode == PIPE_SPRITE_COORD_NONE) {
m += n;
continue;
- }
+ } else
+ if (mode == PIPE_SPRITE_COORD_LOWER_LEFT)
+ origin = 0;
}
/* this is either PointCoord or replaced by sprite coords */
for (c = 0; c < 4; c++) {
- if (!(fp->cfg.io[i].mask & (1 << c)))
+ if (!(fp->cfg.in[i].mask & (1 << c)))
continue;
pntc[m / 8] |= (c + 1) << ((m % 8) * 4);
++m;
}
}
+ return origin;
}
static int
-nv50_sreg4_map(uint32_t *p_map, int mid, uint32_t lin[4],
- struct nv50_sreg4 *fpi, struct nv50_sreg4 *vpo)
+nv50_vec4_map(uint32_t *map32, int mid, uint8_t zval, uint32_t lin[4],
+ struct nv50_sreg4 *fpi, struct nv50_sreg4 *vpo)
{
int c;
uint8_t mv = vpo->mask, mf = fpi->mask, oid = vpo->hw;
- uint8_t *map = (uint8_t *)p_map;
+ uint8_t *map = (uint8_t *)map32;
for (c = 0; c < 4; ++c) {
if (mf & 1) {
if (fpi->linear == TRUE)
lin[mid / 32] |= 1 << (mid % 32);
- map[mid++] = (mv & 1) ? oid : ((c == 3) ? 0x41 : 0x40);
+ if (mv & 1)
+ map[mid] = oid;
+ else
+ map[mid] = (c == 3) ? (zval + 1) : zval;
+ ++mid;
}
oid += mv & 1;
}
void
-nv50_linkage_validate(struct nv50_context *nv50)
+nv50_fp_linkage_validate(struct nv50_context *nv50)
{
struct nouveau_grobj *tesla = nv50->screen->tesla;
struct nv50_program *vp = nv50->vertprog;
struct nv50_program *fp = nv50->fragprog;
struct nouveau_stateobj *so;
- struct nv50_sreg4 dummy, *vpo;
+ struct nv50_sreg4 dummy;
int i, n, c, m = 0;
uint32_t map[16], lin[4], reg[5], pcrd[8];
+ uint8_t zval = 0x40;
+ if (nv50->geomprog) {
+ vp = nv50->geomprog;
+ zval = 0x80;
+ }
memset(map, 0, sizeof(map));
memset(lin, 0, sizeof(lin));
dummy.linear = FALSE;
dummy.mask = 0xf; /* map all components of HPOS */
- m = nv50_sreg4_map(map, m, lin, &dummy, &vp->cfg.io[0]);
+ m = nv50_vec4_map(map, m, zval, lin, &dummy, &vp->cfg.out[0]);
dummy.mask = 0x0;
if (vp->cfg.clpd < 0x40) {
- for (c = 0; c < vp->cfg.clpd_nr; ++c)
- map[m++] = vp->cfg.clpd + c;
+ for (c = 0; c < vp->cfg.clpd_nr; ++c) {
+ map[m / 4] |= (vp->cfg.clpd + c) << ((m % 4) * 8);
+ ++m;
+ }
reg[1] = (m << 8);
}
/* if light_twoside is active, it seems FFC0_ID == BFC0_ID is bad */
if (nv50->rasterizer->pipe.light_twoside) {
- vpo = &vp->cfg.two_side[0];
+ struct nv50_sreg4 *vpo = &vp->cfg.two_side[0];
+ struct nv50_sreg4 *fpi = &fp->cfg.two_side[0];
- m = nv50_sreg4_map(map, m, lin, &fp->cfg.two_side[0], &vpo[0]);
- m = nv50_sreg4_map(map, m, lin, &fp->cfg.two_side[1], &vpo[1]);
+ m = nv50_vec4_map(map, m, zval, lin, &fpi[0], &vpo[0]);
+ m = nv50_vec4_map(map, m, zval, lin, &fpi[1], &vpo[1]);
}
reg[0] += m - 4; /* adjust FFC0 id */
reg[4] |= m << 8; /* set mid where 'normal' FP inputs start */
- for (i = 0; i < fp->cfg.io_nr; i++) {
- ubyte sn = fp->info.input_semantic_name[fp->cfg.io[i].id];
- ubyte si = fp->info.input_semantic_index[fp->cfg.io[i].id];
-
- /* position must be mapped first */
- assert(i == 0 || sn != TGSI_SEMANTIC_POSITION);
-
+ for (i = 0; i < fp->cfg.in_nr; i++) {
/* maybe even remove these from cfg.io */
- if (sn == TGSI_SEMANTIC_POSITION || sn == TGSI_SEMANTIC_FACE)
+ if (fp->cfg.in[i].sn == TGSI_SEMANTIC_POSITION ||
+ fp->cfg.in[i].sn == TGSI_SEMANTIC_FACE)
continue;
- /* VP outputs and vp->cfg.io are in the same order */
- for (n = 0; n < vp->info.num_outputs; ++n) {
- if (vp->info.output_semantic_name[n] == sn &&
- vp->info.output_semantic_index[n] == si)
+ for (n = 0; n < vp->cfg.out_nr; ++n)
+ if (vp->cfg.out[n].sn == fp->cfg.in[i].sn &&
+ vp->cfg.out[n].si == fp->cfg.in[i].si)
break;
- }
- vpo = (n < vp->info.num_outputs) ? &vp->cfg.io[n] : &dummy;
- m = nv50_sreg4_map(map, m, lin, &fp->cfg.io[i], vpo);
+ m = nv50_vec4_map(map, m, zval, lin, &fp->cfg.in[i],
+ (n < vp->cfg.out_nr) ?
+ &vp->cfg.out[n] : &dummy);
}
if (nv50->rasterizer->pipe.point_size_per_vertex) {
reg[3] = (m++ << 4) | 1;
}
- /* now fill the stateobj */
- so = so_new(64, 0);
+ /* now fill the stateobj (at most 28 so_data) */
+ so = so_new(8, 56, 0);
n = (m + 3) / 4;
- so_method(so, tesla, NV50TCL_VP_RESULT_MAP_SIZE, 1);
- so_data (so, m);
- so_method(so, tesla, NV50TCL_VP_RESULT_MAP(0), n);
- so_datap (so, map, n);
+ assert(m <= 32);
+ if (vp->type == PIPE_SHADER_GEOMETRY) {
+ so_method(so, tesla, NV50TCL_GP_RESULT_MAP_SIZE, 1);
+ so_data (so, m);
+ so_method(so, tesla, NV50TCL_GP_RESULT_MAP(0), n);
+ so_datap (so, map, n);
+ } else {
+ so_method(so, tesla, NV50TCL_VP_RESULT_MAP_SIZE, 1);
+ so_data (so, m);
+ so_method(so, tesla, NV50TCL_VP_RESULT_MAP(0), n);
+ so_datap (so, map, n);
+ }
so_method(so, tesla, NV50TCL_MAP_SEMANTIC_0, 4);
so_datap (so, reg, 4);
so_datap (so, lin, 4);
if (nv50->rasterizer->pipe.point_sprite) {
- nv50_pntc_replace(nv50, pcrd, (reg[4] >> 8) & 0xff);
+ so_method(so, tesla, NV50TCL_POINT_SPRITE_CTRL, 1);
+ so_data (so,
+ nv50_pntc_replace(nv50, pcrd, (reg[4] >> 8) & 0xff));
so_method(so, tesla, NV50TCL_POINT_COORD_REPLACE_MAP(0), 8);
so_datap (so, pcrd, 8);
}
- so_ref(so, &nv50->state.programs);
- so_ref(NULL, &so);
+ so_method(so, tesla, NV50TCL_GP_ENABLE, 1);
+ so_data (so, (vp->type == PIPE_SHADER_GEOMETRY) ? 1 : 0);
+
+ so_ref(so, &nv50->state.fp_linkage);
+ so_ref(NULL, &so);
+}
+
+static int
+construct_vp_gp_mapping(uint32_t *map32, int m,
+ struct nv50_program *vp, struct nv50_program *gp)
+{
+ uint8_t *map = (uint8_t *)map32;
+ int i, j, c;
+
+ for (i = 0; i < gp->cfg.in_nr; ++i) {
+ uint8_t oid, mv = 0, mg = gp->cfg.in[i].mask;
+
+ for (j = 0; j < vp->cfg.out_nr; ++j) {
+ if (vp->cfg.out[j].sn == gp->cfg.in[i].sn &&
+ vp->cfg.out[j].si == gp->cfg.in[i].si) {
+ mv = vp->cfg.out[j].mask;
+ oid = vp->cfg.out[j].hw;
+ break;
+ }
+ }
+
+ for (c = 0; c < 4; ++c, mv >>= 1, mg >>= 1) {
+ if (mg & mv & 1)
+ map[m++] = oid;
+ else
+ if (mg & 1)
+ map[m++] = (c == 3) ? 0x41 : 0x40;
+ oid += mv & 1;
+ }
+ }
+ return m;
+}
+
+void
+nv50_gp_linkage_validate(struct nv50_context *nv50)
+{
+ struct nouveau_grobj *tesla = nv50->screen->tesla;
+ struct nouveau_stateobj *so;
+ struct nv50_program *vp = nv50->vertprog;
+ struct nv50_program *gp = nv50->geomprog;
+ uint32_t map[16];
+ int m = 0;
+
+ if (!gp) {
+ so_ref(NULL, &nv50->state.gp_linkage);
+ return;
+ }
+ memset(map, 0, sizeof(map));
+
+ m = construct_vp_gp_mapping(map, m, vp, gp);
+
+ so = so_new(2, 14, 0);
+
+ assert(m <= 32);
+ so_method(so, tesla, NV50TCL_VP_RESULT_MAP_SIZE, 1);
+ so_data (so, m);
+
+ m = (m + 3) / 4;
+ so_method(so, tesla, NV50TCL_VP_RESULT_MAP(0), m);
+ so_datap (so, map, m);
+
+ so_ref(so, &nv50->state.gp_linkage);
+ so_ref(NULL, &so);
}
void