#define NV50_MOD_ABS 2
#define NV50_MOD_SAT 4
-/* arbitrary limits */
-#define MAX_IF_DEPTH 4
-#define MAX_LOOP_DEPTH 4
+/* STACK: Conditionals and loops have to use the (per warp) stack.
+ * Stack entries consist of an entry type (divergent path, join at),
+ * a mask indicating the active threads of the warp, and an address.
+ * MPs can store 12 stack entries internally, if we need more (and
+ * we probably do), we have to create a stack buffer in VRAM.
+ */
+/* impose low limits for now */
+#define NV50_MAX_COND_NESTING 4
+#define NV50_MAX_LOOP_NESTING 3
+
+#define JOIN_ON(e) e; pc->p->exec_tail->inst[1] |= 2
struct nv50_pc {
struct nv50_program *p;
struct nv50_reg *param;
int param_nr;
struct nv50_reg *immd;
- float *immd_buf;
+ uint32_t *immd_buf;
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];
unsigned temp_temp_nr;
struct nv50_reg *iv_p;
struct nv50_reg *iv_c;
- struct nv50_program_exec *if_cond;
- struct nv50_program_exec *if_insn[MAX_IF_DEPTH];
- struct nv50_program_exec *br_join[MAX_IF_DEPTH];
- struct nv50_program_exec *br_loop[MAX_LOOP_DEPTH]; /* for BRK branch */
+ struct nv50_program_exec *if_insn[NV50_MAX_COND_NESTING];
+ struct nv50_program_exec *if_join[NV50_MAX_COND_NESTING];
+ struct nv50_program_exec *loop_brka[NV50_MAX_LOOP_NESTING];
int if_lvl, loop_lvl;
- unsigned loop_pos[MAX_LOOP_DEPTH];
+ unsigned loop_pos[NV50_MAX_LOOP_NESTING];
+
+ unsigned *insn_pos; /* actual program offset of each TGSI insn */
+ boolean in_subroutine;
/* current instruction and total number of insns */
unsigned insn_cur;
unsigned insn_nr;
boolean allow32;
-};
-static INLINE struct nv50_reg *
-reg_instance(struct nv50_pc *pc, struct nv50_reg *reg)
-{
- struct nv50_reg *dup = NULL;
- if (reg) {
- assert(pc->reg_instance_nr < 16);
- dup = &pc->reg_instances[pc->reg_instance_nr++];
- *dup = *reg;
- reg->mod = 0;
- }
- return dup;
-}
+ uint8_t edgeflag_out;
+};
static INLINE void
ctor_reg(struct nv50_reg *reg, unsigned type, int index, int hw)
/* 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;
+ pc->r_addr[i].rhw = -1;
}
static void
assert(0);
}
+static INLINE struct nv50_reg *
+reg_instance(struct nv50_pc *pc, struct nv50_reg *reg)
+{
+ struct nv50_reg *ri;
+
+ assert(pc->reg_instance_nr < 16);
+ ri = &pc->reg_instances[pc->reg_instance_nr++];
+ if (reg) {
+ alloc_reg(pc, reg);
+ *ri = *reg;
+ reg->mod = 0;
+ }
+ return ri;
+}
+
/* XXX: For shaders that aren't executed linearly (e.g. shaders that
* contain loops), we need to assign all hw regs to TGSI TEMPs early,
* lest we risk temp_temps overwriting regs alloc'd "later".
return NULL;
}
-/* Assign the hw of the discarded temporary register src
- * to the tgsi register dst and free src.
- */
-static void
-assimilate_temp(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
-{
- assert(src->index == -1 && src->hw != -1);
-
- if (dst->hw != -1)
- pc->r_temp[dst->hw] = NULL;
- pc->r_temp[src->hw] = dst;
- dst->hw = src->hw;
-
- FREE(src);
-}
-
/* release the hardware resource held by r */
static void
release_hw(struct nv50_pc *pc, struct nv50_reg *r)
kill_temp_temp(struct nv50_pc *pc)
{
int i;
-
+
for (i = 0; i < pc->temp_temp_nr; i++)
free_temp(pc, pc->temp_temp[i]);
pc->temp_temp_nr = 0;
}
static int
-ctor_immd(struct nv50_pc *pc, float x, float y, float z, float w)
+ctor_immd_4u32(struct nv50_pc *pc,
+ uint32_t x, uint32_t y, uint32_t z, uint32_t w)
{
- pc->immd_buf = REALLOC(pc->immd_buf, (pc->immd_nr * 4 * sizeof(float)),
- (pc->immd_nr + 1) * 4 * sizeof(float));
+ unsigned size = pc->immd_nr * 4 * sizeof(uint32_t);
+
+ pc->immd_buf = REALLOC(pc->immd_buf, size, size + 4 * sizeof(uint32_t));
+
pc->immd_buf[(pc->immd_nr * 4) + 0] = x;
pc->immd_buf[(pc->immd_nr * 4) + 1] = y;
pc->immd_buf[(pc->immd_nr * 4) + 2] = z;
pc->immd_buf[(pc->immd_nr * 4) + 3] = w;
-
+
return pc->immd_nr++;
}
+static INLINE int
+ctor_immd_4f32(struct nv50_pc *pc, float x, float y, float z, float w)
+{
+ return ctor_immd_4u32(pc, fui(x), fui(y), fui(z), fui(w));
+}
+
static struct nv50_reg *
alloc_immd(struct nv50_pc *pc, float f)
{
unsigned hw;
for (hw = 0; hw < pc->immd_nr * 4; hw++)
- if (pc->immd_buf[hw] == f)
+ if (pc->immd_buf[hw] == fui(f))
break;
if (hw == pc->immd_nr * 4)
- hw = ctor_immd(pc, f, -f, 0.5 * f, 0) * 4;
+ hw = ctor_immd_4f32(pc, f, -f, 0.5 * f, 0) * 4;
ctor_reg(r, P_IMMD, -1, hw);
return r;
return FALSE;
}
+static boolean
+is_join(struct nv50_program_exec *e)
+{
+ if (is_long(e) && (e->inst[1] & 3) == 2)
+ return TRUE;
+ return FALSE;
+}
+
static INLINE void
set_pred(struct nv50_pc *pc, unsigned pred, unsigned idx,
struct nv50_program_exec *e)
{
+ assert(!is_immd(e));
set_long(pc, e);
e->inst[1] &= ~((0x1f << 7) | (0x3 << 12));
e->inst[1] |= (pred << 7) | (idx << 12);
static INLINE void
set_immd(struct nv50_pc *pc, struct nv50_reg *imm, struct nv50_program_exec *e)
{
- unsigned val;
- float f = pc->immd_buf[imm->hw];
-
- if (imm->mod & NV50_MOD_ABS)
- f = fabsf(f);
- val = fui((imm->mod & NV50_MOD_NEG) ? -f : f);
-
set_long(pc, e);
- /*XXX: can't be predicated - bits overlap.. catch cases where both
- * are required and avoid them. */
+ /* XXX: can't be predicated - bits overlap; cases where both
+ * are required should be avoided by using pc->allow32 */
set_pred(pc, 0, 0, e);
set_pred_wr(pc, 0, 0, e);
e->inst[1] |= 0x00000002 | 0x00000001;
- e->inst[0] |= (val & 0x3f) << 16;
- e->inst[1] |= (val >> 6) << 2;
+ e->inst[0] |= (pc->immd_buf[imm->hw] & 0x3f) << 16;
+ e->inst[1] |= (pc->immd_buf[imm->hw] >> 6) << 2;
}
static INLINE void
static struct nv50_reg *
alloc_addr(struct nv50_pc *pc, struct nv50_reg *ref)
{
- int i;
struct nv50_reg *a_tgsi = NULL, *a = NULL;
+ int i;
+ uint8_t avail = ~pc->addr_alloc;
if (!ref) {
- /* allocate for TGSI address reg */
- 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;
+ /* allocate for TGSI_FILE_ADDRESS */
+ while (avail) {
+ i = ffs(avail) - 1;
- pc->r_addr[i].rhw = -1;
- pc->r_addr[i].index = i;
- return &pc->r_addr[i];
+ 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 has index < 0, it is not reserved for TGSI,
- * and index will be the negative of the TGSI addr index the
- * value in rhw is relative to, or -256 if rhw is an offset
- * from 0. If rhw < 0, the reg has not been initialized.
+ * 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.
*/
- 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 */
+ 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 (ref->hw - pc->r_addr[i].rhw >= 128)
continue;
- if ((ref->acc >= 0 && pc->r_addr[i].index == -256) ||
- (ref->acc < 0 && -pc->r_addr[i].index == ref->index)) {
+ 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];
}
a->rhw = ref->hw & ~0x7f;
a->acc = pc->insn_cur;
- a->index = a_tgsi ? -ref->index : -256;
+ a->index = a_tgsi ? ref->index : -1;
return a;
}
e->inst[1] |= (((src->type == P_IMMD) ? 0 : 1) << 22);
}
+/* Never apply nv50_reg::mod in emit_mov, or carefully check the code !!! */
static void
emit_mov(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
{
if (src->type == P_IMMD || src->type == P_CONST) {
set_long(pc, e);
set_data(pc, src, 0x7f, 9, e);
- e->inst[1] |= 0x20000000; /* src0 const? */
+ e->inst[1] |= 0x20000000; /* mov from c[] */
} else {
if (src->type == P_ATTR) {
set_long(pc, e);
if (is_long(e) && !is_immd(e)) {
e->inst[1] |= 0x04000000; /* 32-bit */
- e->inst[1] |= 0x0000c000; /* "subsubop" 0x3 */
+ e->inst[1] |= 0x0000c000; /* 32-bit c[] load / lane mask 0:1 */
if (!(e->inst[1] & 0x20000000))
- e->inst[1] |= 0x00030000; /* "subsubop" 0xf */
+ e->inst[1] |= 0x00030000; /* lane mask 2:3 */
} else
e->inst[0] |= 0x00008000;
FREE(imm);
}
+/* Assign the hw of the discarded temporary register src
+ * to the tgsi register dst and free src.
+ */
+static void
+assimilate_temp(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
+{
+ assert(src->index == -1 && src->hw != -1);
+
+ if (pc->if_lvl || pc->loop_lvl ||
+ (dst->type != P_TEMP) ||
+ (src->hw < pc->result_nr * 4 &&
+ pc->p->type == PIPE_SHADER_FRAGMENT) ||
+ pc->p->info.opcode_count[TGSI_OPCODE_CAL] ||
+ pc->p->info.opcode_count[TGSI_OPCODE_BRA]) {
+
+ emit_mov(pc, dst, src);
+ free_temp(pc, src);
+ return;
+ }
+
+ if (dst->hw != -1)
+ pc->r_temp[dst->hw] = NULL;
+ pc->r_temp[src->hw] = dst;
+ dst->hw = src->hw;
+
+ FREE(src);
+}
+
+static void
+emit_nop(struct nv50_pc *pc)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0xf0000000;
+ set_long(pc, e);
+ e->inst[1] = 0xe0000000;
+ emit(pc, e);
+}
+
static boolean
check_swap_src_0_1(struct nv50_pc *pc,
struct nv50_reg **s0, struct nv50_reg **s1)
e->inst[1] |= ((src->hw & 127) << 14);
}
+static void
+emit_mov_from_pred(struct nv50_pc *pc, struct nv50_reg *dst, int pred)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ assert(dst->type == P_TEMP);
+ e->inst[1] = 0x20000000 | (pred << 12);
+ set_long(pc, e);
+ set_dst(pc, dst, e);
+
+ emit(pc, e);
+}
+
+static void
+emit_mov_to_pred(struct nv50_pc *pc, int pred, struct nv50_reg *src)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0x000001fc;
+ e->inst[1] = 0xa0000008;
+ set_long(pc, e);
+ set_pred_wr(pc, 1, pred, e);
+ set_src_0_restricted(pc, src, e);
+
+ emit(pc, e);
+}
+
static void
emit_mul(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src0,
struct nv50_reg *src1)
set_dst(pc, dst, e);
set_src_0(pc, src0, e);
if (src1->type == P_IMMD && !is_long(e)) {
- if (src0->mod & NV50_MOD_NEG)
+ if (src0->mod ^ src1->mod)
e->inst[0] |= 0x00008000;
set_immd(pc, src1, e);
} else {
op != TGSI_OPCODE_XOR)
assert(!"invalid bit op");
+ assert(!(src0->mod | src1->mod));
+
if (src1->type == P_IMMD && src0->type == P_TEMP && pc->allow32) {
set_immd(pc, src1, e);
if (op == TGSI_OPCODE_OR)
src2->mod ^= NV50_MOD_NEG;
}
+#define NV50_FLOP_RCP 0
+#define NV50_FLOP_RSQ 2
+#define NV50_FLOP_LG2 3
+#define NV50_FLOP_SIN 4
+#define NV50_FLOP_COS 5
+#define NV50_FLOP_EX2 6
+
+/* rcp, rsqrt, lg2 support neg and abs */
static void
emit_flop(struct nv50_pc *pc, unsigned sub,
struct nv50_reg *dst, struct nv50_reg *src)
struct nv50_program_exec *e = exec(pc);
e->inst[0] |= 0x90000000;
- if (sub) {
+ if (sub || src->mod) {
set_long(pc, e);
e->inst[1] |= (sub << 29);
}
set_dst(pc, dst, e);
+ set_src_0_restricted(pc, src, e);
- if (sub == 0 || sub == 2)
- set_src_0_restricted(pc, src, e);
- else
- set_src_0(pc, src, e);
+ assert(!src->mod || sub < 4);
+
+ if (src->mod & NV50_MOD_NEG)
+ e->inst[1] |= 0x04000000;
+ if (src->mod & NV50_MOD_ABS)
+ e->inst[1] |= 0x00100000;
emit(pc, e);
}
set_long(pc, e);
e->inst[1] |= (6 << 29) | 0x00004000;
+ if (src->mod & NV50_MOD_NEG)
+ e->inst[1] |= 0x04000000;
+ if (src->mod & NV50_MOD_ABS)
+ e->inst[1] |= 0x00100000;
+
emit(pc, e);
}
set_long(pc, e);
e->inst[1] |= (6 << 29);
+ if (src->mod & NV50_MOD_NEG)
+ e->inst[1] |= 0x04000000;
+ if (src->mod & NV50_MOD_ABS)
+ e->inst[1] |= 0x00100000;
+
emit(pc, e);
}
set_src_1(pc, src1, e);
emit(pc, e);
- pc->if_cond = pc->p->exec_tail; /* record for OPCODE_IF */
/* cvt.f32.u32/s32 (?) if we didn't only write the predicate */
if (rdst)
{
struct nv50_reg *temp = alloc_temp(pc, NULL);
- emit_flop(pc, 3, temp, v);
+ emit_flop(pc, NV50_FLOP_LG2, temp, v);
emit_mul(pc, temp, temp, e);
emit_preex2(pc, temp, temp);
- emit_flop(pc, 6, dst, temp);
+ emit_flop(pc, NV50_FLOP_EX2, dst, temp);
free_temp(pc, temp);
}
{
struct nv50_program_exec *e;
const int r_pred = 1;
- unsigned cvn = CVT_F32_F32;
-
- if (src->mod & NV50_MOD_NEG)
- cvn |= CVT_NEG;
- /* write predicate reg */
- emit_cvt(pc, NULL, src, r_pred, CVTOP_RN, cvn);
- /* conditional discard */
e = exec(pc);
- e->inst[0] = 0x00000002;
+ e->inst[0] = 0x00000002; /* discard */
+ 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);
+ }
+
+ emit(pc, e);
+}
+
+static struct nv50_program_exec *
+emit_control_flow(struct nv50_pc *pc, unsigned op, int pred, unsigned cc)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = (op << 28) | 2;
set_long(pc, e);
- set_pred(pc, 0x1 /* LT */, r_pred, e);
+ if (pred >= 0)
+ set_pred(pc, cc, pred, e);
+
emit(pc, e);
+ return e;
+}
+
+static INLINE struct nv50_program_exec *
+emit_breakaddr(struct nv50_pc *pc)
+{
+ return emit_control_flow(pc, 0x4, -1, 0);
+}
+
+static INLINE void
+emit_break(struct nv50_pc *pc, int pred, unsigned cc)
+{
+ emit_control_flow(pc, 0x5, pred, cc);
+}
+
+static INLINE struct nv50_program_exec *
+emit_joinat(struct nv50_pc *pc)
+{
+ return emit_control_flow(pc, 0xa, -1, 0);
+}
+
+static INLINE struct nv50_program_exec *
+emit_branch(struct nv50_pc *pc, int pred, unsigned cc)
+{
+ return emit_control_flow(pc, 0x1, pred, cc);
+}
+
+static INLINE struct nv50_program_exec *
+emit_call(struct nv50_pc *pc, int pred, unsigned cc)
+{
+ return emit_control_flow(pc, 0x2, pred, cc);
+}
+
+static INLINE void
+emit_ret(struct nv50_pc *pc, int pred, unsigned cc)
+{
+ emit_control_flow(pc, 0x3, pred, cc);
+}
+
+#define QOP_ADD 0
+#define QOP_SUBR 1
+#define QOP_SUB 2
+#define QOP_MOV_SRC1 3
+
+/* For a quad of threads / top left, top right, bottom left, bottom right
+ * pixels, do a different operation, and take src0 from a specific thread.
+ */
+static void
+emit_quadop(struct nv50_pc *pc, struct nv50_reg *dst, int wp, int lane_src0,
+ struct nv50_reg *src0, struct nv50_reg *src1, ubyte qop)
+{
+ struct nv50_program_exec *e = exec(pc);
+
+ e->inst[0] = 0xc0000000;
+ e->inst[1] = 0x80000000;
+ set_long(pc, e);
+ e->inst[0] |= lane_src0 << 16;
+ set_src_0(pc, src0, e);
+ set_src_2(pc, src1, e);
+
+ if (wp >= 0)
+ set_pred_wr(pc, 1, wp, e);
+
+ if (dst)
+ set_dst(pc, dst, e);
+ else {
+ e->inst[0] |= 0x000001fc;
+ e->inst[1] |= 0x00000008;
+ }
+
+ e->inst[0] |= (qop & 3) << 20;
+ e->inst[1] |= (qop >> 2) << 22;
+
+ emit(pc, e);
}
static void
if (arg == 4) /* there is no textureProj(samplerCubeShadow) */
emit_mov(pc, t[3], src[3]);
- emit_flop(pc, 0, t[2], t[2]);
+ emit_flop(pc, NV50_FLOP_RCP, t[2], t[2]);
emit_mul(pc, t[0], src[0], t[2]);
emit_mul(pc, t[1], src[1], t[2]);
t[3]->rhw = src[3]->rhw;
emit_interp(pc, t[3], NULL, (mode & INTERP_CENTROID));
- emit_flop(pc, 0, t[3], t[3]);
+ emit_flop(pc, NV50_FLOP_RCP, t[3], t[3]);
for (c = 0; c < dim; ++c) {
t[c]->rhw = src[c]->rhw;
/* XXX: for some reason the blob sometimes uses MAD
* (mad f32 $rX $rY $rZ neg $r63)
*/
- emit_flop(pc, 0, t[3], src[3]);
+ emit_flop(pc, NV50_FLOP_RCP, t[3], src[3]);
for (c = 0; c < dim; ++c)
emit_mul(pc, t[c], src[c], t[3]);
if (arg != dim) /* depth reference value */
}
}
+/* We shouldn't execute TEXLOD if any of the pixels in a quad have
+ * different LOD values, so branch off groups of equal LOD.
+ */
+static void
+emit_texlod_sequence(struct nv50_pc *pc, struct nv50_reg *tlod,
+ struct nv50_reg *src, struct nv50_program_exec *tex)
+{
+ struct nv50_program_exec *join_at;
+ unsigned i, target = pc->p->exec_size + 9 * 2;
+
+ if (pc->p->type != PIPE_SHADER_FRAGMENT) {
+ emit(pc, tex);
+ return;
+ }
+ pc->allow32 = FALSE;
+
+ /* Subtract lod of each pixel from lod of top left pixel, jump
+ * texlod insn if result is 0, then repeat for 2 other pixels.
+ */
+ join_at = emit_joinat(pc);
+ emit_quadop(pc, NULL, 0, 0, tlod, tlod, 0x55);
+ emit_branch(pc, 0, 2)->param.index = target;
+
+ for (i = 1; i < 4; ++i) {
+ emit_quadop(pc, NULL, 0, i, tlod, tlod, 0x55);
+ emit_branch(pc, 0, 2)->param.index = target;
+ }
+
+ emit_mov(pc, tlod, src); /* target */
+ emit(pc, tex); /* texlod */
+
+ join_at->param.index = target + 2 * 2;
+ JOIN_ON(emit_nop(pc)); /* join _after_ tex */
+}
+
+static void
+emit_texbias_sequence(struct nv50_pc *pc, struct nv50_reg *t[4], unsigned arg,
+ struct nv50_program_exec *tex)
+{
+ struct nv50_program_exec *e;
+ struct nv50_reg imm_1248, *t123[4][4], *r_bits = alloc_temp(pc, NULL);
+ int r_pred = 0;
+ unsigned n, c, i, cc[4] = { 0x0a, 0x13, 0x11, 0x10 };
+
+ pc->allow32 = FALSE;
+ ctor_reg(&imm_1248, P_IMMD, -1, ctor_immd_4u32(pc, 1, 2, 4, 8) * 4);
+
+ /* Subtract bias value of thread i from bias values of each thread,
+ * store result in r_pred, and set bit i in r_bits if result was 0.
+ */
+ assert(arg < 4);
+ for (i = 0; i < 4; ++i, ++imm_1248.hw) {
+ emit_quadop(pc, NULL, r_pred, i, t[arg], t[arg], 0x55);
+ emit_mov(pc, r_bits, &imm_1248);
+ set_pred(pc, 2, r_pred, pc->p->exec_tail);
+ }
+ emit_mov_to_pred(pc, r_pred, r_bits);
+
+ /* The lanes of a quad are now grouped by the bit in r_pred they have
+ * set. Put the input values for TEX into a new register set for each
+ * group and execute TEX only for a specific group.
+ * We cannot use the same register set for each group because we need
+ * the derivatives, which are implicitly calculated, to be correct.
+ */
+ for (i = 1; i < 4; ++i) {
+ alloc_temp4(pc, t123[i], 0);
+
+ for (c = 0; c <= arg; ++c)
+ emit_mov(pc, t123[i][c], t[c]);
+
+ *(e = exec(pc)) = *(tex);
+ e->inst[0] &= ~0x01fc;
+ set_dst(pc, t123[i][0], e);
+ set_pred(pc, cc[i], r_pred, e);
+ emit(pc, e);
+ }
+ /* finally TEX on the original regs (where we kept the input) */
+ set_pred(pc, cc[0], r_pred, tex);
+ emit(pc, tex);
+
+ /* put the 3 * n other results into regs for lane 0 */
+ n = popcnt4(((e->inst[0] >> 25) & 0x3) | ((e->inst[1] >> 12) & 0xc));
+ for (i = 1; i < 4; ++i) {
+ for (c = 0; c < n; ++c) {
+ emit_mov(pc, t[c], t123[i][c]);
+ set_pred(pc, cc[i], r_pred, pc->p->exec_tail);
+ }
+ free_temp4(pc, t123[i]);
+ }
+
+ emit_nop(pc);
+ free_temp(pc, r_bits);
+}
+
static void
emit_tex(struct nv50_pc *pc, struct nv50_reg **dst, unsigned mask,
struct nv50_reg **src, unsigned unit, unsigned type,
emit_mov(pc, t[dim], src[2]);
}
- if (bias_lod) {
- assert(arg < 4);
- emit_mov(pc, t[arg++], src[3]);
- e->inst[1] |= (bias_lod < 0) ? 0x20000000 : 0x40000000;
- }
-
- e->inst[0] |= (arg - 1) << 22;
-
e->inst[0] |= (mask & 0x3) << 25;
e->inst[1] |= (mask & 0xc) << 12;
- emit(pc, e);
+ if (!bias_lod) {
+ e->inst[0] |= (arg - 1) << 22;
+ emit(pc, e);
+ } else
+ if (bias_lod < 0) {
+ assert(pc->p->type == PIPE_SHADER_FRAGMENT);
+ e->inst[0] |= arg << 22;
+ e->inst[1] |= 0x20000000; /* texbias */
+ emit_mov(pc, t[arg], src[3]);
+ emit_texbias_sequence(pc, t, arg, e);
+ } else {
+ e->inst[0] |= arg << 22;
+ e->inst[1] |= 0x40000000; /* texlod */
+ emit_mov(pc, t[arg], src[3]);
+ emit_texlod_sequence(pc, t[arg], src[3], e);
+ }
+
#if 1
c = 0;
if (mask & 1) emit_mov(pc, dst[0], t[c++]);
#endif
}
-static void
-emit_branch(struct nv50_pc *pc, int pred, unsigned cc,
- struct nv50_program_exec **join)
-{
- struct nv50_program_exec *e = exec(pc);
-
- if (join) {
- set_long(pc, e);
- e->inst[0] |= 0xa0000002;
- emit(pc, e);
- *join = e;
- e = exec(pc);
- }
-
- set_long(pc, e);
- e->inst[0] |= 0x10000002;
- if (pred >= 0)
- set_pred(pc, cc, pred, e);
- emit(pc, e);
-}
-
-static void
-emit_nop(struct nv50_pc *pc)
-{
- struct nv50_program_exec *e = exec(pc);
-
- e->inst[0] = 0xf0000000;
- set_long(pc, e);
- e->inst[1] = 0xe0000000;
- emit(pc, e);
-}
-
static void
emit_ddx(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
{
assert(src->type == P_TEMP);
- e->inst[0] = 0xc0140000;
- e->inst[1] = 0x89800000;
+ e->inst[0] = (src->mod & NV50_MOD_NEG) ? 0xc0240000 : 0xc0140000;
+ e->inst[1] = (src->mod & NV50_MOD_NEG) ? 0x86400000 : 0x89800000;
set_long(pc, e);
set_dst(pc, dst, e);
set_src_0(pc, src, e);
static void
emit_ddy(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
{
- struct nv50_reg *r = src;
struct nv50_program_exec *e = exec(pc);
assert(src->type == P_TEMP);
- if (!(src->mod & NV50_MOD_NEG)) { /* ! double negation */
- r = alloc_temp(pc, NULL);
- emit_neg(pc, r, src);
- }
-
- e->inst[0] = 0xc0150000;
- e->inst[1] = 0x8a400000;
+ e->inst[0] = (src->mod & NV50_MOD_NEG) ? 0xc0250000 : 0xc0150000;
+ e->inst[1] = (src->mod & NV50_MOD_NEG) ? 0x85800000 : 0x8a400000;
set_long(pc, e);
set_dst(pc, dst, e);
- set_src_0(pc, r, e);
- set_src_2(pc, r, e);
-
- if (r != src)
- free_temp(pc, r);
+ set_src_0(pc, src, e);
+ set_src_2(pc, src, e);
emit(pc, e);
}
negate_supported(const struct tgsi_full_instruction *insn, int i)
{
switch (insn->Instruction.Opcode) {
+ case TGSI_OPCODE_ADD:
+ case TGSI_OPCODE_COS:
+ case TGSI_OPCODE_DDX:
case TGSI_OPCODE_DDY:
case TGSI_OPCODE_DP3:
case TGSI_OPCODE_DP4:
- case TGSI_OPCODE_MUL:
+ case TGSI_OPCODE_EX2:
case TGSI_OPCODE_KIL:
- case TGSI_OPCODE_ADD:
- case TGSI_OPCODE_SUB:
+ case TGSI_OPCODE_LG2:
case TGSI_OPCODE_MAD:
- return TRUE;
+ case TGSI_OPCODE_MUL:
case TGSI_OPCODE_POW:
- if (i == 1)
- return TRUE;
- return FALSE;
+ case TGSI_OPCODE_RCP:
+ case TGSI_OPCODE_RSQ: /* ignored, RSQ = rsqrt(abs(src.x)) */
+ case TGSI_OPCODE_SCS:
+ case TGSI_OPCODE_SIN:
+ case TGSI_OPCODE_SUB:
+ return TRUE;
default:
return FALSE;
}
case TGSI_OPCODE_DST:
return mask & (c ? 0xa : 0x6);
case TGSI_OPCODE_EX2:
+ case TGSI_OPCODE_EXP:
case TGSI_OPCODE_LG2:
+ case TGSI_OPCODE_LOG:
case TGSI_OPCODE_POW:
case TGSI_OPCODE_RCP:
case TGSI_OPCODE_RSQ:
/* Indicate indirection by setting r->acc < 0 and
* use the index field to select the address reg.
*/
- r = MALLOC_STRUCT(nv50_reg);
+ r = reg_instance(pc, NULL);
swz = tgsi_util_get_src_register_swizzle(
&src->Indirect, 0);
ctor_reg(r, P_CONST,
break;
}
+ if (r && r->acc >= 0 && r != temp)
+ return reg_instance(pc, r);
return r;
}
assert(0);
return 0x0;
}
+ case TGSI_OPCODE_EXP:
+ case TGSI_OPCODE_LOG:
case TGSI_OPCODE_LIT:
case TGSI_OPCODE_SCS:
case TGSI_OPCODE_TEX:
if (pc->if_insn[lvl]->next != pc->p->exec_tail)
return FALSE;
+ if (is_immd(pc->p->exec_tail))
+ return FALSE;
/* if ccode == 'true', the BRA is from an ELSE and the predicate
* reg may no longer be valid, since we currently always use $p0
*/
if (has_pred(pc->if_insn[lvl], 0xf))
return FALSE;
- assert(pc->if_insn[lvl] && pc->br_join[lvl]);
+ assert(pc->if_insn[lvl] && pc->if_join[lvl]);
- /* We'll use the exec allocated for JOIN_AT (as we can't easily
- * update prev's next); if exec_tail is BRK, update the pointer.
+ /* We'll use the exec allocated for JOIN_AT (we can't easily
+ * access nv50_program_exec's prev).
*/
- if (pc->loop_lvl && pc->br_loop[pc->loop_lvl - 1] == pc->p->exec_tail)
- pc->br_loop[pc->loop_lvl - 1] = pc->br_join[lvl];
-
pc->p->exec_size -= 4; /* remove JOIN_AT and BRA */
- *pc->br_join[lvl] = *pc->p->exec_tail;
+ *pc->if_join[lvl] = *pc->p->exec_tail;
FREE(pc->if_insn[lvl]);
FREE(pc->p->exec_tail);
- pc->p->exec_tail = pc->br_join[lvl];
+ pc->p->exec_tail = pc->if_join[lvl];
pc->p->exec_tail->next = NULL;
set_pred(pc, 0xd, 0, pc->p->exec_tail);
return TRUE;
}
+static void
+nv50_fp_move_results(struct nv50_pc *pc)
+{
+ struct nv50_reg reg;
+ unsigned i;
+
+ ctor_reg(®, P_TEMP, -1, -1);
+
+ for (i = 0; i < pc->result_nr * 4; ++i) {
+ if (pc->result[i].rhw < 0 || pc->result[i].hw < 0)
+ continue;
+ if (pc->result[i].rhw != pc->result[i].hw) {
+ reg.hw = pc->result[i].rhw;
+ emit_mov(pc, ®, &pc->result[i]);
+ }
+ }
+}
+
static boolean
nv50_program_tx_insn(struct nv50_pc *pc,
const struct tgsi_full_instruction *inst)
for (c = 0; c < 4; c++)
if (src_mask & (1 << c))
- src[i][c] = reg_instance(pc,
- tgsi_src(pc, c, fs, neg_supp));
+ src[i][c] = tgsi_src(pc, c, fs, neg_supp);
}
brdc = temp = pc->r_brdc;
emit_arl(pc, dst[0], temp, 4);
break;
case TGSI_OPCODE_BGNLOOP:
+ pc->loop_brka[pc->loop_lvl] = emit_breakaddr(pc);
pc->loop_pos[pc->loop_lvl++] = pc->p->exec_size;
terminate_mbb(pc);
break;
+ case TGSI_OPCODE_BGNSUB:
+ assert(!pc->in_subroutine);
+ pc->in_subroutine = TRUE;
+ /* probably not necessary, but align to 8 byte boundary */
+ if (!is_long(pc->p->exec_tail))
+ convert_to_long(pc, pc->p->exec_tail);
+ break;
case TGSI_OPCODE_BRK:
- emit_branch(pc, -1, 0, NULL);
assert(pc->loop_lvl > 0);
- pc->br_loop[pc->loop_lvl - 1] = pc->p->exec_tail;
+ emit_break(pc, -1, 0);
+ break;
+ case TGSI_OPCODE_CAL:
+ assert(inst->Label.Label < pc->insn_nr);
+ emit_call(pc, -1, 0)->param.index = inst->Label.Label;
+ /* replaced by actual offset in nv50_program_fixup_insns */
break;
case TGSI_OPCODE_CEIL:
for (c = 0; c < 4; c++) {
set_pred(pc, 0x6, 1, pc->p->exec_tail); /* @NSF */
}
break;
+ case TGSI_OPCODE_CONT:
+ assert(pc->loop_lvl > 0);
+ emit_branch(pc, -1, 0)->param.index =
+ pc->loop_pos[pc->loop_lvl - 1];
+ break;
case TGSI_OPCODE_COS:
if (mask & 8) {
emit_precossin(pc, temp, src[0][3]);
- emit_flop(pc, 5, dst[3], temp);
+ emit_flop(pc, NV50_FLOP_COS, dst[3], temp);
if (!(mask &= 7))
break;
if (temp == dst[3])
temp = brdc = temp_temp(pc);
}
emit_precossin(pc, temp, src[0][0]);
- emit_flop(pc, 5, brdc, temp);
+ emit_flop(pc, NV50_FLOP_COS, brdc, temp);
break;
case TGSI_OPCODE_DDX:
for (c = 0; c < 4; c++) {
emit_mov_immdval(pc, dst[0], 1.0f);
break;
case TGSI_OPCODE_ELSE:
- emit_branch(pc, -1, 0, NULL);
+ emit_branch(pc, -1, 0);
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);
if (nv50_kill_branch(pc) == TRUE)
break;
- if (pc->br_join[pc->if_lvl]) {
- pc->br_join[pc->if_lvl]->param.index = pc->p->exec_size;
- pc->br_join[pc->if_lvl] = NULL;
+ if (pc->if_join[pc->if_lvl]) {
+ pc->if_join[pc->if_lvl]->param.index = pc->p->exec_size;
+ pc->if_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
*/
- emit_nop(pc);
- pc->p->exec_tail->inst[1] |= 2;
+ JOIN_ON(emit_nop(pc));
break;
case TGSI_OPCODE_ENDLOOP:
- 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;
+ emit_branch(pc, -1, 0)->param.index =
+ pc->loop_pos[--pc->loop_lvl];
+ pc->loop_brka[pc->loop_lvl]->param.index = pc->p->exec_size;
terminate_mbb(pc);
break;
+ case TGSI_OPCODE_ENDSUB:
+ assert(pc->in_subroutine);
+ pc->in_subroutine = FALSE;
+ break;
case TGSI_OPCODE_EX2:
emit_preex2(pc, temp, src[0][0]);
- emit_flop(pc, 6, brdc, temp);
+ emit_flop(pc, NV50_FLOP_EX2, brdc, temp);
+ break;
+ case TGSI_OPCODE_EXP:
+ {
+ struct nv50_reg *t[2];
+
+ assert(!temp);
+ t[0] = temp_temp(pc);
+ t[1] = temp_temp(pc);
+
+ if (mask & 0x6)
+ emit_mov(pc, t[0], src[0][0]);
+ if (mask & 0x3)
+ emit_flr(pc, t[1], src[0][0]);
+
+ if (mask & (1 << 1))
+ emit_sub(pc, dst[1], t[0], t[1]);
+ if (mask & (1 << 0)) {
+ emit_preex2(pc, t[1], t[1]);
+ emit_flop(pc, NV50_FLOP_EX2, dst[0], t[1]);
+ }
+ if (mask & (1 << 2)) {
+ emit_preex2(pc, t[0], t[0]);
+ emit_flop(pc, NV50_FLOP_EX2, dst[2], t[0]);
+ }
+ if (mask & (1 << 3))
+ emit_mov_immdval(pc, dst[3], 1.0f);
+ }
break;
case TGSI_OPCODE_FLR:
for (c = 0; c < 4; c++) {
}
break;
case TGSI_OPCODE_IF:
- /* emitting a join_at may not be necessary */
- assert(pc->if_lvl < MAX_IF_DEPTH);
- /* set_pred_wr(pc, 1, 0, pc->if_cond); */
+ assert(pc->if_lvl < NV50_MAX_COND_NESTING);
emit_cvt(pc, NULL, src[0][0], 0, CVTOP_ABS | CVTOP_RN,
CVT_F32_F32);
- emit_branch(pc, 0, 2, &pc->br_join[pc->if_lvl]);
- pc->if_insn[pc->if_lvl++] = pc->p->exec_tail;
+ 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_KIL:
+ assert(src[0][0] && src[0][1] && src[0][2] && src[0][3]);
emit_kil(pc, src[0][0]);
emit_kil(pc, src[0][1]);
emit_kil(pc, src[0][2]);
emit_kil(pc, src[0][3]);
break;
+ case TGSI_OPCODE_KILP:
+ emit_kil(pc, NULL);
+ break;
case TGSI_OPCODE_LIT:
emit_lit(pc, &dst[0], mask, &src[0][0]);
break;
case TGSI_OPCODE_LG2:
- emit_flop(pc, 3, brdc, src[0][0]);
+ emit_flop(pc, NV50_FLOP_LG2, brdc, src[0][0]);
+ break;
+ case TGSI_OPCODE_LOG:
+ {
+ struct nv50_reg *t[2];
+
+ t[0] = temp_temp(pc);
+ if (mask & (1 << 1))
+ t[1] = temp_temp(pc);
+ else
+ t[1] = t[0];
+
+ emit_abs(pc, t[0], src[0][0]);
+ emit_flop(pc, NV50_FLOP_LG2, t[1], t[0]);
+ if (mask & (1 << 2))
+ emit_mov(pc, dst[2], t[1]);
+ emit_flr(pc, t[1], t[1]);
+ if (mask & (1 << 0))
+ emit_mov(pc, dst[0], t[1]);
+ if (mask & (1 << 1)) {
+ t[1]->mod = NV50_MOD_NEG;
+ emit_preex2(pc, t[1], t[1]);
+ t[1]->mod = 0;
+ emit_flop(pc, NV50_FLOP_EX2, t[1], t[1]);
+ emit_mul(pc, dst[1], t[0], t[1]);
+ }
+ if (mask & (1 << 3))
+ emit_mov_immdval(pc, dst[3], 1.0f);
+ }
break;
case TGSI_OPCODE_LRP:
temp = temp_temp(pc);
emit_pow(pc, brdc, src[0][0], src[1][0]);
break;
case TGSI_OPCODE_RCP:
- emit_flop(pc, 0, brdc, src[0][0]);
+ emit_flop(pc, NV50_FLOP_RCP, brdc, src[0][0]);
+ break;
+ case TGSI_OPCODE_RET:
+ if (pc->p->type == PIPE_SHADER_FRAGMENT && !pc->in_subroutine)
+ nv50_fp_move_results(pc);
+ emit_ret(pc, -1, 0);
break;
case TGSI_OPCODE_RSQ:
- emit_flop(pc, 2, brdc, src[0][0]);
+ src[0][0]->mod |= NV50_MOD_ABS;
+ emit_flop(pc, NV50_FLOP_RSQ, brdc, src[0][0]);
break;
case TGSI_OPCODE_SCS:
temp = temp_temp(pc);
if (mask & 3)
emit_precossin(pc, temp, src[0][0]);
if (mask & (1 << 0))
- emit_flop(pc, 5, dst[0], temp);
+ emit_flop(pc, NV50_FLOP_COS, dst[0], temp);
if (mask & (1 << 1))
- emit_flop(pc, 4, dst[1], temp);
+ emit_flop(pc, NV50_FLOP_SIN, dst[1], temp);
if (mask & (1 << 2))
emit_mov_immdval(pc, dst[2], 0.0);
if (mask & (1 << 3))
case TGSI_OPCODE_SIN:
if (mask & 8) {
emit_precossin(pc, temp, src[0][3]);
- emit_flop(pc, 4, dst[3], temp);
+ emit_flop(pc, NV50_FLOP_SIN, dst[3], temp);
if (!(mask &= 7))
break;
if (temp == dst[3])
temp = brdc = temp_temp(pc);
}
emit_precossin(pc, temp, src[0][0]);
- emit_flop(pc, 4, brdc, temp);
+ emit_flop(pc, NV50_FLOP_SIN, brdc, temp);
break;
case TGSI_OPCODE_SLT:
case TGSI_OPCODE_SGE:
emit_mov_immdval(pc, dst[3], 1.0);
break;
case TGSI_OPCODE_END:
+ if (pc->p->type == PIPE_SHADER_FRAGMENT)
+ nv50_fp_move_results(pc);
+
+ /* last insn must be long so it can have the exit bit set */
+ if (!is_long(pc->p->exec_tail))
+ convert_to_long(pc, pc->p->exec_tail);
+ else
+ if (is_immd(pc->p->exec_tail) || is_join(pc->p->exec_tail))
+ emit_nop(pc);
+
+ pc->p->exec_tail->inst[1] |= 1; /* set exit bit */
break;
default:
NOUVEAU_ERR("invalid opcode %d\n", inst->Instruction.Opcode);
}
}
- for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
- for (c = 0; c < 4; c++) {
- if (!src[i][c])
- continue;
- if (src[i][c]->acc < 0 && src[i][c]->type == P_CONST)
- FREE(src[i][c]); /* indirect constant */
- }
- }
-
kill_temp_temp(pc);
pc->reg_instance_nr = 0;
mask = dst->WriteMask;
if (dst->File == TGSI_FILE_TEMPORARY)
- reg = pc->temp;
+ reg = pc->temp;
else
- if (dst->File == TGSI_FILE_OUTPUT)
- reg = pc->result;
+ if (dst->File == TGSI_FILE_OUTPUT) {
+ reg = pc->result;
+
+ if (insn->Instruction.Opcode == TGSI_OPCODE_MOV &&
+ dst->Index == pc->edgeflag_out &&
+ insn->Src[0].Register.File == TGSI_FILE_INPUT)
+ pc->p->cfg.edgeflag_in = insn->Src[0].Register.Index;
+ }
if (reg) {
for (c = 0; c < 4; c++) {
}
pc->r_brdc = NULL;
- if (!deqs)
+ if (!deqs || (!rdep[0] && !rdep[1] && !rdep[2] && !rdep[3]))
return nv50_program_tx_insn(pc, &insn);
deqs = nv50_revdep_reorder(m, rdep);
iv->rhw = popcnt4(pc->p->cfg.regs[1] >> 24) - 1;
emit_interp(pc, iv, NULL, mode & INTERP_CENTROID);
- emit_flop(pc, 0, iv, iv);
+ emit_flop(pc, NV50_FLOP_RCP, iv, iv);
/* XXX: when loading interpolants dynamically, move these
* to the program head, or make sure it can't be skipped.
const struct tgsi_full_immediate *imm =
&tp.FullToken.FullImmediate;
- ctor_immd(pc, imm->u[0].Float,
- imm->u[1].Float,
- imm->u[2].Float,
- imm->u[3].Float);
+ ctor_immd_4f32(pc, imm->u[0].Float,
+ imm->u[1].Float,
+ imm->u[2].Float,
+ imm->u[3].Float);
}
break;
case TGSI_TOKEN_TYPE_DECLARATION:
if (p->cfg.io_nr > first)
p->cfg.io_nr = first;
break;
+ case TGSI_SEMANTIC_EDGEFLAG:
+ pc->edgeflag_out = first;
+ break;
/*
case TGSI_SEMANTIC_CLIP_DISTANCE:
p->cfg.clpd = MIN2(p->cfg.clpd, first);
p->cfg.two_side[0].hw = 0x40;
p->cfg.two_side[1].hw = 0x40;
+ p->cfg.edgeflag_in = pc->edgeflag_out = 0xff;
+
switch (p->type) {
case PIPE_SHADER_VERTEX:
p->cfg.psiz = 0x40;
return TRUE;
}
-static void
-nv50_fp_move_results(struct nv50_pc *pc)
-{
- struct nv50_reg reg;
- unsigned i;
-
- ctor_reg(®, P_TEMP, -1, -1);
-
- for (i = 0; i < pc->result_nr * 4; ++i) {
- if (pc->result[i].rhw < 0 || pc->result[i].hw < 0)
- continue;
- if (pc->result[i].rhw != pc->result[i].hw) {
- reg.hw = pc->result[i].rhw;
- emit_mov(pc, ®, &pc->result[i]);
- }
- }
-}
-
static void
nv50_program_fixup_insns(struct nv50_pc *pc)
{
if (e->param.index >= 0 && !e->param.mask)
bra_list[n++] = e;
- /* last instruction must be long so it can have the exit bit set */
- if (!is_long(pc->p->exec_tail))
- convert_to_long(pc, pc->p->exec_tail);
- /* set exit bit */
- pc->p->exec_tail->inst[1] |= 1;
-
- /* !immd on exit insn simultaneously means !join */
- assert(!is_immd(pc->p->exec_head));
- assert(!is_immd(pc->p->exec_tail));
-
/* Make sure we don't have any single 32 bit instructions. */
for (e = pc->p->exec_head, pos = 0; e; e = e->next) {
pos += is_long(e) ? 2 : 1;
for (i = 0; i < n; ++i)
if (bra_list[i]->param.index >= pos)
bra_list[i]->param.index += 1;
+ for (i = 0; i < pc->insn_nr; ++i)
+ if (pc->insn_pos[i] >= pos)
+ pc->insn_pos[i] += 1;
convert_to_long(pc, e);
++pos;
}
}
FREE(bra_list);
+
+ if (!pc->p->info.opcode_count[TGSI_OPCODE_CAL])
+ return;
+
+ /* fill in CALL offsets */
+ for (e = pc->p->exec_head; e; e = e->next) {
+ if ((e->inst[0] & 2) && (e->inst[0] >> 28) == 0x2)
+ e->param.index = pc->insn_pos[e->param.index];
+ }
}
static boolean
if (ret == FALSE)
goto out_cleanup;
+ pc->insn_pos = MALLOC(pc->insn_nr * sizeof(unsigned));
+
tgsi_parse_init(&parse, pc->p->pipe.tokens);
while (!tgsi_parse_end_of_tokens(&parse)) {
const union tgsi_full_token *tok = &parse.FullToken;
- /* don't allow half insn/immd on first and last instruction */
+ /* previously allow32 was FALSE for first & last instruction */
pc->allow32 = TRUE;
- if (pc->insn_cur == 0 || pc->insn_cur + 2 == pc->insn_nr)
- pc->allow32 = FALSE;
tgsi_parse_token(&parse);
switch (tok->Token.Type) {
case TGSI_TOKEN_TYPE_INSTRUCTION:
+ pc->insn_pos[pc->insn_cur] = pc->p->exec_size;
++pc->insn_cur;
ret = nv50_tgsi_insn(pc, tok);
if (ret == FALSE)
}
}
- if (pc->p->type == PIPE_SHADER_FRAGMENT)
- nv50_fp_move_results(pc);
-
nv50_program_fixup_insns(pc);
p->param_nr = pc->param_nr * 4;
}
static void
-nv50_program_upload_data(struct nv50_context *nv50, float *map,
+nv50_program_upload_data(struct nv50_context *nv50, uint32_t *map,
unsigned start, unsigned count, unsigned cbuf)
{
struct nouveau_channel *chan = nv50->screen->base.channel;
if (p->param_nr) {
unsigned cb;
- float *map = pipe_buffer_map(pscreen, nv50->constbuf[p->type],
- PIPE_BUFFER_USAGE_CPU_READ);
+ uint32_t *map = pipe_buffer_map(pscreen, nv50->constbuf[p->type],
+ PIPE_BUFFER_USAGE_CPU_READ);
if (p->type == PIPE_SHADER_VERTEX)
cb = NV50_CB_PVP;
nv50_program_validate_data(nv50, p);
nv50_program_validate_code(nv50, p);
- so = so_new(13, 2);
+ so = so_new(5, 8, 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);
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_data (so, p->cfg.high_temp);
so_method(so, tesla, NV50TCL_FP_RESULT_COUNT, 1);
so_data (so, p->cfg.high_result);
- so_method(so, tesla, NV50TCL_FP_CTRL_UNK19A8, 1);
+ so_method(so, tesla, NV50TCL_FP_CONTROL, 1);
so_data (so, p->cfg.regs[2]);
so_method(so, tesla, NV50TCL_FP_CTRL_UNK196C, 1);
so_data (so, p->cfg.regs[3]);
}
/* now fill the stateobj */
- so = so_new(64, 0);
+ so = so_new(6, 58, 0);
n = (m + 3) / 4;
so_method(so, tesla, NV50TCL_VP_RESULT_MAP_SIZE, 1);
so_method(so, tesla, NV50TCL_FP_INTERPOLANT_CTRL, 1);
so_data (so, reg[4]);
- so_method(so, tesla, 0x1540, 4);
+ so_method(so, tesla, NV50TCL_NOPERSPECTIVE_BITMAP(0), 4);
so_datap (so, lin, 4);
if (nv50->rasterizer->pipe.point_sprite) {