Hi ! Here are some more changes to the shader code to be applied in
addition to the previous ones.
Well, I've though about getting an account so I could commit some (or
all) of this stuff, probably in splitting some patches up a bit more.
But I'd feel much better doing so (if I was even allowed to) if
someone (darktama) had a look at my code before, preferably with some
comments on how to improve / do / not do things.
Someone testing this stuff and reporting to me whether they see improve-
ments as compared to the code currently in mesa master wouldn't be bad
either.
Well, if no one has any major complaints, and once I feel a little bit
more comfortable with what I've done so far, I'll probably have to
ask/apply for a mesa account as Maarten suggested ...
Thank you, Christoph
commit 78f4b96b9b874e487571b59cec22e683f23d2e3e
Author: chr <c...@echelon.(none)>
Date: Sat May 16 15:33:34 2009 +0200
- Introduce argument negation for MUL, MAD, ADD, DP3, DP4 and POW.
This is now done via a negation bitmask in nv50_pc that is checked
in the emit functions, the 'neg' member of nv50_reg is obsolete
with this.
(maybe I should have used it instead, but I completely forgot
about it and went with this way ...)
- changed emit_sub and emit_msb to use negated emit_add and emit_mad
diff --git a/src/gallium/drivers/nv50/nv50_program.c
b/src/gallium/drivers/nv50/nv50_program.c
index 6e279bd..ed9e20b 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -117,6 +117,7 @@ struct nv50_pc {
unsigned insn_cur;
unsigned insn_nr;
+ unsigned negate;
boolean allow32; /* TRUE when half insns are allowed */
};
@@ -531,6 +532,8 @@ emit_mov(struct nv50_pc *pc, struct nv50_reg *dst, struct
nv50_reg *src)
emit(pc, e);
}
+#define BITS_0_1_SWAPPED(x) ((x & 1) << 1) | ((x & 2) >> 1)
+
static boolean
check_swap_src_0_1(struct nv50_pc *pc,
struct nv50_reg **s0, struct nv50_reg **s1)
@@ -628,14 +631,6 @@ set_src_2(struct nv50_pc *pc, struct nv50_reg *src, struct
nv50_program_exec *e)
e->inst[1] |= (src->hw << 14);
}
-static boolean
-requires_long(struct nv50_program_exec *e, struct nv50_reg *src)
-{
- if (is_long(e) || src->type == P_IMMD || src->type == P_CONST)
- return TRUE;
- return FALSE;
-}
-
static void
emit_mul(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src0,
struct nv50_reg *src1)
@@ -650,6 +645,13 @@ emit_mul(struct nv50_pc *pc, struct nv50_reg *dst, struct
nv50_reg *src0,
set_src_0(pc, src0, e);
set_src_1(pc, src1, e);
+ if (pc->negate & 1) {
+ if (is_long(e))
+ e->inst[1] |= 0x08000000;
+ else
+ e->inst[0] |= 0x00008000;
+ }
+
emit(pc, e);
}
@@ -658,18 +660,25 @@ emit_add(struct nv50_pc *pc, struct nv50_reg *dst,
struct nv50_reg *src0, struct nv50_reg *src1)
{
struct nv50_program_exec *e = exec(pc);
+ unsigned neg = pc->negate;
e->inst[0] |= 0xb0000000;
- check_swap_src_0_1(pc, &src0, &src1);
+ if (check_swap_src_0_1(pc, &src0, &src1))
+ neg = BITS_0_1_SWAPPED(neg);
+ if (!pc->allow32 || neg) {
+ set_long(pc, e);
+ e->inst[1] |= (neg << 26);
+ }
+
set_dst(pc, dst, e);
set_src_0(pc, src0, e);
- if (!is_long(e) && src1->type == P_IMMD && pc->allow32)
- set_immd(pc, src1, e);
- else
- if (requires_long(e, src1))
+ if (src1->type == P_CONST || is_long(e))
set_src_2(pc, src1, e);
else
+ if (src1->type == P_IMMD)
+ set_immd(pc, src1, e);
+ else
set_src_1(pc, src1, e);
emit(pc, e);
@@ -693,25 +702,13 @@ emit_minmax(struct nv50_pc *pc, unsigned sub, struct
nv50_reg *dst,
emit(pc, e);
}
-static void
+static INLINE void
emit_sub(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src0,
struct nv50_reg *src1)
{
- struct nv50_program_exec *e = exec(pc);
-
- e->inst[0] |= 0xb0000000;
-
- set_long(pc, e);
- if (check_swap_src_0_1(pc, &src0, &src1))
- e->inst[1] |= 0x04000000;
- else
- e->inst[1] |= 0x08000000;
-
- set_dst(pc, dst, e);
- set_src_0(pc, src0, e);
- set_src_2(pc, src1, e);
-
- emit(pc, e);
+ pc->negate ^= 2;
+ emit_add(pc, dst, src0, src1);
+ pc->negate ^= 2;
}
static void
@@ -728,26 +725,21 @@ emit_mad(struct nv50_pc *pc, struct nv50_reg *dst, struct
nv50_reg *src0,
set_src_1(pc, src1, e);
set_src_2(pc, src2, e);
+ if (pc->negate & 1)
+ e->inst[1] |= 0x04000000;
+ if (pc->negate & 2)
+ e->inst[1] |= 0x08000000;
+
emit(pc, e);
}
-static void
+static INLINE void
emit_msb(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] |= 0xe0000000;
- set_long(pc, e);
- e->inst[1] |= 0x08000000; /* src0 * src1 - src2 */
-
- 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);
- set_src_2(pc, src2, e);
-
- emit(pc, e);
+ pc->negate ^= 2;
+ emit_mad(pc, dst, src0, src1, src2);
+ pc->negate ^= 2;
}
static void
@@ -1012,8 +1004,15 @@ convert_to_long(struct nv50_pc *pc, struct
nv50_program_exec *e)
if (e->inst[0] & 0x02000000)
q = 0x00020000;
break;
+ case 0xB:
+ /* ADD */
+ m = ~(127 << 16);
+ q = ((e->inst[0] & (~m)) >> 2);
+ break;
case 0xC:
/* MUL */
+ m = ~0x00008000;
+ q = ((e->inst[0] & (~m)) << 12);
break;
case 0x9:
/* RCP */
@@ -1090,8 +1089,33 @@ tgsi_dst(struct nv50_pc *pc, int c, const struct
tgsi_full_dst_register *dst)
return NULL;
}
+/* Returns bit with which to XOR pc->negate on negation, or -1 if negation
+ * is not supported for the instruction. This is used in tgsi_src, and set
+ * to -2 in tx_insn to indicate XOR has already been taken care of in c 0.
+ */
+static int
+negate_supported(const struct tgsi_full_instruction *insn, int i)
+{
+ switch (insn->Instruction.Opcode) {
+ case TGSI_OPCODE_DP3:
+ case TGSI_OPCODE_DP4:
+ case TGSI_OPCODE_MUL:
+ return 0;
+ case TGSI_OPCODE_ADD:
+ case TGSI_OPCODE_SUB:
+ return i;
+ case TGSI_OPCODE_MAD:
+ return (i == 2) ? 1 : 0;
+ case TGSI_OPCODE_POW:
+ return (i == 1) ? 0 : -1;
+ default:
+ return -1;
+ }
+}
+
static struct nv50_reg *
-tgsi_src(struct nv50_pc *pc, int chan, const struct tgsi_full_src_register
*src)
+tgsi_src(struct nv50_pc *pc, int chan, const struct tgsi_full_src_register
*src,
+ int neg)
{
struct nv50_reg *r = NULL;
struct nv50_reg *temp;
@@ -1146,9 +1170,12 @@ tgsi_src(struct nv50_pc *pc, int chan, const struct
tgsi_full_src_register *src)
r = temp;
break;
case TGSI_UTIL_SIGN_TOGGLE:
- temp = temp_temp(pc);
- emit_neg(pc, temp, r);
- r = temp;
+ if (neg == -1) {
+ temp = temp_temp(pc);
+ emit_neg(pc, temp, r);
+ r = temp;
+ } else if (neg >= 0)
+ pc->negate ^= (1 << neg);
break;
case TGSI_UTIL_SIGN_SET:
temp = temp_temp(pc);
@@ -1202,6 +1229,7 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
mask = inst->FullDstRegisters[0].DstRegister.WriteMask;
sat = inst->Instruction.Saturate == TGSI_SAT_ZERO_ONE;
+ pc->negate = 0;
for (c = 0; c < 4; c++) {
if (mask & (1 << c))
@@ -1218,12 +1246,16 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
const struct tgsi_full_src_register *fs =
&inst->FullSrcRegisters[i];
+ int neg = negate_supported(inst, i);
if (fs->SrcRegister.File == TGSI_FILE_SAMPLER)
unit = fs->SrcRegister.Index;
- for (c = 0; c < 4; c++)
- src[i][c] = tgsi_src(pc, c, fs);
+ for (c = 0; c < 4; c++) {
+ src[i][c] = tgsi_src(pc, c, fs, neg);
+ if (neg >= 0)
+ neg = -2; /* already negated */
+ }
}
if (sat) {
commit e611bc623c1526d5c806964287edf31b4b346d0d
Author: chr <c...@echelon.(none)>
Date: Sat May 16 16:50:23 2009 +0200
- Unify moving result from temporary to destination registers.
- Don't do mov and cvt.sat for MOV_SAT, just cvt.sat suffices.
diff --git a/src/gallium/drivers/nv50/nv50_program.c
b/src/gallium/drivers/nv50/nv50_program.c
index ed9e20b..4a03cf5 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -221,7 +221,7 @@ alloc_preferred_temp(struct nv50_pc *pc, int hw)
static void
assimilate_temp(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
{
- assert(dst->index != -1 && src->index == -1 && src->hw != -1);
+ assert(src->index == -1 && src->hw != -1);
if (dst->hw != -1)
pc->r_temp[dst->hw] = NULL;
@@ -1311,22 +1311,12 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
temp = alloc_temp(pc, NULL);
emit_precossin(pc, temp, src[0][0]);
emit_flop(pc, 5, temp, temp);
- for (c = 0; c < 4; c++) {
- if (!(mask & (1 << c)))
- continue;
- emit_mov(pc, dst[c], temp);
- }
break;
case TGSI_OPCODE_DP3:
temp = alloc_temp(pc, NULL);
emit_mul(pc, temp, src[0][0], src[1][0]);
emit_mad(pc, temp, src[0][1], src[1][1], temp);
emit_mad(pc, temp, src[0][2], src[1][2], temp);
- for (c = 0; c < 4; c++) {
- if (!(mask & (1 << c)))
- continue;
- emit_mov(pc, dst[c], temp);
- }
break;
case TGSI_OPCODE_DP4:
temp = alloc_temp(pc, NULL);
@@ -1334,11 +1324,6 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
emit_mad(pc, temp, src[0][1], src[1][1], temp);
emit_mad(pc, temp, src[0][2], src[1][2], temp);
emit_mad(pc, temp, src[0][3], src[1][3], temp);
- for (c = 0; c < 4; c++) {
- if (!(mask & (1 << c)))
- continue;
- emit_mov(pc, dst[c], temp);
- }
break;
case TGSI_OPCODE_DPH:
temp = alloc_temp(pc, NULL);
@@ -1346,11 +1331,6 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
emit_mad(pc, temp, src[0][1], src[1][1], temp);
emit_mad(pc, temp, src[0][2], src[1][2], temp);
emit_add(pc, temp, src[1][3], temp);
- for (c = 0; c < 4; c++) {
- if (!(mask & (1 << c)))
- continue;
- emit_mov(pc, dst[c], temp);
- }
break;
case TGSI_OPCODE_DST:
{
@@ -1370,11 +1350,6 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
temp = alloc_temp(pc, NULL);
emit_preex2(pc, temp, src[0][0]);
emit_flop(pc, 6, temp, temp);
- for (c = 0; c < 4; c++) {
- if (!(mask & (1 << c)))
- continue;
- emit_mov(pc, dst[c], temp);
- }
break;
case TGSI_OPCODE_FLR:
for (c = 0; c < 4; c++) {
@@ -1405,11 +1380,6 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
case TGSI_OPCODE_LG2:
temp = alloc_temp(pc, NULL);
emit_flop(pc, 3, temp, src[0][0]);
- for (c = 0; c < 4; c++) {
- if (!(mask & (1 << c)))
- continue;
- emit_mov(pc, dst[c], temp);
- }
break;
case TGSI_OPCODE_LRP:
temp = alloc_temp(pc, NULL);
@@ -1419,6 +1389,8 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
emit_sub(pc, temp, src[1][c], src[2][c]);
emit_mad(pc, dst[c], temp, src[0][c], src[2][c]);
}
+ free_temp(pc, temp);
+ temp = NULL;
break;
case TGSI_OPCODE_MAD:
for (c = 0; c < 4; c++) {
@@ -1442,6 +1414,13 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
}
break;
case TGSI_OPCODE_MOV:
+ if (sat) {
+ dst[0] = src[0][0];
+ dst[1] = src[0][1];
+ dst[2] = src[0][2];
+ dst[3] = src[0][3];
+ break;
+ }
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
@@ -1458,11 +1437,6 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
case TGSI_OPCODE_POW:
temp = alloc_temp(pc, NULL);
emit_pow(pc, temp, src[0][0], src[1][0]);
- for (c = 0; c < 4; c++) {
- if (!(mask & (1 << c)))
- continue;
- emit_mov(pc, dst[c], temp);
- }
break;
case TGSI_OPCODE_RCP:
for (c = 0; c < 4; c++) {
@@ -1493,6 +1467,8 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
emit_flop(pc, 5, dst[0], temp);
if (mask & (1 << 1))
emit_flop(pc, 4, dst[1], temp);
+ free_temp(pc, temp);
+ temp = NULL;
break;
case TGSI_OPCODE_SGE:
for (c = 0; c < 4; c++) {
@@ -1505,11 +1481,6 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
temp = alloc_temp(pc, NULL);
emit_precossin(pc, temp, src[0][0]);
emit_flop(pc, 4, temp, temp);
- for (c = 0; c < 4; c++) {
- if (!(mask & (1 << c)))
- continue;
- emit_mov(pc, dst[c], temp);
- }
break;
case TGSI_OPCODE_SLT:
for (c = 0; c < 4; c++) {
@@ -1573,8 +1544,24 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
return FALSE;
}
- if (temp)
- free_temp(pc, temp);
+ i = -1;
+ if (temp) {
+ for (c = 0; c < 4; c++) {
+ if (!(mask & (1 << c)))
+ continue;
+ if (i >= 0)
+ emit_mov(pc, dst[c], dst[i]);
+ else if (dst[c]->type == P_TEMP) {
+ assimilate_temp(pc, dst[c], temp);
+ i = c;
+ temp = NULL;
+ } else
+ emit_mov(pc, dst[c], temp);
+ }
+
+ if (temp)
+ free_temp(pc, temp);
+ }
if (sat) {
for (c = 0; c < 4; c++) {
commit 1b3a83bd57065aa015d5251f44bea0b710838581
Author: chr <c...@echelon.(none)>
Date: Sat May 16 16:39:38 2009 +0200
- Make TXP do what it's supposed to (centroid not yet honored, could
store interpolation mode in some per register flag maybe).
- FIX: don't set perspect_load = FALSE if we encounter a centroid
attribute, instead make perspect/centroid_load counters.
diff --git a/src/gallium/drivers/nv50/nv50_program.c
b/src/gallium/drivers/nv50/nv50_program.c
index 3a00c90..3116735 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -1497,14 +1497,33 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
}
break;
case TGSI_OPCODE_TEX:
- case TGSI_OPCODE_TXP: /* XXX: TXP should use w-component as iv on
interp */
+ case TGSI_OPCODE_TXP:
{
struct nv50_reg *t[4];
struct nv50_program_exec *e;
alloc_temp4(pc, t, 0);
- emit_mov(pc, t[0], src[0][0]);
- emit_mov(pc, t[1], src[0][1]);
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_TXP) {
+ if (src[0][0]->type == P_TEMP && src[0][0]->rhw != -1) {
+ /* XXX: centroid is not honored here */
+ t[1]->rhw = src[0][3]->rhw;
+ emit_interp(pc, t[1], NULL, INTERP_LINEAR);
+ emit_flop(pc, 0, t[1], t[1]);
+ t[0]->rhw = src[0][0]->rhw;
+ t[1]->rhw = src[0][1]->rhw;
+ emit_interp(pc, t[0], t[1], INTERP_PERSPECTIVE);
+ emit_interp(pc, t[1], t[1], INTERP_PERSPECTIVE);
+ } else {
+ emit_mov(pc, t[1], src[0][3]);
+ emit_flop(pc, 0, t[1], t[1]);
+ emit_mul(pc, t[0], src[0][0], t[1]);
+ emit_mul(pc, t[1], src[0][1], t[1]);
+ }
+ } else {
+ emit_mov(pc, t[0], src[0][0]);
+ emit_mov(pc, t[1], src[0][1]);
+ }
e = exec(pc);
e->inst[0] = 0xf6400000;
@@ -1514,6 +1533,7 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
set_dst(pc, t[0], e);
emit(pc, e);
+ /* XXX: without these MOVs, it can happen that TEX has no
effect */
if (mask & (1 << 0)) emit_mov(pc, dst[0], t[0]);
if (mask & (1 << 1)) emit_mov(pc, dst[1], t[1]);
if (mask & (1 << 2)) emit_mov(pc, dst[2], t[2]);
@@ -1636,8 +1656,8 @@ nv50_program_tx_prep(struct nv50_pc *pc)
unsigned fcol, bcol, fcrd, depr;
/* record interpolation mode from declaration */
- boolean centroid_load = FALSE;
- boolean perspect_load = FALSE;
+ unsigned centroid_load = 0;
+ unsigned perspect_load = 0;
unsigned interp_mode[32];
/* track register usage for temps and attrs */
@@ -1721,38 +1741,45 @@ nv50_program_tx_prep(struct nv50_pc *pc)
break;
case TGSI_INTERPOLATE_PERSPECTIVE:
mode = INTERP_PERSPECTIVE;
- perspect_load = TRUE;
+ if (!d->Declaration.Centroid)
+ perspect_load++;
break;
default:
mode = INTERP_LINEAR;
break;
}
+ if (d->Declaration.Centroid) {
+ mode |= INTERP_CENTROID;
+ centroid_load++;
+ }
+
if (d->Declaration.Semantic) {
switch (d->Semantic.SemanticName) {
case TGSI_SEMANTIC_POSITION:
fcrd = first;
break;
+ /* XXX: FLAT and LINEAR don't seem to
behave correctly: */
case TGSI_SEMANTIC_COLOR:
fcol = first;
- mode = INTERP_PERSPECTIVE;
- perspect_load = TRUE;
+ if (!(mode &
INTERP_PERSPECTIVE)) {
+ mode &= INTERP_CENTROID;
+ mode |=
INTERP_PERSPECTIVE;
+ perspect_load++;
+ }
break;
case TGSI_SEMANTIC_BCOLOR:
bcol = first;
- mode = INTERP_PERSPECTIVE;
- perspect_load = TRUE;
+ if (!(mode &
INTERP_PERSPECTIVE)) {
+ mode &= INTERP_CENTROID;
+ mode |=
INTERP_PERSPECTIVE;
+ perspect_load++;
+ }
break;
default:
break;
}
}
-
- if (d->Declaration.Centroid) {
- mode |= INTERP_CENTROID;
- centroid_load = TRUE;
- perspect_load = FALSE;
- }
assert(last < 32);
for (i = first; i <= last; i++)
commit 3f66c5d0daf7ef15ddc9f7bc22967d95d52ab2af
Author: chr <c...@echelon.(none)>
Date: Sat May 16 16:43:09 2009 +0200
- Introduce emit_cvt and use it where applicable (flr, abs, sat, and in
set).
- Restructure LIT again, now src == dst case is completely taken care of
there.
- Change emit_kil and add negation support (the generated insn was different
before, but this is how the blob does it here, should look into this).
- Remove unnecessary MALLOC and FREE in program dump ifdef block.
- In alloc_immd(), also put -f and 0.5 * f in the immd buffer, as it is now
this might save some space.
Well, that's a lot of rather unrelated changes, maybe I should break things
up more.
diff --git a/src/gallium/drivers/nv50/nv50_program.c
b/src/gallium/drivers/nv50/nv50_program.c
index 3116735..4dc5676 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -335,7 +335,7 @@ alloc_immd(struct nv50_pc *pc, float f)
break;
if (hw == pc->immd_nr * 4)
- hw = ctor_immd(pc, f, 0, 0, 0) * 4;
+ hw = ctor_immd(pc, f, -f, 0.5f * f, 0) * 4;
r->type = P_IMMD;
r->hw = hw;
@@ -790,6 +790,48 @@ emit_precossin(struct nv50_pc *pc, struct nv50_reg *dst,
struct nv50_reg *src)
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
+
+#define CVT_F32_F32 0xc4
+#define CVT_F32_S32 0x44
+#define CVT_F32_U32 0x64
+#define CVT_S32_F32 0x8c
+#define CVT_S32_S32 0x0c
+#define CVT_R32_F32 0xcc
+
+static void
+emit_cvt(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src,
+ int wp, unsigned cop, unsigned fmt)
+{
+ struct nv50_program_exec *e;
+
+ e = exec(pc);
+ set_long(pc, e);
+
+ e->inst[0] |= 0xa0000000;
+ e->inst[1] |= 0x00004000;
+ e->inst[1] |= (cop << 16);
+ e->inst[1] |= (fmt << 24);
+ set_src_0(pc, src, 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;
+ }
+
+ emit(pc, e);
+}
+
static void
emit_set(struct nv50_pc *pc, unsigned c_op, struct nv50_reg *dst,
struct nv50_reg *src0, struct nv50_reg *src1)
@@ -821,34 +863,16 @@ emit_set(struct nv50_pc *pc, unsigned c_op, struct
nv50_reg *dst,
set_src_1(pc, src1, e);
emit(pc, e);
- /* cvt.f32.u32 */
- e = exec(pc);
- e->inst[0] = 0xa0000001;
- e->inst[1] = 0x64014780;
- set_dst(pc, rdst, e);
- set_src_0(pc, dst, e);
- emit(pc, e);
-
+ emit_cvt(pc, rdst, dst, -1, CVTOP_RN, CVT_F32_U32);
+
if (dst != rdst)
free_temp(pc, dst);
}
-static void
+static INLINE void
emit_flr(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
{
- struct nv50_program_exec *e = exec(pc);
-
- e->inst[0] = 0xa0000000; /* cvt */
- set_long(pc, e);
- e->inst[1] |= (6 << 29); /* cvt */
- e->inst[1] |= 0x08000000; /* integer mode */
- e->inst[1] |= 0x04000000; /* 32 bit */
- e->inst[1] |= ((0x1 << 3)) << 14; /* .rn */
- e->inst[1] |= (1 << 14); /* src .f32 */
- set_dst(pc, dst, e);
- set_src_0(pc, src, e);
-
- emit(pc, e);
+ emit_cvt(pc, dst, src, -1, CVTOP_FLOOR, CVT_R32_F32);
}
static void
@@ -865,21 +889,10 @@ emit_pow(struct nv50_pc *pc, struct nv50_reg *dst,
free_temp(pc, temp);
}
-static void
+static INLINE void
emit_abs(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
{
- struct nv50_program_exec *e = exec(pc);
-
- e->inst[0] = 0xa0000000; /* cvt */
- set_long(pc, e);
- e->inst[1] |= (6 << 29); /* cvt */
- e->inst[1] |= 0x04000000; /* 32 bit */
- e->inst[1] |= (1 << 14); /* src .f32 */
- e->inst[1] |= ((1 << 6) << 14); /* .abs */
- set_dst(pc, dst, e);
- set_src_0(pc, src, e);
-
- emit(pc, e);
+ emit_cvt(pc, dst, src, -1, CVTOP_ABS, CVT_F32_F32);
}
static void
@@ -894,10 +907,7 @@ emit_lit(struct nv50_pc *pc, struct nv50_reg **dst,
unsigned mask,
boolean allow32 = pc->allow32;
if (mask & (3 << 1)) {
- if (mask & (1 << 1))
- tmp[0] = dst[1];
- else
- tmp[0] = temp_temp(pc);
+ tmp[0] = alloc_temp(pc, NULL);
emit_minmax(pc, 4, tmp[0], src[0], zero);
}
@@ -920,6 +930,12 @@ emit_lit(struct nv50_pc *pc, struct nv50_reg **dst,
unsigned mask,
pc->allow32 = allow32;
+ if (mask & (1 << 1))
+ assimilate_temp(pc, dst[1], tmp[0]);
+ else
+ if (mask & (1 << 2))
+ free_temp(pc, tmp[0]);
+
/* do this last, in case src[i,j] == dst[0,3] */
if (mask & (1 << 0))
emit_mov(pc, dst[0], one);
@@ -953,21 +969,16 @@ static void
emit_kil(struct nv50_pc *pc, struct nv50_reg *src)
{
struct nv50_program_exec *e;
- const int r_pred = 1;
- /* Sets predicate reg ? */
- e = exec(pc);
- e->inst[0] = 0xa00001fd;
- e->inst[1] = 0xc4014788;
- set_src_0(pc, src, e);
- set_pred_wr(pc, 1, r_pred, e);
- emit(pc, e);
+ emit_cvt(pc, NULL, src, 0, CVTOP_RN, CVT_F32_F32);
+ if (pc->negate)
+ pc->p->exec_tail->inst[1] |= 0x20000000;
- /* This is probably KILP */
+ /* @p0.lt kil */
e = exec(pc);
- e->inst[0] = 0x000001fe;
set_long(pc, e);
- set_pred(pc, 1 /* LT? */, r_pred, e);
+ e->inst[0] |= 0x00000002;
+ set_pred(pc, 1, 0, e);
emit(pc, e);
}
@@ -1100,6 +1111,7 @@ negate_supported(const struct tgsi_full_instruction
*insn, int i)
case TGSI_OPCODE_DP3:
case TGSI_OPCODE_DP4:
case TGSI_OPCODE_MUL:
+ case TGSI_OPCODE_KIL:
return 0;
case TGSI_OPCODE_ADD:
case TGSI_OPCODE_SUB:
@@ -1281,14 +1293,6 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
rdst[c] = dst[c];
dst[c] = alloc_preferred_temp(pc, rdst[c]->rhw);
}
- } else if (inst->Instruction.Opcode == TGSI_OPCODE_LIT) {
- /* XXX: shouldn't give LIT an extra case here */
- if (src[0][1] == dst[1] ||
- src[0][3] == dst[1]) {
- assimilate = TRUE;
- rdst[1] = dst[1];
- dst[1] = alloc_temp(pc, NULL);
- }
}
i = -1;
@@ -1585,21 +1589,9 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
if (sat) {
for (c = 0; c < 4; c++) {
- struct nv50_program_exec *e;
-
if (!(mask & (1 << c)))
continue;
- e = exec(pc);
-
- e->inst[0] = 0xa0000000; /* cvt */
- set_long(pc, e);
- e->inst[1] |= (6 << 29); /* cvt */
- e->inst[1] |= 0x04000000; /* 32 bit */
- e->inst[1] |= (1 << 14); /* src .f32 */
- e->inst[1] |= ((1 << 5) << 14); /* .sat */
- set_dst(pc, rdst[c], e);
- set_src_0(pc, dst[c], e);
- emit(pc, e);
+ emit_cvt(pc, rdst[c], dst[c], -1, CVTOP_SAT,
CVT_F32_F32);
}
} else if (assimilate) {
for (c = 0; c < 4; c++)
@@ -2272,13 +2264,11 @@ nv50_program_validate_code(struct nv50_context *nv50,
struct nv50_program *p)
#ifdef NV50_PROGRAM_DUMP
NOUVEAU_ERR("-------\n");
- up = ptr = MALLOC(p->exec_size * 4);
for (e = p->exec_head; e; e = e->next) {
NOUVEAU_ERR("0x%08x\n", e->inst[0]);
if (is_long(e))
NOUVEAU_ERR("0x%08x\n", e->inst[1]);
}
- FREE(up);
#endif
up = ptr = MALLOC(p->exec_size * 4);
commit 0bd93153d45bec1616a40b81af28d0c5a1536539
Author: chr <c...@echelon.(none)>
Date: Sat May 16 16:30:59 2009 +0200
Rewrite emit_set.
diff --git a/src/gallium/drivers/nv50/nv50_program.c
b/src/gallium/drivers/nv50/nv50_program.c
index 95e3bdf..c3185a9 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -833,38 +833,45 @@ emit_cvt(struct nv50_pc *pc, struct nv50_reg *dst, struct
nv50_reg *src,
}
static void
-emit_set(struct nv50_pc *pc, unsigned c_op, struct nv50_reg *dst,
+emit_set(struct nv50_pc *pc, unsigned c_op, struct nv50_reg *dst, int wp,
struct nv50_reg *src0, struct nv50_reg *src1)
{
struct nv50_program_exec *e = exec(pc);
- unsigned inv_cop[8] = { 0, 4, 2, 6, 1, 5, 3, 7 };
struct nv50_reg *rdst;
- assert(c_op <= 7);
+ /* This maps TGSI_CC_GT/EQ/LT/GE/LE/NE to condition codes for SET.
+ * Note that conditional execution prefixes are probably different,
+ * and for SET, code 0xE seems to mean 'output condition code'.
+ * XXX: maybe verify these again
+ */
+ static unsigned cop_map[16] = { 0x4, 0x2, 0x1, 0x6, 0x3, 0xd, 0, 0,
+ 0x3,
0xd, 0x6, 0x1, 0x4, 0x2, 0, 0 };
if (check_swap_src_0_1(pc, &src0, &src1))
- c_op = inv_cop[c_op];
+ c_op += 8;
rdst = dst;
if (dst->type != P_TEMP)
dst = alloc_temp(pc, NULL);
- /* set.u32 */
set_long(pc, e);
e->inst[0] |= 0xb0000000;
- e->inst[1] |= (3 << 29);
- e->inst[1] |= (c_op << 14);
- /*XXX: breaks things, .u32 by default?
- * decuda will disasm as .u16 and use .lo/.hi regs, but this
- * doesn't seem to match what the hw actually does.
- inst[1] |= 0x04000000; << breaks things.. .u32 by default?
- */
- set_dst(pc, dst, e);
- set_src_0(pc, src0, e);
- set_src_1(pc, src1, e);
+ e->inst[1] |= 0x60000000;
+ e->inst[1] |= ((c_op < 16) ? cop_map[c_op] : 0xe) << 14;
+
+ if (dst)
+ set_dst(pc, dst, e);
+ else {
+ e->inst[0] |= 0x000001fc;
+ e->inst[1] |= 0x00000008;
+ }
+ if (wp >= 0)
+ set_pred_wr(pc, 1, wp, e);
+
emit(pc, e);
- emit_cvt(pc, rdst, dst, -1, CVTOP_RN, CVT_F32_U32);
-
+ if (dst)
+ emit_cvt(pc, rdst, dst, -1, CVTOP_RN | CVTOP_ABS, CVT_F32_S32);
+
if (dst != rdst)
free_temp(pc, dst);
}
@@ -1478,7 +1485,7 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
- emit_set(pc, 6, dst[c], src[0][c], src[1][c]);
+ emit_set(pc, TGSI_CC_GE, dst[c], -1, src[0][c],
src[1][c]);
}
break;
case TGSI_OPCODE_SIN:
@@ -1490,7 +1497,7 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union
tgsi_full_token *tok)
for (c = 0; c < 4; c++) {
if (!(mask & (1 << c)))
continue;
- emit_set(pc, 1, dst[c], src[0][c], src[1][c]);
+ emit_set(pc, TGSI_CC_LT, dst[c], -1, src[0][c],
src[1][c]);
}
break;
case TGSI_OPCODE_SUB:
_______________________________________________
Nouveau mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/nouveau
