summaryrefslogtreecommitdiff
path: root/src/gallium/auxiliary/tgsi/tgsi_ureg.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/auxiliary/tgsi/tgsi_ureg.c')
-rw-r--r--src/gallium/auxiliary/tgsi/tgsi_ureg.c405
1 files changed, 322 insertions, 83 deletions
diff --git a/src/gallium/auxiliary/tgsi/tgsi_ureg.c b/src/gallium/auxiliary/tgsi/tgsi_ureg.c
index f7096bd8e2..3f752e9352 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_ureg.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_ureg.c
@@ -29,8 +29,10 @@
#include "pipe/p_context.h"
#include "pipe/p_state.h"
#include "tgsi/tgsi_ureg.h"
+#include "tgsi/tgsi_build.h"
#include "tgsi/tgsi_info.h"
#include "tgsi/tgsi_dump.h"
+#include "tgsi/tgsi_sanity.h"
#include "util/u_memory.h"
#include "util/u_math.h"
@@ -45,18 +47,14 @@ union tgsi_any_token {
struct tgsi_immediate imm;
union tgsi_immediate_data imm_data;
struct tgsi_instruction insn;
- struct tgsi_instruction_ext_nv insn_ext_nv;
struct tgsi_instruction_ext_label insn_ext_label;
struct tgsi_instruction_ext_texture insn_ext_texture;
struct tgsi_instruction_ext_predicate insn_ext_predicate;
struct tgsi_src_register src;
- struct tgsi_src_register_ext_swz src_ext_swz;
struct tgsi_src_register_ext_mod src_ext_mod;
struct tgsi_dimension dim;
struct tgsi_dst_register dst;
- struct tgsi_dst_register_ext_concode dst_ext_code;
struct tgsi_dst_register_ext_modulate dst_ext_mod;
- struct tgsi_dst_register_ext_predicate dst_ext_pred;
unsigned value;
};
@@ -70,9 +68,11 @@ struct ureg_tokens {
#define UREG_MAX_INPUT PIPE_MAX_ATTRIBS
#define UREG_MAX_OUTPUT PIPE_MAX_ATTRIBS
+#define UREG_MAX_CONSTANT_RANGE 32
#define UREG_MAX_IMMEDIATE 32
#define UREG_MAX_TEMP 256
#define UREG_MAX_ADDR 2
+#define UREG_MAX_PRED 1
#define DOMAIN_DECL 0
#define DOMAIN_INSN 1
@@ -86,8 +86,10 @@ struct ureg_program
unsigned semantic_name;
unsigned semantic_index;
unsigned interp;
- } input[UREG_MAX_INPUT];
- unsigned nr_inputs;
+ } fs_input[UREG_MAX_INPUT];
+ unsigned nr_fs_inputs;
+
+ unsigned vs_inputs[UREG_MAX_INPUT/32];
struct {
unsigned semantic_name;
@@ -107,9 +109,14 @@ struct ureg_program
unsigned temps_active[UREG_MAX_TEMP / 32];
unsigned nr_temps;
- unsigned nr_addrs;
+ struct {
+ unsigned first;
+ unsigned last;
+ } constant_range[UREG_MAX_CONSTANT_RANGE];
+ unsigned nr_constant_ranges;
- unsigned nr_constants;
+ unsigned nr_addrs;
+ unsigned nr_preds;
unsigned nr_instructions;
struct ureg_tokens domain[2];
@@ -119,6 +126,9 @@ static union tgsi_any_token error_tokens[32];
static void tokens_error( struct ureg_tokens *tokens )
{
+ if (tokens->tokens && tokens->tokens != error_tokens)
+ FREE(tokens->tokens);
+
tokens->tokens = error_tokens;
tokens->size = Elements(error_tokens);
tokens->count = 0;
@@ -130,8 +140,9 @@ static void tokens_expand( struct ureg_tokens *tokens,
{
unsigned old_size = tokens->size * sizeof(unsigned);
- if (tokens->tokens == error_tokens)
- goto fail;
+ if (tokens->tokens == error_tokens) {
+ return;
+ }
while (tokens->count + count > tokens->size) {
tokens->size = (1 << ++tokens->order);
@@ -140,13 +151,9 @@ static void tokens_expand( struct ureg_tokens *tokens,
tokens->tokens = REALLOC(tokens->tokens,
old_size,
tokens->size * sizeof(unsigned));
- if (tokens->tokens == NULL)
- goto fail;
-
- return;
-
-fail:
- tokens_error(tokens);
+ if (tokens->tokens == NULL) {
+ tokens_error(tokens);
+ }
}
static void set_bad( struct ureg_program *ureg )
@@ -196,9 +203,13 @@ ureg_dst_register( unsigned file,
dst.IndirectIndex = 0;
dst.IndirectSwizzle = 0;
dst.Saturate = 0;
+ dst.Predicate = 0;
+ dst.PredNegate = 0;
+ dst.PredSwizzleX = TGSI_SWIZZLE_X;
+ dst.PredSwizzleY = TGSI_SWIZZLE_Y;
+ dst.PredSwizzleZ = TGSI_SWIZZLE_Z;
+ dst.PredSwizzleW = TGSI_SWIZZLE_W;
dst.Index = index;
- dst.Pad1 = 0;
- dst.Pad2 = 0;
return dst;
}
@@ -228,25 +239,25 @@ ureg_src_register( unsigned file,
-static struct ureg_src
-ureg_DECL_input( struct ureg_program *ureg,
- unsigned name,
- unsigned index,
- unsigned interp_mode )
+struct ureg_src
+ureg_DECL_fs_input( struct ureg_program *ureg,
+ unsigned name,
+ unsigned index,
+ unsigned interp_mode )
{
unsigned i;
- for (i = 0; i < ureg->nr_inputs; i++) {
- if (ureg->input[i].semantic_name == name &&
- ureg->input[i].semantic_index == index)
+ for (i = 0; i < ureg->nr_fs_inputs; i++) {
+ if (ureg->fs_input[i].semantic_name == name &&
+ ureg->fs_input[i].semantic_index == index)
goto out;
}
- if (ureg->nr_inputs < UREG_MAX_INPUT) {
- ureg->input[i].semantic_name = name;
- ureg->input[i].semantic_index = index;
- ureg->input[i].interp = interp_mode;
- ureg->nr_inputs++;
+ if (ureg->nr_fs_inputs < UREG_MAX_INPUT) {
+ ureg->fs_input[i].semantic_name = name;
+ ureg->fs_input[i].semantic_index = index;
+ ureg->fs_input[i].interp = interp_mode;
+ ureg->nr_fs_inputs++;
}
else {
set_bad( ureg );
@@ -257,25 +268,14 @@ out:
}
-
-struct ureg_src
-ureg_DECL_fs_input( struct ureg_program *ureg,
- unsigned name,
- unsigned index,
- unsigned interp )
-{
- assert(ureg->processor == TGSI_PROCESSOR_FRAGMENT);
- return ureg_DECL_input( ureg, name, index, interp );
-}
-
-
struct ureg_src
ureg_DECL_vs_input( struct ureg_program *ureg,
- unsigned name,
unsigned index )
{
assert(ureg->processor == TGSI_PROCESSOR_VERTEX);
- return ureg_DECL_input( ureg, name, index, TGSI_INTERPOLATE_CONSTANT );
+
+ ureg->vs_inputs[index/32] |= 1 << (index % 32);
+ return ureg_src_register( TGSI_FILE_INPUT, index );
}
@@ -313,9 +313,57 @@ out:
* value or manage any constant_buffer contents -- that's the
* resposibility of the calling code.
*/
-struct ureg_src ureg_DECL_constant(struct ureg_program *ureg )
+struct ureg_src ureg_DECL_constant(struct ureg_program *ureg,
+ unsigned index )
{
- return ureg_src_register( TGSI_FILE_CONSTANT, ureg->nr_constants++ );
+ unsigned minconst = index, maxconst = index;
+ unsigned i;
+
+ /* Inside existing range?
+ */
+ for (i = 0; i < ureg->nr_constant_ranges; i++) {
+ if (ureg->constant_range[i].first <= index &&
+ ureg->constant_range[i].last >= index)
+ goto out;
+ }
+
+ /* Extend existing range?
+ */
+ for (i = 0; i < ureg->nr_constant_ranges; i++) {
+ if (ureg->constant_range[i].last == index - 1) {
+ ureg->constant_range[i].last = index;
+ goto out;
+ }
+
+ if (ureg->constant_range[i].first == index + 1) {
+ ureg->constant_range[i].first = index;
+ goto out;
+ }
+
+ minconst = MIN2(minconst, ureg->constant_range[i].first);
+ maxconst = MAX2(maxconst, ureg->constant_range[i].last);
+ }
+
+ /* Create new range?
+ */
+ if (ureg->nr_constant_ranges < UREG_MAX_CONSTANT_RANGE) {
+ i = ureg->nr_constant_ranges++;
+ ureg->constant_range[i].first = index;
+ ureg->constant_range[i].last = index;
+ }
+
+ /* Collapse all ranges down to one:
+ */
+ i = 0;
+ ureg->constant_range[0].first = minconst;
+ ureg->constant_range[0].last = maxconst;
+ ureg->nr_constant_ranges = 1;
+
+out:
+ assert(i < ureg->nr_constant_ranges);
+ assert(ureg->constant_range[i].first <= index);
+ assert(ureg->constant_range[i].last >= index);
+ return ureg_src_register( TGSI_FILE_CONSTANT, index );
}
@@ -369,6 +417,19 @@ struct ureg_dst ureg_DECL_address( struct ureg_program *ureg )
return ureg_dst_register( TGSI_FILE_ADDRESS, 0 );
}
+/* Allocate a new predicate register.
+ */
+struct ureg_dst
+ureg_DECL_predicate(struct ureg_program *ureg)
+{
+ if (ureg->nr_preds < UREG_MAX_PRED) {
+ return ureg_dst_register(TGSI_FILE_PREDICATE, ureg->nr_preds++);
+ }
+
+ assert(0);
+ return ureg_dst_register(TGSI_FILE_PREDICATE, 0);
+}
+
/* Allocate a new sampler.
*/
struct ureg_src ureg_DECL_sampler( struct ureg_program *ureg,
@@ -566,17 +627,40 @@ ureg_emit_dst( struct ureg_program *ureg,
}
+static void validate( unsigned opcode,
+ unsigned nr_dst,
+ unsigned nr_src )
+{
+#ifdef DEBUG
+ const struct tgsi_opcode_info *info = tgsi_get_opcode_info( opcode );
+ assert(info);
+ if(info) {
+ assert(nr_dst == info->num_dst);
+ assert(nr_src == info->num_src);
+ }
+#endif
+}
-unsigned
+struct ureg_emit_insn_result
ureg_emit_insn(struct ureg_program *ureg,
unsigned opcode,
boolean saturate,
+ boolean predicate,
+ boolean pred_negate,
+ unsigned pred_swizzle_x,
+ unsigned pred_swizzle_y,
+ unsigned pred_swizzle_z,
+ unsigned pred_swizzle_w,
unsigned num_dst,
unsigned num_src )
{
union tgsi_any_token *out;
+ uint count = predicate ? 2 : 1;
+ struct ureg_emit_insn_result result;
- out = get_tokens( ureg, DOMAIN_INSN, 1 );
+ validate( opcode, num_dst, num_src );
+
+ out = get_tokens( ureg, DOMAIN_INSN, count );
out[0].value = 0;
out[0].insn.Type = TGSI_TOKEN_TYPE_INSTRUCTION;
out[0].insn.NrTokens = 0;
@@ -585,17 +669,34 @@ ureg_emit_insn(struct ureg_program *ureg,
out[0].insn.NumDstRegs = num_dst;
out[0].insn.NumSrcRegs = num_src;
out[0].insn.Padding = 0;
- out[0].insn.Extended = 0;
-
+
+ result.insn_token = ureg->domain[DOMAIN_INSN].count - count;
+
+ if (predicate) {
+ out[0].insn.Extended = 1;
+ out[1].insn_ext_predicate = tgsi_default_instruction_ext_predicate();
+ out[1].insn_ext_predicate.Negate = pred_negate;
+ out[1].insn_ext_predicate.SwizzleX = pred_swizzle_x;
+ out[1].insn_ext_predicate.SwizzleY = pred_swizzle_y;
+ out[1].insn_ext_predicate.SwizzleZ = pred_swizzle_z;
+ out[1].insn_ext_predicate.SwizzleW = pred_swizzle_w;
+
+ result.extended_token = result.insn_token + 1;
+ } else {
+ out[0].insn.Extended = 0;
+
+ result.extended_token = result.insn_token;
+ }
+
ureg->nr_instructions++;
- return ureg->domain[DOMAIN_INSN].count - 1;
+ return result;
}
void
ureg_emit_label(struct ureg_program *ureg,
- unsigned insn_token,
+ unsigned extended_token,
unsigned *label_token )
{
union tgsi_any_token *out, *insn;
@@ -604,9 +705,9 @@ ureg_emit_label(struct ureg_program *ureg,
return;
out = get_tokens( ureg, DOMAIN_INSN, 1 );
- insn = retrieve_token( ureg, DOMAIN_INSN, insn_token );
+ insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
- insn->insn.Extended = 1;
+ insn->token.Extended = 1;
out[0].value = 0;
out[0].insn_ext_label.Type = TGSI_INSTRUCTION_EXT_TYPE_LABEL;
@@ -640,15 +741,15 @@ ureg_fixup_label(struct ureg_program *ureg,
void
ureg_emit_texture(struct ureg_program *ureg,
- unsigned insn_token,
+ unsigned extended_token,
unsigned target )
{
union tgsi_any_token *out, *insn;
out = get_tokens( ureg, DOMAIN_INSN, 1 );
- insn = retrieve_token( ureg, DOMAIN_INSN, insn_token );
+ insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
- insn->insn.Extended = 1;
+ insn->token.Extended = 1;
out[0].value = 0;
out[0].insn_ext_texture.Type = TGSI_INSTRUCTION_EXT_TYPE_TEXTURE;
@@ -675,23 +776,83 @@ ureg_insn(struct ureg_program *ureg,
const struct ureg_src *src,
unsigned nr_src )
{
- unsigned insn, i;
+ struct ureg_emit_insn_result insn;
+ unsigned i;
boolean saturate;
+ boolean predicate;
+ boolean negate;
+ unsigned swizzle[4];
-#ifdef DEBUG
- {
- const struct tgsi_opcode_info *info = tgsi_get_opcode_info( opcode );
- assert(info);
- if(info) {
- assert(nr_dst == info->num_dst);
- assert(nr_src == info->num_src);
- }
+ saturate = nr_dst ? dst[0].Saturate : FALSE;
+ predicate = nr_dst ? dst[0].Predicate : FALSE;
+ if (predicate) {
+ negate = dst[0].PredNegate;
+ swizzle[0] = dst[0].PredSwizzleX;
+ swizzle[1] = dst[0].PredSwizzleY;
+ swizzle[2] = dst[0].PredSwizzleZ;
+ swizzle[3] = dst[0].PredSwizzleW;
}
-#endif
-
+
+ insn = ureg_emit_insn(ureg,
+ opcode,
+ saturate,
+ predicate,
+ negate,
+ swizzle[0],
+ swizzle[1],
+ swizzle[2],
+ swizzle[3],
+ nr_dst,
+ nr_src);
+
+ for (i = 0; i < nr_dst; i++)
+ ureg_emit_dst( ureg, dst[i] );
+
+ for (i = 0; i < nr_src; i++)
+ ureg_emit_src( ureg, src[i] );
+
+ ureg_fixup_insn_size( ureg, insn.insn_token );
+}
+
+void
+ureg_tex_insn(struct ureg_program *ureg,
+ unsigned opcode,
+ const struct ureg_dst *dst,
+ unsigned nr_dst,
+ unsigned target,
+ const struct ureg_src *src,
+ unsigned nr_src )
+{
+ struct ureg_emit_insn_result insn;
+ unsigned i;
+ boolean saturate;
+ boolean predicate;
+ boolean negate;
+ unsigned swizzle[4];
+
saturate = nr_dst ? dst[0].Saturate : FALSE;
+ predicate = nr_dst ? dst[0].Predicate : FALSE;
+ if (predicate) {
+ negate = dst[0].PredNegate;
+ swizzle[0] = dst[0].PredSwizzleX;
+ swizzle[1] = dst[0].PredSwizzleY;
+ swizzle[2] = dst[0].PredSwizzleZ;
+ swizzle[3] = dst[0].PredSwizzleW;
+ }
+
+ insn = ureg_emit_insn(ureg,
+ opcode,
+ saturate,
+ predicate,
+ negate,
+ swizzle[0],
+ swizzle[1],
+ swizzle[2],
+ swizzle[3],
+ nr_dst,
+ nr_src);
- insn = ureg_emit_insn( ureg, opcode, saturate, nr_dst, nr_src );
+ ureg_emit_texture( ureg, insn.extended_token, target );
for (i = 0; i < nr_dst; i++)
ureg_emit_dst( ureg, dst[i] );
@@ -699,7 +860,38 @@ ureg_insn(struct ureg_program *ureg,
for (i = 0; i < nr_src; i++)
ureg_emit_src( ureg, src[i] );
- ureg_fixup_insn_size( ureg, insn );
+ ureg_fixup_insn_size( ureg, insn.insn_token );
+}
+
+
+void
+ureg_label_insn(struct ureg_program *ureg,
+ unsigned opcode,
+ const struct ureg_src *src,
+ unsigned nr_src,
+ unsigned *label_token )
+{
+ struct ureg_emit_insn_result insn;
+ unsigned i;
+
+ insn = ureg_emit_insn(ureg,
+ opcode,
+ FALSE,
+ FALSE,
+ FALSE,
+ TGSI_SWIZZLE_X,
+ TGSI_SWIZZLE_Y,
+ TGSI_SWIZZLE_Z,
+ TGSI_SWIZZLE_W,
+ 0,
+ nr_src);
+
+ ureg_emit_label( ureg, insn.extended_token, label_token );
+
+ for (i = 0; i < nr_src; i++)
+ ureg_emit_src( ureg, src[i] );
+
+ ureg_fixup_insn_size( ureg, insn.insn_token );
}
@@ -777,13 +969,22 @@ static void emit_decls( struct ureg_program *ureg )
{
unsigned i;
- for (i = 0; i < ureg->nr_inputs; i++) {
- emit_decl( ureg,
- TGSI_FILE_INPUT,
- i,
- ureg->input[i].semantic_name,
- ureg->input[i].semantic_index,
- ureg->input[i].interp );
+ if (ureg->processor == TGSI_PROCESSOR_VERTEX) {
+ for (i = 0; i < UREG_MAX_INPUT; i++) {
+ if (ureg->vs_inputs[i/32] & (1 << (i%32))) {
+ emit_decl_range( ureg, TGSI_FILE_INPUT, i, 1 );
+ }
+ }
+ }
+ else {
+ for (i = 0; i < ureg->nr_fs_inputs; i++) {
+ emit_decl( ureg,
+ TGSI_FILE_INPUT,
+ i,
+ ureg->fs_input[i].semantic_name,
+ ureg->fs_input[i].semantic_index,
+ ureg->fs_input[i].interp );
+ }
}
for (i = 0; i < ureg->nr_outputs; i++) {
@@ -801,10 +1002,13 @@ static void emit_decls( struct ureg_program *ureg )
ureg->sampler[i].Index, 1 );
}
- if (ureg->nr_constants) {
- emit_decl_range( ureg,
- TGSI_FILE_CONSTANT,
- 0, ureg->nr_constants );
+ if (ureg->nr_constant_ranges) {
+ for (i = 0; i < ureg->nr_constant_ranges; i++)
+ emit_decl_range( ureg,
+ TGSI_FILE_CONSTANT,
+ ureg->constant_range[i].first,
+ (ureg->constant_range[i].last + 1 -
+ ureg->constant_range[i].first) );
}
if (ureg->nr_temps) {
@@ -819,6 +1023,13 @@ static void emit_decls( struct ureg_program *ureg )
0, ureg->nr_addrs );
}
+ if (ureg->nr_preds) {
+ emit_decl_range(ureg,
+ TGSI_FILE_PREDICATE,
+ 0,
+ ureg->nr_preds);
+ }
+
for (i = 0; i < ureg->nr_immediates; i++) {
emit_immediate( ureg,
ureg->immediate[i].v );
@@ -890,6 +1101,15 @@ const struct tgsi_token *ureg_finalize( struct ureg_program *ureg )
ureg->domain[DOMAIN_DECL].count);
tgsi_dump( tokens, 0 );
}
+
+#if DEBUG
+ if (tokens && !tgsi_sanity_check(tokens)) {
+ debug_printf("tgsi_ureg.c, sanity check failed on generated tokens:\n");
+ tgsi_dump(tokens, 0);
+ assert(0);
+ }
+#endif
+
return tokens;
}
@@ -911,6 +1131,25 @@ void *ureg_create_shader( struct ureg_program *ureg,
}
+const struct tgsi_token *ureg_get_tokens( struct ureg_program *ureg,
+ unsigned *nr_tokens )
+{
+ const struct tgsi_token *tokens;
+
+ ureg_finalize(ureg);
+
+ tokens = &ureg->domain[DOMAIN_DECL].tokens[0].token;
+
+ if (nr_tokens)
+ *nr_tokens = ureg->domain[DOMAIN_DECL].size;
+
+ ureg->domain[DOMAIN_DECL].tokens = 0;
+ ureg->domain[DOMAIN_DECL].size = 0;
+ ureg->domain[DOMAIN_DECL].order = 0;
+ ureg->domain[DOMAIN_DECL].count = 0;
+
+ return tokens;
+}
struct ureg_program *ureg_create( unsigned processor )