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.c403
1 files changed, 294 insertions, 109 deletions
diff --git a/src/gallium/auxiliary/tgsi/tgsi_ureg.c b/src/gallium/auxiliary/tgsi/tgsi_ureg.c
index bf39cf5409..ab557a23f9 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_ureg.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_ureg.c
@@ -37,24 +37,23 @@
#include "util/u_math.h"
union tgsi_any_token {
- struct tgsi_version version;
struct tgsi_header header;
struct tgsi_processor processor;
struct tgsi_token token;
+ struct tgsi_property prop;
+ struct tgsi_property_data prop_data;
struct tgsi_declaration decl;
struct tgsi_declaration_range decl_range;
struct tgsi_declaration_semantic decl_semantic;
struct tgsi_immediate imm;
union tgsi_immediate_data imm_data;
struct tgsi_instruction insn;
- 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_instruction_predicate insn_predicate;
+ struct tgsi_instruction_label insn_label;
+ struct tgsi_instruction_texture insn_texture;
struct tgsi_src_register src;
- struct tgsi_src_register_ext_mod src_ext_mod;
struct tgsi_dimension dim;
struct tgsi_dst_register dst;
- struct tgsi_dst_register_ext_modulate dst_ext_mod;
unsigned value;
};
@@ -67,11 +66,13 @@ struct ureg_tokens {
};
#define UREG_MAX_INPUT PIPE_MAX_ATTRIBS
+#define UREG_MAX_SYSTEM_VALUE 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_LOOP 1
#define UREG_MAX_PRED 1
#define DOMAIN_DECL 0
@@ -92,14 +93,31 @@ struct ureg_program
unsigned vs_inputs[UREG_MAX_INPUT/32];
struct {
+ unsigned index;
+ } gs_input[UREG_MAX_INPUT];
+ unsigned nr_gs_inputs;
+
+ struct {
+ unsigned index;
+ unsigned semantic_name;
+ unsigned semantic_index;
+ } system_value[UREG_MAX_SYSTEM_VALUE];
+ unsigned nr_system_values;
+
+ struct {
unsigned semantic_name;
unsigned semantic_index;
} output[UREG_MAX_OUTPUT];
unsigned nr_outputs;
struct {
- float v[4];
+ union {
+ float f[4];
+ unsigned u[4];
+ int i[4];
+ } value;
unsigned nr;
+ unsigned type;
} immediate[UREG_MAX_IMMEDIATE];
unsigned nr_immediates;
@@ -115,8 +133,11 @@ struct ureg_program
} constant_range[UREG_MAX_CONSTANT_RANGE];
unsigned nr_constant_ranges;
+ unsigned property_gs_input_prim;
+
unsigned nr_addrs;
unsigned nr_preds;
+ unsigned nr_loops;
unsigned nr_instructions;
struct ureg_tokens domain[2];
@@ -225,19 +246,29 @@ ureg_src_register( unsigned file,
src.SwizzleY = TGSI_SWIZZLE_Y;
src.SwizzleZ = TGSI_SWIZZLE_Z;
src.SwizzleW = TGSI_SWIZZLE_W;
- src.Pad = 0;
src.Indirect = 0;
+ src.IndirectFile = TGSI_FILE_NULL;
src.IndirectIndex = 0;
src.IndirectSwizzle = 0;
src.Absolute = 0;
src.Index = index;
src.Negate = 0;
+ src.Dimension = 0;
+ src.DimensionIndex = 0;
return src;
}
+void
+ureg_property_gs_input_prim(struct ureg_program *ureg,
+ unsigned gs_input_prim)
+{
+ ureg->property_gs_input_prim = gs_input_prim;
+}
+
+
struct ureg_src
ureg_DECL_fs_input( struct ureg_program *ureg,
@@ -279,6 +310,41 @@ ureg_DECL_vs_input( struct ureg_program *ureg,
}
+struct ureg_src
+ureg_DECL_gs_input(struct ureg_program *ureg,
+ unsigned index)
+{
+ if (ureg->nr_gs_inputs < UREG_MAX_INPUT) {
+ ureg->gs_input[ureg->nr_gs_inputs].index = index;
+ ureg->nr_gs_inputs++;
+ } else {
+ set_bad(ureg);
+ }
+
+ /* XXX: Add suport for true 2D input registers. */
+ return ureg_src_register(TGSI_FILE_INPUT, index);
+}
+
+
+struct ureg_src
+ureg_DECL_system_value(struct ureg_program *ureg,
+ unsigned index,
+ unsigned semantic_name,
+ unsigned semantic_index)
+{
+ if (ureg->nr_system_values < UREG_MAX_SYSTEM_VALUE) {
+ ureg->system_value[ureg->nr_system_values].index = index;
+ ureg->system_value[ureg->nr_system_values].semantic_name = semantic_name;
+ ureg->system_value[ureg->nr_system_values].semantic_index = semantic_index;
+ ureg->nr_system_values++;
+ } else {
+ set_bad(ureg);
+ }
+
+ return ureg_src_register(TGSI_FILE_SYSTEM_VALUE, index);
+}
+
+
struct ureg_dst
ureg_DECL_output( struct ureg_program *ureg,
unsigned name,
@@ -350,6 +416,7 @@ struct ureg_src ureg_DECL_constant(struct ureg_program *ureg,
i = ureg->nr_constant_ranges++;
ureg->constant_range[i].first = index;
ureg->constant_range[i].last = index;
+ goto out;
}
/* Collapse all ranges down to one:
@@ -417,6 +484,19 @@ struct ureg_dst ureg_DECL_address( struct ureg_program *ureg )
return ureg_dst_register( TGSI_FILE_ADDRESS, 0 );
}
+/* Allocate a new loop register.
+ */
+struct ureg_dst
+ureg_DECL_loop(struct ureg_program *ureg)
+{
+ if (ureg->nr_loops < UREG_MAX_LOOP) {
+ return ureg_dst_register(TGSI_FILE_LOOP, ureg->nr_loops++);
+ }
+
+ assert(0);
+ return ureg_dst_register(TGSI_FILE_LOOP, 0);
+}
+
/* Allocate a new predicate register.
*/
struct ureg_dst
@@ -452,22 +532,22 @@ struct ureg_src ureg_DECL_sampler( struct ureg_program *ureg,
}
-
-
-static int match_or_expand_immediate( const float *v,
- unsigned nr,
- float *v2,
- unsigned *nr2,
- unsigned *swizzle )
+static int
+match_or_expand_immediate( const unsigned *v,
+ unsigned nr,
+ unsigned *v2,
+ unsigned *pnr2,
+ unsigned *swizzle )
{
+ unsigned nr2 = *pnr2;
unsigned i, j;
-
+
*swizzle = 0;
for (i = 0; i < nr; i++) {
boolean found = FALSE;
- for (j = 0; j < *nr2 && !found; j++) {
+ for (j = 0; j < nr2 && !found; j++) {
if (v[i] == v2[j]) {
*swizzle |= j << (i * 2);
found = TRUE;
@@ -475,24 +555,28 @@ static int match_or_expand_immediate( const float *v,
}
if (!found) {
- if (*nr2 >= 4)
+ if (nr2 >= 4) {
return FALSE;
+ }
- v2[*nr2] = v[i];
- *swizzle |= *nr2 << (i * 2);
- (*nr2)++;
+ v2[nr2] = v[i];
+ *swizzle |= nr2 << (i * 2);
+ nr2++;
}
}
+ /* Actually expand immediate only when fully succeeded.
+ */
+ *pnr2 = nr2;
return TRUE;
}
-
-
-struct ureg_src ureg_DECL_immediate( struct ureg_program *ureg,
- const float *v,
- unsigned nr )
+static struct ureg_src
+decl_immediate( struct ureg_program *ureg,
+ const unsigned *v,
+ unsigned nr,
+ unsigned type )
{
unsigned i, j;
unsigned swizzle;
@@ -502,38 +586,111 @@ struct ureg_src ureg_DECL_immediate( struct ureg_program *ureg,
*/
for (i = 0; i < ureg->nr_immediates; i++) {
- if (match_or_expand_immediate( v,
- nr,
- ureg->immediate[i].v,
- &ureg->immediate[i].nr,
- &swizzle ))
+ if (ureg->immediate[i].type != type) {
+ continue;
+ }
+ if (match_or_expand_immediate(v,
+ nr,
+ ureg->immediate[i].value.u,
+ &ureg->immediate[i].nr,
+ &swizzle)) {
goto out;
+ }
}
if (ureg->nr_immediates < UREG_MAX_IMMEDIATE) {
i = ureg->nr_immediates++;
- if (match_or_expand_immediate( v,
- nr,
- ureg->immediate[i].v,
- &ureg->immediate[i].nr,
- &swizzle ))
+ ureg->immediate[i].type = type;
+ if (match_or_expand_immediate(v,
+ nr,
+ ureg->immediate[i].value.u,
+ &ureg->immediate[i].nr,
+ &swizzle)) {
goto out;
+ }
}
- set_bad( ureg );
+ set_bad(ureg);
out:
/* Make sure that all referenced elements are from this immediate.
* Has the effect of making size-one immediates into scalars.
*/
- for (j = nr; j < 4; j++)
+ for (j = nr; j < 4; j++) {
swizzle |= (swizzle & 0x3) << (j * 2);
+ }
+
+ return ureg_swizzle(ureg_src_register(TGSI_FILE_IMMEDIATE, i),
+ (swizzle >> 0) & 0x3,
+ (swizzle >> 2) & 0x3,
+ (swizzle >> 4) & 0x3,
+ (swizzle >> 6) & 0x3);
+}
+
- return ureg_swizzle( ureg_src_register( TGSI_FILE_IMMEDIATE, i ),
- (swizzle >> 0) & 0x3,
- (swizzle >> 2) & 0x3,
- (swizzle >> 4) & 0x3,
- (swizzle >> 6) & 0x3);
+struct ureg_src
+ureg_DECL_immediate( struct ureg_program *ureg,
+ const float *v,
+ unsigned nr )
+{
+ union {
+ float f[4];
+ unsigned u[4];
+ } fu;
+ unsigned int i;
+
+ for (i = 0; i < nr; i++) {
+ fu.f[i] = v[i];
+ }
+
+ return decl_immediate(ureg, fu.u, nr, TGSI_IMM_FLOAT32);
+}
+
+
+struct ureg_src
+ureg_DECL_immediate_uint( struct ureg_program *ureg,
+ const unsigned *v,
+ unsigned nr )
+{
+ return decl_immediate(ureg, v, nr, TGSI_IMM_UINT32);
+}
+
+
+struct ureg_src
+ureg_DECL_immediate_block_uint( struct ureg_program *ureg,
+ const unsigned *v,
+ unsigned nr )
+{
+ uint index;
+ uint i;
+
+ if (ureg->nr_immediates + (nr + 3) / 4 > UREG_MAX_IMMEDIATE) {
+ set_bad(ureg);
+ return ureg_src_register(TGSI_FILE_IMMEDIATE, 0);
+ }
+
+ index = ureg->nr_immediates;
+ ureg->nr_immediates += (nr + 3) / 4;
+
+ for (i = index; i < ureg->nr_immediates; i++) {
+ ureg->immediate[i].type = TGSI_IMM_UINT32;
+ ureg->immediate[i].nr = nr > 4 ? 4 : nr;
+ memcpy(ureg->immediate[i].value.u,
+ &v[(i - index) * 4],
+ ureg->immediate[i].nr * sizeof(uint));
+ nr -= 4;
+ }
+
+ return ureg_src_register(TGSI_FILE_IMMEDIATE, index);
+}
+
+
+struct ureg_src
+ureg_DECL_immediate_int( struct ureg_program *ureg,
+ const int *v,
+ unsigned nr )
+{
+ return decl_immediate(ureg, (const unsigned *)v, nr, TGSI_IMM_INT32);
}
@@ -541,9 +698,7 @@ void
ureg_emit_src( struct ureg_program *ureg,
struct ureg_src src )
{
- unsigned size = (1 +
- (src.Absolute ? 1 : 0) +
- (src.Indirect ? 1 : 0));
+ unsigned size = 1 + (src.Indirect ? 1 : 0) + (src.Dimension ? 1 : 0);
union tgsi_any_token *out = get_tokens( ureg, DOMAIN_INSN, size );
unsigned n = 0;
@@ -560,22 +715,13 @@ ureg_emit_src( struct ureg_program *ureg,
out[n].src.SwizzleW = src.SwizzleW;
out[n].src.Index = src.Index;
out[n].src.Negate = src.Negate;
+ out[0].src.Absolute = src.Absolute;
n++;
-
- if (src.Absolute) {
- out[0].src.Extended = 1;
- out[0].src.Negate = 0;
- out[n].value = 0;
- out[n].src_ext_mod.Type = TGSI_SRC_REGISTER_EXT_TYPE_MOD;
- out[n].src_ext_mod.Absolute = 1;
- out[n].src_ext_mod.Negate = src.Negate;
- n++;
- }
if (src.Indirect) {
out[0].src.Indirect = 1;
out[n].value = 0;
- out[n].src.File = TGSI_FILE_ADDRESS;
+ out[n].src.File = src.IndirectFile;
out[n].src.SwizzleX = src.IndirectSwizzle;
out[n].src.SwizzleY = src.IndirectSwizzle;
out[n].src.SwizzleZ = src.IndirectSwizzle;
@@ -584,6 +730,15 @@ ureg_emit_src( struct ureg_program *ureg,
n++;
}
+ if (src.Dimension) {
+ out[0].src.Dimension = 1;
+ out[n].dim.Indirect = 0;
+ out[n].dim.Dimension = 0;
+ out[n].dim.Padding = 0;
+ out[n].dim.Index = src.DimensionIndex;
+ n++;
+ }
+
assert(n == size);
}
@@ -661,35 +816,27 @@ ureg_emit_insn(struct ureg_program *ureg,
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;
+ out[0].insn = tgsi_default_instruction();
out[0].insn.Opcode = opcode;
out[0].insn.Saturate = saturate;
out[0].insn.NumDstRegs = num_dst;
out[0].insn.NumSrcRegs = num_src;
- out[0].insn.Padding = 0;
result.insn_token = ureg->domain[DOMAIN_INSN].count - count;
+ result.extended_token = result.insn_token;
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;
+ out[0].insn.Predicate = 1;
+ out[1].insn_predicate = tgsi_default_instruction_predicate();
+ out[1].insn_predicate.Negate = pred_negate;
+ out[1].insn_predicate.SwizzleX = pred_swizzle_x;
+ out[1].insn_predicate.SwizzleY = pred_swizzle_y;
+ out[1].insn_predicate.SwizzleZ = pred_swizzle_z;
+ out[1].insn_predicate.SwizzleW = pred_swizzle_w;
}
ureg->nr_instructions++;
-
+
return result;
}
@@ -705,13 +852,11 @@ ureg_emit_label(struct ureg_program *ureg,
return;
out = get_tokens( ureg, DOMAIN_INSN, 1 );
- insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
+ out[0].value = 0;
- insn->token.Extended = 1;
+ insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
+ insn->insn.Label = 1;
- out[0].value = 0;
- out[0].insn_ext_label.Type = TGSI_INSTRUCTION_EXT_TYPE_LABEL;
-
*label_token = ureg->domain[DOMAIN_INSN].count - 1;
}
@@ -734,8 +879,7 @@ ureg_fixup_label(struct ureg_program *ureg,
{
union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_INSN, label_token );
- assert(out->insn_ext_label.Type == TGSI_INSTRUCTION_EXT_TYPE_LABEL);
- out->insn_ext_label.Label = instruction_number;
+ out->insn_label.Label = instruction_number;
}
@@ -749,11 +893,10 @@ ureg_emit_texture(struct ureg_program *ureg,
out = get_tokens( ureg, DOMAIN_INSN, 1 );
insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
- insn->token.Extended = 1;
+ insn->insn.Texture = 1;
out[0].value = 0;
- out[0].insn_ext_texture.Type = TGSI_INSTRUCTION_EXT_TYPE_TEXTURE;
- out[0].insn_ext_texture.Texture = target;
+ out[0].insn_texture.Texture = target;
}
@@ -918,8 +1061,8 @@ static void emit_decl( struct ureg_program *ureg,
out[1].decl_range.Last = index;
out[2].value = 0;
- out[2].decl_semantic.SemanticName = semantic_name;
- out[2].decl_semantic.SemanticIndex = semantic_index;
+ out[2].decl_semantic.Name = semantic_name;
+ out[2].decl_semantic.Index = semantic_index;
}
@@ -944,39 +1087,60 @@ static void emit_decl_range( struct ureg_program *ureg,
out[1].decl_range.Last = first + count - 1;
}
-static void emit_immediate( struct ureg_program *ureg,
- const float *v )
+static void
+emit_immediate( struct ureg_program *ureg,
+ const unsigned *v,
+ unsigned type )
{
union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 5 );
out[0].value = 0;
out[0].imm.Type = TGSI_TOKEN_TYPE_IMMEDIATE;
out[0].imm.NrTokens = 5;
- out[0].imm.DataType = TGSI_IMM_FLOAT32;
+ out[0].imm.DataType = type;
out[0].imm.Padding = 0;
- out[0].imm.Extended = 0;
- out[1].imm_data.Float = v[0];
- out[2].imm_data.Float = v[1];
- out[3].imm_data.Float = v[2];
- out[4].imm_data.Float = v[3];
+ out[1].imm_data.Uint = v[0];
+ out[2].imm_data.Uint = v[1];
+ out[3].imm_data.Uint = v[2];
+ out[4].imm_data.Uint = v[3];
}
+static void
+emit_property(struct ureg_program *ureg,
+ unsigned name,
+ unsigned data)
+{
+ union tgsi_any_token *out = get_tokens(ureg, DOMAIN_DECL, 2);
+ out[0].value = 0;
+ out[0].prop.Type = TGSI_TOKEN_TYPE_PROPERTY;
+ out[0].prop.NrTokens = 2;
+ out[0].prop.PropertyName = name;
+
+ out[1].prop_data.Data = data;
+}
static void emit_decls( struct ureg_program *ureg )
{
unsigned i;
+ if (ureg->property_gs_input_prim != ~0) {
+ assert(ureg->processor == TGSI_PROCESSOR_GEOMETRY);
+
+ emit_property(ureg,
+ TGSI_PROPERTY_GS_INPUT_PRIM,
+ ureg->property_gs_input_prim);
+ }
+
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 {
+ } else if (ureg->processor == TGSI_PROCESSOR_FRAGMENT) {
for (i = 0; i < ureg->nr_fs_inputs; i++) {
emit_decl( ureg,
TGSI_FILE_INPUT,
@@ -985,6 +1149,22 @@ static void emit_decls( struct ureg_program *ureg )
ureg->fs_input[i].semantic_index,
ureg->fs_input[i].interp );
}
+ } else {
+ for (i = 0; i < ureg->nr_gs_inputs; i++) {
+ emit_decl_range(ureg,
+ TGSI_FILE_INPUT,
+ ureg->gs_input[i].index,
+ 1);
+ }
+ }
+
+ for (i = 0; i < ureg->nr_system_values; i++) {
+ emit_decl(ureg,
+ TGSI_FILE_SYSTEM_VALUE,
+ ureg->system_value[i].index,
+ ureg->system_value[i].semantic_name,
+ ureg->system_value[i].semantic_index,
+ TGSI_INTERPOLATE_CONSTANT);
}
for (i = 0; i < ureg->nr_outputs; i++) {
@@ -1023,6 +1203,13 @@ static void emit_decls( struct ureg_program *ureg )
0, ureg->nr_addrs );
}
+ if (ureg->nr_loops) {
+ emit_decl_range(ureg,
+ TGSI_FILE_LOOP,
+ 0,
+ ureg->nr_loops);
+ }
+
if (ureg->nr_preds) {
emit_decl_range(ureg,
TGSI_FILE_PREDICATE,
@@ -1032,7 +1219,8 @@ static void emit_decls( struct ureg_program *ureg )
for (i = 0; i < ureg->nr_immediates; i++) {
emit_immediate( ureg,
- ureg->immediate[i].v );
+ ureg->immediate[i].value.u,
+ ureg->immediate[i].type );
}
}
@@ -1055,26 +1243,22 @@ static void copy_instructions( struct ureg_program *ureg )
static void
fixup_header_size(struct ureg_program *ureg)
{
- union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_DECL, 1 );
+ union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_DECL, 0 );
- out->header.BodySize = ureg->domain[DOMAIN_DECL].count - 3;
+ out->header.BodySize = ureg->domain[DOMAIN_DECL].count - 2;
}
static void
emit_header( struct ureg_program *ureg )
{
- union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 3 );
-
- out[0].version.MajorVersion = 1;
- out[0].version.MinorVersion = 1;
- out[0].version.Padding = 0;
+ union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 2 );
- out[1].header.HeaderSize = 2;
- out[1].header.BodySize = 0;
+ out[0].header.HeaderSize = 2;
+ out[0].header.BodySize = 0;
- out[2].processor.Processor = ureg->processor;
- out[2].processor.Padding = 0;
+ out[1].processor.Processor = ureg->processor;
+ out[1].processor.Padding = 0;
}
@@ -1159,6 +1343,7 @@ struct ureg_program *ureg_create( unsigned processor )
return NULL;
ureg->processor = processor;
+ ureg->property_gs_input_prim = ~0;
return ureg;
}