#include "brw_context.h"
#include "brw_vs.h"
#include "brw_debug.h"
+#include "brw_disasm.h"
+/* Choose one of the 4 vec4's which can be packed into each 16-wide reg.
+ */
+static INLINE struct brw_reg brw_vec4_grf_repeat( GLuint reg, GLuint slot )
+{
+ int nr = reg + slot/2;
+ int subnr = (slot%2) * 4;
+
+ return stride(brw_vec4_grf(nr, subnr), 0, 4, 1);
+}
static struct brw_reg get_tmp( struct brw_vs_compile *c )
}
+static boolean is_position_output( struct brw_vs_compile *c,
+ unsigned vs_output )
+{
+ const struct brw_vertex_shader *vs = c->vp;
+ unsigned semantic = vs->info.output_semantic_name[vs_output];
+ unsigned index = vs->info.output_semantic_index[vs_output];
+
+ return (semantic == TGSI_SEMANTIC_POSITION &&
+ index == 0);
+}
+
+
+static boolean find_output_slot( struct brw_vs_compile *c,
+ unsigned vs_output,
+ unsigned *fs_input_slot )
+{
+ const struct brw_vertex_shader *vs = c->vp;
+ unsigned semantic = vs->info.output_semantic_name[vs_output];
+ unsigned index = vs->info.output_semantic_index[vs_output];
+ unsigned i;
+
+ for (i = 0; i < c->key.fs_signature.nr_inputs; i++) {
+ if (c->key.fs_signature.input[i].semantic == semantic &&
+ c->key.fs_signature.input[i].semantic_index == index) {
+ *fs_input_slot = i;
+ return TRUE;
+ }
+ }
+
+ return FALSE;
+}
+
/**
* Preallocate GRF register before code emit.
*/
static void brw_vs_alloc_regs( struct brw_vs_compile *c )
{
- GLuint i, reg = 0, mrf;
+ GLuint i, reg = 0, subreg = 0, mrf;
int attributes_in_vue;
/* Determine whether to use a real constant buffer or use a block
* works if everything fits in the GRF.
* XXX this heuristic/check may need some fine tuning...
*/
- if (c->vp->info.file_max[TGSI_FILE_CONSTANT] +
- c->vp->info.file_max[TGSI_FILE_IMMEDIATE] +
- c->vp->info.file_max[TGSI_FILE_TEMPORARY] + 21 > BRW_MAX_GRF)
+ if (c->vp->info.file_max[TGSI_FILE_CONSTANT] + 1 +
+ c->vp->info.file_max[TGSI_FILE_IMMEDIATE] + 1 +
+ c->vp->info.file_max[TGSI_FILE_TEMPORARY] + 1 + 21 > BRW_MAX_GRF)
c->vp->use_const_buffer = GL_TRUE;
else {
/* XXX: immediates can go elsewhere if necessary:
*/
- assert(c->vp->info.file_max[TGSI_FILE_IMMEDIATE] +
- c->vp->info.file_max[TGSI_FILE_TEMPORARY] + 21 <= BRW_MAX_GRF);
+ assert(c->vp->info.file_max[TGSI_FILE_IMMEDIATE] + 1 +
+ c->vp->info.file_max[TGSI_FILE_TEMPORARY] + 1 + 21 <= BRW_MAX_GRF);
c->vp->use_const_buffer = GL_FALSE;
}
/* User clip planes from curbe:
*/
if (c->key.nr_userclip) {
- for (i = 0; i < c->key.nr_userclip; i++) {
- c->userplane[i] = stride( brw_vec4_grf(reg+3+i/2, (i%2) * 4), 0, 4, 1);
+ /* Skip over fixed planes: Or never read them into vs unit?
+ */
+ subreg += 6;
+
+ for (i = 0; i < c->key.nr_userclip; i++, subreg++) {
+ c->userplane[i] =
+ stride( brw_vec4_grf(reg+subreg/2, (subreg%2) * 4), 0, 4, 1);
}
/* Deal with curbe alignment:
*/
- reg += ((6 + c->key.nr_userclip + 3) / 4) * 2;
+ subreg = align(subreg, 2);
+ /*reg += ((6 + c->key.nr_userclip + 3) / 4) * 2;*/
}
- /* Vertex program parameters from curbe:
+
+ /* Immediates: always in the curbe.
+ *
+ * XXX: Can try to encode some immediates as brw immediates
+ * XXX: Make sure ureg sets minimal immediate size and respect it
+ * here.
*/
- if (c->vp->use_const_buffer) {
- /* get constants from a real constant buffer */
- c->prog_data.curb_read_length = 0;
- c->prog_data.nr_params = 4; /* XXX 0 causes a bug elsewhere... */
+ for (i = 0; i < c->vp->info.immediate_count; i++, subreg++) {
+ c->regs[TGSI_FILE_IMMEDIATE][i] =
+ stride( brw_vec4_grf(reg+subreg/2, (subreg%2) * 4), 0, 4, 1);
}
- else {
- /* use a section of the GRF for constants */
+ c->prog_data.nr_params = c->vp->info.immediate_count * 4;
+
+
+ /* Vertex constant buffer.
+ *
+ * Constants from the buffer can be either cached in the curbe or
+ * loaded as needed from the actual constant buffer.
+ */
+ if (!c->vp->use_const_buffer) {
GLuint nr_params = c->vp->info.file_max[TGSI_FILE_CONSTANT] + 1;
- for (i = 0; i < nr_params; i++) {
- c->regs[TGSI_FILE_CONSTANT][i] = stride( brw_vec4_grf(reg+i/2, (i%2) * 4), 0, 4, 1);
+
+ for (i = 0; i < nr_params; i++, subreg++) {
+ c->regs[TGSI_FILE_CONSTANT][i] =
+ stride( brw_vec4_grf(reg+subreg/2, (subreg%2) * 4), 0, 4, 1);
}
- reg += (nr_params + 1) / 2;
- c->prog_data.curb_read_length = reg - 1;
- c->prog_data.nr_params = nr_params * 4;
+
+ c->prog_data.nr_params += nr_params * 4;
}
+ /* All regs allocated
+ */
+ reg += (subreg + 1) / 2;
+ c->prog_data.curb_read_length = reg - 1;
+
+
/* Allocate input regs:
*/
c->nr_inputs = c->vp->info.num_inputs;
if (c->nr_inputs == 0)
reg++;
- /* Allocate a GRF and load immediate values by hand with 4 MOVs!!!
- *
- * XXX: Try to encode float immediates as brw immediates
- * XXX: Put immediates into the CURBE.
- * XXX: Make sure ureg sets minimal immediate size and respect it
- * here.
- */
- for (i = 0; i < c->nr_immediates; i++) {
- struct brw_reg r;
- int j;
-
- r = brw_vec8_grf(reg, 0);
-
- for (j = 0; j < 4; j++) {
- brw_MOV(&c->func,
- brw_writemask(r, (1<<j)),
- brw_imm_f(c->immediate[i][j]));
- }
-
- reg++;
- }
/* Allocate outputs. The non-position outputs go straight into message regs.
*/
- c->nr_outputs = 0;
- c->first_output = reg;
- c->first_overflow_output = 0;
+ c->nr_outputs = c->prog_data.nr_outputs;
if (c->chipset.is_igdng)
mrf = 8;
else
mrf = 4;
+
+ if (c->key.fs_signature.nr_inputs > BRW_MAX_MRF) {
+ c->overflow_grf_start = reg;
+ c->overflow_count = c->key.fs_signature.nr_inputs - BRW_MAX_MRF;
+ reg += c->overflow_count;
+ }
+
/* XXX: need to access vertex output semantics here:
*/
- c->nr_outputs = c->prog_data.nr_outputs;
- for (i = 0; i < c->prog_data.nr_outputs; i++) {
- assert(i < Elements(c->regs[TGSI_FILE_OUTPUT]));
+ for (i = 0; i < c->nr_outputs; i++) {
+ unsigned slot;
- /* XXX: Hardwire position to zero:
+ /* XXX: Put output position in slot zero always. Clipper, etc,
+ * need access to this reg.
*/
- if (i == 0) {
- c->regs[TGSI_FILE_OUTPUT][i] = brw_vec8_grf(reg, 0);
+ if (is_position_output(c, i)) {
+ c->regs[TGSI_FILE_OUTPUT][i] = brw_vec8_grf(reg, 0); /* copy to mrf 0 */
reg++;
}
- /* XXX: disable psiz:
- */
- else if (0) {
- c->regs[TGSI_FILE_OUTPUT][i] = brw_vec8_grf(reg, 0);
- reg++;
- mrf++; /* just a placeholder? XXX fix later stages & remove this */
- }
- else if (mrf < 16) {
- c->regs[TGSI_FILE_OUTPUT][i] = brw_message_reg(mrf);
- mrf++;
+ else if (find_output_slot(c, i, &slot)) {
+
+ if (0 /* is_psize_output(c, i) */ ) {
+ /* c->psize_out.grf = reg; */
+ /* c->psize_out.mrf = i; */
+ }
+
+ /* The first (16-4) outputs can go straight into the message regs.
+ */
+ if (slot + mrf < BRW_MAX_MRF) {
+ c->regs[TGSI_FILE_OUTPUT][i] = brw_message_reg(slot + mrf);
+ }
+ else {
+ int grf = c->overflow_grf_start + slot - BRW_MAX_MRF;
+ c->regs[TGSI_FILE_OUTPUT][i] = brw_vec8_grf(grf, 0);
+ }
}
else {
- /* too many vertex results to fit in MRF, use GRF for overflow */
- if (!c->first_overflow_output)
- c->first_overflow_output = i;
- c->regs[TGSI_FILE_OUTPUT][i] = brw_vec8_grf(reg, 0);
- reg++;
+ c->regs[TGSI_FILE_OUTPUT][i] = brw_null_reg();
}
}
}
}
+#if 0
for (i = 0; i < 128; i++) {
if (c->output_regs[i].used_in_src) {
c->output_regs[i].reg = brw_vec8_grf(reg, 0);
reg++;
}
}
+#endif
- c->stack = brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, reg, 0);
- reg += 2;
+ if (c->vp->has_flow_control) {
+ c->stack = brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, reg, 0);
+ reg += 2;
+ }
/* Some opcodes need an internal temporary:
*/
}
+#if 0
/* TODO: relative addressing!
*/
}
}
+#endif
+
/**
* Indirect addressing: get reg[[arg] + offset].
{
struct brw_reg reg;
- if (src->SrcRegister.File == TGSI_FILE_NULL)
+ if (src->Register.File == TGSI_FILE_NULL)
return brw_null_reg();
reg = get_src_reg(c, argIndex,
- src->SrcRegister.File,
- src->SrcRegister.Index,
- src->SrcRegister.Indirect);
+ src->Register.File,
+ src->Register.Index,
+ src->Register.Indirect);
/* Convert 3-bit swizzle to 2-bit.
*/
- reg.dw1.bits.swizzle = BRW_SWIZZLE4(src->SrcRegister.SwizzleX,
- src->SrcRegister.SwizzleY,
- src->SrcRegister.SwizzleZ,
- src->SrcRegister.SwizzleW);
+ reg.dw1.bits.swizzle = BRW_SWIZZLE4(src->Register.SwizzleX,
+ src->Register.SwizzleY,
+ src->Register.SwizzleZ,
+ src->Register.SwizzleW);
- reg.negate = src->SrcRegister.Negate ? 1 : 0;
+ reg.negate = src->Register.Negate ? 1 : 0;
/* XXX: abs, absneg
*/
struct brw_reg pos = c->regs[TGSI_FILE_OUTPUT][VERT_RESULT_HPOS];
struct brw_reg ndc;
int eot;
+ int i;
GLuint len_vertext_header = 2;
- if (c->key.copy_edgeflag) {
- assert(0);
- brw_MOV(p,
- get_reg(c, TGSI_FILE_OUTPUT, 0),
- get_reg(c, TGSI_FILE_INPUT, 0));
- }
-
/* Build ndc coords */
ndc = get_tmp(c);
/* ndc = 1.0 / pos.w */
len_vertext_header = 2;
}
- eot = (c->first_overflow_output == 0);
+ eot = (c->overflow_count == 0);
brw_urb_WRITE(p,
brw_null_reg(), /* dest */
0, /* urb destination offset */
BRW_URB_SWIZZLE_INTERLEAVE);
- if (c->first_overflow_output > 0) {
- /* Not all of the vertex outputs/results fit into the MRF.
- * Move the overflowed attributes from the GRF to the MRF and
- * issue another brw_urb_WRITE().
- */
+ /* Not all of the vertex outputs/results fit into the MRF.
+ * Move the overflowed attributes from the GRF to the MRF and
+ * issue another brw_urb_WRITE().
+ */
+ for (i = 0; i < c->overflow_count; i += BRW_MAX_MRF) {
+ unsigned nr = MIN2(c->overflow_count - i, BRW_MAX_MRF);
+ GLuint j;
+
+ eot = (i + nr >= c->overflow_count);
+
/* XXX I'm not 100% sure about which MRF regs to use here. Starting
* at mrf[4] atm...
*/
- GLuint i, mrf = 0;
- for (i = c->first_overflow_output; i < c->prog_data.nr_outputs; i++) {
- /* move from GRF to MRF */
- brw_MOV(p, brw_message_reg(4+mrf), c->regs[TGSI_FILE_OUTPUT][i]);
- mrf++;
+ for (j = 0; j < nr; j++) {
+ brw_MOV(p, brw_message_reg(4+j),
+ brw_vec8_grf(c->overflow_grf_start + i + j, 0));
}
brw_urb_WRITE(p,
c->r0, /* src */
0, /* allocate */
1, /* used */
- mrf+1, /* msg len */
+ nr+1, /* msg len */
0, /* response len */
- 1, /* eot */
- 1, /* writes complete */
- BRW_MAX_MRF-1, /* urb destination offset */
+ eot, /* eot */
+ eot, /* writes complete */
+ i-1, /* urb destination offset */
BRW_URB_SWIZZLE_INTERLEAVE);
}
}
const struct tgsi_full_instruction *inst)
{
unsigned opcode = inst->Instruction.Opcode;
- unsigned label = inst->InstructionExtLabel.Label;
+ unsigned label = inst->Label.Label;
struct brw_compile *p = &c->func;
struct brw_reg args[3], dst;
GLuint i;
/* Get argument regs.
*/
for (i = 0; i < 3; i++) {
- args[i] = get_arg(c, &inst->FullSrcRegisters[i], i);
+ args[i] = get_arg(c, &inst->Src[i], i);
}
/* Get dest regs. Note that it is possible for a reg to be both
* care needs to be taken emitting multi-operation instructions.
*/
dst = get_dst(c,
- inst->FullDstRegisters[0].DstRegister.File,
- inst->FullDstRegisters[0].DstRegister.Index,
- inst->FullDstRegisters[0].DstRegister.WriteMask);
+ inst->Dst[0].Register.File,
+ inst->Dst[0].Register.Index,
+ inst->Dst[0].Register.WriteMask);
/* XXX: saturate
*/
emit_math1(c, BRW_MATH_FUNCTION_INV, dst, args[0], BRW_MATH_PRECISION_FULL);
break;
case TGSI_OPCODE_RSQ:
- emit_math1(c, BRW_MATH_FUNCTION_RSQ, dst, args[0], BRW_MATH_PRECISION_FULL);
+ emit_math1(c, BRW_MATH_FUNCTION_RSQ, dst,
+ brw_swizzle(args[0], 0,0,0,0), BRW_MATH_PRECISION_FULL);
break;
case TGSI_OPCODE_SEQ:
emit_seq(p, dst, args[0], args[1]);
struct brw_instruction *end_inst, *last_inst;
struct tgsi_parse_context parse;
struct tgsi_full_instruction *inst;
- boolean done = FALSE;
- int i;
if (BRW_DEBUG & DEBUG_VS)
tgsi_dump(c->vp->tokens, 0);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_access_mode(p, BRW_ALIGN_16);
- /* Inputs */
- tgsi_parse_init( &parse, tokens );
- while( !tgsi_parse_end_of_tokens( &parse ) ) {
- tgsi_parse_token( &parse );
-
- switch( parse.FullToken.Token.Type ) {
- case TGSI_TOKEN_TYPE_DECLARATION:
- /* Nothing to do -- using info from tgsi_scan().
- */
- break;
-
- case TGSI_TOKEN_TYPE_IMMEDIATE: {
- static const float id[4] = {0,0,0,1};
- const float *imm = &parse.FullToken.FullImmediate.u[i].Float;
- unsigned size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
-
- for (i = 0; i < size; i++)
- c->immediate[c->nr_immediates][i] = imm[i];
-
- for ( ; i < 4; i++)
- c->immediate[c->nr_immediates][i] = id[i];
-
- c->nr_immediates++;
- break;
- }
-
- case TGSI_TOKEN_TYPE_INSTRUCTION:
- done = 1;
- break;
- }
- }
/* Static register allocation
*/
brw_vs_alloc_regs(c);
- brw_MOV(p, get_addr_reg(c->stack_index), brw_address(c->stack));
+
+ if (c->vp->has_flow_control) {
+ brw_MOV(p, get_addr_reg(c->stack_index), brw_address(c->stack));
+ }
/* Instructions
*/
post_vs_emit(c, end_inst, last_inst);
if (BRW_DEBUG & DEBUG_VS) {
- int i;
-
debug_printf("vs-native:\n");
- for (i = 0; i < p->nr_insn; i++)
- brw_disasm(stderr, &p->store[i]);
- debug_printf("\n");
+ brw_disasm(stderr, p->store, p->nr_insn);
}
}