brw_vs_arg_can_be_immediate(enum prog_opcode opcode, int arg)
{
int opcode_array[] = {
+ [OPCODE_MOV] = 1,
[OPCODE_ADD] = 2,
[OPCODE_CMP] = 3,
+ [OPCODE_DP2] = 2,
[OPCODE_DP3] = 2,
[OPCODE_DP4] = 2,
[OPCODE_DPH] = 2,
c->last_tmp = c->first_tmp;
}
+static int
+get_first_reladdr_output(struct gl_vertex_program *vp)
+{
+ int i;
+ int first_reladdr_output = VERT_RESULT_MAX;
+
+ for (i = 0; i < vp->Base.NumInstructions; i++) {
+ struct prog_instruction *inst = vp->Base.Instructions + i;
+
+ if (inst->DstReg.File == PROGRAM_OUTPUT &&
+ inst->DstReg.RelAddr &&
+ inst->DstReg.Index < first_reladdr_output)
+ first_reladdr_output = inst->DstReg.Index;
+ }
+
+ return first_reladdr_output;
+}
+
+/* Clears the record of which vp_const_buffer elements have been
+ * loaded into our constant buffer registers, for the starts of new
+ * blocks after control flow.
+ */
+static void
+clear_current_const(struct brw_vs_compile *c)
+{
+ unsigned int i;
+
+ if (c->vp->use_const_buffer) {
+ for (i = 0; i < 3; i++) {
+ c->current_const[i].index = -1;
+ }
+ }
+}
/**
* Preallocate GRF register before code emit.
static void brw_vs_alloc_regs( struct brw_vs_compile *c )
{
struct intel_context *intel = &c->func.brw->intel;
- GLuint i, reg = 0, mrf;
+ GLuint i, reg = 0, mrf, j;
int attributes_in_vue;
+ int first_reladdr_output;
+ int max_constant;
+ int constant = 0;
+ int vert_result_reoder[VERT_RESULT_MAX];
+ int bfc = 0;
/* Determine whether to use a real constant buffer or use a block
* of GRF registers for constants. The later is faster but only
/* 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);
- }
+ if (intel->gen >= 6) {
+ for (i = 0; i < c->key.nr_userclip; i++) {
+ c->userplane[i] = stride(brw_vec4_grf(reg + i / 2,
+ (i % 2) * 4), 0, 4, 1);
+ }
+ reg += ALIGN(c->key.nr_userclip, 2) / 2;
+ } else {
+ for (i = 0; i < c->key.nr_userclip; i++) {
+ c->userplane[i] = stride(brw_vec4_grf(reg + (6 + i) / 2,
+ (i % 2) * 4), 0, 4, 1);
+ }
+ reg += (ALIGN(6 + c->key.nr_userclip, 4) / 4) * 2;
+ }
- /* Deal with curbe alignment:
- */
- reg += ((6 + c->key.nr_userclip + 3) / 4) * 2;
}
- /* Vertex program parameters from curbe:
+ /* Assign some (probably all) of the vertex program constants to
+ * the push constant buffer/CURBE.
+ *
+ * There's an obvious limit to the numer of push constants equal to
+ * the number of register available, and that number is smaller
+ * than the minimum maximum number of vertex program parameters, so
+ * support for pull constants is required if we overflow.
+ * Additionally, on gen6 the number of push constants is even
+ * lower.
+ *
+ * When there's relative addressing, we don't know what range of
+ * Mesa IR registers can be accessed. And generally, when relative
+ * addressing is used we also have too many constants to load them
+ * all as push constants. So, we'll just support relative
+ * addressing out of the pull constant buffers, and try to load as
+ * many statically-accessed constants into the push constant buffer
+ * as we can.
*/
- if (c->vp->use_const_buffer) {
- int max_constant = BRW_MAX_GRF - 20 - c->vp->program.Base.NumTemporaries;
- int constant = 0;
-
- /* We've got more constants than we can load with the push
- * mechanism. This is often correlated with reladdr loads where
- * we should probably be using a pull mechanism anyway to avoid
- * excessive reading. However, the pull mechanism is slow in
- * general. So, we try to allocate as many non-reladdr-loaded
- * constants through the push buffer as we can before giving up.
- */
- memset(c->constant_map, -1, c->vp->program.Base.Parameters->NumParameters);
- for (i = 0;
- i < c->vp->program.Base.NumInstructions && constant < max_constant;
- i++) {
- struct prog_instruction *inst = &c->vp->program.Base.Instructions[i];
- int arg;
-
- for (arg = 0; arg < 3 && constant < max_constant; arg++) {
- if ((inst->SrcReg[arg].File != PROGRAM_STATE_VAR &&
- inst->SrcReg[arg].File != PROGRAM_CONSTANT &&
- inst->SrcReg[arg].File != PROGRAM_UNIFORM &&
- inst->SrcReg[arg].File != PROGRAM_ENV_PARAM &&
- inst->SrcReg[arg].File != PROGRAM_LOCAL_PARAM) ||
- inst->SrcReg[arg].RelAddr)
- continue;
-
- if (c->constant_map[inst->SrcReg[arg].Index] == -1) {
- c->constant_map[inst->SrcReg[arg].Index] = constant++;
- }
+ if (intel->gen >= 6) {
+ /* We can only load 32 regs of push constants. */
+ max_constant = 32 * 2 - c->key.nr_userclip;
+ } else {
+ max_constant = BRW_MAX_GRF - 20 - c->vp->program.Base.NumTemporaries;
+ }
+
+ /* constant_map maps from ParameterValues[] index to index in the
+ * push constant buffer, or -1 if it's only in the pull constant
+ * buffer.
+ */
+ memset(c->constant_map, -1, c->vp->program.Base.Parameters->NumParameters);
+ for (i = 0;
+ i < c->vp->program.Base.NumInstructions && constant < max_constant;
+ i++) {
+ struct prog_instruction *inst = &c->vp->program.Base.Instructions[i];
+ int arg;
+
+ for (arg = 0; arg < 3 && constant < max_constant; arg++) {
+ if (inst->SrcReg[arg].File != PROGRAM_STATE_VAR &&
+ inst->SrcReg[arg].File != PROGRAM_CONSTANT &&
+ inst->SrcReg[arg].File != PROGRAM_UNIFORM &&
+ inst->SrcReg[arg].File != PROGRAM_ENV_PARAM &&
+ inst->SrcReg[arg].File != PROGRAM_LOCAL_PARAM) {
+ continue;
+ }
+
+ if (inst->SrcReg[arg].RelAddr) {
+ c->vp->use_const_buffer = GL_TRUE;
+ continue;
}
- }
- for (i = 0; i < constant; i++) {
- c->regs[PROGRAM_STATE_VAR][i] = stride( brw_vec4_grf(reg+i/2,
- (i%2) * 4),
- 0, 4, 1);
+ if (c->constant_map[inst->SrcReg[arg].Index] == -1) {
+ c->constant_map[inst->SrcReg[arg].Index] = constant++;
+ }
}
- reg += (constant + 1) / 2;
- c->prog_data.curb_read_length = reg - 1;
- /* XXX 0 causes a bug elsewhere... */
- c->prog_data.nr_params = MAX2(constant * 4, 4);
}
- else {
- /* use a section of the GRF for constants */
- GLuint nr_params = c->vp->program.Base.Parameters->NumParameters;
- for (i = 0; i < nr_params; i++) {
- c->regs[PROGRAM_STATE_VAR][i] = stride( brw_vec4_grf(reg+i/2, (i%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;
+ /* If we ran out of push constant space, then we'll also upload all
+ * constants through the pull constant buffer so that they can be
+ * accessed no matter what. For relative addressing (the common
+ * case) we need them all in place anyway.
+ */
+ if (constant == max_constant)
+ c->vp->use_const_buffer = GL_TRUE;
+
+ for (i = 0; i < constant; i++) {
+ c->regs[PROGRAM_STATE_VAR][i] = stride(brw_vec4_grf(reg + i / 2,
+ (i % 2) * 4),
+ 0, 4, 1);
}
+ reg += (constant + 1) / 2;
+ c->prog_data.curb_read_length = reg - 1;
+ c->prog_data.nr_params = constant * 4;
+ /* XXX 0 causes a bug elsewhere... */
+ if (intel->gen < 6 && c->prog_data.nr_params == 0)
+ c->prog_data.nr_params = 4;
/* Allocate input regs:
*/
c->first_output = reg;
c->first_overflow_output = 0;
- if (intel->gen >= 6)
- mrf = 4;
- else if (intel->gen == 5)
+ if (intel->gen >= 6) {
+ mrf = 3;
+ if (c->key.nr_userclip)
+ mrf += 2;
+ } else if (intel->gen == 5)
mrf = 8;
else
mrf = 4;
- for (i = 0; i < VERT_RESULT_MAX; i++) {
+ first_reladdr_output = get_first_reladdr_output(&c->vp->program);
+
+ for (i = 0; i < VERT_RESULT_MAX; i++)
+ vert_result_reoder[i] = i;
+
+ /* adjust attribute order in VUE for BFC0/BFC1 on Gen6+ */
+ if (intel->gen >= 6 && c->key.two_side_color) {
+ if ((c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_COL1)) &&
+ (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_BFC1))) {
+ assert(c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_COL0));
+ assert(c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_BFC0));
+ bfc = 2;
+ } else if ((c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_COL0)) &&
+ (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_BFC0)))
+ bfc = 1;
+
+ if (bfc) {
+ for (i = 0; i < bfc; i++) {
+ vert_result_reoder[VERT_RESULT_COL0 + i * 2 + 0] = VERT_RESULT_COL0 + i;
+ vert_result_reoder[VERT_RESULT_COL0 + i * 2 + 1] = VERT_RESULT_BFC0 + i;
+ }
+
+ for (i = VERT_RESULT_COL0 + bfc * 2; i < VERT_RESULT_BFC0 + bfc; i++) {
+ vert_result_reoder[i] = i - bfc;
+ }
+ }
+ }
+
+ for (j = 0; j < VERT_RESULT_MAX; j++) {
+ i = vert_result_reoder[j];
+
if (c->prog_data.outputs_written & BITFIELD64_BIT(i)) {
c->nr_outputs++;
assert(i < Elements(c->regs[PROGRAM_OUTPUT]));
else if (i == VERT_RESULT_PSIZ) {
c->regs[PROGRAM_OUTPUT][i] = brw_vec8_grf(reg, 0);
reg++;
- mrf++; /* just a placeholder? XXX fix later stages & remove this */
}
else {
/* Two restrictions on our compute-to-MRF here. The
* For attributes beyond the compute-to-MRF, we compute to
* GRFs and they will be written in the second URB_WRITE.
*/
- if (mrf < 15) {
+ if (first_reladdr_output > i && mrf < 15) {
c->regs[PROGRAM_OUTPUT][i] = brw_message_reg(mrf);
mrf++;
}
else {
- if (!c->first_overflow_output)
+ if (mrf >= 15 && !c->first_overflow_output)
c->first_overflow_output = i;
c->regs[PROGRAM_OUTPUT][i] = brw_vec8_grf(reg, 0);
reg++;
+ mrf++;
}
}
}
if (c->vp->use_const_buffer) {
for (i = 0; i < 3; i++) {
- c->current_const[i].index = -1;
c->current_const[i].reg = brw_vec8_grf(reg, 0);
reg++;
}
+ clear_current_const(c);
}
for (i = 0; i < 128; i++) {
/* See emit_vertex_write() for where the VUE's overhead on top of the
* attributes comes from.
*/
- if (intel->gen >= 6)
- c->prog_data.urb_entry_size = (attributes_in_vue + 2 + 7) / 8;
- else if (intel->gen == 5)
+ if (intel->gen >= 6) {
+ int header_regs = 2;
+ if (c->key.nr_userclip)
+ header_regs += 2;
+
+ c->prog_data.urb_entry_size = (attributes_in_vue + header_regs + 7) / 8;
+ } else if (intel->gen == 5)
c->prog_data.urb_entry_size = (attributes_in_vue + 6 + 3) / 4;
else
c->prog_data.urb_entry_size = (attributes_in_vue + 2 + 3) / 4;
c->prog_data.total_grf = reg;
- if (INTEL_DEBUG & DEBUG_VS) {
+ if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
printf("%s NumAddrRegs %d\n", __FUNCTION__, c->vp->program.Base.NumAddressRegs);
printf("%s NumTemps %d\n", __FUNCTION__, c->vp->program.Base.NumTemporaries);
printf("%s reg = %d\n", __FUNCTION__, reg);
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
}
+static void emit_sign(struct brw_vs_compile *c,
+ struct brw_reg dst,
+ struct brw_reg arg0)
+{
+ struct brw_compile *p = &c->func;
+
+ brw_MOV(p, dst, brw_imm_f(0));
+
+ brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_L, arg0, brw_imm_f(0));
+ brw_MOV(p, dst, brw_imm_f(-1.0));
+ brw_set_predicate_control(p, BRW_PREDICATE_NONE);
+
+ brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_G, arg0, brw_imm_f(0));
+ brw_MOV(p, dst, brw_imm_f(1.0));
+ brw_set_predicate_control(p, BRW_PREDICATE_NONE);
+}
+
static void emit_max( struct brw_compile *p,
struct brw_reg dst,
struct brw_reg arg0,
struct brw_reg arg1 )
{
- brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_GE, arg0, arg1);
- brw_SEL(p, dst, arg0, arg1);
- brw_set_predicate_control(p, BRW_PREDICATE_NONE);
+ struct intel_context *intel = &p->brw->intel;
+
+ if (intel->gen >= 6) {
+ brw_set_conditionalmod(p, BRW_CONDITIONAL_GE);
+ brw_SEL(p, dst, arg0, arg1);
+ brw_set_conditionalmod(p, BRW_CONDITIONAL_NONE);
+ brw_set_predicate_control(p, BRW_PREDICATE_NONE);
+ } else {
+ brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_GE, arg0, arg1);
+ brw_SEL(p, dst, arg0, arg1);
+ brw_set_predicate_control(p, BRW_PREDICATE_NONE);
+ }
}
static void emit_min( struct brw_compile *p,
struct brw_reg arg0,
struct brw_reg arg1 )
{
- brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_L, arg0, arg1);
- brw_SEL(p, dst, arg0, arg1);
- brw_set_predicate_control(p, BRW_PREDICATE_NONE);
+ struct intel_context *intel = &p->brw->intel;
+
+ if (intel->gen >= 6) {
+ brw_set_conditionalmod(p, BRW_CONDITIONAL_L);
+ brw_SEL(p, dst, arg0, arg1);
+ brw_set_conditionalmod(p, BRW_CONDITIONAL_NONE);
+ brw_set_predicate_control(p, BRW_PREDICATE_NONE);
+ } else {
+ brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_L, arg0, arg1);
+ brw_SEL(p, dst, arg0, arg1);
+ brw_set_predicate_control(p, BRW_PREDICATE_NONE);
+ }
}
+static void emit_arl(struct brw_compile *p,
+ struct brw_reg dst,
+ struct brw_reg src)
+{
+ struct intel_context *intel = &p->brw->intel;
-static void emit_math1( struct brw_vs_compile *c,
- GLuint function,
- struct brw_reg dst,
- struct brw_reg arg0,
- GLuint precision)
+ if (intel->gen >= 6) {
+ struct brw_reg dst_f = retype(dst, BRW_REGISTER_TYPE_F);
+
+ brw_RNDD(p, dst_f, src);
+ brw_MOV(p, dst, dst_f);
+ } else {
+ brw_RNDD(p, dst, src);
+ }
+}
+
+static void emit_math1_gen4(struct brw_vs_compile *c,
+ GLuint function,
+ struct brw_reg dst,
+ struct brw_reg arg0,
+ GLuint precision)
{
/* There are various odd behaviours with SEND on the simulator. In
* addition there are documented issues with the fact that the GEN4
* whether that turns out to be a simulator bug or not:
*/
struct brw_compile *p = &c->func;
- struct intel_context *intel = &p->brw->intel;
struct brw_reg tmp = dst;
- GLboolean need_tmp = (intel->gen < 6 &&
- (dst.dw1.bits.writemask != 0xf ||
- dst.file != BRW_GENERAL_REGISTER_FILE));
+ GLboolean need_tmp = GL_FALSE;
- if (need_tmp)
+ if (dst.file != BRW_GENERAL_REGISTER_FILE ||
+ dst.dw1.bits.writemask != 0xf)
+ need_tmp = GL_TRUE;
+
+ if (need_tmp)
tmp = get_tmp(c);
brw_math(p,
}
}
+static void
+emit_math1_gen6(struct brw_vs_compile *c,
+ GLuint function,
+ struct brw_reg dst,
+ struct brw_reg arg0,
+ GLuint precision)
+{
+ struct brw_compile *p = &c->func;
+ struct brw_reg tmp_src, tmp_dst;
-static void emit_math2( struct brw_vs_compile *c,
+ /* Something is strange on gen6 math in 16-wide mode, though the
+ * docs say it's supposed to work. Punt to using align1 mode,
+ * which doesn't do writemasking and swizzles.
+ */
+ tmp_src = get_tmp(c);
+ tmp_dst = get_tmp(c);
+
+ brw_MOV(p, tmp_src, arg0);
+
+ brw_set_access_mode(p, BRW_ALIGN_1);
+ brw_math(p,
+ tmp_dst,
+ function,
+ BRW_MATH_SATURATE_NONE,
+ 2,
+ tmp_src,
+ BRW_MATH_DATA_SCALAR,
+ precision);
+ brw_set_access_mode(p, BRW_ALIGN_16);
+
+ brw_MOV(p, dst, tmp_dst);
+
+ release_tmp(c, tmp_src);
+ release_tmp(c, tmp_dst);
+}
+
+static void
+emit_math1(struct brw_vs_compile *c,
+ GLuint function,
+ struct brw_reg dst,
+ struct brw_reg arg0,
+ GLuint precision)
+{
+ struct brw_compile *p = &c->func;
+ struct intel_context *intel = &p->brw->intel;
+
+ if (intel->gen >= 6)
+ emit_math1_gen6(c, function, dst, arg0, precision);
+ else
+ emit_math1_gen4(c, function, dst, arg0, precision);
+}
+
+static void emit_math2_gen4( struct brw_vs_compile *c,
GLuint function,
struct brw_reg dst,
struct brw_reg arg0,
GLuint precision)
{
struct brw_compile *p = &c->func;
- struct intel_context *intel = &p->brw->intel;
struct brw_reg tmp = dst;
- GLboolean need_tmp = (intel->gen < 6 &&
- (dst.dw1.bits.writemask != 0xf ||
- dst.file != BRW_GENERAL_REGISTER_FILE));
+ GLboolean need_tmp = GL_FALSE;
+
+ if (dst.file != BRW_GENERAL_REGISTER_FILE ||
+ dst.dw1.bits.writemask != 0xf)
+ need_tmp = GL_TRUE;
if (need_tmp)
tmp = get_tmp(c);
}
}
+static void emit_math2_gen6( struct brw_vs_compile *c,
+ GLuint function,
+ struct brw_reg dst,
+ struct brw_reg arg0,
+ struct brw_reg arg1,
+ GLuint precision)
+{
+ struct brw_compile *p = &c->func;
+ struct brw_reg tmp_src0, tmp_src1, tmp_dst;
+
+ tmp_src0 = get_tmp(c);
+ tmp_src1 = get_tmp(c);
+ tmp_dst = get_tmp(c);
+
+ brw_MOV(p, tmp_src0, arg0);
+ brw_MOV(p, tmp_src1, arg1);
+
+ brw_set_access_mode(p, BRW_ALIGN_1);
+ brw_math2(p,
+ tmp_dst,
+ function,
+ tmp_src0,
+ tmp_src1);
+ brw_set_access_mode(p, BRW_ALIGN_16);
+
+ brw_MOV(p, dst, tmp_dst);
+
+ release_tmp(c, tmp_src0);
+ release_tmp(c, tmp_src1);
+ release_tmp(c, tmp_dst);
+}
+
+static void emit_math2( struct brw_vs_compile *c,
+ GLuint function,
+ struct brw_reg dst,
+ struct brw_reg arg0,
+ struct brw_reg arg1,
+ GLuint precision)
+{
+ struct brw_compile *p = &c->func;
+ struct intel_context *intel = &p->brw->intel;
+
+ if (intel->gen >= 6)
+ emit_math2_gen6(c, function, dst, arg0, arg1, precision);
+ else
+ emit_math2_gen4(c, function, dst, arg0, arg1, precision);
+}
static void emit_exp_noalias( struct brw_vs_compile *c,
struct brw_reg dst,
assert(argIndex < 3);
if (c->current_const[argIndex].index != src->Index) {
- struct brw_reg addrReg = c->regs[PROGRAM_ADDRESS][0];
-
/* Keep track of the last constant loaded in this slot, for reuse. */
c->current_const[argIndex].index = src->Index;
/* need to fetch the constant now */
brw_dp_READ_4_vs(p,
const_reg, /* writeback dest */
- 0, /* oword */
- 0, /* relative indexing? */
- addrReg, /* address register */
16 * src->Index, /* byte offset */
SURF_INDEX_VERT_CONST_BUFFER /* binding table index */
);
}
/* replicate lower four floats into upper half (to get XYZWXYZW) */
- const_reg = stride(const_reg, 0, 4, 0);
+ const_reg = stride(const_reg, 0, 4, 1);
const_reg.subnr = 0;
return const_reg;
{
const struct prog_src_register *src = &inst->SrcReg[argIndex];
struct brw_compile *p = &c->func;
+ struct brw_context *brw = p->brw;
+ struct intel_context *intel = &brw->intel;
struct brw_reg const_reg = c->current_const[argIndex].reg;
- struct brw_reg const2_reg;
- struct brw_reg addrReg = c->regs[PROGRAM_ADDRESS][0];
+ struct brw_reg addr_reg = c->regs[PROGRAM_ADDRESS][0];
+ uint32_t offset;
assert(argIndex < 3);
src->Index, argIndex, c->current_const[argIndex].reg.nr);
#endif
+ if (intel->gen >= 6) {
+ offset = src->Index;
+ } else {
+ struct brw_reg byte_addr_reg = retype(get_tmp(c), BRW_REGISTER_TYPE_D);
+ brw_MUL(p, byte_addr_reg, addr_reg, brw_imm_d(16));
+ addr_reg = byte_addr_reg;
+ offset = 16 * src->Index;
+ }
+
/* fetch the first vec4 */
- brw_dp_READ_4_vs(p,
- const_reg, /* writeback dest */
- 0, /* oword */
- 1, /* relative indexing? */
- addrReg, /* address register */
- 16 * src->Index, /* byte offset */
- SURF_INDEX_VERT_CONST_BUFFER /* binding table index */
- );
- /* second vec4 */
- const2_reg = get_tmp(c);
-
- /* use upper half of address reg for second read */
- addrReg = stride(addrReg, 0, 4, 0);
- addrReg.subnr = 16;
-
- brw_dp_READ_4_vs(p,
- const2_reg, /* writeback dest */
- 1, /* oword */
- 1, /* relative indexing? */
- addrReg, /* address register */
- 16 * src->Index, /* byte offset */
- SURF_INDEX_VERT_CONST_BUFFER
- );
-
- /* merge the two Owords into the constant register */
- /* const_reg[7..4] = const2_reg[7..4] */
- brw_MOV(p,
- suboffset(stride(const_reg, 0, 4, 1), 4),
- suboffset(stride(const2_reg, 0, 4, 1), 4));
- release_tmp(c, const2_reg);
+ brw_dp_READ_4_vs_relative(p,
+ const_reg,
+ addr_reg,
+ offset,
+ SURF_INDEX_VERT_CONST_BUFFER);
return const_reg;
}
*/
static struct brw_reg deref( struct brw_vs_compile *c,
struct brw_reg arg,
- GLint offset)
+ GLint offset,
+ GLuint reg_size )
{
struct brw_compile *p = &c->func;
- struct brw_reg tmp = vec4(get_tmp(c));
+ struct brw_reg tmp = get_tmp(c);
struct brw_reg addr_reg = c->regs[PROGRAM_ADDRESS][0];
- struct brw_reg vp_address = retype(vec1(addr_reg), BRW_REGISTER_TYPE_UW);
- GLuint byte_offset = arg.nr * 32 + arg.subnr + offset * 16;
+ struct brw_reg vp_address = retype(vec1(addr_reg), BRW_REGISTER_TYPE_D);
+ GLuint byte_offset = arg.nr * 32 + arg.subnr + offset * reg_size;
struct brw_reg indirect = brw_vec4_indirect(0,0);
+ struct brw_reg acc = retype(vec1(get_tmp(c)), BRW_REGISTER_TYPE_UW);
+
+ /* Set the vertical stride on the register access so that the first
+ * 4 components come from a0.0 and the second 4 from a0.1.
+ */
+ indirect.vstride = BRW_VERTICAL_STRIDE_ONE_DIMENSIONAL;
{
brw_push_insn_state(p);
brw_set_access_mode(p, BRW_ALIGN_1);
- /* This is pretty clunky - load the address register twice and
- * fetch each 4-dword value in turn. There must be a way to do
- * this in a single pass, but I couldn't get it to work.
- */
- brw_ADD(p, brw_address_reg(0), vp_address, brw_imm_d(byte_offset));
- brw_MOV(p, tmp, indirect);
+ brw_MUL(p, acc, vp_address, brw_imm_uw(reg_size));
+ brw_ADD(p, brw_address_reg(0), acc, brw_imm_uw(byte_offset));
- brw_ADD(p, brw_address_reg(0), suboffset(vp_address, 8), brw_imm_d(byte_offset));
- brw_MOV(p, suboffset(tmp, 4), indirect);
+ brw_MUL(p, acc, suboffset(vp_address, 4), brw_imm_uw(reg_size));
+ brw_ADD(p, brw_address_reg(1), acc, brw_imm_uw(byte_offset));
+
+ brw_MOV(p, tmp, indirect);
brw_pop_insn_state(p);
}
-
+
/* NOTE: tmp not released */
- return vec8(tmp);
+ return tmp;
}
+static void
+move_to_reladdr_dst(struct brw_vs_compile *c,
+ const struct prog_instruction *inst,
+ struct brw_reg val)
+{
+ struct brw_compile *p = &c->func;
+ int reg_size = 32;
+ struct brw_reg addr_reg = c->regs[PROGRAM_ADDRESS][0];
+ struct brw_reg vp_address = retype(vec1(addr_reg), BRW_REGISTER_TYPE_D);
+ struct brw_reg base = c->regs[inst->DstReg.File][inst->DstReg.Index];
+ GLuint byte_offset = base.nr * 32 + base.subnr;
+ struct brw_reg indirect = brw_vec4_indirect(0,0);
+ struct brw_reg acc = retype(vec1(get_tmp(c)), BRW_REGISTER_TYPE_UW);
+
+ /* Because destination register indirect addressing can only use
+ * one index, we'll write each vertex's vec4 value separately.
+ */
+ val.width = BRW_WIDTH_4;
+ val.vstride = BRW_VERTICAL_STRIDE_4;
+
+ brw_push_insn_state(p);
+ brw_set_access_mode(p, BRW_ALIGN_1);
+
+ brw_MUL(p, acc, vp_address, brw_imm_uw(reg_size));
+ brw_ADD(p, brw_address_reg(0), acc, brw_imm_uw(byte_offset));
+ brw_MOV(p, indirect, val);
+
+ brw_MUL(p, acc, suboffset(vp_address, 4), brw_imm_uw(reg_size));
+ brw_ADD(p, brw_address_reg(0), acc,
+ brw_imm_uw(byte_offset + reg_size / 2));
+ brw_MOV(p, indirect, suboffset(val, 4));
+
+ brw_pop_insn_state(p);
+}
/**
* Get brw reg corresponding to the instruction's [argIndex] src reg.
case PROGRAM_INPUT:
case PROGRAM_OUTPUT:
if (relAddr) {
- return deref(c, c->regs[file][0], index);
+ return deref(c, c->regs[file][0], index, 32);
}
else {
assert(c->regs[file][index].nr != 0);
case PROGRAM_UNIFORM:
case PROGRAM_ENV_PARAM:
case PROGRAM_LOCAL_PARAM:
- if (c->vp->use_const_buffer) {
- if (!relAddr && c->constant_map[index] != -1) {
- assert(c->regs[PROGRAM_STATE_VAR][c->constant_map[index]].nr != 0);
- return c->regs[PROGRAM_STATE_VAR][c->constant_map[index]];
- } else if (relAddr)
+ if (!relAddr && c->constant_map[index] != -1) {
+ /* Take from the push constant buffer if possible. */
+ assert(c->regs[PROGRAM_STATE_VAR][c->constant_map[index]].nr != 0);
+ return c->regs[PROGRAM_STATE_VAR][c->constant_map[index]];
+ } else {
+ /* Must be in the pull constant buffer then .*/
+ assert(c->vp->use_const_buffer);
+ if (relAddr)
return get_reladdr_constant(c, inst, argIndex);
else
return get_constant(c, inst, argIndex);
}
- else if (relAddr) {
- return deref(c, c->regs[PROGRAM_STATE_VAR][0], index);
- }
- else {
- assert(c->regs[PROGRAM_STATE_VAR][index].nr != 0);
- return c->regs[PROGRAM_STATE_VAR][index];
- }
case PROGRAM_ADDRESS:
assert(index == 0);
return c->regs[file][index];
}
}
-
-static void emit_arl( struct brw_vs_compile *c,
- struct brw_reg dst,
- struct brw_reg arg0 )
-{
- struct brw_compile *p = &c->func;
- struct brw_reg tmp = dst;
- GLboolean need_tmp = (dst.file != BRW_GENERAL_REGISTER_FILE);
-
- if (need_tmp)
- tmp = get_tmp(c);
-
- brw_RNDD(p, tmp, arg0); /* tmp = round(arg0) */
- brw_MUL(p, dst, tmp, brw_imm_d(16)); /* dst = tmp * 16 */
-
- if (need_tmp)
- release_tmp(c, tmp);
-}
-
-
/**
* Return the brw reg for the given instruction's src argument.
* Will return mangled results for SWZ op. The emit_swz() function
/* Convert 3-bit swizzle to 2-bit.
*/
- reg.dw1.bits.swizzle = BRW_SWIZZLE4(GET_SWZ(src->Swizzle, 0),
- GET_SWZ(src->Swizzle, 1),
- GET_SWZ(src->Swizzle, 2),
- GET_SWZ(src->Swizzle, 3));
-
- /* Note this is ok for non-swizzle instructions:
- */
- reg.negate = src->Negate ? 1 : 0;
+ if (reg.file != BRW_IMMEDIATE_VALUE) {
+ reg.dw1.bits.swizzle = BRW_SWIZZLE4(GET_SWZ(src->Swizzle, 0),
+ GET_SWZ(src->Swizzle, 1),
+ GET_SWZ(src->Swizzle, 2),
+ GET_SWZ(src->Swizzle, 3));
+
+ /* Note this is ok for non-swizzle ARB_vp instructions */
+ reg.negate = src->Negate ? 1 : 0;
+ }
return reg;
}
switch (dst.File) {
case PROGRAM_TEMPORARY:
case PROGRAM_OUTPUT:
- assert(c->regs[dst.File][dst.Index].nr != 0);
- reg = c->regs[dst.File][dst.Index];
+ /* register-indirect addressing is only 1x1, not VxH, for
+ * destination regs. So, for RelAddr we'll return a temporary
+ * for the dest and do a move of the result to the RelAddr
+ * register after the instruction emit.
+ */
+ if (dst.RelAddr) {
+ reg = get_tmp(c);
+ } else {
+ assert(c->regs[dst.File][dst.Index].nr != 0);
+ reg = c->regs[dst.File][dst.Index];
+ }
break;
case PROGRAM_ADDRESS:
assert(dst.Index == 0);
reg = brw_null_reg();
}
+ assert(reg.type != BRW_IMMEDIATE_VALUE);
reg.dw1.bits.writemask = dst.WriteMask;
return reg;
struct brw_compile *p = &c->func;
struct brw_context *brw = p->brw;
struct intel_context *intel = &brw->intel;
- struct brw_reg m0 = brw_message_reg(0);
struct brw_reg pos = c->regs[PROGRAM_OUTPUT][VERT_RESULT_HPOS];
struct brw_reg ndc;
int eot;
GLuint len_vertex_header = 2;
+ int next_mrf, i;
if (c->key.copy_edgeflag) {
brw_MOV(p,
/* Update the header for point size, user clipping flags, and -ve rhw
* workaround.
*/
- if ((c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) ||
- c->key.nr_userclip || brw->has_negative_rhw_bug)
- {
+ if (intel->gen >= 6) {
+ struct brw_reg m1 = brw_message_reg(1);
+
+ /* On gen6, m1 has each value in a separate dword, so we never
+ * need to mess with a temporary for computing the m1 value.
+ */
+ brw_MOV(p, retype(m1, BRW_REGISTER_TYPE_UD), brw_imm_ud(0));
+ if (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) {
+ brw_MOV(p, brw_writemask(m1, WRITEMASK_W),
+ brw_swizzle1(c->regs[PROGRAM_OUTPUT][VERT_RESULT_PSIZ], 0));
+ }
+
+ /* Set the user clip distances in dword 8-15. (m3-4)*/
+ if (c->key.nr_userclip) {
+ for (i = 0; i < c->key.nr_userclip; i++) {
+ struct brw_reg m;
+ if (i < 4)
+ m = brw_message_reg(3);
+ else
+ m = brw_message_reg(4);
+
+ brw_DP4(p, brw_writemask(m, (1 << (i & 7))),pos, c->userplane[i]);
+ }
+ }
+ } else if ((c->prog_data.outputs_written &
+ BITFIELD64_BIT(VERT_RESULT_PSIZ)) ||
+ c->key.nr_userclip || brw->has_negative_rhw_bug) {
struct brw_reg header1 = retype(get_tmp(c), BRW_REGISTER_TYPE_UD);
GLuint i;
if (c->prog_data.outputs_written & BITFIELD64_BIT(VERT_RESULT_PSIZ)) {
struct brw_reg psiz = c->regs[PROGRAM_OUTPUT][VERT_RESULT_PSIZ];
- brw_MUL(p, brw_writemask(header1, WRITEMASK_W), brw_swizzle1(psiz, 0), brw_imm_f(1<<11));
- brw_AND(p, brw_writemask(header1, WRITEMASK_W), header1, brw_imm_ud(0x7ff<<8));
+ brw_MUL(p, brw_writemask(header1, WRITEMASK_W),
+ brw_swizzle1(psiz, 0), brw_imm_f(1<<11));
+ brw_AND(p, brw_writemask(header1, WRITEMASK_W),
+ header1, brw_imm_ud(0x7ff<<8));
}
for (i = 0; i < c->key.nr_userclip; i++) {
* of zeros followed by two sets of NDC coordinates:
*/
brw_set_access_mode(p, BRW_ALIGN_1);
+ brw_set_acc_write_control(p, 0);
/* The VUE layout is documented in Volume 2a. */
if (intel->gen >= 6) {
* dword 0-3 (m1) of the header is indices, point width, clip flags.
* dword 4-7 (m2) is the 4D space position
* dword 8-15 (m3,m4) of the vertex header is the user clip distance if
- * enabled. We don't use it, so skip it.
- * m3 is the first vertex element data we fill, which is the vertex
- * position.
+ * enabled.
+ * m3 or 5 is the first vertex element data we fill, which is
+ * the vertex position.
*/
brw_MOV(p, brw_message_reg(2), pos);
- brw_MOV(p, brw_message_reg(3), pos);
- len_vertex_header = 2;
+ len_vertex_header = 1;
+ if (c->key.nr_userclip > 0)
+ len_vertex_header += 2;
} else if (intel->gen == 5) {
/* There are 20 DWs (D0-D19) in VUE header on Ironlake:
* dword 0-3 (m1) of the header is indices, point width, clip flags.
len_vertex_header = 2;
}
+ /* Move variable-addressed, non-overflow outputs to their MRFs. */
+ next_mrf = 2 + len_vertex_header;
+ for (i = 0; i < VERT_RESULT_MAX; i++) {
+ if (c->first_overflow_output > 0 && i >= c->first_overflow_output)
+ break;
+ if (!(c->prog_data.outputs_written & BITFIELD64_BIT(i)))
+ continue;
+ if (i == VERT_RESULT_PSIZ)
+ continue;
+
+ if (i >= VERT_RESULT_TEX0 &&
+ c->regs[PROGRAM_OUTPUT][i].file == BRW_GENERAL_REGISTER_FILE) {
+ brw_MOV(p, brw_message_reg(next_mrf), c->regs[PROGRAM_OUTPUT][i]);
+ next_mrf++;
+ } else if (c->regs[PROGRAM_OUTPUT][i].file == BRW_MESSAGE_REGISTER_FILE) {
+ next_mrf = c->regs[PROGRAM_OUTPUT][i].nr + 1;
+ }
+ }
+
eot = (c->first_overflow_output == 0);
brw_urb_WRITE(p,
const GLuint nr_insns = c->vp->program.Base.NumInstructions;
GLuint insn, if_depth = 0, loop_depth = 0;
struct brw_instruction *if_inst[MAX_IF_DEPTH], *loop_inst[MAX_LOOP_DEPTH] = { 0 };
+ int if_depth_in_loop[MAX_LOOP_DEPTH];
const struct brw_indirect stack_index = brw_indirect(0, 0);
GLuint index;
GLuint file;
- if (INTEL_DEBUG & DEBUG_VS) {
+ if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
printf("vs-mesa:\n");
- _mesa_print_program(&c->vp->program.Base);
+ _mesa_fprint_program_opt(stdout, &c->vp->program.Base, PROG_PRINT_DEBUG,
+ GL_TRUE);
printf("\n");
}
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_access_mode(p, BRW_ALIGN_16);
+ if_depth_in_loop[loop_depth] = 0;
+
+ brw_set_acc_write_control(p, 1);
for (insn = 0; insn < nr_insns; insn++) {
GLuint i;
const struct prog_instruction *inst = &c->vp->program.Base.Instructions[insn];
struct brw_reg args[3], dst;
GLuint i;
-
+
#if 0
printf("%d: ", insn);
_mesa_print_instruction(inst);
switch (inst->Opcode) {
case OPCODE_ABS:
+ args[0].negate = false;
brw_MOV(p, dst, brw_abs(args[0]));
break;
case OPCODE_ADD:
case OPCODE_COS:
emit_math1(c, BRW_MATH_FUNCTION_COS, dst, args[0], BRW_MATH_PRECISION_FULL);
break;
+ case OPCODE_DP2:
+ brw_DP2(p, dst, args[0], args[1]);
+ break;
case OPCODE_DP3:
brw_DP3(p, dst, args[0], args[1]);
break;
emit_math1(c, BRW_MATH_FUNCTION_EXP, dst, args[0], BRW_MATH_PRECISION_FULL);
break;
case OPCODE_ARL:
- emit_arl(c, dst, args[0]);
+ emit_arl(p, dst, args[0]);
break;
case OPCODE_FLR:
brw_RNDD(p, dst, args[0]);
emit_math1(c, BRW_MATH_FUNCTION_INV, dst, args[0], BRW_MATH_PRECISION_FULL);
break;
case 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_abs(args[0]), BRW_MATH_PRECISION_FULL);
break;
case OPCODE_SEQ:
case OPCODE_SLE:
unalias2(c, dst, args[0], args[1], emit_sle);
break;
+ case OPCODE_SSG:
+ unalias1(c, dst, args[0], emit_sign);
+ break;
case OPCODE_SUB:
brw_ADD(p, dst, args[0], negate(args[1]));
break;
if_inst[if_depth] = brw_IF(p, BRW_EXECUTE_8);
/* Note that brw_IF smashes the predicate_control field. */
if_inst[if_depth]->header.predicate_control = get_predicate(inst);
+ if_depth_in_loop[loop_depth]++;
if_depth++;
break;
case OPCODE_ELSE:
+ clear_current_const(c);
assert(if_depth > 0);
if_inst[if_depth-1] = brw_ELSE(p, if_inst[if_depth-1]);
break;
case OPCODE_ENDIF:
+ clear_current_const(c);
assert(if_depth > 0);
brw_ENDIF(p, if_inst[--if_depth]);
+ if_depth_in_loop[loop_depth]--;
break;
case OPCODE_BGNLOOP:
+ clear_current_const(c);
loop_inst[loop_depth++] = brw_DO(p, BRW_EXECUTE_8);
+ if_depth_in_loop[loop_depth] = 0;
break;
case OPCODE_BRK:
brw_set_predicate_control(p, get_predicate(inst));
- brw_BREAK(p);
+ brw_BREAK(p, if_depth_in_loop[loop_depth]);
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
break;
case OPCODE_CONT:
brw_set_predicate_control(p, get_predicate(inst));
- brw_CONT(p);
+ if (intel->gen >= 6) {
+ brw_CONT_gen6(p, loop_inst[loop_depth - 1]);
+ } else {
+ brw_CONT(p, if_depth_in_loop[loop_depth]);
+ }
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
break;
- case OPCODE_ENDLOOP:
- {
- struct brw_instruction *inst0, *inst1;
- GLuint br = 1;
- loop_depth--;
+ case OPCODE_ENDLOOP: {
+ clear_current_const(c);
+ struct brw_instruction *inst0, *inst1;
+ GLuint br = 1;
+
+ loop_depth--;
- if (intel->gen == 5)
- br = 2;
+ if (intel->gen == 5)
+ br = 2;
- inst0 = inst1 = brw_WHILE(p, loop_inst[loop_depth]);
- /* patch all the BREAK/CONT instructions from last BEGINLOOP */
- while (inst0 > loop_inst[loop_depth]) {
- inst0--;
- if (inst0->header.opcode == BRW_OPCODE_BREAK &&
+ inst0 = inst1 = brw_WHILE(p, loop_inst[loop_depth]);
+
+ if (intel->gen < 6) {
+ /* patch all the BREAK/CONT instructions from last BEGINLOOP */
+ while (inst0 > loop_inst[loop_depth]) {
+ inst0--;
+ if (inst0->header.opcode == BRW_OPCODE_BREAK &&
inst0->bits3.if_else.jump_count == 0) {
- inst0->bits3.if_else.jump_count = br * (inst1 - inst0 + 1);
- inst0->bits3.if_else.pop_count = 0;
- }
- else if (inst0->header.opcode == BRW_OPCODE_CONTINUE &&
- inst0->bits3.if_else.jump_count == 0) {
- inst0->bits3.if_else.jump_count = br * (inst1 - inst0);
- inst0->bits3.if_else.pop_count = 0;
- }
- }
- }
+ inst0->bits3.if_else.jump_count = br * (inst1 - inst0 + 1);
+ } else if (inst0->header.opcode == BRW_OPCODE_CONTINUE &&
+ inst0->bits3.if_else.jump_count == 0) {
+ inst0->bits3.if_else.jump_count = br * (inst1 - inst0);
+ }
+ }
+ }
+ }
break;
+
case OPCODE_BRA:
brw_set_predicate_control(p, get_predicate(inst));
brw_ADD(p, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16));
}
}
+ if (inst->DstReg.RelAddr) {
+ assert(inst->DstReg.File == PROGRAM_TEMPORARY||
+ inst->DstReg.File == PROGRAM_OUTPUT);
+ move_to_reladdr_dst(c, inst, dst);
+ }
+
release_tmps(c);
}
brw_resolve_cals(p);
+ brw_set_uip_jip(p);
brw_optimize(p);
- if (INTEL_DEBUG & DEBUG_VS) {
+ if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
int i;
printf("vs-native:\n");
for (i = 0; i < p->nr_insn; i++)
- brw_disasm(stderr, &p->store[i], intel->gen);
+ brw_disasm(stdout, &p->store[i], intel->gen);
printf("\n");
}
}