#include "main/macros.h"
-#include "shader/program.h"
-#include "shader/prog_parameter.h"
-#include "shader/prog_print.h"
+#include "program/program.h"
+#include "program/prog_parameter.h"
+#include "program/prog_print.h"
#include "brw_context.h"
#include "brw_vs.h"
[OPCODE_DP3] = 2,
[OPCODE_DP4] = 2,
[OPCODE_DPH] = 2,
- [OPCODE_MAX] = 1,
+ [OPCODE_MAX] = 2,
[OPCODE_MIN] = 2,
[OPCODE_MUL] = 2,
[OPCODE_SEQ] = 2,
c->first_overflow_output = 0;
if (intel->gen >= 6)
- mrf = 6;
+ mrf = 4;
else if (intel->gen == 5)
mrf = 8;
else
mrf++; /* just a placeholder? XXX fix later stages & remove this */
}
else {
- if (mrf < 16) {
+ /* Two restrictions on our compute-to-MRF here. The
+ * message length for all SEND messages is restricted to
+ * [1,15], so we can't use mrf 15, as that means a length
+ * of 16.
+ *
+ * Additionally, URB writes are aligned to URB rows, so we
+ * need to put an even number of registers of URB data in
+ * each URB write so that the later write is aligned. A
+ * message length of 15 means 1 message header reg plus 14
+ * regs of URB data.
+ *
+ * For attributes beyond the compute-to-MRF, we compute to
+ * GRFs and they will be written in the second URB_WRITE.
+ */
+ if (mrf < 15) {
c->regs[PROGRAM_OUTPUT][i] = brw_message_reg(mrf);
mrf++;
}
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[PROGRAM_OUTPUT][i] = brw_vec8_grf(reg, 0);
*/
attributes_in_vue = MAX2(c->nr_outputs, c->nr_inputs);
+ /* 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 + 4 + 7) / 8;
+ c->prog_data.urb_entry_size = (attributes_in_vue + 2 + 7) / 8;
else if (intel->gen == 5)
c->prog_data.urb_entry_size = (attributes_in_vue + 6 + 3) / 4;
else
struct brw_reg arg0,
struct brw_reg arg1 )
{
- brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_L, arg0, arg1);
- brw_SEL(p, dst, arg1, arg0);
+ 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);
}
*/
brw_set_access_mode(p, BRW_ALIGN_1);
+ /* The VUE layout is documented in Volume 2a. */
if (intel->gen >= 6) {
- /* There are 16 DWs (D0-D15) in VUE header on Sandybridge:
+ /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge:
* 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.
- * m5 is the first vertex data we fill, which is the vertex 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.
*/
- brw_MOV(p, offset(m0, 2), pos);
- brw_MOV(p, offset(m0, 5), pos);
- len_vertex_header = 4;
+ brw_MOV(p, brw_message_reg(2), pos);
+ brw_MOV(p, brw_message_reg(3), pos);
+ 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.
* m6 is a pad so that the vertex element data is aligned
* m7 is the first vertex data we fill, which is the vertex position.
*/
- brw_MOV(p, offset(m0, 2), ndc);
- brw_MOV(p, offset(m0, 3), pos);
- brw_MOV(p, offset(m0, 7), pos);
+ brw_MOV(p, brw_message_reg(2), ndc);
+ brw_MOV(p, brw_message_reg(3), pos);
+ brw_MOV(p, brw_message_reg(7), pos);
len_vertex_header = 6;
} else {
/* There are 8 dwords in VUE header pre-Ironlake:
* dword 8-11 (m3) is the first vertex data, which we always have be the
* vertex position.
*/
- brw_MOV(p, offset(m0, 2), ndc);
- brw_MOV(p, offset(m0, 3), pos);
+ brw_MOV(p, brw_message_reg(2), ndc);
+ brw_MOV(p, brw_message_reg(3), pos);
len_vertex_header = 2;
}
* Move the overflowed attributes from the GRF to the MRF and
* issue another brw_urb_WRITE().
*/
- /* XXX I'm not 100% sure about which MRF regs to use here. Starting
- * at mrf[4] atm...
- */
- GLuint i, mrf = 0;
+ GLuint i, mrf = 1;
for (i = c->first_overflow_output; i < VERT_RESULT_MAX; i++) {
if (c->prog_data.outputs_written & BITFIELD64_BIT(i)) {
/* move from GRF to MRF */
- brw_MOV(p, brw_message_reg(4+mrf), c->regs[PROGRAM_OUTPUT][i]);
+ brw_MOV(p, brw_message_reg(mrf), c->regs[PROGRAM_OUTPUT][i]);
mrf++;
}
}
brw_urb_WRITE(p,
brw_null_reg(), /* dest */
- 4, /* starting mrf reg nr */
+ 0, /* starting mrf reg nr */
c->r0, /* src */
0, /* allocate */
1, /* used */
- mrf+1, /* msg len */
+ mrf, /* msg len */
0, /* response len */
1, /* eot */
1, /* writes complete */
- BRW_MAX_MRF-1, /* urb destination offset */
+ 14 / 2, /* urb destination offset */
BRW_URB_SWIZZLE_INTERLEAVE);
}
}
printf("vs-native:\n");
for (i = 0; i < p->nr_insn; i++)
- brw_disasm(stderr, &p->store[i]);
+ brw_disasm(stderr, &p->store[i], intel->gen);
printf("\n");
}
}