Bug #18074
---
src/mesa/drivers/dri/i915/i915_context.c | 3 +-
src/mesa/drivers/dri/i915/i915_context.h | 1 +
src/mesa/drivers/dri/i915/i915_fragprog.c | 82 +++++++++++++++++++++++++----
src/mesa/drivers/dri/i915/i915_reg.h | 4 +-
4 files changed, 76 insertions(+), 14 deletions(-)
diff --git a/src/mesa/drivers/dri/i915/i915_context.c
b/src/mesa/drivers/dri/i915/i915_context.c
index 9bff742..010cf0c 100644
--- a/src/mesa/drivers/dri/i915/i915_context.c
+++ b/src/mesa/drivers/dri/i915/i915_context.c
@@ -159,7 +159,8 @@ i915CreateContext(const __GLcontextModes * mesaVis,
* instruction can translate to more than one HW instruction, so
* we'll still have to check and fallback each time.
*/
- ctx->Const.FragmentProgram.MaxNativeTemps = I915_MAX_TEMPORARY;
+ ctx->Const.FragmentProgram.MaxTemps = I915_MAX_TEMPS;
+ ctx->Const.FragmentProgram.MaxNativeTemps = I915_MAX_NATIVE_TEMPS;
ctx->Const.FragmentProgram.MaxNativeAttribs = 11; /* 8 tex, 2 color, fog
*/
ctx->Const.FragmentProgram.MaxNativeParameters = I915_MAX_CONSTANT;
ctx->Const.FragmentProgram.MaxNativeAluInstructions = I915_MAX_ALU_INSN;
diff --git a/src/mesa/drivers/dri/i915/i915_context.h
b/src/mesa/drivers/dri/i915/i915_context.h
index 87bbf5f..a5df3fa 100644
--- a/src/mesa/drivers/dri/i915/i915_context.h
+++ b/src/mesa/drivers/dri/i915/i915_context.h
@@ -177,6 +177,7 @@ struct i915_fragment_program
* it's read. */
GLuint usedRegs[I915_MAX_INSN];
+ GLuint temp_reg_mapping[I915_MAX_TEMPS];
/* Helpers for i915_fragprog.c:
*/
GLuint wpos_tex;
diff --git a/src/mesa/drivers/dri/i915/i915_fragprog.c
b/src/mesa/drivers/dri/i915/i915_fragprog.c
index 46f1740..ef22beb 100644
--- a/src/mesa/drivers/dri/i915/i915_fragprog.c
+++ b/src/mesa/drivers/dri/i915/i915_fragprog.c
@@ -87,11 +87,7 @@ src_vector(struct i915_fragment_program *p,
/* Registers:
*/
case PROGRAM_TEMPORARY:
- if (source->Index >= I915_MAX_TEMPORARY) {
- i915_program_error(p, "Exceeded max temporary reg");
- return 0;
- }
- src = UREG(REG_TYPE_R, source->Index);
+ src = UREG(REG_TYPE_R, p->temp_reg_mapping[source->Index]);
break;
case PROGRAM_INPUT:
switch (source->Index) {
@@ -272,15 +268,17 @@ do {
\
static void calc_live_regs( struct i915_fragment_program *p )
{
const struct gl_fragment_program *program =
p->ctx->FragmentProgram._Current;
- GLuint regsUsed = 0xffff0000;
+ GLuint regsUsed = 0;
GLint i;
/* Work from the front marking regs as live when they're written to. */
for (i = 0; i < program->Base.NumInstructions; i++) {
struct prog_instruction *inst = &program->Base.Instructions[i];
- if (inst->DstReg.File == PROGRAM_TEMPORARY)
+ if (inst->DstReg.File == PROGRAM_TEMPORARY) {
+ assert(inst->DstReg.Index < I915_MAX_TEMPS);
regsUsed |= 1 << inst->DstReg.Index;
+ }
p->usedRegs[i] = regsUsed;
}
@@ -292,20 +290,83 @@ static void calc_live_regs( struct i915_fragment_program
*p )
int a;
for (a = 0; a < opArgs; a++) {
- if (inst->SrcReg[a].File == PROGRAM_TEMPORARY)
+ if (inst->SrcReg[a].File == PROGRAM_TEMPORARY) {
+ assert(inst->SrcReg[a].Index < I915_MAX_TEMPS);
regsUsed |= 1 << inst->SrcReg[a].Index;
+ }
}
p->usedRegs[i] &= regsUsed;
}
}
+/* Returns the set of live hw registers at the moment, by working from
+ * the set of live sw temporaries.
+ */
static GLuint get_live_regs( struct i915_fragment_program *p,
const struct prog_instruction *inst )
{
const struct gl_fragment_program *program =
p->ctx->FragmentProgram._Current;
GLuint nr = inst - program->Base.Instructions;
+ GLuint live_sw_regs = p->usedRegs[nr];
+ GLuint live_hw_regs = 0;
+
+ while (live_sw_regs != 0) {
+ int sw_reg = ffs(live_sw_regs) - 1;
+
+ live_sw_regs &= ~(1 << sw_reg);
+ live_hw_regs |= 1 << p->temp_reg_mapping[sw_reg];
+ }
- return p->usedRegs[nr];
+ /* Return invalid hw regs as live, as the consumer looking for a
+ * temporary will use ffs on the inverse.
+ */
+ return 0xffff0000 | live_hw_regs;
+}
+
+/**
+ * Creates the mapping from ARB program temporary indices to native indices.
+ *
+ * This lets us run some programs which would otherwise exceed our limits on
+ * number of temporaries. Note that this code relies on calc_live_regs
+ * marking live from first write through last read, as otherwise we might
+ * assign it two different hw regs and end up stomping on live stuff.
+ */
+static void
+create_temp_mapping(struct i915_fragment_program *p)
+{
+ const struct gl_fragment_program *program =
p->ctx->FragmentProgram._Current;
+ GLuint avail_hw_regs = 0xffff; /* 1 << I915_MAX_NATIVE_TEMPS - 1 */
+ GLuint assigned_sw_indices = 0;
+ int i;
+
+ for (i = 0; i < program->Base.NumInstructions; i++) {
+ int sw_to_assign;
+
+ /* Get a hardware index for any newly-live registers. */
+ while ((sw_to_assign = ffs(p->usedRegs[i] &
+ ~assigned_sw_indices) - 1) != -1) {
+ int hw_reg = ffs(avail_hw_regs) - 1;
+ if (hw_reg == -1) {
+ i915_program_error(p, "Exceeded max native temporaries");
+ return;
+ }
+
+ avail_hw_regs &= ~(1 << hw_reg);
+ assigned_sw_indices |= 1 << sw_to_assign;
+ p->temp_reg_mapping[sw_to_assign] = hw_reg;
+ }
+
+ /* Restore availability for newly-dead registers */
+ if (i > 0) {
+ int sw_new_dead = p->usedRegs[i - 1] & ~p->usedRegs[i];
+ int sw_to_restore;
+
+ while ((sw_to_restore = ffs(sw_new_dead) - 1) != -1) {
+ avail_hw_regs |= 1 << p->temp_reg_mapping[sw_to_restore];
+ sw_new_dead &= ~(1 << sw_to_restore);
+ }
+ }
+ }
}
/* Possible concerns:
@@ -347,9 +408,8 @@ upload_program(struct i915_fragment_program *p)
return;
}
- /* Not always needed:
- */
calc_live_regs(p);
+ create_temp_mapping(p);
while (1) {
GLuint src0, src1, src2, flags;
diff --git a/src/mesa/drivers/dri/i915/i915_reg.h
b/src/mesa/drivers/dri/i915/i915_reg.h
index 8891e11..b4117eb 100644
--- a/src/mesa/drivers/dri/i915/i915_reg.h
+++ b/src/mesa/drivers/dri/i915/i915_reg.h
@@ -374,8 +374,8 @@
#define I915_MAX_TEX_INSN 32
#define I915_MAX_ALU_INSN 64
#define I915_MAX_DECL_INSN 27
-#define I915_MAX_TEMPORARY 16
-
+#define I915_MAX_NATIVE_TEMPS 16
+#define I915_MAX_TEMPS 32
/* Each instruction is 3 dwords long, though most don't require all
* this space. Maximum of 123 instructions. Smaller maxes per insn
--
1.5.6.5
------------------------------------------------------------------------------
Check out the new SourceForge.net Marketplace.
It is the best place to buy or sell services for
just about anything Open Source.
http://p.sf.net/sfu/Xq1LFB
_______________________________________________
Mesa3d-dev mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/mesa3d-dev
