summaryrefslogtreecommitdiffstats
path: root/src/gallium/drivers/r600
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/drivers/r600')
-rw-r--r--src/gallium/drivers/r600/Makefile1
-rw-r--r--src/gallium/drivers/r600/eg_asm.c27
-rw-r--r--src/gallium/drivers/r600/r600_asm.c1147
-rw-r--r--src/gallium/drivers/r600/r600_asm.h5
-rw-r--r--src/gallium/drivers/r600/r600_pipe.c2
-rw-r--r--src/gallium/drivers/r600/r600_shader.c20
-rw-r--r--src/gallium/drivers/r600/r600_state.c11
-rw-r--r--src/gallium/drivers/r600/r600_state_inlines.h16
-rw-r--r--src/gallium/drivers/r600/r600_texture.c96
-rw-r--r--src/gallium/drivers/r600/r600_video_context.c21
-rw-r--r--src/gallium/drivers/r600/r600_video_context.h11
11 files changed, 1121 insertions, 236 deletions
diff --git a/src/gallium/drivers/r600/Makefile b/src/gallium/drivers/r600/Makefile
index b476b9af3b8..a690b671e49 100644
--- a/src/gallium/drivers/r600/Makefile
+++ b/src/gallium/drivers/r600/Makefile
@@ -17,6 +17,7 @@ C_SOURCES = \
r600_shader.c \
r600_state.c \
r600_texture.c \
+ r600_video_context.c \
r700_asm.c \
evergreen_state.c \
eg_asm.c \
diff --git a/src/gallium/drivers/r600/eg_asm.c b/src/gallium/drivers/r600/eg_asm.c
index 67d742b3760..4f86e3b4c38 100644
--- a/src/gallium/drivers/r600/eg_asm.c
+++ b/src/gallium/drivers/r600/eg_asm.c
@@ -32,12 +32,14 @@
int eg_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
{
unsigned id = cf->id;
+ unsigned end_of_program = bc->cf.prev == &cf->list;
switch (cf->inst) {
case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3):
case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3):
case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3):
case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3):
+ assert(!end_of_program);
bc->bytecode[id++] = S_SQ_CF_ALU_WORD0_ADDR(cf->addr >> 1) |
S_SQ_CF_ALU_WORD0_KCACHE_MODE0(cf->kcache[0].mode) |
S_SQ_CF_ALU_WORD0_KCACHE_BANK0(cf->kcache[0].bank) |
@@ -46,15 +48,16 @@ int eg_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
S_SQ_CF_ALU_WORD1_KCACHE_MODE1(cf->kcache[1].mode) |
S_SQ_CF_ALU_WORD1_KCACHE_ADDR0(cf->kcache[0].addr) |
S_SQ_CF_ALU_WORD1_KCACHE_ADDR1(cf->kcache[1].addr) |
- S_SQ_CF_ALU_WORD1_BARRIER(1) |
- S_SQ_CF_ALU_WORD1_COUNT((cf->ndw / 2) - 1);
+ S_SQ_CF_ALU_WORD1_BARRIER(cf->barrier) |
+ S_SQ_CF_ALU_WORD1_COUNT((cf->ndw / 2) - 1);
break;
case EG_V_SQ_CF_WORD1_SQ_CF_INST_TEX:
case EG_V_SQ_CF_WORD1_SQ_CF_INST_VTX:
bc->bytecode[id++] = S_SQ_CF_WORD0_ADDR(cf->addr >> 1);
bc->bytecode[id++] = S_SQ_CF_WORD1_CF_INST(cf->inst) |
- S_SQ_CF_WORD1_BARRIER(1) |
- S_SQ_CF_WORD1_COUNT((cf->ndw / 4) - 1);
+ S_SQ_CF_WORD1_BARRIER(cf->barrier) |
+ S_SQ_CF_WORD1_COUNT((cf->ndw / 4) - 1) |
+ S_SQ_CF_WORD1_END_OF_PROGRAM(end_of_program);
break;
case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
@@ -62,13 +65,14 @@ int eg_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
S_SQ_CF_ALLOC_EXPORT_WORD0_ELEM_SIZE(cf->output.elem_size) |
S_SQ_CF_ALLOC_EXPORT_WORD0_ARRAY_BASE(cf->output.array_base) |
S_SQ_CF_ALLOC_EXPORT_WORD0_TYPE(cf->output.type);
- bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_X(cf->output.swizzle_x) |
+ bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(cf->output.burst_count - 1) |
+ S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_X(cf->output.swizzle_x) |
S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Y(cf->output.swizzle_y) |
S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Z(cf->output.swizzle_z) |
S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_W(cf->output.swizzle_w) |
- S_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(cf->output.barrier) |
- S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(cf->output.inst) |
- S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(cf->output.end_of_program);
+ S_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(cf->barrier) |
+ S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(cf->inst) |
+ S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(end_of_program);
break;
case EG_V_SQ_CF_WORD1_SQ_CF_INST_JUMP:
case EG_V_SQ_CF_WORD1_SQ_CF_INST_ELSE:
@@ -81,9 +85,10 @@ int eg_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
case EG_V_SQ_CF_WORD1_SQ_CF_INST_RETURN:
bc->bytecode[id++] = S_SQ_CF_WORD0_ADDR(cf->cf_addr >> 1);
bc->bytecode[id++] = S_SQ_CF_WORD1_CF_INST(cf->inst) |
- S_SQ_CF_WORD1_BARRIER(1) |
- S_SQ_CF_WORD1_COND(cf->cond) |
- S_SQ_CF_WORD1_POP_COUNT(cf->pop_count);
+ S_SQ_CF_WORD1_BARRIER(cf->barrier) |
+ S_SQ_CF_WORD1_COND(cf->cond) |
+ S_SQ_CF_WORD1_POP_COUNT(cf->pop_count) |
+ S_SQ_CF_WORD1_END_OF_PROGRAM(end_of_program);
break;
default:
diff --git a/src/gallium/drivers/r600/r600_asm.c b/src/gallium/drivers/r600/r600_asm.c
index a473fb29c4d..61de24b31ae 100644
--- a/src/gallium/drivers/r600/r600_asm.c
+++ b/src/gallium/drivers/r600/r600_asm.c
@@ -35,6 +35,9 @@
#define NUM_OF_CYCLES 3
#define NUM_OF_COMPONENTS 4
+#define PREV_ALU(alu) LIST_ENTRY(struct r600_bc_alu, alu->list.prev, list)
+#define NEXT_ALU(alu) LIST_ENTRY(struct r600_bc_alu, alu->list.next, list)
+
static inline unsigned int r600_bc_get_num_operands(struct r600_bc *bc, struct r600_bc_alu *alu)
{
if(alu->is_op3)
@@ -154,6 +157,7 @@ static struct r600_bc_cf *r600_bc_cf(void)
LIST_INITHEAD(&cf->alu);
LIST_INITHEAD(&cf->vtx);
LIST_INITHEAD(&cf->tex);
+ cf->barrier = 1;
return cf;
}
@@ -242,6 +246,44 @@ static int r600_bc_add_cf(struct r600_bc *bc)
return 0;
}
+static void r600_bc_remove_cf(struct r600_bc *bc, struct r600_bc_cf *cf)
+{
+ struct r600_bc_cf *other;
+ LIST_FOR_EACH_ENTRY(other, &bc->cf, list) {
+ if (other->id > cf->id)
+ other->id -= 2;
+ if (other->cf_addr > cf->id)
+ other->cf_addr -= 2;
+ }
+ LIST_DEL(&cf->list);
+ free(cf);
+}
+
+static void r600_bc_move_cf(struct r600_bc *bc, struct r600_bc_cf *cf, struct r600_bc_cf *next)
+{
+ struct r600_bc_cf *prev = LIST_ENTRY(struct r600_bc_cf, next->list.prev, list);
+ unsigned old_id = cf->id;
+ unsigned new_id = prev->id + 2;
+ struct r600_bc_cf *other;
+
+ if (prev == cf)
+ return; /* position hasn't changed */
+
+ LIST_DEL(&cf->list);
+ LIST_FOR_EACH_ENTRY(other, &bc->cf, list) {
+ if (other->id > old_id)
+ other->id -= 2;
+ if (other->id >= new_id)
+ other->id += 2;
+ if (other->cf_addr > old_id)
+ other->cf_addr -= 2;
+ if (other->cf_addr > new_id)
+ other->cf_addr += 2;
+ }
+ cf->id = new_id;
+ LIST_ADD(&cf->list, &prev->list);
+}
+
int r600_bc_add_output(struct r600_bc *bc, const struct r600_bc_output *output)
{
int r;
@@ -249,28 +291,19 @@ int r600_bc_add_output(struct r600_bc *bc, const struct r600_bc_output *output)
r = r600_bc_add_cf(bc);
if (r)
return r;
- bc->cf_last->inst = output->inst;
+ bc->cf_last->inst = BC_INST(bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT);
memcpy(&bc->cf_last->output, output, sizeof(struct r600_bc_output));
+ bc->cf_last->output.burst_count = 1;
return 0;
}
-/* alu instructions that can ony exits once per group */
-static int is_alu_once_inst(struct r600_bc *bc, struct r600_bc_alu *alu)
+/* alu predicate instructions */
+static int is_alu_pred_inst(struct r600_bc *bc, struct r600_bc_alu *alu)
{
switch (bc->chiprev) {
case CHIPREV_R600:
case CHIPREV_R700:
return !alu->is_op3 && (
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE ||
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT ||
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE ||
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE ||
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_UINT ||
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_UINT ||
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE_INT ||
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_INT ||
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_INT ||
- alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE_INT ||
alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_UINT ||
alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_UINT ||
alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE ||
@@ -298,16 +331,6 @@ static int is_alu_once_inst(struct r600_bc *bc, struct r600_bc_alu *alu)
case CHIPREV_EVERGREEN:
default:
return !alu->is_op3 && (
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE ||
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT ||
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE ||
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE ||
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_UINT ||
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_UINT ||
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE_INT ||
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_INT ||
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_INT ||
- alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE_INT ||
alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_UINT ||
alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_UINT ||
alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE ||
@@ -335,6 +358,46 @@ static int is_alu_once_inst(struct r600_bc *bc, struct r600_bc_alu *alu)
}
}
+/* alu kill instructions */
+static int is_alu_kill_inst(struct r600_bc *bc, struct r600_bc_alu *alu)
+{
+ switch (bc->chiprev) {
+ case CHIPREV_R600:
+ case CHIPREV_R700:
+ return !alu->is_op3 && (
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_UINT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_UINT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE_INT);
+ case CHIPREV_EVERGREEN:
+ default:
+ return !alu->is_op3 && (
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE ||
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT ||
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE ||
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE ||
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_UINT ||
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_UINT ||
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLE_INT ||
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT_INT ||
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE_INT ||
+ alu->inst == EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE_INT);
+ }
+}
+
+/* alu instructions that can ony exits once per group */
+static int is_alu_once_inst(struct r600_bc *bc, struct r600_bc_alu *alu)
+{
+ return is_alu_kill_inst(bc, alu) ||
+ is_alu_pred_inst(bc, alu);
+}
+
static int is_alu_reduction_inst(struct r600_bc *bc, struct r600_bc_alu *alu)
{
switch (bc->chiprev) {
@@ -527,7 +590,7 @@ static void init_bank_swizzle(struct alu_bank_swizzle *bs)
for (i = 0; i < 4; i++)
bs->hw_cfile_elem[i] = -1;
}
-
+
static int reserve_gpr(struct alu_bank_swizzle *bs, unsigned sel, unsigned chan, unsigned cycle)
{
if (bs->hw_gpr[cycle][chan] == -1)
@@ -538,7 +601,7 @@ static int reserve_gpr(struct alu_bank_swizzle *bs, unsigned sel, unsigned chan,
}
return 0;
}
-
+
static int reserve_cfile(struct alu_bank_swizzle *bs, unsigned sel, unsigned chan)
{
int res, resmatch = -1, resempty = -1;
@@ -558,9 +621,9 @@ static int reserve_cfile(struct alu_bank_swizzle *bs, unsigned sel, unsigned cha
// All cfile read ports are used, cannot reference vector element
return -1;
}
- return 0;
+ return 0;
}
-
+
static int is_gpr(unsigned sel)
{
return (sel >= 0 && sel <= 127);
@@ -575,19 +638,19 @@ static int is_cfile(unsigned sel)
(sel > 511 && sel < 4607) || // Kcache before translate
(sel > 127 && sel < 192); // Kcache after translate
}
-
+
static int is_const(int sel)
{
return is_cfile(sel) ||
- (sel >= V_SQ_ALU_SRC_0 &&
+ (sel >= V_SQ_ALU_SRC_0 &&
sel <= V_SQ_ALU_SRC_LITERAL);
}
-
+
static int check_vector(struct r600_bc *bc, struct r600_bc_alu *alu,
struct alu_bank_swizzle *bs, int bank_swizzle)
{
int r, src, num_src, sel, elem, cycle;
-
+
num_src = r600_bc_get_num_operands(bc, alu);
for (src = 0; src < num_src; src++) {
sel = alu->src[src].sel;
@@ -595,7 +658,7 @@ static int check_vector(struct r600_bc *bc, struct r600_bc_alu *alu,
if (is_gpr(sel)) {
cycle = cycle_for_bank_swizzle_vec[bank_swizzle][src];
if (src == 1 && sel == alu->src[0].sel && elem == alu->src[0].chan)
- // Nothing to do; special-case optimization,
+ // Nothing to do; special-case optimization,
// second source uses first source’s reservation
continue;
else {
@@ -612,12 +675,12 @@ static int check_vector(struct r600_bc *bc, struct r600_bc_alu *alu,
}
return 0;
}
-
+
static int check_scalar(struct r600_bc *bc, struct r600_bc_alu *alu,
struct alu_bank_swizzle *bs, int bank_swizzle)
{
int r, src, num_src, const_count, sel, elem, cycle;
-
+
num_src = r600_bc_get_num_operands(bc, alu);
for (const_count = 0, src = 0; src < num_src; ++src) {
sel = alu->src[src].sel;
@@ -626,7 +689,7 @@ static int check_scalar(struct r600_bc *bc, struct r600_bc_alu *alu,
if (const_count >= 2)
// More than two references to a constant in
// transcendental operation.
- return -1;
+ return -1;
else
const_count++;
}
@@ -661,7 +724,7 @@ static int check_and_set_bank_swizzle(struct r600_bc *bc,
struct alu_bank_swizzle bs;
int bank_swizzle[5];
int i, r = 0, forced = 0;
-
+
for (i = 0; i < 5; i++)
if (slots[i] && slots[i]->bank_swizzle_force) {
slots[i]->bank_swizzle = slots[i]->bank_swizzle_force;
@@ -846,7 +909,7 @@ static int merge_inst_groups(struct r600_bc *bc, struct r600_bc_alu *slots[5],
{
struct r600_bc_alu *prev[5];
struct r600_bc_alu *result[5] = { NULL };
-
+
uint32_t literal[4], prev_literal[4];
unsigned nliteral = 0, prev_nliteral = 0;
@@ -890,7 +953,7 @@ static int merge_inst_groups(struct r600_bc *bc, struct r600_bc_alu *slots[5],
return 0;
} else if(!slots[i]) {
continue;
- } else
+ } else
result[i] = slots[i];
// let's check source gprs
@@ -1115,19 +1178,21 @@ int r600_bc_add_alu_type(struct r600_bc *bc, const struct r600_bc_alu *alu, int
if (!bc->cf_last->curr_bs_head) {
bc->cf_last->curr_bs_head = nalu;
}
- /* number of gpr == the last gpr used in any alu */
+ /* replace special constants */
for (i = 0; i < 3; i++) {
- if (nalu->src[i].sel >= bc->ngpr && nalu->src[i].sel < 128) {
- bc->ngpr = nalu->src[i].sel + 1;
- }
if (nalu->src[i].sel == V_SQ_ALU_SRC_LITERAL)
r600_bc_special_constants(
- nalu->src[i].value[nalu->src[i].chan],
+ nalu->src[i].value[nalu->src[i].chan],
&nalu->src[i].sel, &nalu->src[i].neg);
+
+ if (nalu->src[i].sel >= bc->ngpr && nalu->src[i].sel < 128) {
+ bc->ngpr = nalu->src[i].sel + 1;
+ }
}
if (nalu->dst.sel >= bc->ngpr) {
bc->ngpr = nalu->dst.sel + 1;
}
+
LIST_ADDTAIL(&nalu->list, &bc->cf_last->alu);
/* each alu use 2 dwords */
bc->cf_last->ndw += 2;
@@ -1185,6 +1250,16 @@ int r600_bc_add_alu(struct r600_bc *bc, const struct r600_bc_alu *alu)
return r600_bc_add_alu_type(bc, alu, BC_INST(bc, V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU));
}
+static void r600_bc_remove_alu(struct r600_bc_cf *cf, struct r600_bc_alu *alu)
+{
+ if (alu->last && alu->list.prev != &cf->alu) {
+ PREV_ALU(alu)->last = 1;
+ }
+ LIST_DEL(&alu->list);
+ free(alu);
+ cf->ndw -= 2;
+}
+
int r600_bc_add_vtx(struct r600_bc *bc, const struct r600_bc_vtx *vtx)
{
struct r600_bc_vtx *nvtx = r600_bc_vtx();
@@ -1380,16 +1455,61 @@ static int r600_bc_alu_build(struct r600_bc *bc, struct r600_bc_alu *alu, unsign
return 0;
}
-/* common for r600/r700 - eg in eg_asm.c */
-static int r600_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
+enum cf_class
{
- unsigned id = cf->id;
+ CF_CLASS_ALU,
+ CF_CLASS_TEXTURE,
+ CF_CLASS_VERTEX,
+ CF_CLASS_EXPORT,
+ CF_CLASS_OTHER
+};
+static enum cf_class get_cf_class(struct r600_bc_cf *cf)
+{
switch (cf->inst) {
case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3):
case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3):
case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3):
+ case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3):
+ return CF_CLASS_ALU;
+
+ case V_SQ_CF_WORD1_SQ_CF_INST_TEX:
+ return CF_CLASS_TEXTURE;
+
+ case V_SQ_CF_WORD1_SQ_CF_INST_VTX:
+ case V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC:
+ return CF_CLASS_VERTEX;
+
+ case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
+ case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
+ return CF_CLASS_EXPORT;
+
+ case V_SQ_CF_WORD1_SQ_CF_INST_JUMP:
+ case V_SQ_CF_WORD1_SQ_CF_INST_ELSE:
+ case V_SQ_CF_WORD1_SQ_CF_INST_POP:
+ case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_START_NO_AL:
+ case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_END:
+ case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE:
+ case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK:
+ case V_SQ_CF_WORD1_SQ_CF_INST_CALL_FS:
+ case V_SQ_CF_WORD1_SQ_CF_INST_RETURN:
+ return CF_CLASS_OTHER;
+
+ default:
+ R600_ERR("unsupported CF instruction (0x%X)\n", cf->inst);
+ return -EINVAL;
+ }
+}
+
+/* common for r600/r700 - eg in eg_asm.c */
+static int r600_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
+{
+ unsigned id = cf->id;
+ unsigned end_of_program = bc->cf.prev == &cf->list;
+
+ switch (get_cf_class(cf)) {
+ case CF_CLASS_ALU:
+ assert(!end_of_program);
bc->bytecode[id++] = S_SQ_CF_ALU_WORD0_ADDR(cf->addr >> 1) |
S_SQ_CF_ALU_WORD0_KCACHE_MODE0(cf->kcache[0].mode) |
S_SQ_CF_ALU_WORD0_KCACHE_BANK0(cf->kcache[0].bank) |
@@ -1399,46 +1519,39 @@ static int r600_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
S_SQ_CF_ALU_WORD1_KCACHE_MODE1(cf->kcache[1].mode) |
S_SQ_CF_ALU_WORD1_KCACHE_ADDR0(cf->kcache[0].addr) |
S_SQ_CF_ALU_WORD1_KCACHE_ADDR1(cf->kcache[1].addr) |
- S_SQ_CF_ALU_WORD1_BARRIER(1) |
- S_SQ_CF_ALU_WORD1_USES_WATERFALL(bc->chiprev == CHIPREV_R600 ? cf->r6xx_uses_waterfall : 0) |
- S_SQ_CF_ALU_WORD1_COUNT((cf->ndw / 2) - 1);
+ S_SQ_CF_ALU_WORD1_BARRIER(cf->barrier) |
+ S_SQ_CF_ALU_WORD1_USES_WATERFALL(bc->chiprev == CHIPREV_R600 ? cf->r6xx_uses_waterfall : 0) |
+ S_SQ_CF_ALU_WORD1_COUNT((cf->ndw / 2) - 1);
break;
- case V_SQ_CF_WORD1_SQ_CF_INST_TEX:
- case V_SQ_CF_WORD1_SQ_CF_INST_VTX:
- case V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC:
+ case CF_CLASS_TEXTURE:
+ case CF_CLASS_VERTEX:
bc->bytecode[id++] = S_SQ_CF_WORD0_ADDR(cf->addr >> 1);
bc->bytecode[id++] = S_SQ_CF_WORD1_CF_INST(cf->inst) |
- S_SQ_CF_WORD1_BARRIER(1) |
- S_SQ_CF_WORD1_COUNT((cf->ndw / 4) - 1);
+ S_SQ_CF_WORD1_BARRIER(cf->barrier) |
+ S_SQ_CF_WORD1_COUNT((cf->ndw / 4) - 1) |
+ S_SQ_CF_WORD1_END_OF_PROGRAM(end_of_program);
break;
- case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
- case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
+ case CF_CLASS_EXPORT:
bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD0_RW_GPR(cf->output.gpr) |
S_SQ_CF_ALLOC_EXPORT_WORD0_ELEM_SIZE(cf->output.elem_size) |
S_SQ_CF_ALLOC_EXPORT_WORD0_ARRAY_BASE(cf->output.array_base) |
S_SQ_CF_ALLOC_EXPORT_WORD0_TYPE(cf->output.type);
- bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_X(cf->output.swizzle_x) |
+ bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(cf->output.burst_count - 1) |
+ S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_X(cf->output.swizzle_x) |
S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Y(cf->output.swizzle_y) |
S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Z(cf->output.swizzle_z) |
S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_W(cf->output.swizzle_w) |
- S_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(cf->output.barrier) |
- S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(cf->output.inst) |
- S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(cf->output.end_of_program);
+ S_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(cf->barrier) |
+ S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(cf->inst) |
+ S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(end_of_program);
break;
- case V_SQ_CF_WORD1_SQ_CF_INST_JUMP:
- case V_SQ_CF_WORD1_SQ_CF_INST_ELSE:
- case V_SQ_CF_WORD1_SQ_CF_INST_POP:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_START_NO_AL:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_END:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK:
- case V_SQ_CF_WORD1_SQ_CF_INST_CALL_FS:
- case V_SQ_CF_WORD1_SQ_CF_INST_RETURN:
+ case CF_CLASS_OTHER:
bc->bytecode[id++] = S_SQ_CF_WORD0_ADDR(cf->cf_addr >> 1);
bc->bytecode[id++] = S_SQ_CF_WORD1_CF_INST(cf->inst) |
- S_SQ_CF_WORD1_BARRIER(1) |
- S_SQ_CF_WORD1_COND(cf->cond) |
- S_SQ_CF_WORD1_POP_COUNT(cf->pop_count);
+ S_SQ_CF_WORD1_BARRIER(cf->barrier) |
+ S_SQ_CF_WORD1_COND(cf->cond) |
+ S_SQ_CF_WORD1_POP_COUNT(cf->pop_count) |
+ S_SQ_CF_WORD1_END_OF_PROGRAM(end_of_program);
break;
default:
@@ -1448,12 +1561,773 @@ static int r600_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
return 0;
}
+struct gpr_usage_range {
+ int replacement;
+ int32_t start;
+ int32_t end;
+};
+
+struct gpr_usage {
+ unsigned channels:4;
+ int32_t first_write;
+ int32_t last_write[4];
+ unsigned nranges;
+ struct gpr_usage_range *ranges;
+};
+
+static struct gpr_usage_range* add_gpr_usage_range(struct gpr_usage *usage)
+{
+ usage->nranges++;
+ usage->ranges = realloc(usage->ranges, usage->nranges * sizeof(struct gpr_usage_range));
+ if (!usage->ranges)
+ return NULL;
+ return &usage->ranges[usage->nranges-1];
+}
+
+static void notice_gpr_read(struct gpr_usage *usage, int32_t id, unsigned chan)
+{
+ usage->channels |= 1 << chan;
+ usage->first_write = -1;
+ if (!usage->nranges) {
+ struct gpr_usage_range* range = add_gpr_usage_range(usage);
+ range->replacement = -1;
+ range->start = -1;
+ range->end = -1;
+ }
+ if (usage->ranges[usage->nranges-1].end < id)
+ usage->ranges[usage->nranges-1].end = id;
+}
+
+static void notice_gpr_rel_read(struct gpr_usage usage[128], int32_t id, unsigned chan)
+{
+ unsigned i;
+ for (i = 0; i < 128; ++i)
+ notice_gpr_read(&usage[i], id, chan);
+}
+
+static void notice_gpr_last_write(struct gpr_usage *usage, int32_t id, unsigned chan)
+{
+ usage->last_write[chan] = id;
+}
+
+static void notice_gpr_write(struct gpr_usage *usage, int32_t id, unsigned chan,
+ int predicate, int prefered_replacement)
+{
+ int32_t start = usage->first_write != -1 ? usage->first_write : id;
+ usage->channels &= ~(1 << chan);
+ if (usage->channels) {
+ if (usage->first_write == -1)
+ usage->first_write = id;
+ } else if (!usage->nranges || (usage->ranges[usage->nranges-1].start != start && !predicate)) {
+ usage->first_write = start;
+ struct gpr_usage_range* range = add_gpr_usage_range(usage);
+ range->replacement = prefered_replacement;
+ range->start = start;
+ range->end = -1;
+ } else if (usage->ranges[usage->nranges-1].start == start && prefered_replacement != -1) {
+ usage->ranges[usage->nranges-1].replacement = prefered_replacement;
+ }
+ notice_gpr_last_write(usage, id, chan);
+}
+
+static void notice_gpr_rel_last_write(struct gpr_usage usage[128], int32_t id, unsigned chan)
+{
+ unsigned i;
+ for (i = 0; i < 128; ++i)
+ notice_gpr_last_write(&usage[i], id, chan);
+}
+
+static void notice_gpr_rel_write(struct gpr_usage usage[128], int32_t id, unsigned chan)
+{
+ unsigned i;
+ for (i = 0; i < 128; ++i)
+ notice_gpr_write(&usage[i], id, chan, 1, -1);
+}
+
+static void notice_alu_src_gprs(struct r600_bc *bc, struct r600_bc_alu *alu,
+ struct gpr_usage usage[128], int32_t id)
+{
+ unsigned src, num_src;
+
+ num_src = r600_bc_get_num_operands(bc, alu);
+ for (src = 0; src < num_src; ++src) {
+ // constants doesn't matter
+ if (!is_gpr(alu->src[src].sel))
+ continue;
+
+ if (alu->src[src].rel)
+ notice_gpr_rel_read(usage, id, alu->src[src].chan);
+ else
+ notice_gpr_read(&usage[alu->src[src].sel], id, alu->src[src].chan);
+ }
+}
+
+static void notice_alu_dst_gprs(struct r600_bc_alu *alu_first, struct gpr_usage usage[128],
+ int32_t id, int predicate)
+{
+ struct r600_bc_alu *alu;
+ for (alu = alu_first; alu; alu = LIST_ENTRY(struct r600_bc_alu, alu->list.next, list)) {
+ if (alu->dst.write) {
+ if (alu->dst.rel)
+ notice_gpr_rel_write(usage, id, alu->dst.chan);
+ else if (alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV && is_gpr(alu->src[0].sel))
+ notice_gpr_write(&usage[alu->dst.sel], id, alu->dst.chan,
+ predicate, alu->src[0].sel);
+ else
+ notice_gpr_write(&usage[alu->dst.sel], id, alu->dst.chan, predicate, -1);
+ }
+
+ if (alu->last)
+ break;
+ }
+}
+
+static void notice_tex_gprs(struct r600_bc_tex *tex, struct gpr_usage usage[128],
+ int32_t id, int predicate)
+{
+ if (tex->src_rel) {
+ if (tex->src_sel_x < 4)
+ notice_gpr_rel_read(usage, id, tex->src_sel_x);
+ if (tex->src_sel_y < 4)
+ notice_gpr_rel_read(usage, id, tex->src_sel_y);
+ if (tex->src_sel_z < 4)
+ notice_gpr_rel_read(usage, id, tex->src_sel_z);
+ if (tex->src_sel_w < 4)
+ notice_gpr_rel_read(usage, id, tex->src_sel_w);
+ } else {
+ if (tex->src_sel_x < 4)
+ notice_gpr_read(&usage[tex->src_gpr], id, tex->src_sel_x);
+ if (tex->src_sel_y < 4)
+ notice_gpr_read(&usage[tex->src_gpr], id, tex->src_sel_y);
+ if (tex->src_sel_z < 4)
+ notice_gpr_read(&usage[tex->src_gpr], id, tex->src_sel_z);
+ if (tex->src_sel_w < 4)
+ notice_gpr_read(&usage[tex->src_gpr], id, tex->src_sel_w);
+ }
+ if (tex->dst_rel) {
+ if (tex->dst_sel_x != 7)
+ notice_gpr_rel_write(usage, id, 0);
+ if (tex->dst_sel_y != 7)
+ notice_gpr_rel_write(usage, id, 1);
+ if (tex->dst_sel_z != 7)
+ notice_gpr_rel_write(usage, id, 2);
+ if (tex->dst_sel_w != 7)
+ notice_gpr_rel_write(usage, id, 3);
+ } else {
+ if (tex->dst_sel_x != 7)
+ notice_gpr_write(&usage[tex->dst_gpr], id, 0, predicate, -1);
+ if (tex->dst_sel_y != 7)
+ notice_gpr_write(&usage[tex->dst_gpr], id, 1, predicate, -1);
+ if (tex->dst_sel_z != 7)
+ notice_gpr_write(&usage[tex->dst_gpr], id, 2, predicate, -1);
+ if (tex->dst_sel_w != 7)
+ notice_gpr_write(&usage[tex->dst_gpr], id, 3, predicate, -1);
+ }
+}
+
+static void notice_vtx_gprs(struct r600_bc_vtx *vtx, struct gpr_usage usage[128],
+ int32_t id, int predicate)
+{
+ notice_gpr_read(&usage[vtx->src_gpr], id, vtx->src_sel_x);
+
+ if (vtx->dst_sel_x != 7)
+ notice_gpr_write(&usage[vtx->dst_gpr], id, 0, predicate, -1);
+ if (vtx->dst_sel_y != 7)
+ notice_gpr_write(&usage[vtx->dst_gpr], id, 1, predicate, -1);
+ if (vtx->dst_sel_z != 7)
+ notice_gpr_write(&usage[vtx->dst_gpr], id, 2, predicate, -1);
+ if (vtx->dst_sel_w != 7)
+ notice_gpr_write(&usage[vtx->dst_gpr], id, 3, predicate, -1);
+}
+
+static void notice_export_gprs(struct r600_bc_cf *cf, struct gpr_usage usage[128],
+ struct r600_bc_cf *export_cf[128], int32_t export_remap[128])
+{
+ //TODO handle other memory operations
+ struct gpr_usage *output = &usage[cf->output.gpr];
+ int32_t id = (output->last_write[0] + 0x100) & ~0xFF;
+
+ export_cf[cf->output.gpr] = cf;
+ export_remap[cf->output.gpr] = id;
+ if (cf->output.swizzle_x < 4)
+ notice_gpr_read(output, id, cf->output.swizzle_x);
+ if (cf->output.swizzle_y < 4)
+ notice_gpr_read(output, id, cf->output.swizzle_y);
+ if (cf->output.swizzle_z < 4)
+ notice_gpr_read(output, id, cf->output.swizzle_z);
+ if (cf->output.swizzle_w < 4)
+ notice_gpr_read(output, id, cf->output.swizzle_w);
+}
+
+static struct gpr_usage_range *find_src_range(struct gpr_usage *usage, int32_t id)
+{
+ unsigned i;
+ for (i = 0; i < usage->nranges; ++i) {
+ struct gpr_usage_range* range = &usage->ranges[i];
+
+ if (range->start < id && id <= range->end)
+ return range;
+ }
+ return NULL;
+}
+
+static struct gpr_usage_range *find_dst_range(struct gpr_usage *usage, int32_t id)
+{
+ unsigned i;
+ for (i = 0; i < usage->nranges; ++i) {
+ struct gpr_usage_range* range = &usage->ranges[i];
+ int32_t end = range->end;
+
+ if (range->start <= id && (id < end || end == -1))
+ return range;
+ }
+ assert(0); /* should not happen */
+ return NULL;
+}
+
+static int is_barrier_needed(struct gpr_usage *usage, int32_t id, unsigned chan, int32_t last_barrier)
+{
+ if (usage->last_write[chan] != (id & ~0xFF))
+ return usage->last_write[chan] >= last_barrier;
+ else
+ return 0;
+}
+
+static int is_intersection(struct gpr_usage_range* a, struct gpr_usage_range* b)
+{
+ return a->start <= b->end && b->start < a->end;
+}
+
+static int rate_replacement(struct gpr_usage *usage, struct gpr_usage_range* range)
+{
+ unsigned i;
+ int32_t best_start = 0x3FFFFFFF, best_end = 0x3FFFFFFF;
+
+ for (i = 0; i < usage->nranges; ++i) {
+ if (usage->ranges[i].replacement != -1)
+ continue; /* ignore already remapped ranges */
+
+ if (is_intersection(&usage->ranges[i], range))
+ return -1; /* forget it if usages overlap */
+
+ if (range->start >= usage->ranges[i].end)
+ best_start = MIN2(best_start, range->start - usage->ranges[i].end);
+
+ if (range->end != -1 && range->end <= usage->ranges[i].start)
+ best_end = MIN2(best_end, usage->ranges[i].start - range->end);
+ }
+ return best_start + best_end;
+}
+
+static void find_replacement(struct gpr_usage usage[128], unsigned current,
+ struct gpr_usage_range *range, int is_export)
+{
+ unsigned i;
+ int best_gpr = -1, best_rate = 0x7FFFFFFF;
+
+ if (range->replacement != -1 && range->replacement <= current) {
+ struct gpr_usage_range *other = find_src_range(&usage[range->replacement], range->start);
+ if (other && other->replacement != -1)
+ range->replacement = other->replacement;
+ }
+
+ if (range->replacement != -1 && range->replacement < current) {
+ int rate = rate_replacement(&usage[range->replacement], range);
+
+ /* check if prefered replacement can be used */
+ if (rate != -1) {
+ best_rate = rate;
+ best_gpr = range->replacement;
+ }
+ }
+
+ if (best_gpr == -1 && (range->start & ~0xFF) == (range->end & ~0xFF)) {
+ /* register is just used inside one ALU clause */
+ /* try to use clause temporaryis for it */
+ for (i = 127; i > 123; --i) {
+ int rate = rate_replacement(&usage[i], range);
+
+ if (rate == -1) /* can't be used because ranges overlap */
+ continue;
+
+ if (rate < best_rate) {
+ best_rate = rate;
+ best_gpr = i;
+
+ /* can't get better than this */
+ if (rate == 0 || is_export)
+ break;
+ }
+ }
+ }
+
+ if (best_gpr == -1) {
+ for (i = 0; i < current; ++i) {
+ int rate = rate_replacement(&usage[i], range);
+
+ if (rate == -1) /* can't be used because ranges overlap */
+ continue;
+
+ if (rate < best_rate) {
+ best_rate = rate;
+ best_gpr = i;
+
+ /* can't get better than this */
+ if (rate == 0)
+ break;
+ }
+ }
+ }
+
+ range->replacement = best_gpr;
+ if (best_gpr != -1) {
+ struct gpr_usage_range *reservation = add_gpr_usage_range(&usage[best_gpr]);
+ reservation->replacement = -1;
+ reservation->start = range->start;
+ reservation->end = range->end;
+ }
+}
+
+static void find_export_replacement(struct gpr_usage usage[128],
+ struct gpr_usage_range *range, struct r600_bc_cf *current,
+ struct r600_bc_cf *next, int32_t next_id)
+{
+ if (!next || next_id <= range->start || next_id > range->end)
+ return;
+
+ if (current->output.type != next->output.type)
+ return;
+
+ if ((current->output.array_base + 1) != next->output.array_base)
+ return;
+
+ find_src_range(&usage[next->output.gpr], next_id)->replacement = range->replacement + 1;
+}
+
+static void replace_alu_gprs(struct r600_bc *bc, struct r600_bc_alu *alu, struct gpr_usage usage[128],
+ int32_t id, int32_t last_barrier, unsigned *barrier)
+{
+ struct gpr_usage *cur_usage;
+ struct gpr_usage_range *range;
+ unsigned src, num_src;
+
+ num_src = r600_bc_get_num_operands(bc, alu);
+ for (src = 0; src < num_src; ++src) {
+ // constants doesn't matter
+ if (!is_gpr(alu->src[src].sel))
+ continue;
+
+ cur_usage = &usage[alu->src[src].sel];
+ range = find_src_range(cur_usage, id);
+ if (range->replacement != -1)
+ alu->src[src].sel = range->replacement;
+
+ *barrier |= is_barrier_needed(cur_usage, id, alu->src[src].chan, last_barrier);
+ }
+
+ if (alu->dst.write) {
+ cur_usage = &usage[alu->dst.sel];
+ range = find_dst_range(cur_usage, id);
+ if (range->replacement == alu->dst.sel) {
+ if (!alu->is_op3)
+ alu->dst.write = 0;
+ else
+ /*TODO: really check that register 123 is useable */
+ alu->dst.sel = 123;
+ } else if (range->replacement != -1) {
+ alu->dst.sel = range->replacement;
+ }
+ if (alu->dst.rel)
+ notice_gpr_rel_last_write(usage, id, alu->dst.chan);
+ else
+ notice_gpr_last_write(cur_usage, id, alu->dst.chan);
+ }
+}
+
+static void replace_tex_gprs(struct r600_bc_tex *tex, struct gpr_usage usage[128],
+ int32_t id, int32_t last_barrier, unsigned *barrier)
+{
+ struct gpr_usage *cur_usage = &usage[tex->src_gpr];
+ struct gpr_usage_range *range = find_src_range(cur_usage, id);
+
+ if (tex->src_rel) {
+ *barrier = 1;
+ } else {
+ if (tex->src_sel_x < 4)
+ *barrier |= is_barrier_needed(cur_usage, id, tex->src_sel_x, last_barrier);
+ if (tex->src_sel_y < 4)
+ *barrier |= is_barrier_needed(cur_usage, id, tex->src_sel_y, last_barrier);
+ if (tex->src_sel_z < 4)
+ *barrier |= is_barrier_needed(cur_usage, id, tex->src_sel_z, last_barrier);
+ if (tex->src_sel_w < 4)
+ *barrier |= is_barrier_needed(cur_usage, id, tex->src_sel_w, last_barrier);
+ }
+
+ if (range->replacement != -1)
+ tex->src_gpr = range->replacement;
+
+ cur_usage = &usage[tex->dst_gpr];
+ range = find_dst_range(cur_usage, id);
+ if (range->replacement != -1)
+ tex->dst_gpr = range->replacement;
+
+ if (tex->dst_rel) {
+ if (tex->dst_sel_x != 7)
+ notice_gpr_rel_last_write(usage, id, tex->dst_sel_x);
+ if (tex->dst_sel_y != 7)
+ notice_gpr_rel_last_write(usage, id, tex->dst_sel_y);
+ if (tex->dst_sel_z != 7)
+ notice_gpr_rel_last_write(usage, id, tex->dst_sel_z);
+ if (tex->dst_sel_w != 7)
+ notice_gpr_rel_last_write(usage, id, tex->dst_sel_w);
+ } else {
+ if (tex->dst_sel_x != 7)
+ notice_gpr_last_write(cur_usage, id, tex->dst_sel_x);
+ if (tex->dst_sel_y != 7)
+ notice_gpr_last_write(cur_usage, id, tex->dst_sel_y);
+ if (tex->dst_sel_z != 7)
+ notice_gpr_last_write(cur_usage, id, tex->dst_sel_z);
+ if (tex->dst_sel_w != 7)
+ notice_gpr_last_write(cur_usage, id, tex->dst_sel_w);
+ }
+}
+
+static void replace_vtx_gprs(struct r600_bc_vtx *vtx, struct gpr_usage usage[128],
+ int32_t id, int32_t last_barrier, unsigned *barrier)
+{
+ struct gpr_usage *cur_usage = &usage[vtx->src_gpr];
+ struct gpr_usage_range *range = find_src_range(cur_usage, id);
+
+ *barrier |= is_barrier_needed(cur_usage, id, vtx->src_sel_x, last_barrier);
+
+ if (range->replacement != -1)
+ vtx->src_gpr = range->replacement;
+
+ cur_usage = &usage[vtx->dst_gpr];
+ range = find_dst_range(cur_usage, id);
+ if (range->replacement != -1)
+ vtx->dst_gpr = range->replacement;
+
+ if (vtx->dst_sel_x != 7)
+ notice_gpr_last_write(cur_usage, id, vtx->dst_sel_x);
+ if (vtx->dst_sel_y != 7)
+ notice_gpr_last_write(cur_usage, id, vtx->dst_sel_y);
+ if (vtx->dst_sel_z != 7)
+ notice_gpr_last_write(cur_usage, id, vtx->dst_sel_z);
+ if (vtx->dst_sel_w != 7)
+ notice_gpr_last_write(cur_usage, id, vtx->dst_sel_w);
+}
+
+static void replace_export_gprs(struct r600_bc_cf *cf, struct gpr_usage usage[128],
+ int32_t id, int32_t last_barrier)
+{
+ //TODO handle other memory operations
+ struct gpr_usage *cur_usage = &usage[cf->output.gpr];
+ struct gpr_usage_range *range = find_src_range(cur_usage, id);
+
+ cf->barrier = 0;
+ if (cf->output.swizzle_x < 4)
+ cf->barrier |= is_barrier_needed(cur_usage, -1, cf->output.swizzle_x, last_barrier);
+ if (cf->output.swizzle_y < 4)
+ cf->barrier |= is_barrier_needed(cur_usage, -1, cf->output.swizzle_y, last_barrier);
+ if (cf->output.swizzle_z < 4)
+ cf->barrier |= is_barrier_needed(cur_usage, -1, cf->output.swizzle_z, last_barrier);
+ if (cf->output.swizzle_w < 4)
+ cf->barrier |= is_barrier_needed(cur_usage, -1, cf->output.swizzle_w, last_barrier);
+
+ if (range->replacement != -1)
+ cf->output.gpr = range->replacement;
+}
+
+static void optimize_alu_inst(struct r600_bc *bc, struct r600_bc_cf *cf, struct r600_bc_alu *alu)
+{
+ struct r600_bc_alu *alu_next;
+ unsigned chan;
+ unsigned src, num_src;
+
+ /* check if a MOV could be optimized away */
+ if (alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV) {
+
+ /* destination equals source? */
+ if (alu->dst.sel != alu->src[0].sel ||
+ alu->dst.chan != alu->src[0].chan)
+ return;
+
+ /* any special handling for the source? */
+ if (alu->src[0].rel || alu->src[0].neg || alu->src[0].abs)
+ return;
+
+ /* any special handling for destination? */
+ if (alu->dst.rel || alu->dst.clamp)
+ return;
+
+ /* ok find next instruction group and check if ps/pv is used */
+ for (alu_next = alu; !alu_next->last; alu_next = NEXT_ALU(alu_next));
+
+ if (alu_next->list.next != &cf->alu) {
+ chan = is_alu_reduction_inst(bc, alu) ? 0 : alu->dst.chan;
+ for (alu_next = NEXT_ALU(alu_next); alu_next; alu_next = NEXT_ALU(alu_next)) {
+ num_src = r600_bc_get_num_operands(bc, alu_next);
+ for (src = 0; src < num_src; ++src) {
+ if (alu_next->src[src].sel == V_SQ_ALU_SRC_PV &&
+ alu_next->src[src].chan == chan)
+ return;
+
+ if (alu_next->src[src].sel == V_SQ_ALU_SRC_PS)
+ return;
+ }
+
+ if (alu_next->last)
+ break;
+ }
+ }
+
+ r600_bc_remove_alu(cf, alu);
+ }
+}
+
+static void optimize_export_inst(struct r600_bc *bc, struct r600_bc_cf *cf)
+{
+ struct r600_bc_cf *prev = LIST_ENTRY(struct r600_bc_cf, cf->list.prev, list);
+ if (&prev->list == &bc->cf ||
+ prev->inst != cf->inst ||
+ prev->output.type != cf->output.type ||
+ prev->output.elem_size != cf->output.elem_size ||
+ prev->output.swizzle_x != cf->output.swizzle_x ||
+ prev->output.swizzle_y != cf->output.swizzle_y ||
+ prev->output.swizzle_z != cf->output.swizzle_z ||
+ prev->output.swizzle_w != cf->output.swizzle_w)
+ return;
+
+ if ((prev->output.burst_count + cf->output.burst_count) > 16)
+ return;
+
+ if ((prev->output.gpr + prev->output.burst_count) == cf->output.gpr &&
+ (prev->output.array_base + prev->output.burst_count) == cf->output.array_base) {
+
+ prev->output.burst_count += cf->output.burst_count;
+ r600_bc_remove_cf(bc, cf);
+
+ } else if (prev->output.gpr == (cf->output.gpr + cf->output.burst_count) &&
+ prev->output.array_base == (cf->output.array_base + cf->output.burst_count)) {
+
+ cf->output.burst_count += prev->output.burst_count;
+ r600_bc_remove_cf(bc, prev);
+ }
+}
+
+static void r600_bc_optimize(struct r600_bc *bc)
+{
+ struct r600_bc_cf *cf, *next_cf;
+ struct r600_bc_alu *first, *next_alu;
+ struct r600_bc_alu *alu;
+ struct r600_bc_vtx *vtx;
+ struct r600_bc_tex *tex;
+ struct gpr_usage usage[128];
+
+ /* assume that each gpr is exported only once */
+ struct r600_bc_cf *export_cf[128] = { NULL };
+ int32_t export_remap[128];
+
+ int32_t id, barrier[bc->nstack];
+ unsigned i, j, stack, predicate, old_stack;
+
+ memset(&usage, 0, sizeof(usage));
+ for (i = 0; i < 128; ++i) {
+ usage[i].first_write = -1;
+ usage[i].last_write[0] = -1;
+ usage[i].last_write[1] = -1;
+ usage[i].last_write[2] = -1;
+ usage[i].last_write[3] = -1;
+ }
+
+ /* first gather some informations about the gpr usage */
+ id = 0; stack = 0;
+ LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) {
+ switch (get_cf_class(cf)) {
+ case CF_CLASS_ALU:
+ predicate = 0;
+ first = NULL;
+ LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) {
+ if (!first)
+ first = alu;
+ notice_alu_src_gprs(bc, alu, usage, id);
+ if (alu->last) {
+ notice_alu_dst_gprs(first, usage, id, predicate || stack > 0);
+ first = NULL;
+ ++id;
+ }
+ if (is_alu_pred_inst(bc, alu))
+ predicate++;
+ }
+ if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3)
+ stack += predicate;
+ else if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3)
+ stack -= 1;
+ else if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3)
+ stack -= 2;
+ break;
+ case CF_CLASS_TEXTURE:
+ LIST_FOR_EACH_ENTRY(tex, &cf->tex, list) {
+ notice_tex_gprs(tex, usage, id++, stack > 0);
+ }
+ break;
+ case CF_CLASS_VERTEX:
+ LIST_FOR_EACH_ENTRY(vtx, &cf->vtx, list) {
+ notice_vtx_gprs(vtx, usage, id++, stack > 0);
+ }
+ break;
+ case CF_CLASS_EXPORT:
+ notice_export_gprs(cf, usage, export_cf, export_remap);
+ continue; // don't increment id
+ case CF_CLASS_OTHER:
+ switch (cf->inst) {
+ case V_SQ_CF_WORD1_SQ_CF_INST_JUMP:
+ case V_SQ_CF_WORD1_SQ_CF_INST_ELSE:
+ case V_SQ_CF_WORD1_SQ_CF_INST_CALL_FS:
+ break;
+
+ case V_SQ_CF_WORD1_SQ_CF_INST_POP:
+ stack -= cf->pop_count;
+ break;
+
+ default:
+ // TODO implement loop handling
+ goto out;
+ }
+ }
+ id += 0x100;
+ id &= ~0xFF;
+ }
+ assert(stack == 0);
+
+ /* try to optimize gpr usage */
+ for (i = 0; i < 124; ++i) {
+ for (j = 0; j < usage[i].nranges; ++j) {
+ struct gpr_usage_range *range = &usage[i].ranges[j];
+ int is_export = export_cf[i] && export_cf[i + 1] &&
+ range->start < export_remap[i] &&
+ export_remap[i] <= range->end;
+
+ if (range->start == -1)
+ range->replacement = -1;
+ else if (range->end == -1)
+ range->replacement = i;
+ else
+ find_replacement(usage, i, range, is_export);
+
+ if (range->replacement == -1)
+ bc->ngpr = i;
+ else if (range->replacement < i && range->replacement > bc->ngpr)
+ bc->ngpr = range->replacement;
+
+ if (is_export && range->replacement != -1) {
+ find_export_replacement(usage, range, export_cf[i],
+ export_cf[i + 1], export_remap[i + 1]);
+ }
+ }
+ }
+ bc->ngpr++;
+
+ /* apply the changes */
+ for (i = 0; i < 128; ++i) {
+ usage[i].last_write[0] = -1;
+ usage[i].last_write[1] = -1;
+ usage[i].last_write[2] = -1;
+ usage[i].last_write[3] = -1;
+ }
+ barrier[0] = 0;
+ id = 0; stack = 0;
+ LIST_FOR_EACH_ENTRY_SAFE(cf, next_cf, &bc->cf, list) {
+ old_stack = stack;
+ switch (get_cf_class(cf)) {
+ case CF_CLASS_ALU:
+ predicate = 0;
+ first = NULL;
+ cf->barrier = 0;
+ LIST_FOR_EACH_ENTRY_SAFE(alu, next_alu, &cf->alu, list) {
+ replace_alu_gprs(bc, alu, usage, id, barrier[stack], &cf->barrier);
+ if (alu->last)
+ ++id;
+
+ if (is_alu_pred_inst(bc, alu))
+ predicate++;
+
+ if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3)
+ optimize_alu_inst(bc, cf, alu);
+ }
+ if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3)
+ stack += predicate;
+ else if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3)
+ stack -= 1;
+ else if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3)
+ stack -= 2;
+ if (LIST_IS_EMPTY(&cf->alu)) {
+ r600_bc_remove_cf(bc, cf);
+ cf = NULL;
+ }
+ break;
+ case CF_CLASS_TEXTURE:
+ cf->barrier = 0;
+ LIST_FOR_EACH_ENTRY(tex, &cf->tex, list) {
+ replace_tex_gprs(tex, usage, id++, barrier[stack], &cf->barrier);
+ }
+ break;
+ case CF_CLASS_VERTEX:
+ cf->barrier = 0;
+ LIST_FOR_EACH_ENTRY(vtx, &cf->vtx, list) {
+ replace_vtx_gprs(vtx, usage, id++, barrier[stack], &cf->barrier);
+ }
+ break;
+ case CF_CLASS_EXPORT:
+ continue; // don't increment id
+ case CF_CLASS_OTHER:
+ if (cf->inst == V_SQ_CF_WORD1_SQ_CF_INST_POP) {
+ cf->barrier = 0;
+ stack -= cf->pop_count;
+ }
+ break;
+ }
+
+ id &= ~0xFF;
+ if (cf && cf->barrier)
+ barrier[old_stack] = id;
+
+ for (i = old_stack + 1; i <= stack; ++i)
+ barrier[i] = barrier[old_stack];
+
+ id += 0x100;
+ if (stack != 0) /* ensue exports are placed outside of conditional blocks */
+ continue;
+
+ for (i = 0; i < 128; ++i) {
+ if (!export_cf[i] || id < export_remap[i])
+ continue;
+
+ r600_bc_move_cf(bc, export_cf[i], next_cf);
+ replace_export_gprs(export_cf[i], usage, export_remap[i], barrier[stack]);
+ if (export_cf[i]->barrier)
+ barrier[stack] = id - 1;
+ next_cf = LIST_ENTRY(struct r600_bc_cf, export_cf[i]->list.next, list);
+ optimize_export_inst(bc, export_cf[i]);
+ export_cf[i] = NULL;
+ }
+ }
+ assert(stack == 0);
+
+out:
+ for (i = 0; i < 128; ++i) {
+ free(usage[i].ranges);
+ }
+}
+
int r600_bc_build(struct r600_bc *bc)
{
struct r600_bc_cf *cf;
struct r600_bc_alu *alu;
struct r600_bc_vtx *vtx;
struct r600_bc_tex *tex;
+ struct r600_bc_cf *exports[4] = { NULL };
uint32_t literal[4];
unsigned nliteral;
unsigned addr;
@@ -1465,37 +2339,27 @@ int r600_bc_build(struct r600_bc *bc)
bc->nstack = 1;
}
+ r600_bc_optimize(bc);
+
/* first path compute addr of each CF block */
/* addr start after all the CF instructions */
- addr = bc->cf_last->id + 2;
+ addr = LIST_ENTRY(struct r600_bc_cf, bc->cf.prev, list)->id + 2;
LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) {
- switch (cf->inst) {
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3):
+ switch (get_cf_class(cf)) {
+ case CF_CLASS_ALU:
break;
- case V_SQ_CF_WORD1_SQ_CF_INST_TEX:
- case V_SQ_CF_WORD1_SQ_CF_INST_VTX:
- case V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC:
+ case CF_CLASS_TEXTURE:
+ case CF_CLASS_VERTEX:
/* fetch node need to be 16 bytes aligned*/
addr += 3;
addr &= 0xFFFFFFFCUL;
break;
- case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
- case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
- case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
- case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
break;
- case V_SQ_CF_WORD1_SQ_CF_INST_JUMP:
- case V_SQ_CF_WORD1_SQ_CF_INST_ELSE:
- case V_SQ_CF_WORD1_SQ_CF_INST_POP:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_START_NO_AL:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_END:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK:
- case V_SQ_CF_WORD1_SQ_CF_INST_CALL_FS:
- case V_SQ_CF_WORD1_SQ_CF_INST_RETURN:
+ case CF_CLASS_EXPORT:
+ if (cf->inst == BC_INST(bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT))
+ exports[cf->output.type] = cf;
+ break;
+ case CF_CLASS_OTHER:
break;
default:
R600_ERR("unsupported CF instruction (0x%X)\n", cf->inst);
@@ -1505,6 +2369,14 @@ int r600_bc_build(struct r600_bc *bc)
addr += cf->ndw;
bc->ndw = cf->addr + cf->ndw;
}
+
+ /* set export done on last export of each type */
+ for (i = 0; i < 4; ++i) {
+ if (exports[i]) {
+ exports[i]->inst = BC_INST(bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE);
+ }
+ }
+
free(bc->bytecode);
bc->bytecode = calloc(1, bc->ndw * 4);
if (bc->bytecode == NULL)
@@ -1517,11 +2389,8 @@ int r600_bc_build(struct r600_bc *bc)
r = r600_bc_cf_build(bc, cf);
if (r)
return r;
- switch (cf->inst) {
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3):
+ switch (get_cf_class(cf)) {
+ case CF_CLASS_ALU:
nliteral = 0;
memset(literal, 0, sizeof(literal));
LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) {
@@ -1553,8 +2422,7 @@ int r600_bc_build(struct r600_bc *bc)
}
}
break;
- case V_SQ_CF_WORD1_SQ_CF_INST_VTX:
- case V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC:
+ case CF_CLASS_VERTEX:
LIST_FOR_EACH_ENTRY(vtx, &cf->vtx, list) {
r = r600_bc_vtx_build(bc, vtx, addr);
if (r)
@@ -1562,7 +2430,7 @@ int r600_bc_build(struct r600_bc *bc)
addr += 4;
}
break;
- case V_SQ_CF_WORD1_SQ_CF_INST_TEX:
+ case CF_CLASS_TEXTURE:
LIST_FOR_EACH_ENTRY(tex, &cf->tex, list) {
r = r600_bc_tex_build(bc, tex, addr);
if (r)
@@ -1570,19 +2438,8 @@ int r600_bc_build(struct r600_bc *bc)
addr += 4;
}
break;
- case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
- case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
- case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
- case EG_V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_START_NO_AL:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_END:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK:
- case V_SQ_CF_WORD1_SQ_CF_INST_JUMP:
- case V_SQ_CF_WORD1_SQ_CF_INST_ELSE:
- case V_SQ_CF_WORD1_SQ_CF_INST_POP:
- case V_SQ_CF_WORD1_SQ_CF_INST_CALL_FS:
- case V_SQ_CF_WORD1_SQ_CF_INST_RETURN:
+ case CF_CLASS_EXPORT:
+ case CF_CLASS_OTHER:
break;
default:
R600_ERR("unsupported CF instruction (0x%X)\n", cf->inst);
@@ -1658,13 +2515,10 @@ void r600_bc_dump(struct r600_bc *bc)
LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) {
id = cf->id;
- switch (cf->inst) {
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER << 3):
- case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3):
+ switch (get_cf_class(cf)) {
+ case CF_CLASS_ALU:
fprintf(stderr, "%04d %08X ALU ", id, bc->bytecode[id]);
- fprintf(stderr, "ADDR:%d ", cf->addr);
+ fprintf(stderr, "ADDR:%04d ", cf->addr);
fprintf(stderr, "KCACHE_MODE0:%X ", cf->kcache[0].mode);
fprintf(stderr, "KCACHE_BANK0:%X ", cf->kcache[0].bank);
fprintf(stderr, "KCACHE_BANK1:%X\n", cf->kcache[1].bank);
@@ -1674,22 +2528,22 @@ void r600_bc_dump(struct r600_bc *bc)
fprintf(stderr, "KCACHE_MODE1:%X ", cf->kcache[1].mode);
fprintf(stderr, "KCACHE_ADDR0:%X ", cf->kcache[0].addr);
fprintf(stderr, "KCACHE_ADDR1:%X ", cf->kcache[1].addr);
+ fprintf(stderr, "BARRIER:%d ", cf->barrier);
fprintf(stderr, "COUNT:%d\n", cf->ndw / 2);
break;
- case V_SQ_CF_WORD1_SQ_CF_INST_TEX:
- case V_SQ_CF_WORD1_SQ_CF_INST_VTX:
- case V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC:
+ case CF_CLASS_TEXTURE:
+ case CF_CLASS_VERTEX:
fprintf(stderr, "%04d %08X TEX/VTX ", id, bc->bytecode[id]);
- fprintf(stderr, "ADDR:%d\n", cf->addr);
+ fprintf(stderr, "ADDR:%04d\n", cf->addr);
id++;
fprintf(stderr, "%04d %08X TEX/VTX ", id, bc->bytecode[id]);
fprintf(stderr, "INST:%d ", cf->inst);
+ fprintf(stderr, "BARRIER:%d ", cf->barrier);
fprintf(stderr, "COUNT:%d\n", cf->ndw / 4);
break;
- case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT:
- case V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE:
+ case CF_CLASS_EXPORT:
fprintf(stderr, "%04d %08X EXPORT ", id, bc->bytecode[id]);
- fprintf(stderr, "GPR:%X ", cf->output.gpr);
+ fprintf(stderr, "GPR:%d ", cf->output.gpr);
fprintf(stderr, "ELEM_SIZE:%X ", cf->output.elem_size);
fprintf(stderr, "ARRAY_BASE:%X ", cf->output.array_base);
fprintf(stderr, "TYPE:%X\n", cf->output.type);
@@ -1700,25 +2554,18 @@ void r600_bc_dump(struct r600_bc *bc)
fprintf(stderr, "SWIZ_Z:%X ", cf->output.swizzle_z);
fprintf(stderr, "SWIZ_W:%X ", cf->output.swizzle_w);
fprintf(stderr, "SWIZ_W:%X ", cf->output.swizzle_w);
- fprintf(stderr, "BARRIER:%X ", cf->output.barrier);
- fprintf(stderr, "INST:%d ", cf->output.inst);
- fprintf(stderr, "EOP:%X\n", cf->output.end_of_program);
+ fprintf(stderr, "BARRIER:%d ", cf->barrier);
+ fprintf(stderr, "INST:%d ", cf->inst);
+ fprintf(stderr, "BURST_COUNT:%d\n", cf->output.burst_count);
break;
- case V_SQ_CF_WORD1_SQ_CF_INST_JUMP:
- case V_SQ_CF_WORD1_SQ_CF_INST_ELSE:
- case V_SQ_CF_WORD1_SQ_CF_INST_POP:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_START_NO_AL:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_END:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE:
- case V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK:
- case V_SQ_CF_WORD1_SQ_CF_INST_CALL_FS:
- case V_SQ_CF_WORD1_SQ_CF_INST_RETURN:
+ case CF_CLASS_OTHER:
fprintf(stderr, "%04d %08X CF ", id, bc->bytecode[id]);
- fprintf(stderr, "ADDR:%d\n", cf->cf_addr);
+ fprintf(stderr, "ADDR:%04d\n", cf->cf_addr);
id++;
fprintf(stderr, "%04d %08X CF ", id, bc->bytecode[id]);
fprintf(stderr, "INST:%d ", cf->inst);
fprintf(stderr, "COND:%X ", cf->cond);
+ fprintf(stderr, "BARRIER:%d ", cf->barrier);
fprintf(stderr, "POP_COUNT:%X\n", cf->pop_count);
break;
}
@@ -1791,21 +2638,21 @@ void r600_cf_vtx(struct r600_vertex_element *ve, u32 *bytecode, unsigned count)
if (count > 8) {
bytecode[i++] = S_SQ_CF_WORD0_ADDR(8 >> 1);
bytecode[i++] = S_SQ_CF_WORD1_CF_INST(V_SQ_CF_WORD1_SQ_CF_INST_VTX) |
- S_SQ_CF_WORD1_BARRIER(1) |
+ S_SQ_CF_WORD1_BARRIER(0) |
S_SQ_CF_WORD1_COUNT(8 - 1);
bytecode[i++] = S_SQ_CF_WORD0_ADDR(40 >> 1);
bytecode[i++] = S_SQ_CF_WORD1_CF_INST(V_SQ_CF_WORD1_SQ_CF_INST_VTX) |
- S_SQ_CF_WORD1_BARRIER(1) |
+ S_SQ_CF_WORD1_BARRIER(0) |
S_SQ_CF_WORD1_COUNT(count - 8 - 1);
} else {
bytecode[i++] = S_SQ_CF_WORD0_ADDR(8 >> 1);
bytecode[i++] = S_SQ_CF_WORD1_CF_INST(V_SQ_CF_WORD1_SQ_CF_INST_VTX) |
- S_SQ_CF_WORD1_BARRIER(1) |
+ S_SQ_CF_WORD1_BARRIER(0) |
S_SQ_CF_WORD1_COUNT(count - 1);
}
bytecode[i++] = S_SQ_CF_WORD0_ADDR(0);
bytecode[i++] = S_SQ_CF_WORD1_CF_INST(V_SQ_CF_WORD1_SQ_CF_INST_RETURN) |
- S_SQ_CF_WORD1_BARRIER(1);
+ S_SQ_CF_WORD1_BARRIER(0);
rstate = &ve->rstate;
rstate->id = R600_PIPE_STATE_FETCH_SHADER;
@@ -1827,21 +2674,21 @@ void r600_cf_vtx_tc(struct r600_vertex_element *ve, u32 *bytecode, unsigned coun
if (count > 8) {
bytecode[i++] = S_SQ_CF_WORD0_ADDR(8 >> 1);
bytecode[i++] = S_SQ_CF_WORD1_CF_INST(V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC) |
- S_SQ_CF_WORD1_BARRIER(1) |
+ S_SQ_CF_WORD1_BARRIER(0) |
S_SQ_CF_WORD1_COUNT(8 - 1);
bytecode[i++] = S_SQ_CF_WORD0_ADDR(40 >> 1);
bytecode[i++] = S_SQ_CF_WORD1_CF_INST(V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC) |
- S_SQ_CF_WORD1_BARRIER(1) |
+ S_SQ_CF_WORD1_BARRIER(0) |
S_SQ_CF_WORD1_COUNT((count - 8) - 1);
} else {
bytecode[i++] = S_SQ_CF_WORD0_ADDR(8 >> 1);
bytecode[i++] = S_SQ_CF_WORD1_CF_INST(V_SQ_CF_WORD1_SQ_CF_INST_VTX_TC) |
- S_SQ_CF_WORD1_BARRIER(1) |
+ S_SQ_CF_WORD1_BARRIER(0) |
S_SQ_CF_WORD1_COUNT(count - 1);
}
bytecode[i++] = S_SQ_CF_WORD0_ADDR(0);
bytecode[i++] = S_SQ_CF_WORD1_CF_INST(V_SQ_CF_WORD1_SQ_CF_INST_RETURN) |
- S_SQ_CF_WORD1_BARRIER(1);
+ S_SQ_CF_WORD1_BARRIER(0);
rstate = &ve->rstate;
rstate->id = R600_PIPE_STATE_FETCH_SHADER;
diff --git a/src/gallium/drivers/r600/r600_asm.h b/src/gallium/drivers/r600/r600_asm.h
index 278b4466cb0..519245f3af2 100644
--- a/src/gallium/drivers/r600/r600_asm.h
+++ b/src/gallium/drivers/r600/r600_asm.h
@@ -108,15 +108,13 @@ struct r600_bc_vtx {
struct r600_bc_output {
unsigned array_base;
unsigned type;
- unsigned end_of_program;
- unsigned inst;
unsigned elem_size;
unsigned gpr;
unsigned swizzle_x;
unsigned swizzle_y;
unsigned swizzle_z;
unsigned swizzle_w;
- unsigned barrier;
+ unsigned burst_count;
};
struct r600_bc_kcache {
@@ -134,6 +132,7 @@ struct r600_bc_cf {
unsigned cond;
unsigned pop_count;
unsigned cf_addr; /* control flow addr */
+ unsigned barrier;
struct r600_bc_kcache kcache[2];
unsigned r6xx_uses_waterfall;
struct list_head alu;
diff --git a/src/gallium/drivers/r600/r600_pipe.c b/src/gallium/drivers/r600/r600_pipe.c
index 20838e4d98f..ad14dbe14f4 100644
--- a/src/gallium/drivers/r600/r600_pipe.c
+++ b/src/gallium/drivers/r600/r600_pipe.c
@@ -42,6 +42,7 @@
#include "r600_shader.h"
#include "r600_pipe.h"
#include "r600_state_inlines.h"
+#include "r600_video_context.h"
/*
* pipe_context
@@ -483,6 +484,7 @@ struct pipe_screen *r600_screen_create(struct radeon *radeon)
rscreen->screen.get_paramf = r600_get_paramf;
rscreen->screen.is_format_supported = r600_is_format_supported;
rscreen->screen.context_create = r600_create_context;
+ rscreen->screen.video_context_create = r600_video_create;
r600_init_screen_resource_functions(&rscreen->screen);
rscreen->tiling_info = r600_get_tiling_info(radeon);
diff --git a/src/gallium/drivers/r600/r600_shader.c b/src/gallium/drivers/r600/r600_shader.c
index 5d952a0cae7..106852c1082 100644
--- a/src/gallium/drivers/r600/r600_shader.c
+++ b/src/gallium/drivers/r600/r600_shader.c
@@ -497,7 +497,7 @@ int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *s
struct tgsi_full_immediate *immediate;
struct r600_shader_ctx ctx;
struct r600_bc_output output[32];
- unsigned output_done, noutput;
+ unsigned noutput;
unsigned opcode;
int i, r = 0, pos0;
@@ -619,10 +619,8 @@ int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *s
output[i].swizzle_y = 1;
output[i].swizzle_z = 2;
output[i].swizzle_w = 3;
- output[i].barrier = 1;
output[i].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PARAM;
output[i].array_base = i - pos0;
- output[i].inst = BC_INST(ctx.bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT);
switch (ctx.type) {
case TGSI_PROCESSOR_VERTEX:
if (shader->output[i].name == TGSI_SEMANTIC_POSITION) {
@@ -682,10 +680,8 @@ int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *s
output[i].swizzle_y = 1;
output[i].swizzle_z = 2;
output[i].swizzle_w = 3;
- output[i].barrier = 1;
output[i].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PARAM;
output[i].array_base = 0;
- output[i].inst = BC_INST(ctx.bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT);
noutput++;
}
}
@@ -698,22 +694,10 @@ int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *s
output[0].swizzle_y = 7;
output[0].swizzle_z = 7;
output[0].swizzle_w = 7;
- output[0].barrier = 1;
output[0].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL;
output[0].array_base = 0;
- output[0].inst = BC_INST(ctx.bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT);
noutput++;
}
- /* set export done on last export of each type */
- for (i = noutput - 1, output_done = 0; i >= 0; i--) {
- if (i == (noutput - 1)) {
- output[i].end_of_program = 1;
- }
- if (!(output_done & (1 << output[i].type))) {
- output_done |= (1 << output[i].type);
- output[i].inst = BC_INST(ctx.bc, V_SQ_CF_ALLOC_EXPORT_WORD1_SQ_CF_INST_EXPORT_DONE);
- }
- }
/* add output to bytecode */
for (i = 0; i < noutput; i++) {
r = r600_bc_add_output(ctx.bc, &output[i]);
@@ -747,7 +731,7 @@ static int tgsi_src(struct r600_shader_ctx *ctx,
memset(r600_src, 0, sizeof(struct r600_bc_alu_src));
r600_src->neg = tgsi_src->Register.Negate;
r600_src->abs = tgsi_src->Register.Absolute;
- if (tgsi_src->Register.File == TGSI_FILE_IMMEDIATE) {
+ if (tgsi_src->Register.File == TGSI_FILE_IMMEDIATE) {
int index;
if((tgsi_src->Register.SwizzleX == tgsi_src->Register.SwizzleY) &&
(tgsi_src->Register.SwizzleX == tgsi_src->Register.SwizzleZ) &&
diff --git a/src/gallium/drivers/r600/r600_state.c b/src/gallium/drivers/r600/r600_state.c
index d678f42ad67..de2668cee16 100644
--- a/src/gallium/drivers/r600/r600_state.c
+++ b/src/gallium/drivers/r600/r600_state.c
@@ -925,6 +925,17 @@ static void r600_cb(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
desc = util_format_description(rtex->resource.base.b.format);
if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
ntype = V_0280A0_NUMBER_SRGB;
+ else if (desc->layout == UTIL_FORMAT_LAYOUT_PLAIN) {
+ switch(desc->channel[0].type) {
+ case UTIL_FORMAT_TYPE_UNSIGNED:
+ ntype = V_0280A0_NUMBER_UNORM;
+ break;
+
+ case UTIL_FORMAT_TYPE_SIGNED:
+ ntype = V_0280A0_NUMBER_SNORM;
+ break;
+ }
+ }
format = r600_translate_colorformat(rtex->resource.base.b.format);
swap = r600_translate_colorswap(rtex->resource.base.b.format);
diff --git a/src/gallium/drivers/r600/r600_state_inlines.h b/src/gallium/drivers/r600/r600_state_inlines.h
index d5dabdc69b6..a0ec493fc85 100644
--- a/src/gallium/drivers/r600/r600_state_inlines.h
+++ b/src/gallium/drivers/r600/r600_state_inlines.h
@@ -310,6 +310,7 @@ static inline uint32_t r600_translate_colorswap(enum pipe_format format)
return V_0280A0_SWAP_STD;
case PIPE_FORMAT_R16_UNORM:
+ case PIPE_FORMAT_R16_SNORM:
return V_0280A0_SWAP_STD;
/* 32-bit buffers. */
@@ -360,7 +361,7 @@ static inline uint32_t r600_translate_colorswap(enum pipe_format format)
// return FMT_16_16_16_16_FLOAT;
/* 128-bit buffers. */
- case PIPE_FORMAT_R32G32B32A32_FLOAT:
+ //case PIPE_FORMAT_R32G32B32A32_FLOAT:
// return FMT_32_32_32_32_FLOAT;
return 0;
default:
@@ -401,6 +402,7 @@ static INLINE uint32_t r600_translate_colorformat(enum pipe_format format)
return V_0280A0_COLOR_8_8;
case PIPE_FORMAT_R16_UNORM:
+ case PIPE_FORMAT_R16_SNORM:
return V_0280A0_COLOR_16;
/* 32-bit buffers. */
@@ -433,8 +435,8 @@ static INLINE uint32_t r600_translate_colorformat(enum pipe_format format)
case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
return V_0280A0_COLOR_24_8;
- case PIPE_FORMAT_R32_FLOAT:
- return V_0280A0_COLOR_32_FLOAT;
+ //case PIPE_FORMAT_R32_FLOAT:
+ // return V_0280A0_COLOR_32_FLOAT;
case PIPE_FORMAT_R16G16_FLOAT:
return V_0280A0_COLOR_16_16_FLOAT;
@@ -465,10 +467,10 @@ static INLINE uint32_t r600_translate_colorformat(enum pipe_format format)
return V_0280A0_COLOR_32_32;
/* 128-bit buffers. */
- case PIPE_FORMAT_R32G32B32_FLOAT:
- return V_0280A0_COLOR_32_32_32_FLOAT;
- case PIPE_FORMAT_R32G32B32A32_FLOAT:
- return V_0280A0_COLOR_32_32_32_32_FLOAT;
+ //case PIPE_FORMAT_R32G32B32_FLOAT:
+ // return V_0280A0_COLOR_32_32_32_FLOAT;
+ //case PIPE_FORMAT_R32G32B32A32_FLOAT:
+ // return V_0280A0_COLOR_32_32_32_32_FLOAT;
/* YUV buffers. */
case PIPE_FORMAT_UYVY:
diff --git a/src/gallium/drivers/r600/r600_texture.c b/src/gallium/drivers/r600/r600_texture.c
index e2745624575..1f4f453c091 100644
--- a/src/gallium/drivers/r600/r600_texture.c
+++ b/src/gallium/drivers/r600/r600_texture.c
@@ -321,6 +321,48 @@ r600_texture_create_object(struct pipe_screen *screen,
return rtex;
}
+/* Figure out whether u_blitter will fallback to a transfer operation.
+ * If so, don't use a staging resource.
+ */
+static boolean permit_hardware_blit(struct pipe_screen *screen,
+ const struct pipe_resource *res)
+{
+ unsigned bind;
+
+ if (util_format_is_depth_or_stencil(res->format))
+ bind = PIPE_BIND_DEPTH_STENCIL;
+ else
+ bind = PIPE_BIND_RENDER_TARGET;
+
+ /* See r600_resource_copy_region: there is something wrong
+ * with depth resource copies at the moment so avoid them for
+ * now.
+ */
+ if (util_format_get_component_bits(res->format,
+ UTIL_FORMAT_COLORSPACE_ZS,
+ 0) != 0)
+ return FALSE;
+
+ if (!screen->is_format_supported(screen,
+ res->format,
+ res->target,
+ res->nr_samples,
+ bind, 0))
+ return FALSE;
+
+ if (!screen->is_format_supported(screen,
+ res->format,
+ res->target,
+ res->nr_samples,
+ PIPE_BIND_SAMPLER_VIEW, 0))
+ return FALSE;
+
+ if (res->usage == PIPE_USAGE_STREAM)
+ return FALSE;
+
+ return TRUE;
+}
+
struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
const struct pipe_resource *templ)
{
@@ -332,7 +374,7 @@ struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
if (force_tiling == -1)
force_tiling = debug_get_bool_option("R600_FORCE_TILING", FALSE);
- if (force_tiling) {
+ if (force_tiling && permit_hardware_blit(screen, templ)) {
if (!(templ->flags & R600_RESOURCE_FLAG_TRANSFER) &&
!(templ->bind & PIPE_BIND_SCANOUT)) {
array_mode = V_038000_ARRAY_2D_TILED_THIN1;
@@ -485,46 +527,6 @@ static INLINE unsigned u_box_volume( const struct pipe_box *box )
return box->width * box->depth * box->height;
};
-
-/* Figure out whether u_blitter will fallback to a transfer operation.
- * If so, don't use a staging resource.
- */
-static boolean permit_hardware_blit(struct pipe_screen *screen,
- struct pipe_resource *res)
-{
- unsigned bind;
-
- if (util_format_is_depth_or_stencil(res->format))
- bind = PIPE_BIND_DEPTH_STENCIL;
- else
- bind = PIPE_BIND_RENDER_TARGET;
-
- /* See r600_resource_copy_region: there is something wrong
- * with depth resource copies at the moment so avoid them for
- * now.
- */
- if (util_format_get_component_bits(res->format,
- UTIL_FORMAT_COLORSPACE_ZS,
- 0) != 0)
- return FALSE;
-
- if (!screen->is_format_supported(screen,
- res->format,
- res->target,
- res->nr_samples,
- bind, 0))
- return FALSE;
-
- if (!screen->is_format_supported(screen,
- res->format,
- res->target,
- res->nr_samples,
- PIPE_BIND_SAMPLER_VIEW, 0))
- return FALSE;
-
- return TRUE;
-}
-
struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
struct pipe_resource *texture,
unsigned level,
@@ -794,7 +796,7 @@ static unsigned r600_get_swizzle_combined(const unsigned char *swizzle_format,
/* texture format translate */
uint32_t r600_translate_texformat(enum pipe_format format,
- const unsigned char *swizzle_view,
+ const unsigned char *swizzle_view,
uint32_t *word4_p, uint32_t *yuv_format_p)
{
uint32_t result = 0, word4 = 0, yuv_format = 0;
@@ -848,7 +850,7 @@ uint32_t r600_translate_texformat(enum pipe_format format,
break;
}
goto out_unknown; /* TODO */
-
+
case UTIL_FORMAT_COLORSPACE_SRGB:
word4 |= S_038010_FORCE_DEGAMMA(1);
if (format == PIPE_FORMAT_L8A8_SRGB || format == PIPE_FORMAT_L8_SRGB)
@@ -864,7 +866,7 @@ uint32_t r600_translate_texformat(enum pipe_format format,
static int r600_enable_s3tc = -1;
if (r600_enable_s3tc == -1)
- r600_enable_s3tc =
+ r600_enable_s3tc =
debug_get_bool_option("R600_ENABLE_S3TC", FALSE);
if (!r600_enable_s3tc)
@@ -887,7 +889,7 @@ uint32_t r600_translate_texformat(enum pipe_format format,
}
- for (i = 0; i < desc->nr_channels; i++) {
+ for (i = 0; i < desc->nr_channels; i++) {
if (desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) {
word4 |= sign_bit[i];
}
@@ -901,7 +903,7 @@ uint32_t r600_translate_texformat(enum pipe_format format,
for (i = 1; i < desc->nr_channels; i++) {
uniform = uniform && desc->channel[0].size == desc->channel[i].size;
}
-
+
/* Non-uniform formats. */
if (!uniform) {
switch(desc->nr_channels) {
@@ -1019,7 +1021,7 @@ uint32_t r600_translate_texformat(enum pipe_format format,
goto out_word4;
}
}
-
+
}
out_word4:
if (word4_p)
diff --git a/src/gallium/drivers/r600/r600_video_context.c b/src/gallium/drivers/r600/r600_video_context.c
new file mode 100644
index 00000000000..b3885db0f55
--- /dev/null
+++ b/src/gallium/drivers/r600/r600_video_context.c
@@ -0,0 +1,21 @@
+#include "r600_video_context.h"
+#include <softpipe/sp_video_context.h>
+
+struct pipe_video_context *
+r600_video_create(struct pipe_screen *screen, enum pipe_video_profile profile,
+ enum pipe_video_chroma_format chroma_format,
+ unsigned width, unsigned height, void *priv)
+{
+ struct pipe_context *pipe;
+
+ assert(screen);
+
+ pipe = screen->context_create(screen, priv);
+ if (!pipe)
+ return NULL;
+
+ return sp_video_create_ex(pipe, profile, chroma_format, width, height,
+ VL_MPEG12_MC_RENDERER_BUFFER_PICTURE,
+ true,
+ PIPE_FORMAT_VUYX);
+}
diff --git a/src/gallium/drivers/r600/r600_video_context.h b/src/gallium/drivers/r600/r600_video_context.h
new file mode 100644
index 00000000000..bda33a00d44
--- /dev/null
+++ b/src/gallium/drivers/r600/r600_video_context.h
@@ -0,0 +1,11 @@
+#ifndef __R600_VIDEO_CONTEXT_H__
+#define __R600_VIDEO_CONTEXT_H__
+
+#include <pipe/p_video_context.h>
+
+struct pipe_video_context *
+r600_video_create(struct pipe_screen *screen, enum pipe_video_profile profile,
+ enum pipe_video_chroma_format chroma_format,
+ unsigned width, unsigned height, void *priv);
+
+#endif
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-18 16:28:13 +0000