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.c2007
-rw-r--r--src/gallium/drivers/r600/r600_asm.h19
-rw-r--r--src/gallium/drivers/r600/r600_pipe.c2
-rw-r--r--src/gallium/drivers/r600/r600_shader.c388
-rw-r--r--src/gallium/drivers/r600/r600_shader.h2
-rw-r--r--src/gallium/drivers/r600/r600_sq.h2
-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.c3
-rw-r--r--src/gallium/drivers/r600/r600_video_context.c21
-rw-r--r--src/gallium/drivers/r600/r600_video_context.h11
-rw-r--r--src/gallium/drivers/r600/r700_asm.c11
14 files changed, 1909 insertions, 612 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 c44506c7eba..1881e633d54 100644
--- a/src/gallium/drivers/r600/eg_asm.c
+++ b/src/gallium/drivers/r600/eg_asm.c
@@ -32,10 +32,12 @@
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_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) |
@@ -44,15 +46,16 @@ 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:
@@ -60,13 +63,14 @@ 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:
@@ -79,9 +83,10 @@ 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 326724520b3..b15758adc33 100644
--- a/src/gallium/drivers/r600/r600_asm.c
+++ b/src/gallium/drivers/r600/r600_asm.c
@@ -32,6 +32,12 @@
#include "r600_formats.h"
#include "r600d.h"
+#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_alu *alu)
{
if(alu->is_op3)
@@ -45,10 +51,10 @@ static inline unsigned int r600_bc_get_num_operands(struct r600_bc_alu *alu)
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGT:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLGE:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_KILLNE:
- case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL:
+ case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MIN:
- case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETE:
+ case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETE:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETNE:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGT:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SETGE:
@@ -61,7 +67,7 @@ static inline unsigned int r600_bc_get_num_operands(struct r600_bc_alu *alu)
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_CUBE:
return 2;
- case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV:
+ case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_FLOOR:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FRACT:
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR:
@@ -76,7 +82,7 @@ static inline unsigned int r600_bc_get_num_operands(struct r600_bc_alu *alu)
case V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_COS:
return 1;
default: R600_ERR(
- "Need instruction operand number for 0x%x.\n", alu->inst);
+ "Need instruction operand number for 0x%x.\n", alu->inst);
};
return 3;
@@ -94,6 +100,7 @@ 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;
}
@@ -104,7 +111,6 @@ static struct r600_bc_alu *r600_bc_alu(void)
if (alu == NULL)
return NULL;
LIST_INITHEAD(&alu->list);
- LIST_INITHEAD(&alu->bs_list);
return alu;
}
@@ -183,6 +189,44 @@ 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;
@@ -190,227 +234,620 @@ 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;
}
-const unsigned bank_swizzle_vec[8] = {SQ_ALU_VEC_210, //000
- SQ_ALU_VEC_120, //001
- SQ_ALU_VEC_102, //010
+/* alu predicate instructions */
+static int is_alu_pred_inst(struct r600_bc_alu *alu)
+{
+ return !alu->is_op3 && (
+ 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 ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_INV ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_POP ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_CLR ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SET_RESTORE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_PUSH ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_PUSH ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_PUSH ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_PUSH ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETE_PUSH_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGT_PUSH_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETGE_PUSH_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETNE_PUSH_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETLT_PUSH_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_PRED_SETLE_PUSH_INT);
+}
- SQ_ALU_VEC_201, //011
- SQ_ALU_VEC_012, //100
- SQ_ALU_VEC_021, //101
+/* alu kill instructions */
+static int is_alu_kill_inst(struct r600_bc_alu *alu)
+{
+ 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);
+}
- SQ_ALU_VEC_012, //110
- SQ_ALU_VEC_012}; //111
+/* alu instructions that can ony exits once per group */
+static int is_alu_once_inst(struct r600_bc_alu *alu)
+{
+ return is_alu_kill_inst(alu) ||
+ is_alu_pred_inst(alu);
+}
-const unsigned bank_swizzle_scl[8] = {SQ_ALU_SCL_210, //000
- SQ_ALU_SCL_122, //001
- SQ_ALU_SCL_122, //010
+static int is_alu_reduction_inst(struct r600_bc_alu *alu)
+{
+ return !alu->is_op3 && (
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_CUBE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4 ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4_IEEE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MAX4);
+}
+
+static int is_alu_mova_inst(struct r600_bc_alu *alu)
+{
+ return !alu->is_op3 && (
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_FLOOR ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT);
+}
+
+/* alu instructions that can only execute on the vector unit */
+static int is_alu_vec_unit_inst(struct r600_bc_alu *alu)
+{
+ return is_alu_reduction_inst(alu) ||
+ is_alu_mova_inst(alu);
+}
+
+/* alu instructions that can only execute on the trans unit */
+static int is_alu_trans_unit_inst(struct r600_bc_alu *alu)
+{
+ if(!alu->is_op3)
+ return alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ASHR_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_INT_TO_FLT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHL_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHR_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULHI_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULHI_UINT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MULLO_UINT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_INT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_UINT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_UINT_TO_FLT ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_COS ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_CLAMPED ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LOG_IEEE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_CLAMPED ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_FF ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIPSQRT_CLAMPED ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIPSQRT_FF ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIPSQRT_IEEE ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SIN ||
+ alu->inst == V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SQRT_IEEE;
+ else
+ return alu->inst == V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MUL_LIT ||
+ alu->inst == V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MUL_LIT_D2 ||
+ alu->inst == V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MUL_LIT_M2 ||
+ alu->inst == V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MUL_LIT_M4;
+}
+
+/* alu instructions that can execute on any unit */
+static int is_alu_any_unit_inst(struct r600_bc_alu *alu)
+{
+ return !is_alu_vec_unit_inst(alu) &&
+ !is_alu_trans_unit_inst(alu);
+}
+
+static int assign_alu_units(struct r600_bc_alu *alu_first, struct r600_bc_alu *assignment[5])
+{
+ struct r600_bc_alu *alu;
+ unsigned i, chan, trans;
+
+ for (i = 0; i < 5; i++)
+ assignment[i] = NULL;
+
+ for (alu = alu_first; alu; alu = NEXT_ALU(alu)) {
+ chan = alu->dst.chan;
+ if (is_alu_trans_unit_inst(alu))
+ trans = 1;
+ else if (is_alu_vec_unit_inst(alu))
+ trans = 0;
+ else if (assignment[chan])
+ trans = 1; // assume ALU_INST_PREFER_VECTOR
+ else
+ trans = 0;
- SQ_ALU_SCL_221, //011
- SQ_ALU_SCL_212, //100
- SQ_ALU_SCL_122, //101
+ if (trans) {
+ if (assignment[4]) {
+ assert(0); //ALU.Trans has already been allocated
+ return -1;
+ }
+ assignment[4] = alu;
+ } else {
+ if (assignment[chan]) {
+ assert(0); //ALU.chan has already been allocated
+ return -1;
+ }
+ assignment[chan] = alu;
+ }
- SQ_ALU_SCL_122, //110
- SQ_ALU_SCL_122}; //111
+ if (alu->last)
+ break;
+ }
+ return 0;
+}
-static int init_gpr(struct r600_bc_alu *alu)
+struct alu_bank_swizzle {
+ int hw_gpr[NUM_OF_CYCLES][NUM_OF_COMPONENTS];
+ int hw_cfile_addr[4];
+ int hw_cfile_elem[4];
+};
+
+const unsigned cycle_for_bank_swizzle_vec[][3] = {
+ [SQ_ALU_VEC_012] = { 0, 1, 2 },
+ [SQ_ALU_VEC_021] = { 0, 2, 1 },
+ [SQ_ALU_VEC_120] = { 1, 2, 0 },
+ [SQ_ALU_VEC_102] = { 1, 0, 2 },
+ [SQ_ALU_VEC_201] = { 2, 0, 1 },
+ [SQ_ALU_VEC_210] = { 2, 1, 0 }
+};
+
+const unsigned cycle_for_bank_swizzle_scl[][3] = {
+ [SQ_ALU_SCL_210] = { 2, 1, 0 },
+ [SQ_ALU_SCL_122] = { 1, 2, 2 },
+ [SQ_ALU_SCL_212] = { 2, 1, 2 },
+ [SQ_ALU_SCL_221] = { 2, 2, 1 }
+};
+
+static void init_bank_swizzle(struct alu_bank_swizzle *bs)
{
- int cycle, component;
+ int i, cycle, component;
/* set up gpr use */
for (cycle = 0; cycle < NUM_OF_CYCLES; cycle++)
for (component = 0; component < NUM_OF_COMPONENTS; component++)
- alu->hw_gpr[cycle][component] = -1;
- return 0;
+ bs->hw_gpr[cycle][component] = -1;
+ for (i = 0; i < 4; i++)
+ bs->hw_cfile_addr[i] = -1;
+ for (i = 0; i < 4; i++)
+ bs->hw_cfile_elem[i] = -1;
}
-#if 0
-static int reserve_gpr(struct r600_bc_alu *alu, unsigned sel, unsigned chan, unsigned cycle)
+static int reserve_gpr(struct alu_bank_swizzle *bs, unsigned sel, unsigned chan, unsigned cycle)
{
- if (alu->hw_gpr[cycle][chan] < 0)
- alu->hw_gpr[cycle][chan] = sel;
- else if (alu->hw_gpr[cycle][chan] != (int)sel) {
- R600_ERR("Another scalar operation has already used GPR read port for channel\n");
+ if (bs->hw_gpr[cycle][chan] == -1)
+ bs->hw_gpr[cycle][chan] = sel;
+ else if (bs->hw_gpr[cycle][chan] != (int)sel) {
+ // Another scalar operation has already used GPR read port for channel
return -1;
}
return 0;
}
-static int cycle_for_scalar_bank_swizzle(const int swiz, const int sel, unsigned *p_cycle)
-{
- int table[3];
- int ret = 0;
- switch (swiz) {
- case SQ_ALU_SCL_210:
- table[0] = 2; table[1] = 1; table[2] = 0;
- *p_cycle = table[sel];
- break;
- case SQ_ALU_SCL_122:
- table[0] = 1; table[1] = 2; table[2] = 2;
- *p_cycle = table[sel];
- break;
- case SQ_ALU_SCL_212:
- table[0] = 2; table[1] = 1; table[2] = 2;
- *p_cycle = table[sel];
- break;
- case SQ_ALU_SCL_221:
- table[0] = 2; table[1] = 2; table[2] = 1;
- *p_cycle = table[sel];
- break;
- break;
- default:
- R600_ERR("bad scalar bank swizzle value\n");
- ret = -1;
- break;
+static int reserve_cfile(struct alu_bank_swizzle *bs, unsigned sel, unsigned chan)
+{
+ int res, resmatch = -1, resempty = -1;
+ for (res = 3; res >= 0; --res) {
+ if (bs->hw_cfile_addr[res] == -1)
+ resempty = res;
+ else if (bs->hw_cfile_addr[res] == sel &&
+ bs->hw_cfile_elem[res] == chan)
+ resmatch = res;
}
- return ret;
-}
-
-static int cycle_for_vector_bank_swizzle(const int swiz, const int sel, unsigned *p_cycle)
-{
- int table[3];
- int ret;
-
- switch (swiz) {
- case SQ_ALU_VEC_012:
- table[0] = 0; table[1] = 1; table[2] = 2;
- *p_cycle = table[sel];
- break;
- case SQ_ALU_VEC_021:
- table[0] = 0; table[1] = 2; table[2] = 1;
- *p_cycle = table[sel];
- break;
- case SQ_ALU_VEC_120:
- table[0] = 1; table[1] = 2; table[2] = 0;
- *p_cycle = table[sel];
- break;
- case SQ_ALU_VEC_102:
- table[0] = 1; table[1] = 0; table[2] = 2;
- *p_cycle = table[sel];
- break;
- case SQ_ALU_VEC_201:
- table[0] = 2; table[1] = 0; table[2] = 1;
- *p_cycle = table[sel];
- break;
- case SQ_ALU_VEC_210:
- table[0] = 2; table[1] = 1; table[2] = 0;
- *p_cycle = table[sel];
- break;
- default:
- R600_ERR("bad vector bank swizzle value\n");
- ret = -1;
- break;
+ if (resmatch != -1)
+ return 0; // Read for this scalar element already reserved, nothing to do here.
+ else if (resempty != -1) {
+ bs->hw_cfile_addr[resempty] = sel;
+ bs->hw_cfile_elem[resempty] = chan;
+ } else {
+ // All cfile read ports are used, cannot reference vector element
+ return -1;
}
- return ret;
+ return 0;
}
+static int is_gpr(unsigned sel)
+{
+ return (sel >= 0 && sel <= 127);
+}
+static int is_cfile(unsigned sel)
+{
+ return (sel > 255 && sel < 512);
+}
-static void update_chan_counter(struct r600_bc_alu *alu, int *chan_counter)
+/* CB constants start at 512, and get translated to a kcache index when ALU
+ * clauses are constructed. Note that we handle kcache constants the same way
+ * as (the now gone) cfile constants, is that really required? */
+static int is_cb_const(int sel)
{
- int num_src;
- int i;
- int channel_swizzle;
+ if (sel > 511 && sel < 4607)
+ return 1;
+ return 0;
+}
- num_src = r600_bc_get_num_operands(alu);
+static int is_const(int sel)
+{
+ return is_cfile(sel) ||
+ is_cb_const(sel) ||
+ (sel >= V_SQ_ALU_SRC_0 &&
+ sel <= V_SQ_ALU_SRC_LITERAL);
+}
+
+static int check_vector(struct r600_bc_alu *alu, struct alu_bank_swizzle *bs, int bank_swizzle)
+{
+ int r, src, num_src, sel, elem, cycle;
- for (i = 0; i < num_src; i++) {
- channel_swizzle = alu->src[i].chan;
- if ((alu->src[i].sel > 0 && alu->src[i].sel < 128) && channel_swizzle <= 3)
- chan_counter[channel_swizzle]++;
+ num_src = r600_bc_get_num_operands(alu);
+ for (src = 0; src < num_src; src++) {
+ sel = alu->src[src].sel;
+ elem = alu->src[src].chan;
+ 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,
+ // second source uses first source’s reservation
+ continue;
+ else {
+ r = reserve_gpr(bs, sel, elem, cycle);
+ if (r)
+ return r;
+ }
+ } else if (is_cfile(sel)) {
+ r = reserve_cfile(bs, sel, elem);
+ if (r)
+ return r;
+ }
+ // No restrictions on PV, PS, literal or special constants
}
+ return 0;
}
-/* we need something like this I think - but this is bogus */
-int check_read_slots(struct r600_bc *bc, struct r600_bc_alu *alu_first)
+static int check_scalar(struct r600_bc_alu *alu, struct alu_bank_swizzle *bs, int bank_swizzle)
{
- struct r600_bc_alu *alu;
- int chan_counter[4] = { 0 };
-
- update_chan_counter(alu_first, chan_counter);
+ int r, src, num_src, const_count, sel, elem, cycle;
- LIST_FOR_EACH_ENTRY(alu, &alu_first->bs_list, bs_list) {
- update_chan_counter(alu, chan_counter);
+ num_src = r600_bc_get_num_operands(alu);
+ for (const_count = 0, src = 0; src < num_src; ++src) {
+ sel = alu->src[src].sel;
+ elem = alu->src[src].chan;
+ if (is_const(sel)) { // Any constant, including literal and inline constants
+ if (const_count >= 2)
+ // More than two references to a constant in
+ // transcendental operation.
+ return -1;
+ else
+ const_count++;
+ }
+ if (is_cfile(sel)) {
+ r = reserve_cfile(bs, sel, elem);
+ if (r)
+ return r;
+ }
}
-
- if (chan_counter[0] > 3 ||
- chan_counter[1] > 3 ||
- chan_counter[2] > 3 ||
- chan_counter[3] > 3) {
- R600_ERR("needed to split instruction for input ran out of banks %x %d %d %d %d\n",
- alu_first->inst, chan_counter[0], chan_counter[1], chan_counter[2], chan_counter[3]);
- return -1;
+ for (src = 0; src < num_src; ++src) {
+ sel = alu->src[src].sel;
+ elem = alu->src[src].chan;
+ if (is_gpr(sel)) {
+ cycle = cycle_for_bank_swizzle_scl[bank_swizzle][src];
+ if (cycle < const_count)
+ // Cycle for GPR load conflicts with
+ // constant load in transcendental operation.
+ return -1;
+ r = reserve_gpr(bs, sel, elem, cycle);
+ if (r)
+ return r;
+ }
+ // Constants already processed
+ // No restrictions on PV, PS
}
return 0;
}
-#endif
-/* CB constants start at 512, and get translated to a kcache index when ALU
- * clauses are constructed. Note that we handle kcache constants the same way
- * as (the now gone) cfile constants, is that really required? */
-static int is_const(int sel)
+static int check_and_set_bank_swizzle(struct r600_bc_alu *slots[5])
{
- if (sel > 511 && sel < 4607)
- return 1;
- return 0;
+ 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;
+ forced = 1;
+ }
+
+ if (forced)
+ return 0;
+
+ // just check every possible combination of bank swizzle
+ // not very efficent, but works on the first try in most of the cases
+ for (i = 0; i < 4; i++)
+ bank_swizzle[i] = SQ_ALU_VEC_012;
+ bank_swizzle[4] = SQ_ALU_SCL_210;
+ while(bank_swizzle[4] <= SQ_ALU_SCL_221) {
+ init_bank_swizzle(&bs);
+ for (i = 0; i < 4; i++) {
+ if (slots[i]) {
+ r = check_vector(slots[i], &bs, bank_swizzle[i]);
+ if (r)
+ break;
+ }
+ }
+ if (!r && slots[4]) {
+ r = check_scalar(slots[4], &bs, bank_swizzle[4]);
+ }
+ if (!r) {
+ for (i = 0; i < 5; i++) {
+ if (slots[i])
+ slots[i]->bank_swizzle = bank_swizzle[i];
+ }
+ return 0;
+ }
+
+ for (i = 0; i < 5; i++) {
+ bank_swizzle[i]++;
+ if (bank_swizzle[i] <= SQ_ALU_VEC_210)
+ break;
+ else
+ bank_swizzle[i] = SQ_ALU_VEC_012;
+ }
+ }
+
+ // couldn't find a working swizzle
+ return -1;
}
-static int check_scalar(struct r600_bc *bc, struct r600_bc_alu *alu)
+static int replace_gpr_with_pv_ps(struct r600_bc_alu *slots[5], struct r600_bc_alu *alu_prev)
{
- unsigned swizzle_key;
+ struct r600_bc_alu *prev[5];
+ int gpr[5], chan[5];
+ int i, j, r, src, num_src;
- if (alu->bank_swizzle_force) {
- alu->bank_swizzle = alu->bank_swizzle_force;
- return 0;
+ r = assign_alu_units(alu_prev, prev);
+ if (r)
+ return r;
+
+ for (i = 0; i < 5; ++i) {
+ if(prev[i] && prev[i]->dst.write && !prev[i]->dst.rel) {
+ gpr[i] = prev[i]->dst.sel;
+ if (is_alu_reduction_inst(prev[i]))
+ chan[i] = 0;
+ else
+ chan[i] = prev[i]->dst.chan;
+ } else
+ gpr[i] = -1;
+ }
+
+ for (i = 0; i < 5; ++i) {
+ struct r600_bc_alu *alu = slots[i];
+ if(!alu)
+ continue;
+
+ num_src = r600_bc_get_num_operands(alu);
+ for (src = 0; src < num_src; ++src) {
+ if (!is_gpr(alu->src[src].sel) || alu->src[src].rel)
+ continue;
+
+ if (alu->src[src].sel == gpr[4] &&
+ alu->src[src].chan == chan[4]) {
+ alu->src[src].sel = V_SQ_ALU_SRC_PS;
+ alu->src[src].chan = 0;
+ continue;
+ }
+
+ for (j = 0; j < 4; ++j) {
+ if (alu->src[src].sel == gpr[j] &&
+ alu->src[src].chan == j) {
+ alu->src[src].sel = V_SQ_ALU_SRC_PV;
+ alu->src[src].chan = chan[j];
+ break;
+ }
+ }
+ }
}
- swizzle_key = (is_const(alu->src[0].sel) ? 4 : 0 ) +
- (is_const(alu->src[1].sel) ? 2 : 0 ) +
- (is_const(alu->src[2].sel) ? 1 : 0 );
- alu->bank_swizzle = bank_swizzle_scl[swizzle_key];
return 0;
}
-static int check_vector(struct r600_bc *bc, struct r600_bc_alu *alu)
+void r600_bc_special_constants(u32 value, unsigned *sel, unsigned *neg)
{
- unsigned swizzle_key;
-
- if (alu->bank_swizzle_force) {
- alu->bank_swizzle = alu->bank_swizzle_force;
- return 0;
+ switch(value) {
+ case 0:
+ *sel = V_SQ_ALU_SRC_0;
+ break;
+ case 1:
+ *sel = V_SQ_ALU_SRC_1_INT;
+ break;
+ case -1:
+ *sel = V_SQ_ALU_SRC_M_1_INT;
+ break;
+ case 0x3F800000: // 1.0f
+ *sel = V_SQ_ALU_SRC_1;
+ break;
+ case 0x3F000000: // 0.5f
+ *sel = V_SQ_ALU_SRC_0_5;
+ break;
+ case 0xBF800000: // -1.0f
+ *sel = V_SQ_ALU_SRC_1;
+ *neg ^= 1;
+ break;
+ case 0xBF000000: // -0.5f
+ *sel = V_SQ_ALU_SRC_0_5;
+ *neg ^= 1;
+ break;
+ default:
+ *sel = V_SQ_ALU_SRC_LITERAL;
+ break;
}
- swizzle_key = (is_const(alu->src[0].sel) ? 4 : 0 ) +
- (is_const(alu->src[1].sel) ? 2 : 0 ) +
- (is_const(alu->src[2].sel) ? 1 : 0 );
+}
- alu->bank_swizzle = bank_swizzle_vec[swizzle_key];
+/* compute how many literal are needed */
+static int r600_bc_alu_nliterals(struct r600_bc_alu *alu, uint32_t literal[4], unsigned *nliteral)
+{
+ unsigned num_src = r600_bc_get_num_operands(alu);
+ unsigned i, j;
+
+ for (i = 0; i < num_src; ++i) {
+ if (alu->src[i].sel == V_SQ_ALU_SRC_LITERAL) {
+ uint32_t value = alu->src[i].value[alu->src[i].chan];
+ unsigned found = 0;
+ for (j = 0; j < *nliteral; ++j) {
+ if (literal[j] == value) {
+ found = 1;
+ break;
+ }
+ }
+ if (!found) {
+ if (*nliteral >= 4)
+ return -EINVAL;
+ literal[(*nliteral)++] = value;
+ }
+ }
+ }
return 0;
}
-static int check_and_set_bank_swizzle(struct r600_bc *bc, struct r600_bc_alu *alu_first)
+static void r600_bc_alu_adjust_literals(struct r600_bc_alu *alu, uint32_t literal[4], unsigned nliteral)
{
- struct r600_bc_alu *alu = NULL;
- int num_instr = 1;
+ unsigned num_src = r600_bc_get_num_operands(alu);
+ unsigned i, j;
+
+ for (i = 0; i < num_src; ++i) {
+ if (alu->src[i].sel == V_SQ_ALU_SRC_LITERAL) {
+ uint32_t value = alu->src[i].value[alu->src[i].chan];
+ for (j = 0; j < nliteral; ++j) {
+ if (literal[j] == value) {
+ alu->src[i].chan = j;
+ break;
+ }
+ }
+ }
+ }
+}
- init_gpr(alu_first);
+static int merge_inst_groups(struct r600_bc *bc, struct r600_bc_alu *slots[5], struct r600_bc_alu *alu_prev)
+{
+ struct r600_bc_alu *prev[5];
+ struct r600_bc_alu *result[5] = { NULL };
+
+ uint32_t literal[4];
+ unsigned nliteral = 0;
+
+ int i, j, r, src, num_src;
+ int num_once_inst = 0;
+
+ r = assign_alu_units(alu_prev, prev);
+ if (r)
+ return r;
- LIST_FOR_EACH_ENTRY(alu, &alu_first->bs_list, bs_list) {
- num_instr++;
+ for (i = 0; i < 5; ++i) {
+ /* check number of literals */
+ if (prev[i] && r600_bc_alu_nliterals(prev[i], literal, &nliteral))
+ return 0;
+ if (slots[i] && r600_bc_alu_nliterals(slots[i], literal, &nliteral))
+ return 0;
+
+ // let's check used slots
+ if (prev[i] && !slots[i]) {
+ result[i] = prev[i];
+ num_once_inst += is_alu_once_inst(prev[i]);
+ continue;
+ } else if (prev[i] && slots[i]) {
+ if (result[4] == NULL && prev[4] == NULL && slots[4] == NULL) {
+ // trans unit is still free try to use it
+ if (is_alu_any_unit_inst(slots[i])) {
+ result[i] = prev[i];
+ result[4] = slots[i];
+ } else if (is_alu_any_unit_inst(prev[i])) {
+ result[i] = slots[i];
+ result[4] = prev[i];
+ } else
+ return 0;
+ } else
+ return 0;
+ } else if(!slots[i]) {
+ continue;
+ } else
+ result[i] = slots[i];
+
+ // let's check source gprs
+ struct r600_bc_alu *alu = slots[i];
+ num_once_inst += is_alu_once_inst(alu);
+
+ num_src = r600_bc_get_num_operands(alu);
+ for (src = 0; src < num_src; ++src) {
+ // constants doesn't matter
+ if (!is_gpr(alu->src[src].sel))
+ continue;
+
+ for (j = 0; j < 5; ++j) {
+ if (!prev[j] || !prev[j]->dst.write)
+ continue;
+
+ // if it's relative then we can't determin which gpr is really used
+ if (prev[j]->dst.chan == alu->src[src].chan &&
+ (prev[j]->dst.sel == alu->src[src].sel ||
+ prev[j]->dst.rel || alu->src[src].rel))
+ return 0;
+ }
+ }
}
- if (num_instr == 1) {
- check_scalar(bc, alu_first);
-
- } else {
-/* check_read_slots(bc, bc->cf_last->curr_bs_head);*/
- check_vector(bc, alu_first);
- LIST_FOR_EACH_ENTRY(alu, &alu_first->bs_list, bs_list) {
- check_vector(bc, alu);
+ /* more than one PRED_ or KILL_ ? */
+ if (num_once_inst > 1)
+ return 0;
+
+ /* check if the result can still be swizzlet */
+ r = check_and_set_bank_swizzle(result);
+ if (r)
+ return 0;
+
+ /* looks like everything worked out right, apply the changes */
+
+ /* sort instructions */
+ for (i = 0; i < 5; ++i) {
+ slots[i] = result[i];
+ if (result[i]) {
+ LIST_DEL(&result[i]->list);
+ result[i]->last = 0;
+ LIST_ADDTAIL(&result[i]->list, &bc->cf_last->alu);
}
}
+
+ /* determine new last instruction */
+ LIST_ENTRY(struct r600_bc_alu, bc->cf_last->alu.prev, list)->last = 1;
+
+ /* determine new first instruction */
+ for (i = 0; i < 5; ++i) {
+ if (result[i]) {
+ bc->cf_last->curr_bs_head = result[i];
+ break;
+ }
+ }
+
+ bc->cf_last->prev_bs_head = bc->cf_last->prev2_bs_head;
+ bc->cf_last->prev2_bs_head = NULL;
+
return 0;
}
@@ -537,18 +974,30 @@ int r600_bc_add_alu_type(struct r600_bc *bc, const struct r600_bc_alu *alu, int
if (nalu == NULL)
return -ENOMEM;
memcpy(nalu, alu, sizeof(struct r600_bc_alu));
- nalu->nliteral = 0;
+
+ if (bc->cf_last != NULL && bc->cf_last->inst != (type << 3)) {
+ /* check if we could add it anyway */
+ if (bc->cf_last->inst == (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3) &&
+ type == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE) {
+ LIST_FOR_EACH_ENTRY(lalu, &bc->cf_last->alu, list) {
+ if (lalu->predicate) {
+ bc->force_add_cf = 1;
+ break;
+ }
+ }
+ } else
+ bc->force_add_cf = 1;
+ }
/* cf can contains only alu or only vtx or only tex */
- if (bc->cf_last == NULL || bc->cf_last->inst != (type << 3) ||
- bc->force_add_cf) {
+ if (bc->cf_last == NULL || bc->force_add_cf) {
r = r600_bc_add_cf(bc);
if (r) {
free(nalu);
return r;
}
- bc->cf_last->inst = (type << 3);
}
+ bc->cf_last->inst = (type << 3);
/* Setup the kcache for this ALU instruction. This will start a new
* ALU clause if needed. */
@@ -559,37 +1008,18 @@ 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;
- LIST_INITHEAD(&nalu->bs_list);
- } else {
- LIST_ADDTAIL(&nalu->bs_list, &bc->cf_last->curr_bs_head->bs_list);
}
- /* at most 128 slots, one add alu can add 4 slots + 4 constants(2 slots)
+ /* at most 128 slots, one add alu can add 5 slots + 4 constants(2 slots)
* worst case */
if (nalu->last && (bc->cf_last->ndw >> 1) >= 120) {
bc->force_add_cf = 1;
}
- /* 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;
- }
- /* compute how many literal are needed
- * either 2 or 4 literals
- */
- if (nalu->src[i].sel == 253) {
- if (((nalu->src[i].chan + 2) & 0x6) > nalu->nliteral) {
- nalu->nliteral = (nalu->src[i].chan + 2) & 0x6;
- }
- }
- }
- if (!LIST_IS_EMPTY(&bc->cf_last->alu)) {
- lalu = LIST_ENTRY(struct r600_bc_alu, bc->cf_last->alu.prev, list);
- if (!lalu->last && lalu->nliteral > nalu->nliteral) {
- nalu->nliteral = lalu->nliteral;
- }
- }
- if (nalu->dst.sel >= bc->ngpr) {
- bc->ngpr = nalu->dst.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].sel, &nalu->src[i].neg);
}
LIST_ADDTAIL(&nalu->list, &bc->cf_last->alu);
/* each alu use 2 dwords */
@@ -598,7 +1028,29 @@ int r600_bc_add_alu_type(struct r600_bc *bc, const struct r600_bc_alu *alu, int
/* process cur ALU instructions for bank swizzle */
if (nalu->last) {
- check_and_set_bank_swizzle(bc, bc->cf_last->curr_bs_head);
+ struct r600_bc_alu *slots[5];
+ r = assign_alu_units(bc->cf_last->curr_bs_head, slots);
+ if (r)
+ return r;
+
+ if (bc->cf_last->prev_bs_head) {
+ r = merge_inst_groups(bc, slots, bc->cf_last->prev_bs_head);
+ if (r)
+ return r;
+ }
+
+ if (bc->cf_last->prev_bs_head) {
+ r = replace_gpr_with_pv_ps(slots, bc->cf_last->prev_bs_head);
+ if (r)
+ return r;
+ }
+
+ r = check_and_set_bank_swizzle(slots);
+ if (r)
+ return r;
+
+ bc->cf_last->prev2_bs_head = bc->cf_last->prev_bs_head;
+ bc->cf_last->prev_bs_head = bc->cf_last->curr_bs_head;
bc->cf_last->curr_bs_head = NULL;
}
return 0;
@@ -609,42 +1061,14 @@ 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));
}
-int r600_bc_add_literal(struct r600_bc *bc, const u32 *value)
+static void r600_bc_remove_alu(struct r600_bc_cf *cf, struct r600_bc_alu *alu)
{
- struct r600_bc_alu *alu;
-
- if (bc->cf_last == NULL) {
- return 0;
- }
- if (bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_TEX) {
- return 0;
- }
- /* all same on EG */
- if (bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_JUMP ||
- bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_ELSE ||
- bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_LOOP_START_NO_AL ||
- bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK ||
- bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_LOOP_CONTINUE ||
- bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_LOOP_END ||
- bc->cf_last->inst == V_SQ_CF_WORD1_SQ_CF_INST_POP) {
- return 0;
- }
- /* same on EG */
- if (((bc->cf_last->inst != (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3)) &&
- (bc->cf_last->inst != (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3))) ||
- LIST_IS_EMPTY(&bc->cf_last->alu)) {
- R600_ERR("last CF is not ALU (%p)\n", bc->cf_last);
- return -EINVAL;
- }
- alu = LIST_ENTRY(struct r600_bc_alu, bc->cf_last->alu.prev, list);
- if (!alu->last || !alu->nliteral || alu->literal_added) {
- return 0;
+ if (alu->last && alu->list.prev != &cf->alu) {
+ PREV_ALU(alu)->last = 1;
}
- memcpy(alu->value, value, 4 * 4);
- bc->cf_last->ndw += alu->nliteral;
- bc->ndw += alu->nliteral;
- alu->literal_added = 1;
- return 0;
+ LIST_DEL(&alu->list);
+ free(alu);
+ cf->ndw -= 2;
}
int r600_bc_add_vtx(struct r600_bc *bc, const struct r600_bc_vtx *vtx)
@@ -672,7 +1096,7 @@ int r600_bc_add_vtx(struct r600_bc *bc, const struct r600_bc_vtx *vtx)
/* each fetch use 4 dwords */
bc->cf_last->ndw += 4;
bc->ndw += 4;
- if ((bc->ndw / 4) > 7)
+ if ((bc->cf_last->ndw / 4) > 7)
bc->force_add_cf = 1;
return 0;
}
@@ -701,7 +1125,7 @@ int r600_bc_add_tex(struct r600_bc *bc, const struct r600_bc_tex *tex)
/* each texture fetch use 4 dwords */
bc->cf_last->ndw += 4;
bc->ndw += 4;
- if ((bc->ndw / 4) > 7)
+ if ((bc->cf_last->ndw / 4) > 7)
bc->force_add_cf = 1;
return 0;
}
@@ -797,8 +1221,6 @@ static int r600_bc_tex_build(struct r600_bc *bc, struct r600_bc_tex *tex, unsign
/* r600 only, r700/eg bits in r700_asm.c */
static int r600_bc_alu_build(struct r600_bc *bc, struct r600_bc_alu *alu, unsigned id)
{
- unsigned i;
-
/* don't replace gpr by pv or ps for destination register */
bc->bytecode[id++] = S_SQ_ALU_WORD0_SRC0_SEL(alu->src[0].sel) |
S_SQ_ALU_WORD0_SRC0_REL(alu->src[0].rel) |
@@ -829,30 +1251,70 @@ static int r600_bc_alu_build(struct r600_bc *bc, struct r600_bc_alu *alu, unsign
S_SQ_ALU_WORD1_OP2_SRC0_ABS(alu->src[0].abs) |
S_SQ_ALU_WORD1_OP2_SRC1_ABS(alu->src[1].abs) |
S_SQ_ALU_WORD1_OP2_WRITE_MASK(alu->dst.write) |
+ S_SQ_ALU_WORD1_OP2_OMOD(alu->omod) |
S_SQ_ALU_WORD1_OP2_ALU_INST(alu->inst) |
S_SQ_ALU_WORD1_BANK_SWIZZLE(alu->bank_swizzle) |
S_SQ_ALU_WORD1_OP2_UPDATE_EXECUTE_MASK(alu->predicate) |
S_SQ_ALU_WORD1_OP2_UPDATE_PRED(alu->predicate);
}
- if (alu->last) {
- if (alu->nliteral && !alu->literal_added) {
- R600_ERR("Bug in ALU processing for instruction 0x%08x, literal not added correctly\n", alu->inst);
- }
- for (i = 0; i < alu->nliteral; i++) {
- bc->bytecode[id++] = alu->value[i];
- }
- }
return 0;
}
+enum cf_class
+{
+ 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_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 (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):
+ 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) |
@@ -862,46 +1324,39 @@ 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:
@@ -911,14 +1366,776 @@ 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_alu *alu, struct gpr_usage usage[128], int32_t id)
+{
+ unsigned src, num_src;
+
+ num_src = r600_bc_get_num_operands(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_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(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_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(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(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(alu, usage, id);
+ if (alu->last) {
+ notice_alu_dst_gprs(first, usage, id, predicate || stack > 0);
+ first = NULL;
+ ++id;
+ }
+ if (is_alu_pred_inst(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(alu, usage, id, barrier[stack], &cf->barrier);
+ if (alu->last)
+ ++id;
+
+ if (is_alu_pred_inst(alu))
+ predicate++;
+
+ if (cf->inst == V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3)
+ optimize_alu_inst(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;
- int r;
+ int i, r;
if (bc->callstack[0].max > 0)
bc->nstack = ((bc->callstack[0].max + 3) >> 2) + 2;
@@ -926,35 +2143,37 @@ 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_PUSH_BEFORE << 3):
+ switch (get_cf_class(cf)) {
+ case CF_CLASS_ALU:
+ nliteral = 0;
+ LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) {
+ r = r600_bc_alu_nliterals(alu, literal, &nliteral);
+ if (r)
+ return r;
+ if (alu->last) {
+ cf->ndw += align(nliteral, 2);
+ nliteral = 0;
+ }
+ }
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);
@@ -964,6 +2183,14 @@ 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)
@@ -976,10 +2203,14 @@ 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_PUSH_BEFORE << 3):
+ switch (get_cf_class(cf)) {
+ case CF_CLASS_ALU:
+ nliteral = 0;
LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) {
+ r = r600_bc_alu_nliterals(alu, literal, &nliteral);
+ if (r)
+ return r;
+ r600_bc_alu_adjust_literals(alu, literal, nliteral);
switch(bc->chiprev) {
case CHIPREV_R600:
r = r600_bc_alu_build(bc, alu, addr);
@@ -996,12 +2227,14 @@ int r600_bc_build(struct r600_bc *bc)
return r;
addr += 2;
if (alu->last) {
- addr += alu->nliteral;
+ for (i = 0; i < align(nliteral, 2); ++i) {
+ bc->bytecode[addr++] = literal[i];
+ }
+ nliteral = 0;
}
}
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)
@@ -1009,7 +2242,7 @@ 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)
@@ -1017,19 +2250,8 @@ 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);
@@ -1077,7 +2299,14 @@ void r600_bc_clear(struct r600_bc *bc)
void r600_bc_dump(struct r600_bc *bc)
{
- unsigned i;
+ struct r600_bc_cf *cf;
+ struct r600_bc_alu *alu;
+ struct r600_bc_vtx *vtx;
+ struct r600_bc_tex *tex;
+
+ unsigned i, id;
+ uint32_t literal[4];
+ unsigned nliteral;
char chip = '6';
switch (bc->chiprev) {
@@ -1092,11 +2321,124 @@ void r600_bc_dump(struct r600_bc *bc)
chip = '6';
break;
}
- fprintf(stderr, "bytecode %d dw -----------------------\n", bc->ndw);
+ fprintf(stderr, "bytecode %d dw -- %d gprs -----------------------\n", bc->ndw, bc->ngpr);
fprintf(stderr, " %c\n", chip);
- for (i = 0; i < bc->ndw; i++) {
- fprintf(stderr, "0x%08X\n", bc->bytecode[i]);
+
+ LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) {
+ id = cf->id;
+
+ switch (get_cf_class(cf)) {
+ case CF_CLASS_ALU:
+ fprintf(stderr, "%04d %08X ALU ", id, bc->bytecode[id]);
+ fprintf(stderr, "ADDR:%04d ", cf->addr);
+ fprintf(stderr, "KCACHE_MODE0:%X ", cf->kcache0_mode);
+ fprintf(stderr, "KCACHE_BANK0:%X ", cf->kcache0_bank);
+ fprintf(stderr, "KCACHE_BANK1:%X\n", cf->kcache1_bank);
+ id++;
+ fprintf(stderr, "%04d %08X ALU ", id, bc->bytecode[id]);
+ fprintf(stderr, "INST:%d ", cf->inst);
+ fprintf(stderr, "KCACHE_MODE1:%X ", cf->kcache1_mode);
+ fprintf(stderr, "KCACHE_ADDR0:%X ", cf->kcache0_addr);
+ fprintf(stderr, "KCACHE_ADDR1:%X ", cf->kcache1_addr);
+ fprintf(stderr, "BARRIER:%d ", cf->barrier);
+ fprintf(stderr, "COUNT:%d\n", cf->ndw / 2);
+ break;
+ case CF_CLASS_TEXTURE:
+ case CF_CLASS_VERTEX:
+ fprintf(stderr, "%04d %08X TEX/VTX ", id, bc->bytecode[id]);
+ 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 CF_CLASS_EXPORT:
+ fprintf(stderr, "%04d %08X EXPORT ", id, bc->bytecode[id]);
+ 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);
+ id++;
+ fprintf(stderr, "%04d %08X EXPORT ", id, bc->bytecode[id]);
+ fprintf(stderr, "SWIZ_X:%X ", cf->output.swizzle_x);
+ fprintf(stderr, "SWIZ_Y:%X ", cf->output.swizzle_y);
+ 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:%d ", cf->barrier);
+ fprintf(stderr, "INST:%d ", cf->inst);
+ fprintf(stderr, "BURST_COUNT:%d\n", cf->output.burst_count);
+ break;
+ case CF_CLASS_OTHER:
+ fprintf(stderr, "%04d %08X CF ", id, bc->bytecode[id]);
+ 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;
+ }
+
+ id = cf->addr;
+ nliteral = 0;
+ LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) {
+ r600_bc_alu_nliterals(alu, literal, &nliteral);
+
+ fprintf(stderr, "%04d %08X ", id, bc->bytecode[id]);
+ fprintf(stderr, "SRC0(SEL:%d ", alu->src[0].sel);
+ fprintf(stderr, "REL:%d ", alu->src[0].rel);
+ fprintf(stderr, "CHAN:%d ", alu->src[0].chan);
+ fprintf(stderr, "NEG:%d) ", alu->src[0].neg);
+ fprintf(stderr, "SRC1(SEL:%d ", alu->src[1].sel);
+ fprintf(stderr, "REL:%d ", alu->src[1].rel);
+ fprintf(stderr, "CHAN:%d ", alu->src[1].chan);
+ fprintf(stderr, "NEG:%d) ", alu->src[1].neg);
+ fprintf(stderr, "LAST:%d)\n", alu->last);
+ id++;
+ fprintf(stderr, "%04d %08X %c ", id, bc->bytecode[id], alu->last ? '*' : ' ');
+ fprintf(stderr, "INST:%d ", alu->inst);
+ fprintf(stderr, "DST(SEL:%d ", alu->dst.sel);
+ fprintf(stderr, "CHAN:%d ", alu->dst.chan);
+ fprintf(stderr, "REL:%d ", alu->dst.rel);
+ fprintf(stderr, "CLAMP:%d) ", alu->dst.clamp);
+ fprintf(stderr, "BANK_SWIZZLE:%d ", alu->bank_swizzle);
+ if (alu->is_op3) {
+ fprintf(stderr, "SRC2(SEL:%d ", alu->src[2].sel);
+ fprintf(stderr, "REL:%d ", alu->src[2].rel);
+ fprintf(stderr, "CHAN:%d ", alu->src[2].chan);
+ fprintf(stderr, "NEG:%d)\n", alu->src[2].neg);
+ } else {
+ fprintf(stderr, "SRC0_ABS:%d ", alu->src[0].abs);
+ fprintf(stderr, "SRC1_ABS:%d ", alu->src[1].abs);
+ fprintf(stderr, "WRITE_MASK:%d ", alu->dst.write);
+ fprintf(stderr, "OMOD:%d ", alu->omod);
+ fprintf(stderr, "EXECUTE_MASK:%d ", alu->predicate);
+ fprintf(stderr, "UPDATE_PRED:%d\n", alu->predicate);
+ }
+
+ id++;
+ if (alu->last) {
+ for (i = 0; i < nliteral; i++, id++) {
+ float *f = (float*)(bc->bytecode + id);
+ fprintf(stderr, "%04d %08X %f\n", id, bc->bytecode[id], *f);
+ }
+ id += nliteral & 1;
+ nliteral = 0;
+ }
+ }
+
+ LIST_FOR_EACH_ENTRY(tex, &cf->tex, list) {
+ //TODO
+ }
+
+ LIST_FOR_EACH_ENTRY(vtx, &cf->vtx, list) {
+ //TODO
+ }
}
+
fprintf(stderr, "--------------------------------------\n");
}
@@ -1108,21 +2450,21 @@ 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;
@@ -1144,21 +2486,21 @@ 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;
@@ -1308,31 +2650,6 @@ out_unknown:
R600_ERR("unsupported vertex format %s\n", util_format_name(pformat));
}
-static void r600_bc(unsigned ndw, unsigned chiprev, u32 *bytecode)
-{
- unsigned i;
- char chip = '6';
-
- switch (chiprev) {
- case 1:
- chip = '7';
- break;
- case 2:
- chip = 'E';
- break;
- case 0:
- default:
- chip = '6';
- break;
- }
- fprintf(stderr, "bytecode %d dw -----------------------\n", ndw);
- fprintf(stderr, " %c\n", chip);
- for (i = 0; i < ndw; i++) {
- fprintf(stderr, "0x%08X\n", bytecode[i]);
- }
- fprintf(stderr, "--------------------------------------\n");
-}
-
int r600_vertex_elements_build_fetch_shader(struct r600_pipe_context *rctx, struct r600_vertex_element *ve)
{
unsigned ndw, i;
diff --git a/src/gallium/drivers/r600/r600_asm.h b/src/gallium/drivers/r600/r600_asm.h
index a5504ad39f4..519245f3af2 100644
--- a/src/gallium/drivers/r600/r600_asm.h
+++ b/src/gallium/drivers/r600/r600_asm.h
@@ -25,9 +25,6 @@
#include "util/u_double_list.h"
-#define NUM_OF_CYCLES 3
-#define NUM_OF_COMPONENTS 4
-
struct r600_vertex_element;
struct r600_pipe_context;
@@ -37,6 +34,7 @@ struct r600_bc_alu_src {
unsigned neg;
unsigned abs;
unsigned rel;
+ u32 *value;
};
struct r600_bc_alu_dst {
@@ -49,19 +47,15 @@ struct r600_bc_alu_dst {
struct r600_bc_alu {
struct list_head list;
- struct list_head bs_list; /* bank swizzle list */
struct r600_bc_alu_src src[3];
struct r600_bc_alu_dst dst;
unsigned inst;
unsigned last;
unsigned is_op3;
unsigned predicate;
- unsigned nliteral;
- unsigned literal_added;
unsigned bank_swizzle;
unsigned bank_swizzle_force;
- u32 value[4];
- int hw_gpr[NUM_OF_CYCLES][NUM_OF_COMPONENTS];
+ unsigned omod;
};
struct r600_bc_tex {
@@ -114,15 +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 {
@@ -140,6 +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;
@@ -147,6 +140,8 @@ struct r600_bc_cf {
struct list_head vtx;
struct r600_bc_output output;
struct r600_bc_alu *curr_bs_head;
+ struct r600_bc_alu *prev_bs_head;
+ struct r600_bc_alu *prev2_bs_head;
};
#define FC_NONE 0
@@ -198,13 +193,13 @@ void eg_cf_vtx(struct r600_vertex_element *ve, u32 *bytecode, unsigned count);
int r600_bc_init(struct r600_bc *bc, enum radeon_family family);
void r600_bc_clear(struct r600_bc *bc);
int r600_bc_add_alu(struct r600_bc *bc, const struct r600_bc_alu *alu);
-int r600_bc_add_literal(struct r600_bc *bc, const u32 *value);
int r600_bc_add_vtx(struct r600_bc *bc, const struct r600_bc_vtx *vtx);
int r600_bc_add_tex(struct r600_bc *bc, const struct r600_bc_tex *tex);
int r600_bc_add_output(struct r600_bc *bc, const struct r600_bc_output *output);
int r600_bc_build(struct r600_bc *bc);
int r600_bc_add_cfinst(struct r600_bc *bc, int inst);
int r600_bc_add_alu_type(struct r600_bc *bc, const struct r600_bc_alu *alu, int type);
+void r600_bc_special_constants(u32 value, unsigned *sel, unsigned *neg);
void r600_bc_dump(struct r600_bc *bc);
void r600_cf_vtx(struct r600_vertex_element *ve, u32 *bytecode, unsigned count);
void r600_cf_vtx_tc(struct r600_vertex_element *ve, u32 *bytecode, unsigned count);
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 51ff7db3612..95367d7c536 100644
--- a/src/gallium/drivers/r600/r600_shader.c
+++ b/src/gallium/drivers/r600/r600_shader.c
@@ -225,21 +225,23 @@ int r600_pipe_shader(struct pipe_context *ctx, struct r600_pipe_shader *shader)
return 0;
}
-int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *shader);
+int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *shader, u32 **literals);
int r600_pipe_shader_create(struct pipe_context *ctx, struct r600_pipe_shader *shader, const struct tgsi_token *tokens)
{
struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+ u32 *literals;
int r;
//fprintf(stderr, "--------------------------------------------------------------\n");
//tgsi_dump(tokens, 0);
shader->shader.family = r600_get_family(rctx->radeon);
- r = r600_shader_from_tgsi(tokens, &shader->shader);
+ r = r600_shader_from_tgsi(tokens, &shader->shader, &literals);
if (r) {
R600_ERR("translation from TGSI failed !\n");
return r;
}
r = r600_bc_build(&shader->shader.bc);
+ free(literals);
if (r) {
R600_ERR("building bytecode failed !\n");
return r;
@@ -272,7 +274,6 @@ struct r600_shader_ctx {
struct r600_shader_tgsi_instruction *inst_info;
struct r600_bc *bc;
struct r600_shader *shader;
- u32 value[4];
u32 *literals;
u32 nliterals;
u32 max_driver_temp_used;
@@ -481,12 +482,12 @@ static int evergreen_gpr_count(struct r600_shader_ctx *ctx)
return ctx->num_interp_gpr;
}
-int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *shader)
+int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *shader, u32 **literals)
{
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;
@@ -546,7 +547,7 @@ int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *s
* kcache banks later. */
ctx.file_offset[TGSI_FILE_CONSTANT] = 512;
- ctx.file_offset[TGSI_FILE_IMMEDIATE] = 253;
+ ctx.file_offset[TGSI_FILE_IMMEDIATE] = V_SQ_ALU_SRC_LITERAL;
ctx.temp_reg = ctx.file_offset[TGSI_FILE_TEMPORARY] +
ctx.info.file_count[TGSI_FILE_TEMPORARY];
@@ -589,9 +590,6 @@ int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *s
r = ctx.inst_info->process(&ctx);
if (r)
goto out_err;
- r = r600_bc_add_literal(ctx.bc, ctx.value);
- if (r)
- goto out_err;
break;
case TGSI_TOKEN_TYPE_PROPERTY:
break;
@@ -611,10 +609,8 @@ 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) {
@@ -674,10 +670,8 @@ 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++;
}
}
@@ -690,29 +684,17 @@ 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]);
if (r)
goto out_err;
}
- free(ctx.literals);
+ *literals = ctx.literals;
tgsi_parse_free(&ctx.parse);
return 0;
out_err:
@@ -736,22 +718,29 @@ static int tgsi_src(struct r600_shader_ctx *ctx,
const struct tgsi_full_src_register *tgsi_src,
struct r600_bc_alu_src *r600_src)
{
- int index;
memset(r600_src, 0, sizeof(struct r600_bc_alu_src));
- r600_src->sel = tgsi_src->Register.Index;
+ r600_src->neg = tgsi_src->Register.Negate;
+ r600_src->abs = tgsi_src->Register.Absolute;
if (tgsi_src->Register.File == TGSI_FILE_IMMEDIATE) {
- r600_src->sel = 0;
+ int index;
+ if((tgsi_src->Register.SwizzleX == tgsi_src->Register.SwizzleY) &&
+ (tgsi_src->Register.SwizzleX == tgsi_src->Register.SwizzleZ) &&
+ (tgsi_src->Register.SwizzleX == tgsi_src->Register.SwizzleW)) {
+
+ index = tgsi_src->Register.Index * 4 + tgsi_src->Register.SwizzleX;
+ r600_bc_special_constants(ctx->literals[index], &r600_src->sel, &r600_src->neg);
+ if (r600_src->sel != V_SQ_ALU_SRC_LITERAL)
+ return 0;
+ }
index = tgsi_src->Register.Index;
- ctx->value[0] = ctx->literals[index * 4 + 0];
- ctx->value[1] = ctx->literals[index * 4 + 1];
- ctx->value[2] = ctx->literals[index * 4 + 2];
- ctx->value[3] = ctx->literals[index * 4 + 3];
+ r600_src->sel = V_SQ_ALU_SRC_LITERAL;
+ r600_src->value = ctx->literals + index * 4;
+ } else {
+ if (tgsi_src->Register.Indirect)
+ r600_src->rel = V_SQ_REL_RELATIVE;
+ r600_src->sel = tgsi_src->Register.Index;
+ r600_src->sel += ctx->file_offset[tgsi_src->Register.File];
}
- if (tgsi_src->Register.Indirect)
- r600_src->rel = V_SQ_REL_RELATIVE;
- r600_src->neg = tgsi_src->Register.Negate;
- r600_src->abs = tgsi_src->Register.Absolute;
- r600_src->sel += ctx->file_offset[tgsi_src->Register.File];
return 0;
}
@@ -839,18 +828,19 @@ static int tgsi_split_literal_constant(struct r600_shader_ctx *ctx, struct r600_
int i, j, k, nliteral, r;
for (i = 0, nliteral = 0; i < inst->Instruction.NumSrcRegs; i++) {
- if (inst->Src[i].Register.File == TGSI_FILE_IMMEDIATE) {
+ if (r600_src[i].sel == V_SQ_ALU_SRC_LITERAL) {
nliteral++;
}
}
for (i = 0, j = nliteral - 1; i < inst->Instruction.NumSrcRegs; i++) {
- if (j > 0 && inst->Src[i].Register.File == TGSI_FILE_IMMEDIATE) {
+ if (j > 0 && r600_src[i].sel == V_SQ_ALU_SRC_LITERAL) {
int treg = r600_get_temp(ctx);
for (k = 0; k < 4; k++) {
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
alu.src[0].sel = r600_src[i].sel;
alu.src[0].chan = k;
+ alu.src[0].value = r600_src[i].value;
alu.dst.sel = treg;
alu.dst.chan = k;
alu.dst.write = 1;
@@ -860,9 +850,6 @@ static int tgsi_split_literal_constant(struct r600_shader_ctx *ctx, struct r600_
if (r)
return r;
}
- r = r600_bc_add_literal(ctx->bc, &ctx->literals[inst->Src[i].Register.Index * 4]);
- if (r)
- return r;
r600_src[i].sel = treg;
j--;
}
@@ -870,19 +857,25 @@ static int tgsi_split_literal_constant(struct r600_shader_ctx *ctx, struct r600_
return 0;
}
-static int tgsi_op2_s(struct r600_shader_ctx *ctx, int swap)
+static int tgsi_last_instruction(unsigned writemask)
{
- struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
- struct r600_bc_alu_src r600_src[3];
- struct r600_bc_alu alu;
- int i, j, r;
- int lasti = 0;
+ int i, lasti = 0;
for (i = 0; i < 4; i++) {
- if (inst->Dst[0].Register.WriteMask & (1 << i)) {
+ if (writemask & (1 << i)) {
lasti = i;
}
}
+ return lasti;
+}
+
+static int tgsi_op2_s(struct r600_shader_ctx *ctx, int swap)
+{
+ struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
+ struct r600_bc_alu_src r600_src[3];
+ struct r600_bc_alu alu;
+ int i, j, r;
+ int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
r = tgsi_split_constant(ctx, r600_src);
if (r)
@@ -951,12 +944,14 @@ static int tgsi_op2_swap(struct r600_shader_ctx *ctx)
static int tgsi_setup_trig(struct r600_shader_ctx *ctx,
struct r600_bc_alu_src r600_src[3])
{
+ static float half_inv_pi = 1.0 /(3.1415926535 * 2);
+ static float double_pi = 3.1415926535 * 2;
+ static float neg_pi = -3.1415926535;
+
struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
int r;
- uint32_t lit_vals[4];
struct r600_bc_alu alu;
- memset(lit_vals, 0, 4*4);
r = tgsi_split_constant(ctx, r600_src);
if (r)
return r;
@@ -964,9 +959,6 @@ static int tgsi_setup_trig(struct r600_shader_ctx *ctx,
if (r)
return r;
- lit_vals[0] = fui(1.0 /(3.1415926535 * 2));
- lit_vals[1] = fui(0.5f);
-
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD);
alu.is_op3 = 1;
@@ -980,15 +972,13 @@ static int tgsi_setup_trig(struct r600_shader_ctx *ctx,
alu.src[1].sel = V_SQ_ALU_SRC_LITERAL;
alu.src[1].chan = 0;
- alu.src[2].sel = V_SQ_ALU_SRC_LITERAL;
+ alu.src[1].value = (uint32_t *)&half_inv_pi;
+ alu.src[2].sel = V_SQ_ALU_SRC_0_5;
alu.src[2].chan = 1;
alu.last = 1;
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, lit_vals);
- if (r)
- return r;
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FRACT);
@@ -1004,14 +994,6 @@ static int tgsi_setup_trig(struct r600_shader_ctx *ctx,
if (r)
return r;
- if (ctx->bc->chiprev == CHIPREV_R600) {
- lit_vals[0] = fui(3.1415926535897f * 2.0f);
- lit_vals[1] = fui(-3.1415926535897f);
- } else {
- lit_vals[0] = fui(1.0f);
- lit_vals[1] = fui(-0.5f);
- }
-
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD);
alu.is_op3 = 1;
@@ -1027,13 +1009,20 @@ static int tgsi_setup_trig(struct r600_shader_ctx *ctx,
alu.src[1].chan = 0;
alu.src[2].sel = V_SQ_ALU_SRC_LITERAL;
alu.src[2].chan = 1;
+
+ if (ctx->bc->chiprev == CHIPREV_R600) {
+ alu.src[1].value = (uint32_t *)&double_pi;
+ alu.src[2].value = (uint32_t *)&neg_pi;
+ } else {
+ alu.src[1].sel = V_SQ_ALU_SRC_1;
+ alu.src[2].sel = V_SQ_ALU_SRC_0_5;
+ alu.src[2].neg = 1;
+ }
+
alu.last = 1;
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, lit_vals);
- if (r)
- return r;
return 0;
}
@@ -1043,7 +1032,7 @@ static int tgsi_trig(struct r600_shader_ctx *ctx)
struct r600_bc_alu_src r600_src[3];
struct r600_bc_alu alu;
int i, r;
- int lasti = 0;
+ int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
r = tgsi_setup_trig(ctx, r600_src);
if (r)
@@ -1063,10 +1052,6 @@ static int tgsi_trig(struct r600_shader_ctx *ctx)
return r;
/* replicate result */
- for (i = 0; i < 4; i++) {
- if (inst->Dst[0].Register.WriteMask & (1 << i))
- lasti = i;
- }
for (i = 0; i < lasti + 1; i++) {
if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
continue;
@@ -1153,10 +1138,6 @@ static int tgsi_scs(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
-
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
/* dst.w = 1.0; */
@@ -1177,10 +1158,6 @@ static int tgsi_scs(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
-
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
return 0;
@@ -1216,9 +1193,6 @@ static int tgsi_kill(struct r600_shader_ctx *ctx)
if (r)
return r;
}
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
/* kill must be last in ALU */
ctx->bc->force_add_cf = 1;
@@ -1281,10 +1255,6 @@ static int tgsi_lit(struct r600_shader_ctx *ctx)
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
-
if (inst->Dst[0].Register.WriteMask & (1 << 2))
{
int chan;
@@ -1303,10 +1273,6 @@ static int tgsi_lit(struct r600_shader_ctx *ctx)
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
-
chan = alu.dst.chan;
sel = alu.dst.sel;
@@ -1329,9 +1295,6 @@ static int tgsi_lit(struct r600_shader_ctx *ctx)
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
/* dst.z = exp(tmp.x) */
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
@@ -1375,9 +1338,6 @@ static int tgsi_rsq(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
/* replicate result */
return tgsi_helper_tempx_replicate(ctx);
}
@@ -1426,9 +1386,6 @@ static int tgsi_trans_srcx_replicate(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
/* replicate result */
return tgsi_helper_tempx_replicate(ctx);
}
@@ -1452,9 +1409,6 @@ static int tgsi_pow(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc,ctx->value);
- if (r)
- return r;
/* b * LOG2(a) */
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
@@ -1469,9 +1423,6 @@ static int tgsi_pow(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc,ctx->value);
- if (r)
- return r;
/* POW(a,b) = EXP2(b * LOG2(a))*/
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
@@ -1482,9 +1433,6 @@ static int tgsi_pow(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc,ctx->value);
- if (r)
- return r;
return tgsi_helper_tempx_replicate(ctx);
}
@@ -1524,9 +1472,6 @@ static int tgsi_ssg(struct r600_shader_ctx *ctx)
if (r)
return r;
}
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
/* dst = (-tmp > 0 ? -1 : tmp) */
for (i = 0; i < 4; i++) {
@@ -1561,9 +1506,6 @@ static int tgsi_helper_copy(struct r600_shader_ctx *ctx, struct tgsi_full_instru
struct r600_bc_alu alu;
int i, r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
for (i = 0; i < 4; i++) {
memset(&alu, 0, sizeof(struct r600_bc_alu));
if (!(inst->Dst[0].Register.WriteMask & (1 << i))) {
@@ -1593,6 +1535,7 @@ static int tgsi_op3(struct r600_shader_ctx *ctx)
struct r600_bc_alu_src r600_src[3];
struct r600_bc_alu alu;
int i, j, r;
+ int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
r = tgsi_split_constant(ctx, r600_src);
if (r)
@@ -1600,26 +1543,32 @@ static int tgsi_op3(struct r600_shader_ctx *ctx)
r = tgsi_split_literal_constant(ctx, r600_src);
if (r)
return r;
- /* do it in 2 step as op3 doesn't support writemask */
- for (i = 0; i < 4; i++) {
+ for (i = 0; i < lasti + 1; i++) {
+ if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
+ continue;
+
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = ctx->inst_info->r600_opcode;
for (j = 0; j < inst->Instruction.NumSrcRegs; j++) {
alu.src[j] = r600_src[j];
alu.src[j].chan = tgsi_chan(&inst->Src[j], i);
}
- alu.dst.sel = ctx->temp_reg;
+
+ r = tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
+ if (r)
+ return r;
+
alu.dst.chan = i;
alu.dst.write = 1;
alu.is_op3 = 1;
- if (i == 3) {
+ if (i == lasti) {
alu.last = 1;
}
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
}
- return tgsi_helper_copy(ctx, inst);
+ return 0;
}
static int tgsi_dp(struct r600_shader_ctx *ctx)
@@ -1642,9 +1591,13 @@ static int tgsi_dp(struct r600_shader_ctx *ctx)
alu.src[j] = r600_src[j];
alu.src[j].chan = tgsi_chan(&inst->Src[j], i);
}
- alu.dst.sel = ctx->temp_reg;
+
+ r = tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
+ if (r)
+ return r;
+
alu.dst.chan = i;
- alu.dst.write = 1;
+ alu.dst.write = (inst->Dst[0].Register.WriteMask >> i) & 1;
/* handle some special cases */
switch (ctx->inst_info->tgsi_opcode) {
case TGSI_OPCODE_DP2:
@@ -1676,19 +1629,21 @@ static int tgsi_dp(struct r600_shader_ctx *ctx)
if (r)
return r;
}
- return tgsi_helper_copy(ctx, inst);
+ return 0;
}
static int tgsi_tex(struct r600_shader_ctx *ctx)
{
+ static float one_point_five = 1.5f;
struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
struct r600_bc_tex tex;
struct r600_bc_alu alu;
unsigned src_gpr;
int r, i;
int opcode;
- boolean src_not_temp = inst->Src[0].Register.File != TGSI_FILE_TEMPORARY;
- uint32_t lit_vals[4];
+ boolean src_not_temp =
+ inst->Src[0].Register.File != TGSI_FILE_TEMPORARY &&
+ inst->Src[0].Register.File != TGSI_FILE_INPUT;
src_gpr = ctx->file_offset[inst->Src[0].Register.File] + inst->Src[0].Register.Index;
@@ -1837,6 +1792,7 @@ static int tgsi_tex(struct r600_shader_ctx *ctx)
alu.src[2].sel = V_SQ_ALU_SRC_LITERAL;
alu.src[2].chan = 0;
+ alu.src[2].value = (u32*)&one_point_five;
alu.dst.sel = ctx->temp_reg;
alu.dst.chan = 1;
@@ -1847,11 +1803,6 @@ static int tgsi_tex(struct r600_shader_ctx *ctx)
if (r)
return r;
- lit_vals[0] = fui(1.5f);
-
- r = r600_bc_add_literal(ctx->bc, lit_vals);
- if (r)
- return r;
src_not_temp = FALSE;
src_gpr = ctx->temp_reg;
}
@@ -1924,6 +1875,7 @@ static int tgsi_lrp(struct r600_shader_ctx *ctx)
struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
struct r600_bc_alu_src r600_src[3];
struct r600_bc_alu alu;
+ int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
unsigned i;
int r;
@@ -1933,8 +1885,40 @@ static int tgsi_lrp(struct r600_shader_ctx *ctx)
r = tgsi_split_literal_constant(ctx, r600_src);
if (r)
return r;
+
+ /* optimize if it's just an equal balance */
+ if(r600_src[0].sel == V_SQ_ALU_SRC_0_5) {
+ for (i = 0; i < lasti + 1; i++) {
+ if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
+ continue;
+
+ memset(&alu, 0, sizeof(struct r600_bc_alu));
+ alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD);
+ alu.src[0] = r600_src[1];
+ alu.src[0].chan = tgsi_chan(&inst->Src[1], i);
+ alu.src[1] = r600_src[2];
+ alu.src[1].chan = tgsi_chan(&inst->Src[2], i);
+ alu.omod = 3;
+ r = tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
+ if (r)
+ return r;
+
+ alu.dst.chan = i;
+ if (i == lasti) {
+ alu.last = 1;
+ }
+ r = r600_bc_add_alu(ctx->bc, &alu);
+ if (r)
+ return r;
+ }
+ return 0;
+ }
+
/* 1 - src0 */
- for (i = 0; i < 4; i++) {
+ for (i = 0; i < lasti + 1; i++) {
+ if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
+ continue;
+
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ADD);
alu.src[0].sel = V_SQ_ALU_SRC_1;
@@ -1944,7 +1928,7 @@ static int tgsi_lrp(struct r600_shader_ctx *ctx)
alu.src[1].neg = 1;
alu.dst.sel = ctx->temp_reg;
alu.dst.chan = i;
- if (i == 3) {
+ if (i == lasti) {
alu.last = 1;
}
alu.dst.write = 1;
@@ -1952,12 +1936,12 @@ static int tgsi_lrp(struct r600_shader_ctx *ctx)
if (r)
return r;
}
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
/* (1 - src0) * src2 */
- for (i = 0; i < 4; i++) {
+ for (i = 0; i < lasti + 1; i++) {
+ if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
+ continue;
+
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
alu.src[0].sel = ctx->temp_reg;
@@ -1966,7 +1950,7 @@ static int tgsi_lrp(struct r600_shader_ctx *ctx)
alu.src[1].chan = tgsi_chan(&inst->Src[2], i);
alu.dst.sel = ctx->temp_reg;
alu.dst.chan = i;
- if (i == 3) {
+ if (i == lasti) {
alu.last = 1;
}
alu.dst.write = 1;
@@ -1974,12 +1958,12 @@ static int tgsi_lrp(struct r600_shader_ctx *ctx)
if (r)
return r;
}
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
/* src0 * src1 + (1 - src0) * src2 */
- for (i = 0; i < 4; i++) {
+ for (i = 0; i < lasti + 1; i++) {
+ if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
+ continue;
+
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_MULADD);
alu.is_op3 = 1;
@@ -1989,16 +1973,20 @@ static int tgsi_lrp(struct r600_shader_ctx *ctx)
alu.src[1].chan = tgsi_chan(&inst->Src[1], i);
alu.src[2].sel = ctx->temp_reg;
alu.src[2].chan = i;
- alu.dst.sel = ctx->temp_reg;
+
+ r = tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
+ if (r)
+ return r;
+
alu.dst.chan = i;
- if (i == 3) {
+ if (i == lasti) {
alu.last = 1;
}
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
}
- return tgsi_helper_copy(ctx, inst);
+ return 0;
}
static int tgsi_cmp(struct r600_shader_ctx *ctx)
@@ -2006,8 +1994,8 @@ static int tgsi_cmp(struct r600_shader_ctx *ctx)
struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
struct r600_bc_alu_src r600_src[3];
struct r600_bc_alu alu;
- int use_temp = 0;
int i, r;
+ int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
r = tgsi_split_constant(ctx, r600_src);
if (r)
@@ -2016,10 +2004,10 @@ static int tgsi_cmp(struct r600_shader_ctx *ctx)
if (r)
return r;
- if (inst->Dst[0].Register.WriteMask != 0xf)
- use_temp = 1;
+ for (i = 0; i < lasti + 1; i++) {
+ if (!(inst->Dst[0].Register.WriteMask & (1 << i)))
+ continue;
- for (i = 0; i < 4; i++) {
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP3_SQ_OP3_INST_CNDGE);
alu.src[0] = r600_src[0];
@@ -2031,24 +2019,19 @@ static int tgsi_cmp(struct r600_shader_ctx *ctx)
alu.src[2] = r600_src[1];
alu.src[2].chan = tgsi_chan(&inst->Src[1], i);
- if (use_temp)
- alu.dst.sel = ctx->temp_reg;
- else {
- r = tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
- if (r)
- return r;
- }
+ r = tgsi_dst(ctx, &inst->Dst[0], i, &alu.dst);
+ if (r)
+ return r;
+
alu.dst.chan = i;
alu.dst.write = 1;
alu.is_op3 = 1;
- if (i == 3)
+ if (i == lasti)
alu.last = 1;
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
}
- if (use_temp)
- return tgsi_helper_copy(ctx, inst);
return 0;
}
@@ -2115,10 +2098,6 @@ static int tgsi_xpd(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
-
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
for (i = 0; i < 4; i++) {
@@ -2176,10 +2155,6 @@ static int tgsi_xpd(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
-
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
if (use_temp)
return tgsi_helper_copy(ctx, inst);
@@ -2212,10 +2187,6 @@ static int tgsi_exp(struct r600_shader_ctx *ctx)
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
-
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
alu.src[0].sel = ctx->temp_reg;
alu.src[0].chan = 0;
@@ -2227,10 +2198,6 @@ static int tgsi_exp(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
-
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
/* result.y = tmp - floor(tmp); */
@@ -2256,9 +2223,6 @@ static int tgsi_exp(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
/* result.z = RoughApprox2ToX(tmp);*/
@@ -2279,9 +2243,6 @@ static int tgsi_exp(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
/* result.w = 1.0;*/
@@ -2299,9 +2260,6 @@ static int tgsi_exp(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
return tgsi_helper_copy(ctx, inst);
}
@@ -2331,10 +2289,6 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
-
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR);
alu.src[0].sel = ctx->temp_reg;
alu.src[0].chan = 0;
@@ -2347,10 +2301,6 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
-
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
/* result.y = src.x / (2 ^ floor(log2(src.x))); */
@@ -2373,10 +2323,6 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
-
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR);
@@ -2392,10 +2338,6 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
-
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_EXP_IEEE);
@@ -2411,10 +2353,6 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
-
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE);
@@ -2430,10 +2368,6 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
if (r)
return r;
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
-
memset(&alu, 0, sizeof(struct r600_bc_alu));
alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MUL);
@@ -2455,10 +2389,6 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
-
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
/* result.z = log2(src);*/
@@ -2480,10 +2410,6 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
-
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
/* result.w = 1.0; */
@@ -2502,10 +2428,6 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
r = r600_bc_add_alu(ctx->bc, &alu);
if (r)
return r;
-
- r = r600_bc_add_literal(ctx->bc, ctx->value);
- if (r)
- return r;
}
return tgsi_helper_copy(ctx, inst);
@@ -2660,9 +2582,25 @@ static int emit_logic_pred(struct r600_shader_ctx *ctx, int opcode)
static int pops(struct r600_shader_ctx *ctx, int pops)
{
- r600_bc_add_cfinst(ctx->bc, CTX_INST(V_SQ_CF_WORD1_SQ_CF_INST_POP));
- ctx->bc->cf_last->pop_count = pops;
- ctx->bc->cf_last->cf_addr = ctx->bc->cf_last->id + 2;
+ int alu_pop = 3;
+ if (ctx->bc->cf_last) {
+ if (ctx->bc->cf_last->inst == CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU) << 3)
+ alu_pop = 0;
+ else if (ctx->bc->cf_last->inst == CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER) << 3)
+ alu_pop = 1;
+ }
+ alu_pop += pops;
+ if (alu_pop == 1) {
+ ctx->bc->cf_last->inst = CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP_AFTER) << 3;
+ ctx->bc->force_add_cf = 1;
+ } else if (alu_pop == 2) {
+ ctx->bc->cf_last->inst = CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_POP2_AFTER) << 3;
+ ctx->bc->force_add_cf = 1;
+ } else {
+ r600_bc_add_cfinst(ctx->bc, CTX_INST(V_SQ_CF_WORD1_SQ_CF_INST_POP));
+ ctx->bc->cf_last->pop_count = pops;
+ ctx->bc->cf_last->cf_addr = ctx->bc->cf_last->id + 2;
+ }
return 0;
}
diff --git a/src/gallium/drivers/r600/r600_shader.h b/src/gallium/drivers/r600/r600_shader.h
index 35b0331525a..935dd6fe3ab 100644
--- a/src/gallium/drivers/r600/r600_shader.h
+++ b/src/gallium/drivers/r600/r600_shader.h
@@ -47,6 +47,6 @@ struct r600_shader {
boolean uses_kill;
};
-int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *shader);
+int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *shader, u32 **literals);
#endif
diff --git a/src/gallium/drivers/r600/r600_sq.h b/src/gallium/drivers/r600/r600_sq.h
index d812bfd1fed..56ed35e8b32 100644
--- a/src/gallium/drivers/r600/r600_sq.h
+++ b/src/gallium/drivers/r600/r600_sq.h
@@ -191,6 +191,8 @@
#define V_SQ_ALU_SRC_M_1_INT 0x000000FB
#define V_SQ_ALU_SRC_0_5 0x000000FC
#define V_SQ_ALU_SRC_LITERAL 0x000000FD
+#define V_SQ_ALU_SRC_PV 0x000000FE
+#define V_SQ_ALU_SRC_PS 0x000000FF
#define V_SQ_ALU_SRC_PARAM_BASE 0x000001C0
#define S_SQ_ALU_WORD0_SRC0_REL(x) (((x) & 0x1) << 9)
#define G_SQ_ALU_WORD0_SRC0_REL(x) (((x) >> 9) & 0x1)
diff --git a/src/gallium/drivers/r600/r600_state.c b/src/gallium/drivers/r600/r600_state.c
index 96b02d72b94..9572ff9a1a2 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 d994196e19d..39ca0a74f3f 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..6add92e6d4c 100644
--- a/src/gallium/drivers/r600/r600_texture.c
+++ b/src/gallium/drivers/r600/r600_texture.c
@@ -562,7 +562,8 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
use_staging_texture = TRUE;
if (!permit_hardware_blit(ctx->screen, texture) ||
- (texture->flags & R600_RESOURCE_FLAG_TRANSFER))
+ (texture->flags & R600_RESOURCE_FLAG_TRANSFER) ||
+ (texture->usage == PIPE_USAGE_STREAM))
use_staging_texture = FALSE;
trans = CALLOC_STRUCT(r600_transfer);
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
diff --git a/src/gallium/drivers/r600/r700_asm.c b/src/gallium/drivers/r600/r700_asm.c
index 892dee86baf..a7f2f54736e 100644
--- a/src/gallium/drivers/r600/r700_asm.c
+++ b/src/gallium/drivers/r600/r700_asm.c
@@ -29,8 +29,6 @@
int r700_bc_alu_build(struct r600_bc *bc, struct r600_bc_alu *alu, unsigned id)
{
- unsigned i;
-
bc->bytecode[id++] = S_SQ_ALU_WORD0_SRC0_SEL(alu->src[0].sel) |
S_SQ_ALU_WORD0_SRC0_REL(alu->src[0].rel) |
S_SQ_ALU_WORD0_SRC0_CHAN(alu->src[0].chan) |
@@ -61,18 +59,11 @@ int r700_bc_alu_build(struct r600_bc *bc, struct r600_bc_alu *alu, unsigned id)
S_SQ_ALU_WORD1_OP2_SRC0_ABS(alu->src[0].abs) |
S_SQ_ALU_WORD1_OP2_SRC1_ABS(alu->src[1].abs) |
S_SQ_ALU_WORD1_OP2_WRITE_MASK(alu->dst.write) |
+ S_SQ_ALU_WORD1_OP2_OMOD(alu->omod) |
S_SQ_ALU_WORD1_OP2_ALU_INST(alu->inst) |
S_SQ_ALU_WORD1_BANK_SWIZZLE(alu->bank_swizzle) |
S_SQ_ALU_WORD1_OP2_UPDATE_EXECUTE_MASK(alu->predicate) |
S_SQ_ALU_WORD1_OP2_UPDATE_PRED(alu->predicate);
}
- if (alu->last) {
- if (alu->nliteral && !alu->literal_added) {
- R600_ERR("Bug in ALU processing for instruction 0x%08x, literal not added correctly\n", alu->inst);
- }
- for (i = 0; i < alu->nliteral; i++) {
- bc->bytecode[id++] = alu->value[i];
- }
- }
return 0;
}
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-20 05:57:15 +0000