GLuint msg_control,
GLuint msg_type,
GLuint msg_length,
+ GLboolean header_present,
GLuint pixel_scoreboard_clear,
GLuint response_length,
GLuint end_of_thread,
insn->bits3.dp_render_cache.pixel_scoreboard_clear = pixel_scoreboard_clear;
insn->bits3.dp_render_cache.msg_type = msg_type;
insn->bits3.dp_render_cache.send_commit_msg = send_commit_msg;
- insn->bits3.dp_render_cache.header_present = 0; /* XXX */
+ insn->bits3.dp_render_cache.header_present = header_present;
insn->bits3.dp_render_cache.response_length = response_length;
insn->bits3.dp_render_cache.msg_length = msg_length;
insn->bits3.dp_render_cache.end_of_thread = end_of_thread;
insn->bits3.dp_write_gen5.pixel_scoreboard_clear = pixel_scoreboard_clear;
insn->bits3.dp_write_gen5.msg_type = msg_type;
insn->bits3.dp_write_gen5.send_commit_msg = send_commit_msg;
- insn->bits3.dp_write_gen5.header_present = 1;
+ insn->bits3.dp_write_gen5.header_present = header_present;
insn->bits3.dp_write_gen5.response_length = response_length;
insn->bits3.dp_write_gen5.msg_length = msg_length;
insn->bits3.dp_write_gen5.end_of_thread = end_of_thread;
assert(eot == 0);
brw_set_src1(insn, brw_imm_d(0));
- if (intel->gen == 5) {
+ if (intel->gen >= 5) {
insn->bits3.sampler_gen5.binding_table_index = binding_table_index;
insn->bits3.sampler_gen5.sampler = sampler;
insn->bits3.sampler_gen5.msg_type = msg_type;
insn->bits3.sampler_gen5.response_length = response_length;
insn->bits3.sampler_gen5.msg_length = msg_length;
insn->bits3.sampler_gen5.end_of_thread = eot;
- insn->bits2.send_gen5.sfid = BRW_MESSAGE_TARGET_SAMPLER;
- insn->bits2.send_gen5.end_of_thread = eot;
+ if (intel->gen >= 6)
+ insn->header.destreg__conditionalmod = BRW_MESSAGE_TARGET_SAMPLER;
+ else {
+ insn->bits2.send_gen5.sfid = BRW_MESSAGE_TARGET_SAMPLER;
+ insn->bits2.send_gen5.end_of_thread = eot;
+ }
} else if (intel->is_g4x) {
insn->bits3.sampler_g4x.binding_table_index = binding_table_index;
insn->bits3.sampler_g4x.sampler = sampler;
return brw_alu2(p, BRW_OPCODE_##OP, dest, src0, src1); \
}
+/* Rounding operations (other than RNDD) require two instructions - the first
+ * stores a rounded value (possibly the wrong way) in the dest register, but
+ * also sets a per-channel "increment bit" in the flag register. A predicated
+ * add of 1.0 fixes dest to contain the desired result.
+ */
+#define ROUND(OP) \
+void brw_##OP(struct brw_compile *p, \
+ struct brw_reg dest, \
+ struct brw_reg src) \
+{ \
+ struct brw_instruction *rnd, *add; \
+ rnd = next_insn(p, BRW_OPCODE_##OP); \
+ brw_set_dest(rnd, dest); \
+ brw_set_src0(rnd, src); \
+ rnd->header.destreg__conditionalmod = 0x7; /* turn on round-increments */ \
+ \
+ add = brw_ADD(p, dest, dest, brw_imm_f(1.0f)); \
+ add->header.predicate_control = BRW_PREDICATE_NORMAL; \
+}
+
ALU1(MOV)
ALU2(SEL)
ALU2(ASR)
ALU1(FRC)
ALU1(RNDD)
-ALU1(RNDZ)
ALU2(MAC)
ALU2(MACH)
ALU1(LZD)
ALU2(LINE)
ALU2(PLN)
+
+ROUND(RNDZ)
+ROUND(RNDE)
+
+
struct brw_instruction *brw_ADD(struct brw_compile *p,
struct brw_reg dest,
struct brw_reg src0,
*/
struct brw_instruction *brw_IF(struct brw_compile *p, GLuint execute_size)
{
+ struct intel_context *intel = &p->brw->intel;
struct brw_instruction *insn;
if (p->single_program_flow) {
/* Override the defaults for this instruction:
*/
- brw_set_dest(insn, brw_ip_reg());
- brw_set_src0(insn, brw_ip_reg());
- brw_set_src1(insn, brw_imm_d(0x0));
+ if (intel->gen < 6) {
+ brw_set_dest(insn, brw_ip_reg());
+ brw_set_src0(insn, brw_ip_reg());
+ brw_set_src1(insn, brw_imm_d(0x0));
+ } else {
+ brw_set_dest(insn, brw_imm_w(0));
+ insn->bits1.branch_gen6.jump_count = 0;
+ brw_set_src0(insn, retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
+ brw_set_src1(insn, retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
+ }
insn->header.execution_size = execute_size;
insn->header.compression_control = BRW_COMPRESSION_NONE;
return insn;
}
+struct brw_instruction *
+brw_IF_gen6(struct brw_compile *p, uint32_t conditional,
+ struct brw_reg src0, struct brw_reg src1)
+{
+ struct brw_instruction *insn;
+
+ insn = next_insn(p, BRW_OPCODE_IF);
+
+ brw_set_dest(insn, brw_imm_w(0));
+ insn->header.execution_size = BRW_EXECUTE_8;
+ insn->bits1.branch_gen6.jump_count = 0;
+ brw_set_src0(insn, src0);
+ brw_set_src1(insn, src1);
+
+ assert(insn->header.compression_control == BRW_COMPRESSION_NONE);
+ assert(insn->header.predicate_control == BRW_PREDICATE_NONE);
+ insn->header.destreg__conditionalmod = conditional;
+
+ if (!p->single_program_flow)
+ insn->header.thread_control = BRW_THREAD_SWITCH;
+
+ return insn;
+}
struct brw_instruction *brw_ELSE(struct brw_compile *p,
struct brw_instruction *if_insn)
struct brw_instruction *insn;
GLuint br = 1;
- if (intel->gen == 5)
+ /* jump count is for 64bit data chunk each, so one 128bit
+ instruction requires 2 chunks. */
+ if (intel->gen >= 5)
br = 2;
if (p->single_program_flow) {
insn = next_insn(p, BRW_OPCODE_ELSE);
}
- brw_set_dest(insn, brw_ip_reg());
- brw_set_src0(insn, brw_ip_reg());
- brw_set_src1(insn, brw_imm_d(0x0));
+ if (intel->gen < 6) {
+ brw_set_dest(insn, brw_ip_reg());
+ brw_set_src0(insn, brw_ip_reg());
+ brw_set_src1(insn, brw_imm_d(0x0));
+ } else {
+ brw_set_dest(insn, brw_imm_w(0));
+ insn->bits1.branch_gen6.jump_count = 0;
+ brw_set_src0(insn, retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
+ brw_set_src1(insn, retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
+ }
insn->header.compression_control = BRW_COMPRESSION_NONE;
insn->header.execution_size = if_insn->header.execution_size;
} else {
assert(if_insn->header.opcode == BRW_OPCODE_IF);
- if_insn->bits3.if_else.jump_count = br * (insn - if_insn);
- if_insn->bits3.if_else.pop_count = 0;
- if_insn->bits3.if_else.pad0 = 0;
+ if (intel->gen < 6) {
+ if_insn->bits3.if_else.jump_count = br * (insn - if_insn);
+ if_insn->bits3.if_else.pop_count = 0;
+ if_insn->bits3.if_else.pad0 = 0;
+ } else {
+ if_insn->bits1.branch_gen6.jump_count = br * (insn - if_insn + 1);
+ }
}
return insn;
struct intel_context *intel = &p->brw->intel;
GLuint br = 1;
- if (intel->gen == 5)
+ if (intel->gen >= 5)
br = 2;
if (p->single_program_flow) {
} else {
struct brw_instruction *insn = next_insn(p, BRW_OPCODE_ENDIF);
- brw_set_dest(insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
- brw_set_src0(insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
- brw_set_src1(insn, brw_imm_d(0x0));
+ if (intel->gen < 6) {
+ brw_set_dest(insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
+ brw_set_src0(insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
+ brw_set_src1(insn, brw_imm_d(0x0));
+ } else {
+ brw_set_dest(insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_W));
+ brw_set_src0(insn, retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
+ brw_set_src1(insn, retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
+ }
insn->header.compression_control = BRW_COMPRESSION_NONE;
insn->header.execution_size = patch_insn->header.execution_size;
insn->header.mask_control = BRW_MASK_ENABLE;
insn->header.thread_control = BRW_THREAD_SWITCH;
- assert(patch_insn->bits3.if_else.jump_count == 0);
+ if (intel->gen < 6)
+ assert(patch_insn->bits3.if_else.jump_count == 0);
+ else
+ assert(patch_insn->bits1.branch_gen6.jump_count == 0);
/* Patch the if or else instructions to point at this or the next
* instruction respectively.
*/
if (patch_insn->header.opcode == BRW_OPCODE_IF) {
- /* Automagically turn it into an IFF:
- */
- patch_insn->header.opcode = BRW_OPCODE_IFF;
- patch_insn->bits3.if_else.jump_count = br * (insn - patch_insn + 1);
- patch_insn->bits3.if_else.pop_count = 0;
- patch_insn->bits3.if_else.pad0 = 0;
- } else if (patch_insn->header.opcode == BRW_OPCODE_ELSE) {
- patch_insn->bits3.if_else.jump_count = br * (insn - patch_insn + 1);
- patch_insn->bits3.if_else.pop_count = 1;
- patch_insn->bits3.if_else.pad0 = 0;
+ if (intel->gen < 6) {
+ /* Turn it into an IFF, which means no mask stack operations for
+ * all-false and jumping past the ENDIF.
+ */
+ patch_insn->header.opcode = BRW_OPCODE_IFF;
+ patch_insn->bits3.if_else.jump_count = br * (insn - patch_insn + 1);
+ patch_insn->bits3.if_else.pop_count = 0;
+ patch_insn->bits3.if_else.pad0 = 0;
+ } else {
+ /* As of gen6, there is no IFF and IF must point to the ENDIF. */
+ patch_insn->bits1.branch_gen6.jump_count = br * (insn - patch_insn);
+ }
} else {
- assert(0);
+ assert(patch_insn->header.opcode == BRW_OPCODE_ELSE);
+ if (intel->gen < 6) {
+ /* BRW_OPCODE_ELSE pre-gen6 should point just past the
+ * matching ENDIF.
+ */
+ patch_insn->bits3.if_else.jump_count = br * (insn - patch_insn + 1);
+ patch_insn->bits3.if_else.pop_count = 1;
+ patch_insn->bits3.if_else.pad0 = 0;
+ } else {
+ /* BRW_OPCODE_ELSE on gen6 should point to the matching ENDIF. */
+ patch_insn->bits1.branch_gen6.jump_count = br * (insn - patch_insn);
+ }
}
/* Also pop item off the stack in the endif instruction:
*/
- insn->bits3.if_else.jump_count = 0;
- insn->bits3.if_else.pop_count = 1;
- insn->bits3.if_else.pad0 = 0;
+ if (intel->gen < 6) {
+ insn->bits3.if_else.jump_count = 0;
+ insn->bits3.if_else.pop_count = 1;
+ insn->bits3.if_else.pad0 = 0;
+ } else {
+ insn->bits1.branch_gen6.jump_count = 2;
+ }
}
}
struct brw_instruction *insn;
GLuint br = 1;
- if (intel->gen == 5)
+ if (intel->gen >= 5)
br = 2;
if (p->single_program_flow)
struct brw_instruction *landing = &p->store[p->nr_insn];
GLuint jmpi = 1;
- if (intel->gen == 5)
+ if (intel->gen >= 5)
jmpi = 2;
assert(jmp_insn->header.opcode == BRW_OPCODE_JMPI);
if (intel->gen >= 6) {
struct brw_instruction *insn = next_insn(p, BRW_OPCODE_MATH);
+ assert(dest.file == BRW_GENERAL_REGISTER_FILE);
+ assert(src.file == BRW_GENERAL_REGISTER_FILE);
+
+ assert(dest.hstride == BRW_HORIZONTAL_STRIDE_1);
+ assert(src.hstride == BRW_HORIZONTAL_STRIDE_1);
+
+ if (function != BRW_MATH_FUNCTION_INT_DIV_QUOTIENT &&
+ function != BRW_MATH_FUNCTION_INT_DIV_QUOTIENT_AND_REMAINDER) {
+ assert(src.type == BRW_REGISTER_TYPE_F);
+ }
+
/* Math is the same ISA format as other opcodes, except that CondModifier
* becomes FC[3:0] and ThreadCtrl becomes FC[5:4].
*/
struct brw_instruction *insn = next_insn(p, BRW_OPCODE_MATH);
assert(intel->gen >= 6);
+ (void) intel;
+
+
+ assert(dest.file == BRW_GENERAL_REGISTER_FILE);
+ assert(src0.file == BRW_GENERAL_REGISTER_FILE);
+ assert(src1.file == BRW_GENERAL_REGISTER_FILE);
+
+ assert(dest.hstride == BRW_HORIZONTAL_STRIDE_1);
+ assert(src0.hstride == BRW_HORIZONTAL_STRIDE_1);
+ assert(src1.hstride == BRW_HORIZONTAL_STRIDE_1);
+
+ if (function != BRW_MATH_FUNCTION_INT_DIV_QUOTIENT &&
+ function != BRW_MATH_FUNCTION_INT_DIV_QUOTIENT_AND_REMAINDER) {
+ assert(src0.type == BRW_REGISTER_TYPE_F);
+ assert(src1.type == BRW_REGISTER_TYPE_F);
+ }
/* Math is the same ISA format as other opcodes, except that CondModifier
* becomes FC[3:0] and ThreadCtrl becomes FC[5:4].
BRW_DATAPORT_OWORD_BLOCK_4_OWORDS, /* msg_control */
BRW_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE, /* msg_type */
msg_length,
+ GL_TRUE, /* header_present */
0, /* pixel scoreboard */
send_commit_msg, /* response_length */
0, /* eot */
struct intel_context *intel = &p->brw->intel;
struct brw_instruction *insn;
GLuint msg_control, msg_type;
+ GLboolean header_present = GL_TRUE;
insn = next_insn(p, BRW_OPCODE_SEND);
insn->header.predicate_control = 0; /* XXX */
insn->header.compression_control = BRW_COMPRESSION_NONE;
if (intel->gen >= 6) {
+ if (msg_length == 4)
+ header_present = GL_FALSE;
+
/* headerless version, just submit color payload */
src0 = brw_message_reg(msg_reg_nr);
msg_control,
msg_type,
msg_length,
+ header_present,
1, /* pixel scoreboard */
response_length,
eot,
GLuint header_present,
GLuint simd_mode)
{
+ struct intel_context *intel = &p->brw->intel;
GLboolean need_stall = 0;
if (writemask == 0) {
}
{
- struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
+ struct brw_instruction *insn;
+ /* Sandybridge doesn't have the implied move for SENDs,
+ * and the first message register index comes from src0.
+ */
+ if (intel->gen >= 6) {
+ brw_push_insn_state(p);
+ brw_set_mask_control( p, BRW_MASK_DISABLE );
+ /* m1 contains header? */
+ brw_MOV(p, brw_message_reg(msg_reg_nr), src0);
+ brw_pop_insn_state(p);
+ src0 = brw_message_reg(msg_reg_nr);
+ }
+
+ insn = next_insn(p, BRW_OPCODE_SEND);
insn->header.predicate_control = 0; /* XXX */
insn->header.compression_control = BRW_COMPRESSION_NONE;
- insn->header.destreg__conditionalmod = msg_reg_nr;
+ if (intel->gen < 6)
+ insn->header.destreg__conditionalmod = msg_reg_nr;
brw_set_dest(insn, dest);
brw_set_src0(insn, src0);
GLuint response_length,
GLboolean eot)
{
- struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
+ struct intel_context *intel = &p->brw->intel;
+ struct brw_instruction *insn;
+
+ /* Sandybridge doesn't have the implied move for SENDs,
+ * and the first message register index comes from src0.
+ */
+ if (intel->gen >= 6) {
+ brw_push_insn_state(p);
+ brw_set_mask_control( p, BRW_MASK_DISABLE );
+ brw_MOV(p, retype(brw_message_reg(msg_reg_nr), BRW_REGISTER_TYPE_UD),
+ retype(src0, BRW_REGISTER_TYPE_UD));
+ brw_pop_insn_state(p);
+ src0 = brw_message_reg(msg_reg_nr);
+ }
+ insn = next_insn(p, BRW_OPCODE_SEND);
brw_set_dest(insn, dest);
brw_set_src0(insn, src0);
brw_set_src1(insn, brw_imm_d(0));
- insn->header.destreg__conditionalmod = msg_reg_nr;
+ if (intel->gen < 6)
+ insn->header.destreg__conditionalmod = msg_reg_nr;
brw_set_ff_sync_message(p->brw,
insn,
projects / mesa.git / blobdiff