*/
unsigned
brw_reg_type_to_hw_type(const struct brw_device_info *devinfo,
- enum brw_reg_type type, unsigned file)
+ enum brw_reg_type type, enum brw_reg_file file)
{
if (file == BRW_IMMEDIATE_VALUE) {
- const static int imm_hw_types[] = {
+ static const int imm_hw_types[] = {
[BRW_REGISTER_TYPE_UD] = BRW_HW_REG_TYPE_UD,
[BRW_REGISTER_TYPE_D] = BRW_HW_REG_TYPE_D,
[BRW_REGISTER_TYPE_UW] = BRW_HW_REG_TYPE_UW,
return imm_hw_types[type];
} else {
/* Non-immediate registers */
- const static int hw_types[] = {
+ static const int hw_types[] = {
[BRW_REGISTER_TYPE_UD] = BRW_HW_REG_TYPE_UD,
[BRW_REGISTER_TYPE_D] = BRW_HW_REG_TYPE_D,
[BRW_REGISTER_TYPE_UW] = BRW_HW_REG_TYPE_UW,
{
const struct brw_device_info *devinfo = p->devinfo;
- if (dest.file != BRW_ARCHITECTURE_REGISTER_FILE &&
- dest.file != BRW_MESSAGE_REGISTER_FILE)
+ if (dest.file == BRW_MESSAGE_REGISTER_FILE)
+ assert((dest.nr & ~BRW_MRF_COMPR4) < BRW_MAX_MRF(devinfo->gen));
+ else if (dest.file != BRW_ARCHITECTURE_REGISTER_FILE)
assert(dest.nr < 128);
gen7_convert_mrf_to_grf(p, &dest);
brw_inst_set_dst_hstride(devinfo, inst, dest.hstride);
} else {
brw_inst_set_dst_da16_subreg_nr(devinfo, inst, dest.subnr / 16);
- brw_inst_set_da16_writemask(devinfo, inst, dest.dw1.bits.writemask);
+ brw_inst_set_da16_writemask(devinfo, inst, dest.writemask);
if (dest.file == BRW_GENERAL_REGISTER_FILE ||
dest.file == BRW_MESSAGE_REGISTER_FILE) {
- assert(dest.dw1.bits.writemask != 0);
+ assert(dest.writemask != 0);
}
/* From the Ivybridge PRM, Vol 4, Part 3, Section 5.2.4.1:
* Although Dst.HorzStride is a don't care for Align16, HW needs
*/
if (brw_inst_access_mode(devinfo, inst) == BRW_ALIGN_1) {
brw_inst_set_dst_ia1_addr_imm(devinfo, inst,
- dest.dw1.bits.indirect_offset);
+ dest.indirect_offset);
if (dest.hstride == BRW_HORIZONTAL_STRIDE_0)
dest.hstride = BRW_HORIZONTAL_STRIDE_1;
brw_inst_set_dst_hstride(devinfo, inst, dest.hstride);
} else {
brw_inst_set_dst_ia16_addr_imm(devinfo, inst,
- dest.dw1.bits.indirect_offset);
+ dest.indirect_offset);
/* even ignored in da16, still need to set as '01' */
brw_inst_set_dst_hstride(devinfo, inst, 1);
}
/* Generators should set a default exec_size of either 8 (SIMD4x2 or SIMD8)
* or 16 (SIMD16), as that's normally correct. However, when dealing with
* small registers, we automatically reduce it to match the register size.
+ *
+ * In platforms that support fp64 we can emit instructions with a width of
+ * 4 that need two SIMD8 registers and an exec_size of 8 or 16. In these
+ * cases we need to make sure that these instructions have their exec sizes
+ * set properly when they are emitted and we can't rely on this code to fix
+ * it.
*/
- if (dest.width < BRW_EXECUTE_8)
+ bool fix_exec_size;
+ if (devinfo->gen >= 6)
+ fix_exec_size = dest.width < BRW_EXECUTE_4;
+ else
+ fix_exec_size = dest.width < BRW_EXECUTE_8;
+
+ if (fix_exec_size)
brw_inst_set_exec_size(devinfo, inst, dest.width);
}
reg.file == BRW_ARF_NULL)
return;
+ /* From the IVB PRM Vol. 4, Pt. 3, Section 3.3.3.5:
+ *
+ * "Swizzling is not allowed when an accumulator is used as an implicit
+ * source or an explicit source in an instruction."
+ */
+ if (reg.file == BRW_ARCHITECTURE_REGISTER_FILE &&
+ reg.nr == BRW_ARF_ACCUMULATOR)
+ assert(reg.swizzle == BRW_SWIZZLE_XYZW);
+
assert(reg.hstride >= 0 && reg.hstride < ARRAY_SIZE(hstride_for_reg));
hstride = hstride_for_reg[reg.hstride];
{
const struct brw_device_info *devinfo = p->devinfo;
- if (reg.file != BRW_ARCHITECTURE_REGISTER_FILE)
+ if (reg.file == BRW_MESSAGE_REGISTER_FILE)
+ assert((reg.nr & ~BRW_MRF_COMPR4) < BRW_MAX_MRF(devinfo->gen));
+ else if (reg.file != BRW_ARCHITECTURE_REGISTER_FILE)
assert(reg.nr < 128);
gen7_convert_mrf_to_grf(p, ®);
brw_inst_set_src0_address_mode(devinfo, inst, reg.address_mode);
if (reg.file == BRW_IMMEDIATE_VALUE) {
- brw_inst_set_imm_ud(devinfo, inst, reg.dw1.ud);
+ if (reg.type == BRW_REGISTER_TYPE_DF)
+ brw_inst_set_imm_df(devinfo, inst, reg.df);
+ else
+ brw_inst_set_imm_ud(devinfo, inst, reg.ud);
/* The Bspec's section titled "Non-present Operands" claims that if src0
* is an immediate that src1's type must be the same as that of src0.
* The GM45 instruction compaction tables do not contain mapped meanings
* so it's not clear whether it has the restriction. We'll assume it was
* lifted on SNB. (FINISHME: decode the GM45 tables and check.)
+ *
+ * Don't do any of this for 64-bit immediates, since the src1 fields
+ * overlap with the immediate and setting them would overwrite the
+ * immediate we set.
*/
- brw_inst_set_src1_reg_file(devinfo, inst, BRW_ARCHITECTURE_REGISTER_FILE);
- if (devinfo->gen < 6) {
- brw_inst_set_src1_reg_type(devinfo, inst,
- brw_inst_src0_reg_type(devinfo, inst));
- } else {
- brw_inst_set_src1_reg_type(devinfo, inst, BRW_HW_REG_TYPE_UD);
+ if (type_sz(reg.type) < 8) {
+ brw_inst_set_src1_reg_file(devinfo, inst,
+ BRW_ARCHITECTURE_REGISTER_FILE);
+ if (devinfo->gen < 6) {
+ brw_inst_set_src1_reg_type(devinfo, inst,
+ brw_inst_src0_reg_type(devinfo, inst));
+ } else {
+ brw_inst_set_src1_reg_type(devinfo, inst, BRW_HW_REG_TYPE_UD);
+ }
}
/* Compacted instructions only have 12-bits (plus 1 for the other 20)
brw_inst_set_src0_ia_subreg_nr(devinfo, inst, reg.subnr);
if (brw_inst_access_mode(devinfo, inst) == BRW_ALIGN_1) {
- brw_inst_set_src0_ia1_addr_imm(devinfo, inst, reg.dw1.bits.indirect_offset);
+ brw_inst_set_src0_ia1_addr_imm(devinfo, inst, reg.indirect_offset);
} else {
- brw_inst_set_src0_ia_subreg_nr(devinfo, inst, reg.dw1.bits.indirect_offset);
+ brw_inst_set_src0_ia16_addr_imm(devinfo, inst, reg.indirect_offset);
}
}
}
} else {
brw_inst_set_src0_da16_swiz_x(devinfo, inst,
- BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_X));
+ BRW_GET_SWZ(reg.swizzle, BRW_CHANNEL_X));
brw_inst_set_src0_da16_swiz_y(devinfo, inst,
- BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_Y));
+ BRW_GET_SWZ(reg.swizzle, BRW_CHANNEL_Y));
brw_inst_set_src0_da16_swiz_z(devinfo, inst,
- BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_Z));
+ BRW_GET_SWZ(reg.swizzle, BRW_CHANNEL_Z));
brw_inst_set_src0_da16_swiz_w(devinfo, inst,
- BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_W));
+ BRW_GET_SWZ(reg.swizzle, BRW_CHANNEL_W));
/* This is an oddity of the fact we're using the same
* descriptions for registers in align_16 as align_1:
if (reg.file != BRW_ARCHITECTURE_REGISTER_FILE)
assert(reg.nr < 128);
+ /* From the IVB PRM Vol. 4, Pt. 3, Section 3.3.3.5:
+ *
+ * "Accumulator registers may be accessed explicitly as src0
+ * operands only."
+ */
+ assert(reg.file != BRW_ARCHITECTURE_REGISTER_FILE ||
+ reg.nr != BRW_ARF_ACCUMULATOR);
+
gen7_convert_mrf_to_grf(p, ®);
assert(reg.file != BRW_MESSAGE_REGISTER_FILE);
assert(brw_inst_src0_reg_file(devinfo, inst) != BRW_IMMEDIATE_VALUE);
if (reg.file == BRW_IMMEDIATE_VALUE) {
- brw_inst_set_imm_ud(devinfo, inst, reg.dw1.ud);
+ /* two-argument instructions can only use 32-bit immediates */
+ assert(type_sz(reg.type) < 8);
+ brw_inst_set_imm_ud(devinfo, inst, reg.ud);
} else {
/* This is a hardware restriction, which may or may not be lifted
* in the future:
}
} else {
brw_inst_set_src1_da16_swiz_x(devinfo, inst,
- BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_X));
+ BRW_GET_SWZ(reg.swizzle, BRW_CHANNEL_X));
brw_inst_set_src1_da16_swiz_y(devinfo, inst,
- BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_Y));
+ BRW_GET_SWZ(reg.swizzle, BRW_CHANNEL_Y));
brw_inst_set_src1_da16_swiz_z(devinfo, inst,
- BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_Z));
+ BRW_GET_SWZ(reg.swizzle, BRW_CHANNEL_Z));
brw_inst_set_src1_da16_swiz_w(devinfo, inst,
- BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_W));
+ BRW_GET_SWZ(reg.swizzle, BRW_CHANNEL_W));
/* This is an oddity of the fact we're using the same
* descriptions for registers in align_16 as align_1:
* \b before filling out any message-specific data. Callers can
* choose not to fill in irrelevant bits; they will be zero.
*/
-static void
+void
brw_set_message_descriptor(struct brw_codegen *p,
brw_inst *inst,
enum brw_message_target sfid,
const struct brw_device_info *devinfo = p->devinfo;
assert(num_regs == 1 || num_regs == 2 || num_regs == 4 ||
(devinfo->gen >= 8 && num_regs == 8));
+ const unsigned block_size = (devinfo->gen >= 8 ? _mesa_logbase2(num_regs) :
+ num_regs - 1);
+
brw_set_message_descriptor(p, inst, GEN7_SFID_DATAPORT_DATA_CACHE,
mlen, rlen, header_present, false);
brw_inst_set_dp_category(devinfo, inst, 1); /* Scratch Block Read/Write msgs */
brw_inst_set_scratch_read_write(devinfo, inst, write);
brw_inst_set_scratch_type(devinfo, inst, dword);
brw_inst_set_scratch_invalidate_after_read(devinfo, inst, invalidate_after_read);
- brw_inst_set_scratch_block_size(devinfo, inst, ffs(num_regs) - 1);
+ brw_inst_set_scratch_block_size(devinfo, inst, block_size);
brw_inst_set_scratch_addr_offset(devinfo, inst, addr_offset);
}
brw_alu2(struct brw_codegen *p, unsigned opcode,
struct brw_reg dest, struct brw_reg src0, struct brw_reg src1)
{
+ /* 64-bit immediates are only supported on 1-src instructions */
+ assert(src0.file != BRW_IMMEDIATE_VALUE || type_sz(src0.type) <= 4);
+ assert(src1.file != BRW_IMMEDIATE_VALUE || type_sz(src1.type) <= 4);
+
brw_inst *insn = next_insn(p, opcode);
brw_set_dest(p, insn, dest);
brw_set_src0(p, insn, src0);
static int
get_3src_subreg_nr(struct brw_reg reg)
{
- if (reg.vstride == BRW_VERTICAL_STRIDE_0) {
- assert(brw_is_single_value_swizzle(reg.dw1.bits.swizzle));
- return reg.subnr / 4 + BRW_GET_SWZ(reg.dw1.bits.swizzle, 0);
- } else {
- return reg.subnr / 4;
- }
+ /* Normally, SubRegNum is in bytes (0..31). However, 3-src instructions
+ * use 32-bit units (components 0..7). Since they only support F/D/UD
+ * types, this doesn't lose any flexibility, but uses fewer bits.
+ */
+ return reg.subnr / 4;
}
static brw_inst *
dest.file == BRW_MESSAGE_REGISTER_FILE);
assert(dest.nr < 128);
assert(dest.address_mode == BRW_ADDRESS_DIRECT);
- assert(dest.type == BRW_REGISTER_TYPE_F ||
- dest.type == BRW_REGISTER_TYPE_D ||
+ assert(dest.type == BRW_REGISTER_TYPE_F ||
+ dest.type == BRW_REGISTER_TYPE_DF ||
+ dest.type == BRW_REGISTER_TYPE_D ||
dest.type == BRW_REGISTER_TYPE_UD);
if (devinfo->gen == 6) {
brw_inst_set_3src_dst_reg_file(devinfo, inst,
}
brw_inst_set_3src_dst_reg_nr(devinfo, inst, dest.nr);
brw_inst_set_3src_dst_subreg_nr(devinfo, inst, dest.subnr / 16);
- brw_inst_set_3src_dst_writemask(devinfo, inst, dest.dw1.bits.writemask);
+ brw_inst_set_3src_dst_writemask(devinfo, inst, dest.writemask);
assert(src0.file == BRW_GENERAL_REGISTER_FILE);
assert(src0.address_mode == BRW_ADDRESS_DIRECT);
assert(src0.nr < 128);
- brw_inst_set_3src_src0_swizzle(devinfo, inst, src0.dw1.bits.swizzle);
+ brw_inst_set_3src_src0_swizzle(devinfo, inst, src0.swizzle);
brw_inst_set_3src_src0_subreg_nr(devinfo, inst, get_3src_subreg_nr(src0));
brw_inst_set_3src_src0_reg_nr(devinfo, inst, src0.nr);
brw_inst_set_3src_src0_abs(devinfo, inst, src0.abs);
assert(src1.file == BRW_GENERAL_REGISTER_FILE);
assert(src1.address_mode == BRW_ADDRESS_DIRECT);
assert(src1.nr < 128);
- brw_inst_set_3src_src1_swizzle(devinfo, inst, src1.dw1.bits.swizzle);
+ brw_inst_set_3src_src1_swizzle(devinfo, inst, src1.swizzle);
brw_inst_set_3src_src1_subreg_nr(devinfo, inst, get_3src_subreg_nr(src1));
brw_inst_set_3src_src1_reg_nr(devinfo, inst, src1.nr);
brw_inst_set_3src_src1_abs(devinfo, inst, src1.abs);
assert(src2.file == BRW_GENERAL_REGISTER_FILE);
assert(src2.address_mode == BRW_ADDRESS_DIRECT);
assert(src2.nr < 128);
- brw_inst_set_3src_src2_swizzle(devinfo, inst, src2.dw1.bits.swizzle);
+ brw_inst_set_3src_src2_swizzle(devinfo, inst, src2.swizzle);
brw_inst_set_3src_src2_subreg_nr(devinfo, inst, get_3src_subreg_nr(src2));
brw_inst_set_3src_src2_reg_nr(devinfo, inst, src2.nr);
brw_inst_set_3src_src2_abs(devinfo, inst, src2.abs);
brw_inst_set_3src_src_type(devinfo, inst, BRW_3SRC_TYPE_F);
brw_inst_set_3src_dst_type(devinfo, inst, BRW_3SRC_TYPE_F);
break;
+ case BRW_REGISTER_TYPE_DF:
+ brw_inst_set_3src_src_type(devinfo, inst, BRW_3SRC_TYPE_DF);
+ brw_inst_set_3src_dst_type(devinfo, inst, BRW_3SRC_TYPE_DF);
+ break;
case BRW_REGISTER_TYPE_D:
brw_inst_set_3src_src_type(devinfo, inst, BRW_3SRC_TYPE_D);
brw_inst_set_3src_dst_type(devinfo, inst, BRW_3SRC_TYPE_D);
brw_inst_set_3src_src_type(devinfo, inst, BRW_3SRC_TYPE_UD);
brw_inst_set_3src_dst_type(devinfo, inst, BRW_3SRC_TYPE_UD);
break;
+ default:
+ unreachable("not reached");
}
}
struct brw_reg src1, \
struct brw_reg src2) \
{ \
- assert(dest.type == BRW_REGISTER_TYPE_F); \
- assert(src0.type == BRW_REGISTER_TYPE_F); \
- assert(src1.type == BRW_REGISTER_TYPE_F); \
- assert(src2.type == BRW_REGISTER_TYPE_F); \
+ assert(dest.type == BRW_REGISTER_TYPE_F || \
+ dest.type == BRW_REGISTER_TYPE_DF); \
+ if (dest.type == BRW_REGISTER_TYPE_F) { \
+ assert(src0.type == BRW_REGISTER_TYPE_F); \
+ assert(src1.type == BRW_REGISTER_TYPE_F); \
+ assert(src2.type == BRW_REGISTER_TYPE_F); \
+ } else if (dest.type == BRW_REGISTER_TYPE_DF) { \
+ assert(src0.type == BRW_REGISTER_TYPE_DF); \
+ assert(src1.type == BRW_REGISTER_TYPE_DF); \
+ assert(src2.type == BRW_REGISTER_TYPE_DF); \
+ } \
return brw_alu3(p, BRW_OPCODE_##OP, dest, src0, src1, src2); \
}
void brw_NOP(struct brw_codegen *p)
{
brw_inst *insn = next_insn(p, BRW_OPCODE_NOP);
- brw_set_dest(p, insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
- brw_set_src0(p, insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
+ brw_inst_set_exec_size(p->devinfo, insn, BRW_EXECUTE_1);
+ brw_set_dest(p, insn, retype(brw_vec1_grf(0,0), BRW_REGISTER_TYPE_UD));
+ brw_set_src0(p, insn, retype(brw_vec1_grf(0,0), BRW_REGISTER_TYPE_UD));
brw_set_src1(p, insn, brw_imm_ud(0x0));
}
static void
push_loop_stack(struct brw_codegen *p, brw_inst *inst)
{
- if (p->loop_stack_array_size < p->loop_stack_depth) {
+ if (p->loop_stack_array_size <= (p->loop_stack_depth + 1)) {
p->loop_stack_array_size *= 2;
p->loop_stack = reralloc(p->mem_ctx, p->loop_stack, int,
p->loop_stack_array_size);
*
* When the matching 'else' instruction is reached (presumably by
* countdown of the instruction count patched in by our ELSE/ENDIF
- * functions), the relevent flags are inverted.
+ * functions), the relevant flags are inverted.
*
* When the matching 'endif' instruction is reached, the flags are
* popped off. If the stack is now empty, normal execution resumes.
insn = next_insn(p, BRW_OPCODE_IF);
brw_set_dest(p, insn, brw_imm_w(0));
- brw_inst_set_exec_size(devinfo, insn, p->compressed ? BRW_EXECUTE_16
- : BRW_EXECUTE_8);
+ brw_inst_set_exec_size(devinfo, insn,
+ brw_inst_exec_size(devinfo, p->current));
brw_inst_set_gen6_jump_count(devinfo, insn, 0);
brw_set_src0(p, insn, src0);
brw_set_src1(p, insn, src1);
emit_endif = false;
/*
- * A single next_insn() may change the base adress of instruction store
+ * A single next_insn() may change the base address of instruction store
* memory(p->store), so call it first before referencing the instruction
* store pointer from an index
*/
}
if (devinfo->gen < 6) {
- brw_set_dest(p, insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
- brw_set_src0(p, insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
+ brw_set_dest(p, insn, retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
+ brw_set_src0(p, insn, retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
brw_set_src1(p, insn, brw_imm_d(0x0));
} else if (devinfo->gen == 6) {
brw_set_dest(p, insn, brw_imm_w(0));
p->if_depth_in_loop[p->loop_stack_depth]);
}
brw_inst_set_qtr_control(devinfo, insn, BRW_COMPRESSION_NONE);
- brw_inst_set_exec_size(devinfo, insn, p->compressed ? BRW_EXECUTE_16
- : BRW_EXECUTE_8);
+ brw_inst_set_exec_size(devinfo, insn,
+ brw_inst_exec_size(devinfo, p->current));
return insn;
}
p->if_depth_in_loop[p->loop_stack_depth]);
}
brw_inst_set_qtr_control(devinfo, insn, BRW_COMPRESSION_NONE);
- brw_inst_set_exec_size(devinfo, insn, p->compressed ? BRW_EXECUTE_16
- : BRW_EXECUTE_8);
+ brw_inst_set_exec_size(devinfo, insn,
+ brw_inst_exec_size(devinfo, p->current));
return insn;
}
brw_set_src1(p, insn, brw_imm_d(0x0)); /* UIP and JIP, updated later. */
}
- if (p->compressed) {
- brw_inst_set_exec_size(devinfo, insn, BRW_EXECUTE_16);
- } else {
- brw_inst_set_qtr_control(devinfo, insn, BRW_COMPRESSION_NONE);
- brw_inst_set_exec_size(devinfo, insn, BRW_EXECUTE_8);
- }
+ brw_inst_set_qtr_control(devinfo, insn, BRW_COMPRESSION_NONE);
+ brw_inst_set_exec_size(devinfo, insn,
+ brw_inst_exec_size(devinfo, p->current));
return insn;
}
brw_set_src1(p, insn, retype(brw_null_reg(), BRW_REGISTER_TYPE_D));
}
- brw_inst_set_exec_size(devinfo, insn, p->compressed ? BRW_EXECUTE_16
- : BRW_EXECUTE_8);
+ brw_inst_set_exec_size(devinfo, insn,
+ brw_inst_exec_size(devinfo, p->current));
+
} else {
if (p->single_program_flow) {
insn = next_insn(p, BRW_OPCODE_ADD);
assert(dest.hstride == BRW_HORIZONTAL_STRIDE_1);
if (devinfo->gen == 6) {
- assert(src0.hstride == BRW_HORIZONTAL_STRIDE_1);
- assert(src1.hstride == BRW_HORIZONTAL_STRIDE_1);
+ assert(has_scalar_region(src0) ||
+ src0.hstride == BRW_HORIZONTAL_STRIDE_1);
+ assert(has_scalar_region(src1) ||
+ src1.hstride == BRW_HORIZONTAL_STRIDE_1);
}
if (function == BRW_MATH_FUNCTION_INT_DIV_QUOTIENT ||
brw_set_src1(p, insn, src1);
}
+/**
+ * Return the right surface index to access the thread scratch space using
+ * stateless dataport messages.
+ */
+unsigned
+brw_scratch_surface_idx(const struct brw_codegen *p)
+{
+ /* The scratch space is thread-local so IA coherency is unnecessary. */
+ if (p->devinfo->gen >= 8)
+ return GEN8_BTI_STATELESS_NON_COHERENT;
+ else
+ return BRW_BTI_STATELESS;
+}
/**
* Write a block of OWORDs (half a GRF each) from the scratch buffer,
unsigned offset)
{
const struct brw_device_info *devinfo = p->devinfo;
- uint32_t msg_control, msg_type;
- int mlen;
+ uint32_t msg_type;
if (devinfo->gen >= 6)
offset /= 16;
mrf = retype(mrf, BRW_REGISTER_TYPE_UD);
- if (num_regs == 1) {
- msg_control = BRW_DATAPORT_OWORD_BLOCK_2_OWORDS;
- mlen = 2;
- } else {
- msg_control = BRW_DATAPORT_OWORD_BLOCK_4_OWORDS;
- mlen = 3;
- }
+ const unsigned mlen = 1 + num_regs;
+ const unsigned msg_control =
+ (num_regs == 1 ? BRW_DATAPORT_OWORD_BLOCK_2_OWORDS :
+ num_regs == 2 ? BRW_DATAPORT_OWORD_BLOCK_4_OWORDS :
+ num_regs == 4 ? BRW_DATAPORT_OWORD_BLOCK_8_OWORDS : 0);
+ assert(msg_control);
/* Set up the message header. This is g0, with g0.2 filled with
* the offset. We don't want to leave our offset around in g0 or
struct brw_reg src_header = retype(brw_vec8_grf(0, 0),
BRW_REGISTER_TYPE_UW);
- if (brw_inst_qtr_control(devinfo, insn) != BRW_COMPRESSION_NONE) {
- brw_inst_set_qtr_control(devinfo, insn, BRW_COMPRESSION_NONE);
+ brw_inst_set_compression(devinfo, insn, false);
+
+ if (brw_inst_exec_size(devinfo, insn) >= 16)
src_header = vec16(src_header);
- }
+
assert(brw_inst_pred_control(devinfo, insn) == BRW_PREDICATE_NONE);
if (devinfo->gen < 6)
brw_inst_set_base_mrf(devinfo, insn, mrf.nr);
brw_set_dp_write_message(p,
insn,
- 255, /* binding table index (255=stateless) */
+ brw_scratch_surface_idx(p),
msg_control,
msg_type,
mlen,
unsigned offset)
{
const struct brw_device_info *devinfo = p->devinfo;
- uint32_t msg_control;
- int rlen;
if (devinfo->gen >= 6)
offset /= 16;
}
dest = retype(dest, BRW_REGISTER_TYPE_UW);
- if (num_regs == 1) {
- msg_control = BRW_DATAPORT_OWORD_BLOCK_2_OWORDS;
- rlen = 1;
- } else {
- msg_control = BRW_DATAPORT_OWORD_BLOCK_4_OWORDS;
- rlen = 2;
- }
+ const unsigned rlen = num_regs;
+ const unsigned msg_control =
+ (num_regs == 1 ? BRW_DATAPORT_OWORD_BLOCK_2_OWORDS :
+ num_regs == 2 ? BRW_DATAPORT_OWORD_BLOCK_4_OWORDS :
+ num_regs == 4 ? BRW_DATAPORT_OWORD_BLOCK_8_OWORDS : 0);
+ assert(msg_control);
{
brw_push_insn_state(p);
brw_inst *insn = next_insn(p, BRW_OPCODE_SEND);
assert(brw_inst_pred_control(devinfo, insn) == 0);
- brw_inst_set_qtr_control(devinfo, insn, BRW_COMPRESSION_NONE);
+ brw_inst_set_compression(devinfo, insn, false);
brw_set_dest(p, insn, dest); /* UW? */
if (devinfo->gen >= 6) {
brw_set_dp_read_message(p,
insn,
- 255, /* binding table index (255=stateless) */
+ brw_scratch_surface_idx(p),
msg_control,
BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ, /* msg_type */
BRW_DATAPORT_READ_TARGET_RENDER_CACHE,
brw_inst *insn = next_insn(p, BRW_OPCODE_SEND);
assert(brw_inst_pred_control(devinfo, insn) == BRW_PREDICATE_NONE);
- brw_inst_set_qtr_control(devinfo, insn, BRW_COMPRESSION_NONE);
brw_set_dest(p, insn, retype(dest, BRW_REGISTER_TYPE_UW));
/* The HW requires that the header is present; this is to get the g0.5
void brw_fb_WRITE(struct brw_codegen *p,
- int dispatch_width,
struct brw_reg payload,
struct brw_reg implied_header,
unsigned msg_control,
unsigned msg_type;
struct brw_reg dest, src0;
- if (dispatch_width == 16)
+ if (brw_inst_exec_size(devinfo, p->current) >= BRW_EXECUTE_16)
dest = retype(vec16(brw_null_reg()), BRW_REGISTER_TYPE_UW);
else
dest = retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW);
} else {
insn = next_insn(p, BRW_OPCODE_SEND);
}
- brw_inst_set_qtr_control(devinfo, insn, BRW_COMPRESSION_NONE);
+ brw_inst_set_compression(devinfo, insn, false);
if (devinfo->gen >= 6) {
/* headerless version, just submit color payload */
* are allowed in SIMD16 mode and they could not work without SecHalf. For
* these reasons, we allow BRW_COMPRESSION_2NDHALF here.
*/
- if (brw_inst_qtr_control(devinfo, insn) != BRW_COMPRESSION_2NDHALF)
- brw_inst_set_qtr_control(devinfo, insn, BRW_COMPRESSION_NONE);
+ brw_inst_set_compression(devinfo, insn, false);
if (devinfo->gen < 6)
brw_inst_set_base_mrf(devinfo, insn, msg_reg_nr);
if (sampler_index.file == BRW_IMMEDIATE_VALUE) {
const int sampler_state_size = 16; /* 16 bytes */
- uint32_t sampler = sampler_index.dw1.ud;
+ uint32_t sampler = sampler_index.ud;
if (sampler >= 16) {
assert(devinfo->is_haswell || devinfo->gen >= 8);
insn = next_insn(p, BRW_OPCODE_SEND);
- assert(msg_length < BRW_MAX_MRF);
+ assert(msg_length < BRW_MAX_MRF(devinfo->gen));
brw_set_dest(p, insn, dest);
brw_set_src0(p, insn, src0);
struct brw_reg desc)
{
const struct brw_device_info *devinfo = p->devinfo;
- struct brw_inst *send, *setup;
+ struct brw_inst *send;
+ int setup;
+
+ dst = retype(dst, BRW_REGISTER_TYPE_UW);
assert(desc.type == BRW_REGISTER_TYPE_UD);
+ /* We hold on to the setup instruction (the SEND in the direct case, the OR
+ * in the indirect case) by its index in the instruction store. The
+ * pointer returned by next_insn() may become invalid if emitting the SEND
+ * in the indirect case reallocs the store.
+ */
+
if (desc.file == BRW_IMMEDIATE_VALUE) {
- setup = send = next_insn(p, BRW_OPCODE_SEND);
+ setup = p->nr_insn;
+ send = next_insn(p, BRW_OPCODE_SEND);
brw_set_src1(p, send, desc);
} else {
* caller can specify additional descriptor bits with the usual
* brw_set_*_message() helper functions.
*/
- setup = brw_OR(p, addr, desc, brw_imm_ud(0));
+ setup = p->nr_insn;
+ brw_OR(p, addr, desc, brw_imm_ud(0));
brw_pop_insn_state(p);
brw_set_src1(p, send, addr);
}
+ if (dst.width < BRW_EXECUTE_8)
+ brw_inst_set_exec_size(devinfo, send, dst.width);
+
brw_set_dest(p, send, dst);
brw_set_src0(p, send, retype(payload, BRW_REGISTER_TYPE_UD));
brw_inst_set_sfid(devinfo, send, sfid);
- return setup;
+ return &p->store[setup];
+}
+
+static struct brw_inst *
+brw_send_indirect_surface_message(struct brw_codegen *p,
+ unsigned sfid,
+ struct brw_reg dst,
+ struct brw_reg payload,
+ struct brw_reg surface,
+ unsigned message_len,
+ unsigned response_len,
+ bool header_present)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ struct brw_inst *insn;
+
+ if (surface.file != BRW_IMMEDIATE_VALUE) {
+ struct brw_reg addr = retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD);
+
+ brw_push_insn_state(p);
+ brw_set_default_access_mode(p, BRW_ALIGN_1);
+ brw_set_default_mask_control(p, BRW_MASK_DISABLE);
+ brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
+
+ /* Mask out invalid bits from the surface index to avoid hangs e.g. when
+ * some surface array is accessed out of bounds.
+ */
+ insn = brw_AND(p, addr,
+ suboffset(vec1(retype(surface, BRW_REGISTER_TYPE_UD)),
+ BRW_GET_SWZ(surface.swizzle, 0)),
+ brw_imm_ud(0xff));
+
+ brw_pop_insn_state(p);
+
+ surface = addr;
+ }
+
+ insn = brw_send_indirect_message(p, sfid, dst, payload, surface);
+ brw_inst_set_mlen(devinfo, insn, message_len);
+ brw_inst_set_rlen(devinfo, insn, response_len);
+ brw_inst_set_header_present(devinfo, insn, header_present);
+
+ return insn;
+}
+
+static bool
+while_jumps_before_offset(const struct brw_device_info *devinfo,
+ brw_inst *insn, int while_offset, int start_offset)
+{
+ int scale = 16 / brw_jump_scale(devinfo);
+ int jip = devinfo->gen == 6 ? brw_inst_gen6_jump_count(devinfo, insn)
+ : brw_inst_jip(devinfo, insn);
+ return while_offset + jip * scale <= start_offset;
}
+
static int
brw_find_next_block_end(struct brw_codegen *p, int start_offset)
{
void *store = p->store;
const struct brw_device_info *devinfo = p->devinfo;
+ int depth = 0;
+
for (offset = next_offset(devinfo, store, start_offset);
offset < p->next_insn_offset;
offset = next_offset(devinfo, store, offset)) {
brw_inst *insn = store + offset;
switch (brw_inst_opcode(devinfo, insn)) {
+ case BRW_OPCODE_IF:
+ depth++;
+ break;
case BRW_OPCODE_ENDIF:
- case BRW_OPCODE_ELSE:
+ if (depth == 0)
+ return offset;
+ depth--;
+ break;
case BRW_OPCODE_WHILE:
+ /* If the while doesn't jump before our instruction, it's the end
+ * of a sibling do...while loop. Ignore it.
+ */
+ if (!while_jumps_before_offset(devinfo, insn, offset, start_offset))
+ continue;
+ /* fallthrough */
+ case BRW_OPCODE_ELSE:
case BRW_OPCODE_HALT:
- return offset;
+ if (depth == 0)
+ return offset;
}
}
{
const struct brw_device_info *devinfo = p->devinfo;
int offset;
- int scale = 16 / brw_jump_scale(devinfo);
void *store = p->store;
assert(devinfo->gen >= 6);
brw_inst *insn = store + offset;
if (brw_inst_opcode(devinfo, insn) == BRW_OPCODE_WHILE) {
- int jip = devinfo->gen == 6 ? brw_inst_gen6_jump_count(devinfo, insn)
- : brw_inst_jip(devinfo, insn);
- if (offset + jip * scale <= start_offset)
+ if (while_jumps_before_offset(devinfo, insn, offset, start_offset))
return offset;
}
}
bool has_simd4x2,
bool has_simd16)
{
- if (has_simd4x2 && brw_inst_access_mode(p->devinfo, p->current) == BRW_ALIGN_16)
+ if (has_simd4x2 &&
+ brw_inst_access_mode(p->devinfo, p->current) == BRW_ALIGN_16)
return 1;
- else if (has_simd16 && p->compressed)
+ else if (has_simd16 &&
+ brw_inst_exec_size(p->devinfo, p->current) == BRW_EXECUTE_16)
return 2 * num_channels;
else
return num_channels;
brw_set_dp_untyped_atomic_message(struct brw_codegen *p,
brw_inst *insn,
unsigned atomic_op,
- unsigned bind_table_index,
- unsigned msg_length,
- unsigned response_length,
- bool header_present)
+ bool response_expected)
{
const struct brw_device_info *devinfo = p->devinfo;
-
unsigned msg_control =
atomic_op | /* Atomic Operation Type: BRW_AOP_* */
- (response_length ? 1 << 5 : 0); /* Return data expected */
+ (response_expected ? 1 << 5 : 0); /* Return data expected */
if (devinfo->gen >= 8 || devinfo->is_haswell) {
- brw_set_message_descriptor(p, insn, HSW_SFID_DATAPORT_DATA_CACHE_1,
- msg_length, response_length,
- header_present, false);
-
-
- if (brw_inst_access_mode(devinfo, insn) == BRW_ALIGN_1) {
- if (brw_inst_exec_size(devinfo, insn) != BRW_EXECUTE_16)
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_exec_size(devinfo, p->current) == BRW_EXECUTE_8)
msg_control |= 1 << 4; /* SIMD8 mode */
brw_inst_set_dp_msg_type(devinfo, insn,
HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP_SIMD4X2);
}
} else {
- brw_set_message_descriptor(p, insn, GEN7_SFID_DATAPORT_DATA_CACHE,
- msg_length, response_length,
- header_present, false);
-
- brw_inst_set_dp_msg_type(devinfo, insn, GEN7_DATAPORT_DC_UNTYPED_ATOMIC_OP);
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ GEN7_DATAPORT_DC_UNTYPED_ATOMIC_OP);
- if (brw_inst_exec_size(devinfo, insn) != BRW_EXECUTE_16)
+ if (brw_inst_exec_size(devinfo, p->current) == BRW_EXECUTE_8)
msg_control |= 1 << 4; /* SIMD8 mode */
}
- brw_inst_set_binding_table_index(devinfo, insn, bind_table_index);
brw_inst_set_dp_msg_control(devinfo, insn, msg_control);
}
void
brw_untyped_atomic(struct brw_codegen *p,
- struct brw_reg dest,
+ struct brw_reg dst,
struct brw_reg payload,
+ struct brw_reg surface,
unsigned atomic_op,
- unsigned bind_table_index,
unsigned msg_length,
bool response_expected)
{
const struct brw_device_info *devinfo = p->devinfo;
+ const unsigned sfid = (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN7_SFID_DATAPORT_DATA_CACHE);
const bool align1 = brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1;
/* Mask out unused components -- This is especially important in Align16
* mode on generations that don't have native support for SIMD4x2 atomics,
* uninitialized Y, Z and W coordinates of the payload.
*/
const unsigned mask = align1 ? WRITEMASK_XYZW : WRITEMASK_X;
- brw_inst *insn = brw_next_insn(p, BRW_OPCODE_SEND);
-
- brw_set_dest(p, insn, retype(brw_writemask(dest, mask),
- BRW_REGISTER_TYPE_UD));
- brw_set_src0(p, insn, retype(payload, BRW_REGISTER_TYPE_UD));
- brw_set_src1(p, insn, brw_imm_d(0));
- brw_set_dp_untyped_atomic_message(
- p, insn, atomic_op, bind_table_index, msg_length,
+ struct brw_inst *insn = brw_send_indirect_surface_message(
+ p, sfid, brw_writemask(dst, mask), payload, surface, msg_length,
brw_surface_payload_size(p, response_expected,
devinfo->gen >= 8 || devinfo->is_haswell, true),
align1);
+
+ brw_set_dp_untyped_atomic_message(
+ p, insn, atomic_op, response_expected);
}
static void
brw_set_dp_untyped_surface_read_message(struct brw_codegen *p,
- brw_inst *insn,
- unsigned bind_table_index,
- unsigned msg_length,
- unsigned response_length,
- unsigned num_channels,
- bool header_present)
+ struct brw_inst *insn,
+ unsigned num_channels)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ /* Set mask of 32-bit channels to drop. */
+ unsigned msg_control = 0xf & (0xf << num_channels);
+
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_exec_size(devinfo, p->current) == BRW_EXECUTE_16)
+ msg_control |= 1 << 4; /* SIMD16 mode */
+ else
+ msg_control |= 2 << 4; /* SIMD8 mode */
+ }
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_READ :
+ GEN7_DATAPORT_DC_UNTYPED_SURFACE_READ));
+ brw_inst_set_dp_msg_control(devinfo, insn, msg_control);
+}
+
+void
+brw_untyped_surface_read(struct brw_codegen *p,
+ struct brw_reg dst,
+ struct brw_reg payload,
+ struct brw_reg surface,
+ unsigned msg_length,
+ unsigned num_channels)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ const unsigned sfid = (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN7_SFID_DATAPORT_DATA_CACHE);
+ struct brw_inst *insn = brw_send_indirect_surface_message(
+ p, sfid, dst, payload, surface, msg_length,
+ brw_surface_payload_size(p, num_channels, true, true),
+ false);
+
+ brw_set_dp_untyped_surface_read_message(
+ p, insn, num_channels);
+}
+
+static void
+brw_set_dp_untyped_surface_write_message(struct brw_codegen *p,
+ struct brw_inst *insn,
+ unsigned num_channels)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ /* Set mask of 32-bit channels to drop. */
+ unsigned msg_control = 0xf & (0xf << num_channels);
+
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_exec_size(devinfo, p->current) == BRW_EXECUTE_16)
+ msg_control |= 1 << 4; /* SIMD16 mode */
+ else
+ msg_control |= 2 << 4; /* SIMD8 mode */
+ } else {
+ if (devinfo->gen >= 8 || devinfo->is_haswell)
+ msg_control |= 0 << 4; /* SIMD4x2 mode */
+ else
+ msg_control |= 2 << 4; /* SIMD8 mode */
+ }
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_WRITE :
+ GEN7_DATAPORT_DC_UNTYPED_SURFACE_WRITE);
+ brw_inst_set_dp_msg_control(devinfo, insn, msg_control);
+}
+
+void
+brw_untyped_surface_write(struct brw_codegen *p,
+ struct brw_reg payload,
+ struct brw_reg surface,
+ unsigned msg_length,
+ unsigned num_channels)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ const unsigned sfid = (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN7_SFID_DATAPORT_DATA_CACHE);
+ const bool align1 = brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1;
+ /* Mask out unused components -- See comment in brw_untyped_atomic(). */
+ const unsigned mask = devinfo->gen == 7 && !devinfo->is_haswell && !align1 ?
+ WRITEMASK_X : WRITEMASK_XYZW;
+ struct brw_inst *insn = brw_send_indirect_surface_message(
+ p, sfid, brw_writemask(brw_null_reg(), mask),
+ payload, surface, msg_length, 0, align1);
+
+ brw_set_dp_untyped_surface_write_message(
+ p, insn, num_channels);
+}
+
+static void
+brw_set_dp_typed_atomic_message(struct brw_codegen *p,
+ struct brw_inst *insn,
+ unsigned atomic_op,
+ bool response_expected)
{
const struct brw_device_info *devinfo = p->devinfo;
- const unsigned dispatch_width =
- (brw_inst_exec_size(devinfo, insn) == BRW_EXECUTE_16 ? 16 : 8);
+ unsigned msg_control =
+ atomic_op | /* Atomic Operation Type: BRW_AOP_* */
+ (response_expected ? 1 << 5 : 0); /* Return data expected */
if (devinfo->gen >= 8 || devinfo->is_haswell) {
- brw_set_message_descriptor(p, insn, HSW_SFID_DATAPORT_DATA_CACHE_1,
- msg_length, response_length,
- header_present, false);
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 1 << 4; /* Use high 8 slots of the sample mask */
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP);
+ } else {
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP_SIMD4X2);
+ }
+ } else {
brw_inst_set_dp_msg_type(devinfo, insn,
- HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_READ);
+ GEN7_DATAPORT_RC_TYPED_ATOMIC_OP);
+
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 1 << 4; /* Use high 8 slots of the sample mask */
+ }
+
+ brw_inst_set_dp_msg_control(devinfo, insn, msg_control);
+}
+
+void
+brw_typed_atomic(struct brw_codegen *p,
+ struct brw_reg dst,
+ struct brw_reg payload,
+ struct brw_reg surface,
+ unsigned atomic_op,
+ unsigned msg_length,
+ bool response_expected) {
+ const struct brw_device_info *devinfo = p->devinfo;
+ const unsigned sfid = (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN6_SFID_DATAPORT_RENDER_CACHE);
+ const bool align1 = (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1);
+ /* Mask out unused components -- See comment in brw_untyped_atomic(). */
+ const unsigned mask = align1 ? WRITEMASK_XYZW : WRITEMASK_X;
+ struct brw_inst *insn = brw_send_indirect_surface_message(
+ p, sfid, brw_writemask(dst, mask), payload, surface, msg_length,
+ brw_surface_payload_size(p, response_expected,
+ devinfo->gen >= 8 || devinfo->is_haswell, false),
+ true);
+
+ brw_set_dp_typed_atomic_message(
+ p, insn, atomic_op, response_expected);
+}
+
+static void
+brw_set_dp_typed_surface_read_message(struct brw_codegen *p,
+ struct brw_inst *insn,
+ unsigned num_channels)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ /* Set mask of unused channels. */
+ unsigned msg_control = 0xf & (0xf << num_channels);
+
+ if (devinfo->gen >= 8 || devinfo->is_haswell) {
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 2 << 4; /* Use high 8 slots of the sample mask */
+ else
+ msg_control |= 1 << 4; /* Use low 8 slots of the sample mask */
+ }
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_READ);
} else {
- brw_set_message_descriptor(p, insn, GEN7_SFID_DATAPORT_DATA_CACHE,
- msg_length, response_length,
- header_present, false);
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 1 << 5; /* Use high 8 slots of the sample mask */
+ }
brw_inst_set_dp_msg_type(devinfo, insn,
- GEN7_DATAPORT_DC_UNTYPED_SURFACE_READ);
+ GEN7_DATAPORT_RC_TYPED_SURFACE_READ);
}
- /* Set mask of 32-bit channels to drop. */
- unsigned msg_control = (0xf & (0xf << num_channels));
+ brw_inst_set_dp_msg_control(devinfo, insn, msg_control);
+}
- if (brw_inst_access_mode(devinfo, insn) == BRW_ALIGN_1) {
- if (dispatch_width == 16)
- msg_control |= 1 << 4; /* SIMD16 mode */
- else
- msg_control |= 2 << 4; /* SIMD8 mode */
+void
+brw_typed_surface_read(struct brw_codegen *p,
+ struct brw_reg dst,
+ struct brw_reg payload,
+ struct brw_reg surface,
+ unsigned msg_length,
+ unsigned num_channels)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ const unsigned sfid = (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN6_SFID_DATAPORT_RENDER_CACHE);
+ struct brw_inst *insn = brw_send_indirect_surface_message(
+ p, sfid, dst, payload, surface, msg_length,
+ brw_surface_payload_size(p, num_channels,
+ devinfo->gen >= 8 || devinfo->is_haswell, false),
+ true);
+
+ brw_set_dp_typed_surface_read_message(
+ p, insn, num_channels);
+}
+
+static void
+brw_set_dp_typed_surface_write_message(struct brw_codegen *p,
+ struct brw_inst *insn,
+ unsigned num_channels)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ /* Set mask of unused channels. */
+ unsigned msg_control = 0xf & (0xf << num_channels);
+
+ if (devinfo->gen >= 8 || devinfo->is_haswell) {
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 2 << 4; /* Use high 8 slots of the sample mask */
+ else
+ msg_control |= 1 << 4; /* Use low 8 slots of the sample mask */
+ }
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_WRITE);
+
+ } else {
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 1 << 5; /* Use high 8 slots of the sample mask */
+ }
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ GEN7_DATAPORT_RC_TYPED_SURFACE_WRITE);
}
- brw_inst_set_binding_table_index(devinfo, insn, bind_table_index);
brw_inst_set_dp_msg_control(devinfo, insn, msg_control);
}
void
-brw_untyped_surface_read(struct brw_codegen *p,
- struct brw_reg dest,
- struct brw_reg mrf,
- unsigned bind_table_index,
- unsigned msg_length,
- unsigned num_channels)
+brw_typed_surface_write(struct brw_codegen *p,
+ struct brw_reg payload,
+ struct brw_reg surface,
+ unsigned msg_length,
+ unsigned num_channels)
{
const struct brw_device_info *devinfo = p->devinfo;
- brw_inst *insn = next_insn(p, BRW_OPCODE_SEND);
+ const unsigned sfid = (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN6_SFID_DATAPORT_RENDER_CACHE);
+ const bool align1 = (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1);
+ /* Mask out unused components -- See comment in brw_untyped_atomic(). */
+ const unsigned mask = (devinfo->gen == 7 && !devinfo->is_haswell && !align1 ?
+ WRITEMASK_X : WRITEMASK_XYZW);
+ struct brw_inst *insn = brw_send_indirect_surface_message(
+ p, sfid, brw_writemask(brw_null_reg(), mask),
+ payload, surface, msg_length, 0, true);
- brw_set_dest(p, insn, retype(dest, BRW_REGISTER_TYPE_UD));
- brw_set_src0(p, insn, retype(mrf, BRW_REGISTER_TYPE_UD));
- brw_set_dp_untyped_surface_read_message(
- p, insn, bind_table_index, msg_length,
- brw_surface_payload_size(p, num_channels, true, true),
- num_channels, brw_inst_access_mode(devinfo, insn) == BRW_ALIGN_1);
+ brw_set_dp_typed_surface_write_message(
+ p, insn, num_channels);
+}
+
+static void
+brw_set_memory_fence_message(struct brw_codegen *p,
+ struct brw_inst *insn,
+ enum brw_message_target sfid,
+ bool commit_enable)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+
+ brw_set_message_descriptor(p, insn, sfid,
+ 1 /* message length */,
+ (commit_enable ? 1 : 0) /* response length */,
+ true /* header present */,
+ false);
+
+ switch (sfid) {
+ case GEN6_SFID_DATAPORT_RENDER_CACHE:
+ brw_inst_set_dp_msg_type(devinfo, insn, GEN7_DATAPORT_RC_MEMORY_FENCE);
+ break;
+ case GEN7_SFID_DATAPORT_DATA_CACHE:
+ brw_inst_set_dp_msg_type(devinfo, insn, GEN7_DATAPORT_DC_MEMORY_FENCE);
+ break;
+ default:
+ unreachable("Not reached");
+ }
+
+ if (commit_enable)
+ brw_inst_set_dp_msg_control(devinfo, insn, 1 << 5);
+}
+
+void
+brw_memory_fence(struct brw_codegen *p,
+ struct brw_reg dst)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ const bool commit_enable = devinfo->gen == 7 && !devinfo->is_haswell;
+ struct brw_inst *insn;
+
+ brw_push_insn_state(p);
+ brw_set_default_mask_control(p, BRW_MASK_DISABLE);
+ brw_set_default_exec_size(p, BRW_EXECUTE_1);
+ dst = vec1(dst);
+
+ /* Set dst as destination for dependency tracking, the MEMORY_FENCE
+ * message doesn't write anything back.
+ */
+ insn = next_insn(p, BRW_OPCODE_SEND);
+ dst = retype(dst, BRW_REGISTER_TYPE_UW);
+ brw_set_dest(p, insn, dst);
+ brw_set_src0(p, insn, dst);
+ brw_set_memory_fence_message(p, insn, GEN7_SFID_DATAPORT_DATA_CACHE,
+ commit_enable);
+
+ if (devinfo->gen == 7 && !devinfo->is_haswell) {
+ /* IVB does typed surface access through the render cache, so we need to
+ * flush it too. Use a different register so both flushes can be
+ * pipelined by the hardware.
+ */
+ insn = next_insn(p, BRW_OPCODE_SEND);
+ brw_set_dest(p, insn, offset(dst, 1));
+ brw_set_src0(p, insn, offset(dst, 1));
+ brw_set_memory_fence_message(p, insn, GEN6_SFID_DATAPORT_RENDER_CACHE,
+ commit_enable);
+
+ /* Now write the response of the second message into the response of the
+ * first to trigger a pipeline stall -- This way future render and data
+ * cache messages will be properly ordered with respect to past data and
+ * render cache messages.
+ */
+ brw_MOV(p, dst, offset(dst, 1));
+ }
+
+ brw_pop_insn_state(p);
}
void
struct brw_reg mrf,
bool noperspective,
unsigned mode,
- unsigned data,
+ struct brw_reg data,
unsigned msg_length,
unsigned response_length)
{
const struct brw_device_info *devinfo = p->devinfo;
- struct brw_inst *insn = next_insn(p, BRW_OPCODE_SEND);
+ struct brw_inst *insn;
+ const uint16_t exec_size = brw_inst_exec_size(devinfo, p->current);
- brw_set_dest(p, insn, dest);
- brw_set_src0(p, insn, mrf);
- brw_set_message_descriptor(p, insn, GEN7_SFID_PIXEL_INTERPOLATOR,
- msg_length, response_length,
- false /* header is never present for PI */,
- false);
-
- brw_inst_set_pi_simd_mode(
- devinfo, insn, brw_inst_exec_size(devinfo, insn) == BRW_EXECUTE_16);
+ /* brw_send_indirect_message will automatically use a direct send message
+ * if data is actually immediate.
+ */
+ insn = brw_send_indirect_message(p,
+ GEN7_SFID_PIXEL_INTERPOLATOR,
+ dest,
+ mrf,
+ vec1(data));
+ brw_inst_set_mlen(devinfo, insn, msg_length);
+ brw_inst_set_rlen(devinfo, insn, response_length);
+
+ brw_inst_set_pi_simd_mode(devinfo, insn, exec_size == BRW_EXECUTE_16);
brw_inst_set_pi_slot_group(devinfo, insn, 0); /* zero unless 32/64px dispatch */
brw_inst_set_pi_nopersp(devinfo, insn, noperspective);
brw_inst_set_pi_message_type(devinfo, insn, mode);
- brw_inst_set_pi_message_data(devinfo, insn, data);
+}
+
+void
+brw_find_live_channel(struct brw_codegen *p, struct brw_reg dst)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ const unsigned exec_size = 1 << brw_inst_exec_size(devinfo, p->current);
+ const unsigned qtr_control = brw_inst_qtr_control(devinfo, p->current);
+ brw_inst *inst;
+
+ assert(devinfo->gen >= 7);
+
+ brw_push_insn_state(p);
+
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ brw_set_default_mask_control(p, BRW_MASK_DISABLE);
+
+ if (devinfo->gen >= 8) {
+ /* Getting the first active channel index is easy on Gen8: Just find
+ * the first bit set in the mask register. The same register exists
+ * on HSW already but it reads back as all ones when the current
+ * instruction has execution masking disabled, so it's kind of
+ * useless.
+ */
+ inst = brw_FBL(p, vec1(dst),
+ retype(brw_mask_reg(0), BRW_REGISTER_TYPE_UD));
+
+ /* Quarter control has the effect of magically shifting the value of
+ * this register so you'll get the first active channel relative to
+ * the specified quarter control as result.
+ */
+ } else {
+ const struct brw_reg flag = brw_flag_reg(1, 0);
+
+ brw_MOV(p, retype(flag, BRW_REGISTER_TYPE_UD), brw_imm_ud(0));
+
+ /* Run enough instructions returning zero with execution masking and
+ * a conditional modifier enabled in order to get the full execution
+ * mask in f1.0. We could use a single 32-wide move here if it
+ * weren't because of the hardware bug that causes channel enables to
+ * be applied incorrectly to the second half of 32-wide instructions
+ * on Gen7.
+ */
+ const unsigned lower_size = MIN2(16, exec_size);
+ for (unsigned i = 0; i < exec_size / lower_size; i++) {
+ inst = brw_MOV(p, retype(brw_null_reg(), BRW_REGISTER_TYPE_UW),
+ brw_imm_uw(0));
+ brw_inst_set_mask_control(devinfo, inst, BRW_MASK_ENABLE);
+ brw_inst_set_group(devinfo, inst, lower_size * i + 8 * qtr_control);
+ brw_inst_set_cond_modifier(devinfo, inst, BRW_CONDITIONAL_Z);
+ brw_inst_set_flag_reg_nr(devinfo, inst, 1);
+ brw_inst_set_exec_size(devinfo, inst, cvt(lower_size) - 1);
+ }
+
+ /* Find the first bit set in the exec_size-wide portion of the flag
+ * register that was updated by the last sequence of MOV
+ * instructions.
+ */
+ const enum brw_reg_type type = brw_int_type(exec_size / 8, false);
+ brw_FBL(p, vec1(dst), byte_offset(retype(flag, type), qtr_control));
+ }
+ } else {
+ brw_set_default_mask_control(p, BRW_MASK_DISABLE);
+
+ if (devinfo->gen >= 8) {
+ /* In SIMD4x2 mode the first active channel index is just the
+ * negation of the first bit of the mask register.
+ */
+ inst = brw_AND(p, brw_writemask(dst, WRITEMASK_X),
+ negate(retype(brw_mask_reg(0), BRW_REGISTER_TYPE_UD)),
+ brw_imm_ud(1));
+
+ } else {
+ /* Overwrite the destination without and with execution masking to
+ * find out which of the channels is active.
+ */
+ brw_push_insn_state(p);
+ brw_set_default_exec_size(p, BRW_EXECUTE_4);
+ brw_MOV(p, brw_writemask(vec4(dst), WRITEMASK_X),
+ brw_imm_ud(1));
+
+ inst = brw_MOV(p, brw_writemask(vec4(dst), WRITEMASK_X),
+ brw_imm_ud(0));
+ brw_pop_insn_state(p);
+ brw_inst_set_mask_control(devinfo, inst, BRW_MASK_ENABLE);
+ }
+ }
+
+ brw_pop_insn_state(p);
+}
+
+void
+brw_broadcast(struct brw_codegen *p,
+ struct brw_reg dst,
+ struct brw_reg src,
+ struct brw_reg idx)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ const bool align1 = brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1;
+ brw_inst *inst;
+
+ brw_push_insn_state(p);
+ brw_set_default_mask_control(p, BRW_MASK_DISABLE);
+ brw_set_default_exec_size(p, align1 ? BRW_EXECUTE_1 : BRW_EXECUTE_4);
+
+ assert(src.file == BRW_GENERAL_REGISTER_FILE &&
+ src.address_mode == BRW_ADDRESS_DIRECT);
+
+ if ((src.vstride == 0 && (src.hstride == 0 || !align1)) ||
+ idx.file == BRW_IMMEDIATE_VALUE) {
+ /* Trivial, the source is already uniform or the index is a constant.
+ * We will typically not get here if the optimizer is doing its job, but
+ * asserting would be mean.
+ */
+ const unsigned i = idx.file == BRW_IMMEDIATE_VALUE ? idx.ud : 0;
+ brw_MOV(p, dst,
+ (align1 ? stride(suboffset(src, i), 0, 1, 0) :
+ stride(suboffset(src, 4 * i), 0, 4, 1)));
+ } else {
+ if (align1) {
+ const struct brw_reg addr =
+ retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD);
+ const unsigned offset = src.nr * REG_SIZE + src.subnr;
+ /* Limit in bytes of the signed indirect addressing immediate. */
+ const unsigned limit = 512;
+
+ brw_push_insn_state(p);
+ brw_set_default_mask_control(p, BRW_MASK_DISABLE);
+ brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
+
+ /* Take into account the component size and horizontal stride. */
+ assert(src.vstride == src.hstride + src.width);
+ brw_SHL(p, addr, vec1(idx),
+ brw_imm_ud(_mesa_logbase2(type_sz(src.type)) +
+ src.hstride - 1));
+
+ /* We can only address up to limit bytes using the indirect
+ * addressing immediate, account for the difference if the source
+ * register is above this limit.
+ */
+ if (offset >= limit)
+ brw_ADD(p, addr, addr, brw_imm_ud(offset - offset % limit));
+
+ brw_pop_insn_state(p);
+
+ /* Use indirect addressing to fetch the specified component. */
+ brw_MOV(p, dst,
+ retype(brw_vec1_indirect(addr.subnr, offset % limit),
+ src.type));
+ } else {
+ /* In SIMD4x2 mode the index can be either zero or one, replicate it
+ * to all bits of a flag register,
+ */
+ inst = brw_MOV(p,
+ brw_null_reg(),
+ stride(brw_swizzle(idx, BRW_SWIZZLE_XXXX), 4, 4, 1));
+ brw_inst_set_pred_control(devinfo, inst, BRW_PREDICATE_NONE);
+ brw_inst_set_cond_modifier(devinfo, inst, BRW_CONDITIONAL_NZ);
+ brw_inst_set_flag_reg_nr(devinfo, inst, 1);
+
+ /* and use predicated SEL to pick the right channel. */
+ inst = brw_SEL(p, dst,
+ stride(suboffset(src, 4), 4, 4, 1),
+ stride(src, 4, 4, 1));
+ brw_inst_set_pred_control(devinfo, inst, BRW_PREDICATE_NORMAL);
+ brw_inst_set_flag_reg_nr(devinfo, inst, 1);
+ }
+ }
+
+ brw_pop_insn_state(p);
}
/**
struct brw_reg payload,
uint32_t surf_index)
{
+ const unsigned sfid = (p->devinfo->gen >= 8 || p->devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN7_SFID_DATAPORT_DATA_CACHE);
assert(p->devinfo->gen >= 7);
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
+ brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
- brw_pop_insn_state(p);
/* We use brw_vec1_reg and unmasked because we want to increment the given
* offset only once.
BRW_ARF_NULL, 0));
brw_set_src0(p, send, brw_vec1_reg(payload.file,
payload.nr, 0));
- brw_set_dp_untyped_atomic_message(p, send, BRW_AOP_ADD, surf_index,
- 2 /* message length */,
- 0 /* response length */,
- false /* header present */);
+ brw_set_src1(p, send, brw_imm_ud(0));
+ brw_set_message_descriptor(p, send, sfid, 2, 0, false, false);
+ brw_inst_set_binding_table_index(p->devinfo, send, surf_index);
+ brw_set_dp_untyped_atomic_message(p, send, BRW_AOP_ADD, false);
+
+ brw_pop_insn_state(p);
+}
+
+
+/**
+ * Emit the SEND message for a barrier
+ */
+void
+brw_barrier(struct brw_codegen *p, struct brw_reg src)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ struct brw_inst *inst;
+
+ assert(devinfo->gen >= 7);
+
+ inst = next_insn(p, BRW_OPCODE_SEND);
+ brw_set_dest(p, inst, retype(brw_null_reg(), BRW_REGISTER_TYPE_UW));
+ brw_set_src0(p, inst, src);
+ brw_set_src1(p, inst, brw_null_reg());
+
+ brw_set_message_descriptor(p, inst, BRW_SFID_MESSAGE_GATEWAY,
+ 1 /* msg_length */,
+ 0 /* response_length */,
+ false /* header_present */,
+ false /* end_of_thread */);
+
+ brw_inst_set_gateway_notify(devinfo, inst, 1);
+ brw_inst_set_gateway_subfuncid(devinfo, inst,
+ BRW_MESSAGE_GATEWAY_SFID_BARRIER_MSG);
+
+ brw_inst_set_mask_control(devinfo, inst, BRW_MASK_DISABLE);
+}
+
+
+/**
+ * Emit the wait instruction for a barrier
+ */
+void
+brw_WAIT(struct brw_codegen *p)
+{
+ const struct brw_device_info *devinfo = p->devinfo;
+ struct brw_inst *insn;
+
+ struct brw_reg src = brw_notification_reg();
+
+ insn = next_insn(p, BRW_OPCODE_WAIT);
+ brw_set_dest(p, insn, src);
+ brw_set_src0(p, insn, src);
+ brw_set_src1(p, insn, brw_null_reg());
+
+ brw_inst_set_exec_size(devinfo, insn, BRW_EXECUTE_1);
+ brw_inst_set_mask_control(devinfo, insn, BRW_MASK_DISABLE);
}
projects / mesa.git / blobdiff