void
vec4_tes_visitor::nir_emit_intrinsic(nir_intrinsic_instr *instr)
{
const struct brw_tes_prog_data *tes_prog_data =
(const struct brw_tes_prog_data *) prog_data;
switch (instr->intrinsic) {
case nir_intrinsic_load_tess_coord:
/* gl_TessCoord is part of the payload in g1 channels 0-2 and 4-6. */
emit(MOV(get_nir_dest(instr->dest, BRW_REGISTER_TYPE_F),
src_reg(brw_vec8_grf(1, 0))));
break;
case nir_intrinsic_load_tess_level_outer:
if (tes_prog_data->domain == BRW_TESS_DOMAIN_ISOLINE) {
emit(MOV(get_nir_dest(instr->dest, BRW_REGISTER_TYPE_F),
swizzle(src_reg(ATTR, 1, glsl_type::vec4_type),
BRW_SWIZZLE_ZWZW)));
} else {
emit(MOV(get_nir_dest(instr->dest, BRW_REGISTER_TYPE_F),
swizzle(src_reg(ATTR, 1, glsl_type::vec4_type),
BRW_SWIZZLE_WZYX)));
}
break;
case nir_intrinsic_load_tess_level_inner:
if (tes_prog_data->domain == BRW_TESS_DOMAIN_QUAD) {
emit(MOV(get_nir_dest(instr->dest, BRW_REGISTER_TYPE_F),
swizzle(src_reg(ATTR, 0, glsl_type::vec4_type),
BRW_SWIZZLE_WZYX)));
} else {
emit(MOV(get_nir_dest(instr->dest, BRW_REGISTER_TYPE_F),
src_reg(ATTR, 1, glsl_type::float_type)));
}
break;
case nir_intrinsic_load_primitive_id:
emit(TES_OPCODE_GET_PRIMITIVE_ID,
get_nir_dest(instr->dest, BRW_REGISTER_TYPE_UD));
break;
case nir_intrinsic_load_input:
case nir_intrinsic_load_per_vertex_input: {
src_reg indirect_offset = get_indirect_offset(instr);
unsigned imm_offset = instr->const_index[0];
src_reg header = input_read_header;
bool is_64bit = nir_dest_bit_size(instr->dest) == 64;
unsigned first_component = nir_intrinsic_component(instr);
if (is_64bit)
first_component /= 2;
if (indirect_offset.file != BAD_FILE) {
header = src_reg(this, glsl_type::uvec4_type);
emit(TES_OPCODE_ADD_INDIRECT_URB_OFFSET, dst_reg(header),
input_read_header, indirect_offset);
} else {
/* Arbitrarily only push up to 24 vec4 slots worth of data,
* which is 12 registers (since each holds 2 vec4 slots).
*/
const unsigned max_push_slots = 24;
if (imm_offset < max_push_slots) {
const glsl_type *src_glsl_type =
is_64bit ? glsl_type::dvec4_type : glsl_type::ivec4_type;
src_reg src = src_reg(ATTR, imm_offset, src_glsl_type);
src.swizzle = BRW_SWZ_COMP_INPUT(first_component);
const brw_reg_type dst_reg_type =
is_64bit ? BRW_REGISTER_TYPE_DF : BRW_REGISTER_TYPE_D;
emit(MOV(get_nir_dest(instr->dest, dst_reg_type), src));
prog_data->urb_read_length =
MAX2(prog_data->urb_read_length,
DIV_ROUND_UP(imm_offset + (is_64bit ? 2 : 1), 2));
break;
}
}
if (!is_64bit) {
dst_reg temp(this, glsl_type::ivec4_type);
vec4_instruction *read =
emit(VEC4_OPCODE_URB_READ, temp, src_reg(header));
read->offset = imm_offset;
read->urb_write_flags = BRW_URB_WRITE_PER_SLOT_OFFSET;
src_reg src = src_reg(temp);
src.swizzle = BRW_SWZ_COMP_INPUT(first_component);
/* Copy to target. We might end up with some funky writemasks landing
* in here, but we really don't want them in the above pseudo-ops.
*/
dst_reg dst = get_nir_dest(instr->dest, BRW_REGISTER_TYPE_D);
dst.writemask = brw_writemask_for_size(instr->num_components);
emit(MOV(dst, src));
} else {
/* For 64-bit we need to load twice as many 32-bit components, and for
* dvec3/4 we need to emit 2 URB Read messages
*/
dst_reg temp(this, glsl_type::dvec4_type);
dst_reg temp_d = retype(temp, BRW_REGISTER_TYPE_D);
vec4_instruction *read =
emit(VEC4_OPCODE_URB_READ, temp_d, src_reg(header));
read->offset = imm_offset;
read->urb_write_flags = BRW_URB_WRITE_PER_SLOT_OFFSET;
if (instr->num_components > 2) {
read = emit(VEC4_OPCODE_URB_READ, byte_offset(temp_d, REG_SIZE),
src_reg(header));
read->offset = imm_offset + 1;
read->urb_write_flags = BRW_URB_WRITE_PER_SLOT_OFFSET;
}
src_reg temp_as_src = src_reg(temp);
temp_as_src.swizzle = BRW_SWZ_COMP_INPUT(first_component);
dst_reg shuffled(this, glsl_type::dvec4_type);
shuffle_64bit_data(shuffled, temp_as_src, false);
dst_reg dst = get_nir_dest(instr->dest, BRW_REGISTER_TYPE_DF);
dst.writemask = brw_writemask_for_size(instr->num_components);
emit(MOV(dst, src_reg(shuffled)));
}
break;
}
default:
vec4_visitor::nir_emit_intrinsic(instr);
}
}
static void merge_edgeflags( struct brw_clip_compile *c )
{
struct brw_compile *p = &c->func;
struct brw_reg tmp0 = get_element_ud(c->reg.tmp0, 0);
brw_AND(p, tmp0, get_element_ud(c->reg.R0, 2), brw_imm_ud(PRIM_MASK));
brw_CMP(p,
vec1(brw_null_reg()),
BRW_CONDITIONAL_EQ,
tmp0,
brw_imm_ud(_3DPRIM_POLYGON));
/* Get away with using reg.vertex because we know that this is not
* a _3DPRIM_TRISTRIP_REVERSE:
*/
brw_IF(p, BRW_EXECUTE_1);
{
brw_set_conditionalmod(p, BRW_CONDITIONAL_EQ);
brw_AND(p, vec1(brw_null_reg()), get_element_ud(c->reg.R0, 2), brw_imm_ud(1<<8));
brw_MOV(p, byte_offset(c->reg.vertex[0],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_EDGE)),
brw_imm_f(0));
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
brw_set_conditionalmod(p, BRW_CONDITIONAL_EQ);
brw_AND(p, vec1(brw_null_reg()), get_element_ud(c->reg.R0, 2), brw_imm_ud(1<<9));
brw_MOV(p, byte_offset(c->reg.vertex[2],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_EDGE)),
brw_imm_f(0));
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
}
brw_ENDIF(p);
}
static void copy_bfc( struct brw_clip_compile *c )
{
struct brw_compile *p = &c->func;
GLuint conditional;
/* Do we have any colors to copy?
*/
if (!(brw_clip_have_varying(c, VARYING_SLOT_COL0) &&
brw_clip_have_varying(c, VARYING_SLOT_BFC0)) &&
!(brw_clip_have_varying(c, VARYING_SLOT_COL1) &&
brw_clip_have_varying(c, VARYING_SLOT_BFC1)))
return;
/* In some wierd degnerate cases we can end up testing the
* direction twice, once for culling and once for bfc copying. Oh
* well, that's what you get for setting wierd GL state.
*/
if (c->key.copy_bfc_ccw)
conditional = BRW_CONDITIONAL_GE;
else
conditional = BRW_CONDITIONAL_L;
brw_CMP(p,
vec1(brw_null_reg()),
conditional,
get_element(c->reg.dir, 2),
brw_imm_f(0));
brw_IF(p, BRW_EXECUTE_1);
{
GLuint i;
for (i = 0; i < 3; i++) {
if (brw_clip_have_varying(c, VARYING_SLOT_COL0) &&
brw_clip_have_varying(c, VARYING_SLOT_BFC0))
brw_MOV(p,
byte_offset(c->reg.vertex[i],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_COL0)),
byte_offset(c->reg.vertex[i],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_BFC0)));
if (brw_clip_have_varying(c, VARYING_SLOT_COL1) &&
brw_clip_have_varying(c, VARYING_SLOT_BFC1))
brw_MOV(p,
byte_offset(c->reg.vertex[i],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_COL1)),
byte_offset(c->reg.vertex[i],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_BFC1)));
}
}
brw_ENDIF(p);
}
/* Distribute flatshaded attributes from provoking vertex prior to
* clipping.
*/
void brw_clip_copy_flatshaded_attributes( struct brw_clip_compile *c,
GLuint to, GLuint from )
{
struct brw_codegen *p = &c->func;
for (int i = 0; i < c->vue_map.num_slots; i++) {
if (c->key.interpolation_mode.mode[i] == INTERP_QUALIFIER_FLAT) {
brw_MOV(p,
byte_offset(c->reg.vertex[to], brw_vue_slot_to_offset(i)),
byte_offset(c->reg.vertex[from], brw_vue_slot_to_offset(i)));
}
}
}
/* If flatshading, distribute color from provoking vertex prior to
* clipping.
*/
void brw_clip_copy_colors( struct brw_clip_compile *c,
unsigned to, unsigned from )
{
#if 0
struct brw_compile *p = &c->func;
if (c->offset[VERT_RESULT_COL0])
brw_MOV(p,
byte_offset(c->reg.vertex[to], c->offset[VERT_RESULT_COL0]),
byte_offset(c->reg.vertex[from], c->offset[VERT_RESULT_COL0]));
if (c->offset[VERT_RESULT_COL1])
brw_MOV(p,
byte_offset(c->reg.vertex[to], c->offset[VERT_RESULT_COL1]),
byte_offset(c->reg.vertex[from], c->offset[VERT_RESULT_COL1]));
if (c->offset[VERT_RESULT_BFC0])
brw_MOV(p,
byte_offset(c->reg.vertex[to], c->offset[VERT_RESULT_BFC0]),
byte_offset(c->reg.vertex[from], c->offset[VERT_RESULT_BFC0]));
if (c->offset[VERT_RESULT_BFC1])
brw_MOV(p,
byte_offset(c->reg.vertex[to], c->offset[VERT_RESULT_BFC1]),
byte_offset(c->reg.vertex[from], c->offset[VERT_RESULT_BFC1]));
#else
#warning "disabled"
#endif
}
/* If flatshading, distribute color from provoking vertex prior to
* clipping.
*/
void brw_clip_copy_colors( struct brw_clip_compile *c,
GLuint to, GLuint from )
{
struct brw_compile *p = &c->func;
if (c->offset[VERT_RESULT_COL0])
brw_MOV(p,
byte_offset(c->reg.vertex[to], c->offset[VERT_RESULT_COL0]),
byte_offset(c->reg.vertex[from], c->offset[VERT_RESULT_COL0]));
if (c->offset[VERT_RESULT_COL1])
brw_MOV(p,
byte_offset(c->reg.vertex[to], c->offset[VERT_RESULT_COL1]),
byte_offset(c->reg.vertex[from], c->offset[VERT_RESULT_COL1]));
if (c->offset[VERT_RESULT_BFC0])
brw_MOV(p,
byte_offset(c->reg.vertex[to], c->offset[VERT_RESULT_BFC0]),
byte_offset(c->reg.vertex[from], c->offset[VERT_RESULT_BFC0]));
if (c->offset[VERT_RESULT_BFC1])
brw_MOV(p,
byte_offset(c->reg.vertex[to], c->offset[VERT_RESULT_BFC1]),
byte_offset(c->reg.vertex[from], c->offset[VERT_RESULT_BFC1]));
}
void brw_copy_from_indirect(struct brw_compile *p,
struct brw_reg dst,
struct brw_indirect ptr,
GLuint count)
{
GLuint i;
dst = vec4(dst);
for (i = 0; i < count; i++)
{
GLuint delta = i*32;
brw_MOV(p, byte_offset(dst, delta), deref_4f(ptr, delta));
brw_MOV(p, byte_offset(dst, delta+16), deref_4f(ptr, delta+16));
}
}
void brw_copy8(struct brw_compile *p,
struct brw_reg dst,
struct brw_reg src,
GLuint count)
{
GLuint i;
dst = vec8(dst);
src = vec8(src);
for (i = 0; i < count; i++)
{
GLuint delta = i*32;
brw_MOV(p, byte_offset(dst, delta), byte_offset(src, delta));
}
}
std::pair<bool, bool>
MemoryType::checkAndMergeMemType(Location addr, const RsType* type)
{
const size_t ofs = byte_offset(startAddress, addr, blockSize(), 0);
return checkAndMergeMemType(ofs, type);
}
void brw_copy8(struct brw_codegen *p,
struct brw_reg dst,
struct brw_reg src,
unsigned count)
{
unsigned i;
dst = vec8(dst);
src = vec8(src);
for (i = 0; i < count; i++)
{
unsigned delta = i*32;
brw_MOV(p, byte_offset(dst, delta), byte_offset(src, delta));
}
}
void brw_copy_from_indirect(struct brw_codegen *p,
struct brw_reg dst,
struct brw_indirect ptr,
unsigned count)
{
unsigned i;
dst = vec4(dst);
for (i = 0; i < count; i++)
{
unsigned delta = i*32;
brw_MOV(p, byte_offset(dst, delta), deref_4f(ptr, delta));
brw_MOV(p, byte_offset(dst, delta+16), deref_4f(ptr, delta+16));
}
}
/* This is performed against the original triangles, so no indirection
* required:
BZZZT!
*/
static void compute_tri_direction( struct brw_clip_compile *c )
{
struct brw_compile *p = &c->func;
struct brw_reg e = c->reg.tmp0;
struct brw_reg f = c->reg.tmp1;
GLuint hpos_offset = brw_vert_result_to_offset(&c->vue_map,
VARYING_SLOT_POS);
struct brw_reg v0 = byte_offset(c->reg.vertex[0], hpos_offset);
struct brw_reg v1 = byte_offset(c->reg.vertex[1], hpos_offset);
struct brw_reg v2 = byte_offset(c->reg.vertex[2], hpos_offset);
struct brw_reg v0n = get_tmp(c);
struct brw_reg v1n = get_tmp(c);
struct brw_reg v2n = get_tmp(c);
/* Convert to NDC.
* NOTE: We can't modify the original vertex coordinates,
* as it may impact further operations.
* So, we have to keep normalized coordinates in temp registers.
*
* TBD-KC
* Try to optimize unnecessary MOV's.
*/
brw_MOV(p, v0n, v0);
brw_MOV(p, v1n, v1);
brw_MOV(p, v2n, v2);
brw_clip_project_position(c, v0n);
brw_clip_project_position(c, v1n);
brw_clip_project_position(c, v2n);
/* Calculate the vectors of two edges of the triangle:
*/
brw_ADD(p, e, v0n, negate(v2n));
brw_ADD(p, f, v1n, negate(v2n));
/* Take their crossproduct:
*/
brw_set_access_mode(p, BRW_ALIGN_16);
brw_MUL(p, vec4(brw_null_reg()), brw_swizzle(e, 1,2,0,3), brw_swizzle(f,2,0,1,3));
brw_MAC(p, vec4(e), negate(brw_swizzle(e, 2,0,1,3)), brw_swizzle(f,1,2,0,3));
brw_set_access_mode(p, BRW_ALIGN_1);
brw_MUL(p, c->reg.dir, c->reg.dir, vec4(e));
}
void brw_copy4(struct brw_compile *p,
struct brw_reg dst,
struct brw_reg src,
unsigned count)
{
unsigned i;
dst = vec4(dst);
src = vec4(src);
for (i = 0; i < count; i++)
{
unsigned delta = i*32;
brw_MOV(p, byte_offset(dst, delta), byte_offset(src, delta));
brw_MOV(p, byte_offset(dst, delta+16), byte_offset(src, delta+16));
}
}
static void copy_bfc( struct brw_clip_compile *c )
{
struct brw_compile *p = &c->func;
struct brw_instruction *ccw;
GLuint conditional;
/* Do we have any colors to copy?
*/
if (!(c->offset[VERT_RESULT_COL0] && c->offset[VERT_RESULT_BFC0]) &&
!(c->offset[VERT_RESULT_COL1] && c->offset[VERT_RESULT_BFC1]))
return;
/* In some wierd degnerate cases we can end up testing the
* direction twice, once for culling and once for bfc copying. Oh
* well, that's what you get for setting wierd GL state.
*/
if (c->key.copy_bfc_ccw)
conditional = BRW_CONDITIONAL_GE;
else
conditional = BRW_CONDITIONAL_L;
brw_CMP(p,
vec1(brw_null_reg()),
conditional,
get_element(c->reg.dir, 2),
brw_imm_f(0));
ccw = brw_IF(p, BRW_EXECUTE_1);
{
GLuint i;
for (i = 0; i < 3; i++) {
if (c->offset[VERT_RESULT_COL0] && c->offset[VERT_RESULT_BFC0])
brw_MOV(p,
byte_offset(c->reg.vertex[i], c->offset[VERT_RESULT_COL0]),
byte_offset(c->reg.vertex[i], c->offset[VERT_RESULT_BFC0]));
if (c->offset[VERT_RESULT_COL1] && c->offset[VERT_RESULT_BFC1])
brw_MOV(p,
byte_offset(c->reg.vertex[i], c->offset[VERT_RESULT_COL1]),
byte_offset(c->reg.vertex[i], c->offset[VERT_RESULT_BFC1]));
}
}
brw_ENDIF(p, ccw);
}
/// @copydoc block_partitioning::bytes_used(uint32_t) const
uint32_t bytes_used(uint32_t block_id) const
{
assert(block_id < m_total_blocks);
uint32_t offset = byte_offset(block_id);
assert(offset < m_object_size);
uint32_t remaining = m_object_size - offset;
uint32_t the_block_size = block_size(block_id);
return std::min(remaining, the_block_size);
}
bool MemoryType::isInitialized(Location addr, size_t len) const
{
const size_t offset = byte_offset(startAddress, addr, blockSize(), 0);
const size_t limit = offset + len;
for (size_t i = offset; i < limit; ++i)
{
if ( !byteInitialized(i) ) return false;
}
return true;
}
static int pack_double (grib_accessor* a, const double* val, size_t *len)
{
grib_accessor_ibmfloat* self = (grib_accessor_ibmfloat*)a;
int ret = 0;
unsigned long i = 0;
unsigned long rlen = *len;
size_t buflen = 0;
unsigned char *buf = NULL;
long off = 0;
if(*len < 1)
{
grib_context_log(a->parent->h->context, GRIB_LOG_ERROR, " wrong size for %s it pack at least 1 values ", a->name , rlen );
*len = 0;
return GRIB_ARRAY_TOO_SMALL;
}
if (rlen == 1){
/*
double x = 0;
grib_nearest_smaller_ibm_float(val[0],&x);
double y = grib_long_to_ibm(grib_ibm_to_long(val[0]));
printf("IBMFLOAT val=%.20f nearest_smaller_ibm_float=%.20f long_to_ibm=%.20f\n",val[0],x ,y);
*/
off = byte_offset(a)*8;
ret = grib_encode_unsigned_long(a->parent->h->buffer->data,grib_ibm_to_long(val[0]), &off, 32);
if (*len > 1) grib_context_log(a->parent->h->context, GRIB_LOG_WARNING, "grib_accessor_unsigned : Trying to pack %d values in a scalar %s, packing first value", *len, a->name );
if (ret == GRIB_SUCCESS) len[0] = 1;
return ret;
}
buflen = rlen*4;
buf = (unsigned char*)grib_context_malloc(a->parent->h->context,buflen);
for(i=0; i < rlen;i++){
grib_encode_unsigned_longb(buf,grib_ibm_to_long(val[i]), &off, 32);
}
ret = grib_set_long_internal(a->parent->h,grib_arguments_get_name(a->parent->h,self->arg,0),rlen);
if(ret == GRIB_SUCCESS)
grib_buffer_replace(a, buf, buflen,1,1);
else
*len = 0;
grib_context_free(a->parent->h->context,buf);
a->length = byte_count(a);
return ret;
}
/* This is performed against the original triangles, so no indirection
* required:
BZZZT!
*/
static void compute_tri_direction( struct brw_clip_compile *c )
{
struct brw_compile *p = &c->func;
struct brw_reg e = c->reg.tmp0;
struct brw_reg f = c->reg.tmp1;
struct brw_reg v0 = byte_offset(c->reg.vertex[0], c->offset[VERT_RESULT_HPOS]);
struct brw_reg v1 = byte_offset(c->reg.vertex[1], c->offset[VERT_RESULT_HPOS]);
struct brw_reg v2 = byte_offset(c->reg.vertex[2], c->offset[VERT_RESULT_HPOS]);
/* Calculate the vectors of two edges of the triangle:
*/
brw_ADD(p, e, v0, negate(v2));
brw_ADD(p, f, v1, negate(v2));
/* Take their crossproduct:
*/
brw_set_access_mode(p, BRW_ALIGN_16);
brw_MUL(p, vec4(brw_null_reg()), brw_swizzle(e, 1,2,0,3), brw_swizzle(f,2,0,1,3));
brw_MAC(p, vec4(e), negate(brw_swizzle(e, 2,0,1,3)), brw_swizzle(f,1,2,0,3));
brw_set_access_mode(p, BRW_ALIGN_1);
brw_MUL(p, c->reg.dir, c->reg.dir, vec4(e));
}
// declared static by preceeding declaration
std::ptrdiff_t byte_offset(Address base, Address elem, size_t blocksize, size_t blockofs)
{
ez::unused(blocksize), ez::unused(blockofs);
#if WITH_UPC
if (blocksize > 0)
{
base.local -= blockofs;
return byte_offset(base, elem, blocksize) - blockofs;
}
#endif /* WITH_UPC */
assert(blocksize == 0 && blockofs == 0);
return elem.local - base.local;
}
/* If flatshading, distribute color from provoking vertex prior to
* clipping.
*/
void brw_clip_copy_colors( struct brw_clip_compile *c,
GLuint to, GLuint from )
{
struct brw_compile *p = &c->func;
if (brw_clip_have_varying(c, VARYING_SLOT_COL0))
brw_MOV(p,
byte_offset(c->reg.vertex[to],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_COL0)),
byte_offset(c->reg.vertex[from],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_COL0)));
if (brw_clip_have_varying(c, VARYING_SLOT_COL1))
brw_MOV(p,
byte_offset(c->reg.vertex[to],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_COL1)),
byte_offset(c->reg.vertex[from],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_COL1)));
if (brw_clip_have_varying(c, VARYING_SLOT_BFC0))
brw_MOV(p,
byte_offset(c->reg.vertex[to],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_BFC0)),
byte_offset(c->reg.vertex[from],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_BFC0)));
if (brw_clip_have_varying(c, VARYING_SLOT_BFC1))
brw_MOV(p,
byte_offset(c->reg.vertex[to],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_BFC1)),
byte_offset(c->reg.vertex[from],
brw_varying_to_offset(&c->vue_map,
VARYING_SLOT_BFC1)));
}
std::pair<MemoryType*, bool>
MemoryManager::checkWrite(Location addr, size_t size, const RsType* t)
{
if ( diagnostics::message(diagnostics::memory) )
{
std::stringstream msg;
msg << " ++ checkWrite: " << addr << " size: " << size;
RuntimeSystem::instance().printMessage(msg.str());
}
bool statuschange = false;
MemoryType* mt = ::checkLocation(*this, addr, size, RuntimeViolation::INVALID_WRITE);
// the address has to be inside the block
assert(mt && mt->containsMemArea(addr, size));
const long blocksize = mt->blockSize();
const size_t ofs = byte_offset(mt->beginAddress(), addr, blocksize, 0);
if (t)
{
if ( diagnostics::message(diagnostics::memory) )
{
RuntimeSystem& rs = RuntimeSystem::instance();
std::stringstream msg;
msg << "++ found memory addr: " << *mt << " for " << addr << " with size " << ToString(size);
rs.printMessage(msg.str());
}
statuschange = mt->registerMemType(addr, ofs, t);
}
const bool initmod = mt->initialize(ofs, size);
if ( diagnostics::message(diagnostics::memory) )
{
RuntimeSystem::instance().printMessage(" ++ checkWrite done.");
}
return std::make_pair(mt, initmod || statuschange);
}
static
std::ptrdiff_t byte_offset(Address base, Address elem, size_t blocksize)
{
assert(blocksize != 0);
// adjust base so that it is aligned on a blocksize address
const size_t baseaddr = reinterpret_cast<size_t>(base.local);
const size_t basescale = baseaddr / blocksize;
const size_t baseshift = baseaddr % blocksize;
// adjust elem by the same amount
const size_t elemaddr = reinterpret_cast<size_t>(elem.local - baseshift);
const size_t elemscale = elemaddr / blocksize;
const size_t elemphase = elemaddr % blocksize;
// use simpler calculation offset calculation
const std::ptrdiff_t scaleofs = byte_offset(basescale, base.thread_id, elemscale, elem.thread_id);
// unscale and add back the elem's phase
return scaleofs * blocksize + elemphase;
}
void
vec4_tcs_visitor::emit_urb_write(const src_reg &value,
unsigned writemask,
unsigned base_offset,
const src_reg &indirect_offset)
{
if (writemask == 0)
return;
src_reg message(this, glsl_type::uvec4_type, 2);
vec4_instruction *inst;
inst = emit(TCS_OPCODE_SET_OUTPUT_URB_OFFSETS, dst_reg(message),
brw_imm_ud(writemask), indirect_offset);
inst->force_writemask_all = true;
inst = emit(MOV(byte_offset(dst_reg(retype(message, value.type)), REG_SIZE),
value));
inst->force_writemask_all = true;
inst = emit(TCS_OPCODE_URB_WRITE, dst_null_f(), message);
inst->offset = base_offset;
inst->mlen = 2;
inst->base_mrf = -1;
}
void brw_clip_tri_alloc_regs( struct brw_clip_compile *c,
GLuint nr_verts )
{
GLuint i = 0,j;
/* Register usage is static, precompute here:
*/
c->reg.R0 = retype(brw_vec8_grf(i, 0), BRW_REGISTER_TYPE_UD); i++;
if (c->key.nr_userclip) {
c->reg.fixed_planes = brw_vec4_grf(i, 0);
i += (6 + c->key.nr_userclip + 1) / 2;
c->prog_data.curb_read_length = (6 + c->key.nr_userclip + 1) / 2;
}
else
c->prog_data.curb_read_length = 0;
/* Payload vertices plus space for more generated vertices:
*/
for (j = 0; j < nr_verts; j++) {
c->reg.vertex[j] = brw_vec4_grf(i, 0);
i += c->nr_regs;
}
if (c->key.nr_attrs & 1) {
for (j = 0; j < 3; j++) {
GLuint delta = c->key.nr_attrs*16 + 32;
if (c->chipset.is_igdng)
delta = c->key.nr_attrs * 16 + 32 * 3;
brw_MOV(&c->func, byte_offset(c->reg.vertex[j], delta), brw_imm_f(0));
}
}
c->reg.t = brw_vec1_grf(i, 0);
c->reg.loopcount = retype(brw_vec1_grf(i, 1), BRW_REGISTER_TYPE_D);
c->reg.nr_verts = retype(brw_vec1_grf(i, 2), BRW_REGISTER_TYPE_UD);
c->reg.planemask = retype(brw_vec1_grf(i, 3), BRW_REGISTER_TYPE_UD);
c->reg.plane_equation = brw_vec4_grf(i, 4);
i++;
c->reg.dpPrev = brw_vec1_grf(i, 0); /* fixme - dp4 will clobber r.1,2,3 */
c->reg.dp = brw_vec1_grf(i, 4);
i++;
c->reg.inlist = brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, i, 0);
i++;
c->reg.outlist = brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, i, 0);
i++;
c->reg.freelist = brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, i, 0);
i++;
if (!c->key.nr_userclip) {
c->reg.fixed_planes = brw_vec8_grf(i, 0);
i++;
}
if (c->key.do_unfilled) {
c->reg.dir = brw_vec4_grf(i, 0);
c->reg.offset = brw_vec4_grf(i, 4);
i++;
c->reg.tmp0 = brw_vec4_grf(i, 0);
c->reg.tmp1 = brw_vec4_grf(i, 4);
i++;
}
if (c->need_ff_sync) {
c->reg.ff_sync = retype(brw_vec1_grf(i, 0), BRW_REGISTER_TYPE_UD);
i++;
}
c->first_tmp = i;
c->last_tmp = i;
c->prog_data.urb_read_length = c->nr_regs; /* ? */
c->prog_data.total_grf = i;
}
static long next_offset(grib_accessor* a) {
return byte_offset(a)+a->length;
}
bool MemoryType::containsMemArea(Location queryAddress, size_t len) const
{
std::ptrdiff_t ofs = byte_offset(startAddress, queryAddress, blockSize(), 0);
return (ofs >= 0) && ((ofs+len) <= getSize());
}
bool
MemoryManager::checkIfSameChunk(Location addr1, Location addr2, size_t typeSize, RuntimeViolation::Type violation) const
{
RuntimeViolation::Type access_violation = violation;
if (access_violation != RuntimeViolation::NONE) access_violation = RuntimeViolation::INVALID_READ;
const MemoryType* mem1 = checkLocation(addr1, typeSize, access_violation);
const bool sameChunk = (mem1 && mem1->containsMemArea(addr2, typeSize));
// check that addr1 and addr2 point to the same base location
if (! sameChunk)
{
const MemoryType* mem2 = checkLocation(addr2, typeSize, access_violation);
// the error is skipped, if a chunk is not available b/c pointer errors
// should be recorded only when the out-of-bounds memory is accessed, but
// not when the pointer is moved out of bounds.
const bool skiperr = (violation == RuntimeViolation::NONE && !(mem1 && mem2));
if (!skiperr)
{
assert(mem1 && mem2 && mem1 != mem2);
RuntimeSystem& rs = RuntimeSystem::instance();
std::stringstream ss;
ss << "Pointer changed allocation block from " << addr1 << " to " << addr2 << std::endl;
rs.violationHandler( violation, ss.str() );
}
return false;
}
// so far we know that addr1, addr2 have been allocated within the same chunk
// now, test for same array chunks
const size_t totalblocksize = mem1->blockSize();
const Location memloc = mem1->beginAddress();
const std::ptrdiff_t ofs1 = byte_offset(memloc, addr1, totalblocksize, 0);
const std::ptrdiff_t ofs2 = byte_offset(memloc, addr2, totalblocksize, 0);
//~ std::cerr << *mem1 << std::endl;
//~ std::cerr << "a = " << addr1 << std::endl;
//~ std::cerr << "b = " << addr2 << std::endl;
//~ std::cerr << "ts = " << typeSize << std::endl;
//~ std::cerr << "tbs = " << totalblocksize << std::endl;
// \pp as far as I understand, getTypeAt returns the innermost type
// that overlaps a certain memory region [addr1, typeSize).
// The current implementation stops at the highest level of array...
const RsType* type1 = mem1 -> getTypeAt( ofs1, typeSize );
const RsType* type2 = mem1 -> getTypeAt( ofs2, typeSize );
std::auto_ptr<const RsType> guard1(NULL);
std::auto_ptr<const RsType> guard2(NULL);
if( !type1 ) {
//~ std::cerr << "type1! " << std::endl;
type1 = mem1->computeCompoundTypeAt( ofs1, typeSize );
guard1.reset(type1);
}
if( !type2 ) {
//~ std::cerr << "type2! " << std::endl;
type2 = mem1->computeCompoundTypeAt( ofs2, typeSize );
guard2.reset(type2);
}
assert( type1 && type2 );
//~ std::cerr << "-- T1 = " << typeid(*type1).name() << std::endl;
//~ std::cerr << "-- T2 = " << typeid(*type2).name() << std::endl;
bool accessOK = type1->isConsistentWith( *type2 );
const RsArrayType* array = (accessOK ? dynamic_cast< const RsArrayType* >( type1 ) : 0);
//~ std::cerr << "accOK: " << accessOK << std::endl;
if (array)
{
// \pp in order to calculate the last memory location of elem, we need to
// know whether its type is an array type, and whether this is an
// immediate (there is no user defined type or pointer in between)
// subarray of the containing array.
const size_t blockofs = 0; /* \todo calculate offset from addr1 */
// \pp \todo
// not sure why bounds checking is based on a relative address (addr1)
// and not on absolute boundaries of the chunk where this address
// is located...
// e.g., addr2 < array(addr).lb || array(addr).ub < (addr2+typeSize)
// - a reason is that we want to bounds check within a multi-dimensional
// array; a sub-array might not start at the same address as the
// allocated chunk (which is always at [0][0]...)
// \bug this works fine for arrays, but it does not seem to be OK for
// pointers; in which case addr1 might point in the middle of a chunk.
const Location arrlb = addr1;
const std::ptrdiff_t arrlen = array->getByteSize();
const Location elemlo = addr2;
Location elemhi = addr2;
const RsType* memtype = mem1->getTypeAt( 0, mem1->getSize() );
const size_t basesize = basetypeSize(memtype);
if (typeSize > basesize)
{
// elem is another array
// assert( rs->testing() || dynamic_cast< const RsArrayType* >(type2) );
const long elemblockofs = 0; // \todo calculate the ofs of elem's address (is this needed?)
elemhi = ::add(elemhi, typeSize-1, totalblocksize, elemblockofs);
}
else
{
elemhi.local += (typeSize-1);
}
// \pp \note [arrlb, arrub] and [elemlo, elemhi] (therefore typeSize - 1)
// arrub < elemhi this holds for UPC
// in contrast, using a range [elemlb, elemub) causes problems
// for UPC where the elemub can be larger than arrub
// for last elements on threads other than 0.
// e.g., shared int x[THREADS];
// arrub == &x@thread(0) + sizeof(int)
// writing to x[1] has the following upper bound
// elemub == &x@thread(1) + sizeof(int)
// consequently arrub < elemub :(
#if EXTENDEDDBG
std::cerr << " arrlb = " << arrlb << "( lb + " << arrlen << " [" << totalblocksize << "] )" << std::endl;
std::cerr << " elmlb = " << elemlo << std::endl;
std::cerr << " elmub = " << elemhi << " ( lb + " << typeSize << " )" << std::endl;
std::cerr << " ofs(elemlb) >= 0 : " << byte_offset(arrlb, elemlo, totalblocksize, blockofs) << std::endl;
std::cerr << " ofs(elemub) < sz(arr) : " << (byte_offset(arrlb, elemhi, totalblocksize, blockofs)) << std::endl;
#endif /* EXTENDEDDBG */
// the offset of the element in respect to the array has to be positive
// and the offset of the elements last byte has to be less than the
// array length.
accessOK = (byte_offset(arrlb, elemlo, totalblocksize, blockofs) >= 0);
accessOK &= (byte_offset(arrlb, elemhi, totalblocksize, blockofs) < arrlen);
// the array element might be before it [ -1 ]
// ... or after the array [ N ]...
// was: accessOK = ! ( (elemlo < arrlb) || (arrub < elemhi) )
/* was:
if !( addr1 + array -> getByteSize() >= addr2 + typeSize // the array element might be after the array [ N ]...
&& addr1 <= addr2 // ... or before it [ -1 ]
* (12) (8)
)
{
// out of bounds error (e.g. int[2][3], ref [0][3])
consistent = false;
}
*/
}
if (!accessOK)
failNotSameChunk( *type1, *type2, addr1, addr2, *mem1, violation );
return accessOK;
}
static long next_offset(grib_accessor* a){
return byte_offset(a)+byte_count(a);
}
void brw_clip_tri_alloc_regs( struct brw_clip_compile *c,
GLuint nr_verts )
{
struct intel_context *intel = &c->func.brw->intel;
GLuint i = 0,j;
/* Register usage is static, precompute here:
*/
c->reg.R0 = retype(brw_vec8_grf(i, 0), BRW_REGISTER_TYPE_UD); i++;
if (c->key.nr_userclip) {
c->reg.fixed_planes = brw_vec4_grf(i, 0);
i += (6 + c->key.nr_userclip + 1) / 2;
c->prog_data.curb_read_length = (6 + c->key.nr_userclip + 1) / 2;
}
else
c->prog_data.curb_read_length = 0;
/* Payload vertices plus space for more generated vertices:
*/
for (j = 0; j < nr_verts; j++) {
c->reg.vertex[j] = brw_vec4_grf(i, 0);
i += c->nr_regs;
}
if (c->vue_map.num_slots % 2) {
/* The VUE has an odd number of slots so the last register is only half
* used. Fill the second half with zero.
*/
for (j = 0; j < 3; j++) {
GLuint delta = brw_vue_slot_to_offset(c->vue_map.num_slots);
brw_MOV(&c->func, byte_offset(c->reg.vertex[j], delta), brw_imm_f(0));
}
}
c->reg.t = brw_vec1_grf(i, 0);
c->reg.loopcount = retype(brw_vec1_grf(i, 1), BRW_REGISTER_TYPE_D);
c->reg.nr_verts = retype(brw_vec1_grf(i, 2), BRW_REGISTER_TYPE_UD);
c->reg.planemask = retype(brw_vec1_grf(i, 3), BRW_REGISTER_TYPE_UD);
c->reg.plane_equation = brw_vec4_grf(i, 4);
i++;
c->reg.dpPrev = brw_vec1_grf(i, 0); /* fixme - dp4 will clobber r.1,2,3 */
c->reg.dp = brw_vec1_grf(i, 4);
i++;
c->reg.inlist = brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, i, 0);
i++;
c->reg.outlist = brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, i, 0);
i++;
c->reg.freelist = brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, i, 0);
i++;
if (!c->key.nr_userclip) {
c->reg.fixed_planes = brw_vec8_grf(i, 0);
i++;
}
if (c->key.do_unfilled) {
c->reg.dir = brw_vec4_grf(i, 0);
c->reg.offset = brw_vec4_grf(i, 4);
i++;
c->reg.tmp0 = brw_vec4_grf(i, 0);
c->reg.tmp1 = brw_vec4_grf(i, 4);
i++;
}
if (intel->needs_ff_sync) {
c->reg.ff_sync = retype(brw_vec1_grf(i, 0), BRW_REGISTER_TYPE_UD);
i++;
}
c->first_tmp = i;
c->last_tmp = i;
c->prog_data.urb_read_length = c->nr_regs; /* ? */
c->prog_data.total_grf = i;
}