static void emit_pixel_xy(struct brw_compile *p,
const struct brw_reg *dst,
GLuint mask,
const struct brw_reg *arg0)
{
struct brw_reg r1 = brw_vec1_grf(1, 0);
struct brw_reg r1_uw = retype(r1, BRW_REGISTER_TYPE_UW);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
/* Calculate pixel centers by adding 1 or 0 to each of the
* micro-tile coordinates passed in r1.
*/
if (mask & WRITEMASK_X) {
brw_ADD(p,
vec16(retype(dst[0], BRW_REGISTER_TYPE_UW)),
stride(suboffset(r1_uw, 4), 2, 4, 0),
brw_imm_v(0x10101010));
}
if (mask & WRITEMASK_Y) {
brw_ADD(p,
vec16(retype(dst[1], BRW_REGISTER_TYPE_UW)),
stride(suboffset(r1_uw,5), 2, 4, 0),
brw_imm_v(0x11001100));
}
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
}
static void brw_wm_xy(struct brw_compile *p, int dw)
{
struct brw_reg r1 = brw_vec1_grf(1, 0);
struct brw_reg r1_uw = __retype_uw(r1);
struct brw_reg x_uw, y_uw;
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
if (dw == 16) {
x_uw = brw_uw16_grf(30, 0);
y_uw = brw_uw16_grf(28, 0);
} else {
x_uw = brw_uw8_grf(30, 0);
y_uw = brw_uw8_grf(28, 0);
}
brw_ADD(p,
x_uw,
__stride(__suboffset(r1_uw, 4), 2, 4, 0),
brw_imm_v(0x10101010));
brw_ADD(p,
y_uw,
__stride(__suboffset(r1_uw, 5), 2, 4, 0),
brw_imm_v(0x11001100));
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
brw_ADD(p, brw_vec8_grf(X16, 0), vec8(x_uw), brw_negate(r1));
brw_ADD(p, brw_vec8_grf(Y16, 0), vec8(y_uw), brw_negate(__suboffset(r1, 1)));
}
void emit_math1(struct brw_wm_compile *c,
GLuint function,
const struct brw_reg *dst,
GLuint mask,
const struct brw_reg *arg0)
{
struct brw_compile *p = &c->func;
struct intel_context *intel = &p->brw->intel;
int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
GLuint saturate = ((mask & SATURATE) ?
BRW_MATH_SATURATE_SATURATE :
BRW_MATH_SATURATE_NONE);
struct brw_reg src;
if (!(mask & WRITEMASK_XYZW))
return; /* Do not emit dead code */
assert(is_power_of_two(mask & WRITEMASK_XYZW));
if (intel->gen >= 6 && ((arg0[0].hstride == BRW_HORIZONTAL_STRIDE_0 ||
arg0[0].file != BRW_GENERAL_REGISTER_FILE) ||
arg0[0].negate || arg0[0].abs)) {
/* Gen6 math requires that source and dst horizontal stride be 1,
* and that the argument be in the GRF.
*
* The hardware ignores source modifiers (negate and abs) on math
* instructions, so we also move to a temp to set those up.
*/
src = dst[dst_chan];
brw_MOV(p, src, arg0[0]);
} else {
src = arg0[0];
}
/* Send two messages to perform all 16 operations:
*/
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_math(p,
dst[dst_chan],
function,
saturate,
2,
src,
BRW_MATH_DATA_VECTOR,
BRW_MATH_PRECISION_FULL);
if (c->dispatch_width == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_math(p,
offset(dst[dst_chan],1),
function,
saturate,
3,
sechalf(src),
BRW_MATH_DATA_VECTOR,
BRW_MATH_PRECISION_FULL);
}
brw_pop_insn_state(p);
}
static void brw_fb_write(struct brw_compile *p, int dw)
{
struct brw_instruction *insn;
unsigned msg_control, msg_type, msg_len;
struct brw_reg src0;
bool header;
if (dw == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE;
msg_len = 8;
} else {
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01;
msg_len = 4;
}
if (p->gen < 060) {
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_MOV(p, brw_message_reg(1), brw_vec8_grf(1, 0));
brw_pop_insn_state(p);
msg_len += 2;
}
/* The execution mask is ignored for render target writes. */
insn = brw_next_insn(p, BRW_OPCODE_SEND);
insn->header.predicate_control = 0;
insn->header.compression_control = BRW_COMPRESSION_NONE;
if (p->gen >= 060) {
msg_type = GEN6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE;
src0 = brw_message_reg(2);
header = false;
} else {
insn->header.destreg__conditionalmod = 0;
msg_type = BRW_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE;
src0 = __retype_uw(brw_vec8_grf(0, 0));
header = true;
}
brw_set_dest(p, insn, null_result(dw));
brw_set_src0(p, insn, src0);
brw_set_dp_write_message(p, insn, 0,
msg_control, msg_type, msg_len,
header, true, 0, true, false);
}
/* Kill pixel - set execution mask to zero for those pixels which
* fail.
*/
static void emit_kil( struct brw_wm_compile *c,
struct brw_reg *arg0)
{
struct brw_compile *p = &c->func;
struct intel_context *intel = &p->brw->intel;
struct brw_reg pixelmask;
GLuint i, j;
if (intel->gen >= 6)
pixelmask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
else
pixelmask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
for (i = 0; i < 4; i++) {
/* Check if we've already done the comparison for this reg
* -- common when someone does KIL TEMP.wwww.
*/
for (j = 0; j < i; j++) {
if (memcmp(&arg0[j], &arg0[i], sizeof(arg0[0])) == 0)
break;
}
if (j != i)
continue;
brw_push_insn_state(p);
brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_GE, arg0[i], brw_imm_f(0));
brw_set_predicate_control_flag_value(p, 0xff);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_AND(p, pixelmask, brw_flag_reg(), pixelmask);
brw_pop_insn_state(p);
}
}
static void fire_fb_write( struct brw_wm_compile *c,
GLuint base_reg,
GLuint nr,
GLuint target,
GLuint eot )
{
struct brw_compile *p = &c->func;
/* Pass through control information:
*/
/* mov (8) m1.0<1>:ud r1.0<8;8,1>:ud { Align1 NoMask } */
{
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE); /* ? */
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p,
brw_message_reg(base_reg + 1),
brw_vec8_grf(1, 0));
brw_pop_insn_state(p);
}
/* Send framebuffer write message: */
/* send (16) null.0<1>:uw m0 r0.0<8;8,1>:uw 0x85a04000:ud { Align1 EOT } */
brw_fb_WRITE(p,
retype(vec16(brw_null_reg()), BRW_REGISTER_TYPE_UW),
base_reg,
retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW),
target,
nr,
0,
eot);
}
void
brw_init_compile(struct brw_context *brw, struct brw_compile *p, void *mem_ctx)
{
p->brw = brw;
/*
* Set the initial instruction store array size to 1024, if found that
* isn't enough, then it will double the store size at brw_next_insn()
* until out of memory.
*/
p->store_size = 1024;
p->store = rzalloc_array(mem_ctx, struct brw_instruction, p->store_size);
p->nr_insn = 0;
p->current = p->stack;
p->compressed = false;
memset(p->current, 0, sizeof(p->current[0]));
p->mem_ctx = mem_ctx;
/* Some defaults?
*/
brw_set_mask_control(p, BRW_MASK_ENABLE); /* what does this do? */
brw_set_saturate(p, 0);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_predicate_control_flag_value(p, 0xff);
/* Set up control flow stack */
p->if_stack_depth = 0;
p->if_stack_array_size = 16;
p->if_stack = rzalloc_array(mem_ctx, int, p->if_stack_array_size);
p->loop_stack_depth = 0;
p->loop_stack_array_size = 16;
p->loop_stack = rzalloc_array(mem_ctx, int, p->loop_stack_array_size);
p->if_depth_in_loop = rzalloc_array(mem_ctx, int, p->loop_stack_array_size);
}
static void brw_wm_write(struct brw_compile *p, int dw, int src)
{
int n;
if (dw == 8 && p->gen >= 060) {
/* XXX pixel execution mask? */
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p, brw_message_reg(2), brw_vec8_grf(src+0, 0));
brw_MOV(p, brw_message_reg(3), brw_vec8_grf(src+1, 0));
brw_MOV(p, brw_message_reg(4), brw_vec8_grf(src+2, 0));
brw_MOV(p, brw_message_reg(5), brw_vec8_grf(src+3, 0));
goto done;
}
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
for (n = 0; n < 4; n++) {
if (p->gen >= 060) {
brw_MOV(p,
brw_message_reg(2 + 2*n),
brw_vec8_grf(src + 2*n, 0));
} else if (p->gen >= 045 && dw == 16) {
brw_MOV(p,
brw_message_reg(2 + n + BRW_MRF_COMPR4),
brw_vec8_grf(src + 2*n, 0));
} else {
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p,
brw_message_reg(2 + n),
brw_vec8_grf(src + 2*n, 0));
if (dw == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_MOV(p,
brw_message_reg(2 + n + 4),
brw_vec8_grf(src + 2*n+1, 0));
}
}
}
done:
brw_fb_write(p, dw);
}
void brw_init_compile( struct brw_compile *p )
{
p->nr_insn = 0;
p->current = p->stack;
memset(p->current, 0, sizeof(p->current[0]));
/* Some defaults?
*/
brw_set_mask_control(p, BRW_MASK_ENABLE); /* what does this do? */
brw_set_saturate(p, 0);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_predicate_control_flag_value(p, 0xff);
}
/**
* Extended math function, float[16].
* Use 2 send instructions.
*/
void brw_math_16( struct brw_compile *p,
struct brw_reg dest,
GLuint function,
GLuint saturate,
GLuint msg_reg_nr,
struct brw_reg src,
GLuint precision )
{
struct brw_instruction *insn;
GLuint msg_length = (function == BRW_MATH_FUNCTION_POW) ? 2 : 1;
GLuint response_length = (function == BRW_MATH_FUNCTION_SINCOS) ? 2 : 1;
/* First instruction:
*/
brw_push_insn_state(p);
brw_set_predicate_control_flag_value(p, 0xff);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
insn = next_insn(p, BRW_OPCODE_SEND);
insn->header.destreg__conditionalmod = msg_reg_nr;
brw_set_dest(insn, dest);
brw_set_src0(insn, src);
brw_set_math_message(p->brw,
insn,
msg_length, response_length,
function,
BRW_MATH_INTEGER_UNSIGNED,
precision,
saturate,
BRW_MATH_DATA_VECTOR);
/* Second instruction:
*/
insn = next_insn(p, BRW_OPCODE_SEND);
insn->header.compression_control = BRW_COMPRESSION_2NDHALF;
insn->header.destreg__conditionalmod = msg_reg_nr+1;
brw_set_dest(insn, offset(dest,1));
brw_set_src0(insn, src);
brw_set_math_message(p->brw,
insn,
msg_length, response_length,
function,
BRW_MATH_INTEGER_UNSIGNED,
precision,
saturate,
BRW_MATH_DATA_VECTOR);
brw_pop_insn_state(p);
}
/**
* Read a float[4] vector from the data port Data Cache (const buffer).
* Location (in buffer) should be a multiple of 16.
* Used for fetching shader constants.
* If relAddr is true, we'll do an indirect fetch using the address register.
*/
void brw_dp_READ_4( struct brw_compile *p,
struct brw_reg dest,
GLboolean relAddr,
GLuint location,
GLuint bind_table_index )
{
/* XXX: relAddr not implemented */
GLuint msg_reg_nr = 1;
{
struct brw_reg b;
brw_push_insn_state(p);
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_mask_control(p, BRW_MASK_DISABLE);
/* Setup MRF[1] with location/offset into const buffer */
b = brw_message_reg(msg_reg_nr);
b = retype(b, BRW_REGISTER_TYPE_UD);
/* XXX I think we're setting all the dwords of MRF[1] to 'location'.
* when the docs say only dword[2] should be set. Hmmm. But it works.
*/
brw_MOV(p, b, brw_imm_ud(location));
brw_pop_insn_state(p);
}
{
struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
insn->header.predicate_control = BRW_PREDICATE_NONE;
insn->header.compression_control = BRW_COMPRESSION_NONE;
insn->header.destreg__conditionalmod = msg_reg_nr;
insn->header.mask_control = BRW_MASK_DISABLE;
/* cast dest to a uword[8] vector */
dest = retype(vec8(dest), BRW_REGISTER_TYPE_UW);
brw_set_dest(insn, dest);
brw_set_src0(insn, brw_null_reg());
brw_set_dp_read_message(p->brw,
insn,
bind_table_index,
0, /* msg_control (0 means 1 Oword) */
BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ, /* msg_type */
0, /* source cache = data cache */
1, /* msg_length */
1, /* response_length (1 Oword) */
0); /* eot */
}
}
static void emit_aa( struct brw_wm_compile *c,
struct brw_reg *arg1,
GLuint reg )
{
struct brw_compile *p = &c->func;
GLuint comp = c->aa_dest_stencil_reg / 2;
GLuint off = c->aa_dest_stencil_reg % 2;
struct brw_reg aa = offset(arg1[comp], off);
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE); /* ?? */
brw_MOV(p, brw_message_reg(reg), aa);
brw_pop_insn_state(p);
}
static void brw_wm_affine_st(struct brw_compile *p, int dw,
int channel, int msg)
{
int uv;
if (dw == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
uv = p->gen >= 060 ? 6 : 3;
} else {
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
uv = p->gen >= 060 ? 4 : 3;
}
uv += 2*channel;
msg++;
if (p->gen >= 060) {
brw_PLN(p,
brw_message_reg(msg),
brw_vec1_grf(uv, 0),
brw_vec8_grf(2, 0));
msg += dw/8;
brw_PLN(p,
brw_message_reg(msg),
brw_vec1_grf(uv, 4),
brw_vec8_grf(2, 0));
} else {
struct brw_reg r = brw_vec1_grf(uv, 0);
brw_LINE(p, brw_null_reg(), __suboffset(r, 0), brw_vec8_grf(X16, 0));
brw_MAC(p, brw_message_reg(msg), __suboffset(r, 1), brw_vec8_grf(Y16, 0));
msg += dw/8;
brw_LINE(p, brw_null_reg(), __suboffset(r, 4), brw_vec8_grf(X16, 0));
brw_MAC(p, brw_message_reg(msg), __suboffset(r, 5), brw_vec8_grf(Y16, 0));
}
}
static void fire_fb_write( struct brw_wm_compile *c,
GLuint base_reg,
GLuint nr,
GLuint target,
GLuint eot )
{
struct brw_compile *p = &c->func;
struct intel_context *intel = &p->brw->intel;
uint32_t msg_control;
/* Pass through control information:
*
* Gen6 has done m1 mov in emit_fb_write() for current SIMD16 case.
*/
/* mov (8) m1.0<1>:ud r1.0<8;8,1>:ud { Align1 NoMask } */
if (intel->gen < 6)
{
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE); /* ? */
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p,
brw_message_reg(base_reg + 1),
brw_vec8_grf(1, 0));
brw_pop_insn_state(p);
}
if (c->dispatch_width == 16)
msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE;
else
msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01;
/* Send framebuffer write message: */
/* send (16) null.0<1>:uw m0 r0.0<8;8,1>:uw 0x85a04000:ud { Align1 EOT } */
brw_fb_WRITE(p,
c->dispatch_width,
base_reg,
retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW),
msg_control,
target,
nr,
0,
eot,
true);
}
/* Kill pixel - set execution mask to zero for those pixels which
* fail.
*/
static void emit_kil( struct brw_wm_compile *c,
struct brw_reg *arg0)
{
struct brw_compile *p = &c->func;
struct brw_reg r0uw = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
GLuint i;
/* XXX - usually won't need 4 compares!
*/
for (i = 0; i < 4; i++) {
brw_push_insn_state(p);
brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_GE, arg0[i], brw_imm_f(0));
brw_set_predicate_control_flag_value(p, 0xff);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_AND(p, r0uw, brw_flag_reg(), r0uw);
brw_pop_insn_state(p);
}
}
static void emit_math2( struct brw_compile *p,
GLuint function,
const struct brw_reg *dst,
GLuint mask,
const struct brw_reg *arg0,
const struct brw_reg *arg1)
{
if (!(mask & WRITEMASK_XYZW))
return; /* Do not emit dead code*/
assert((mask & WRITEMASK_XYZW) == WRITEMASK_X);
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p, brw_message_reg(2), arg0[0]);
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_MOV(p, brw_message_reg(4), sechalf(arg0[0]));
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p, brw_message_reg(3), arg1[0]);
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_MOV(p, brw_message_reg(5), sechalf(arg1[0]));
/* Send two messages to perform all 16 operations:
*/
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_math(p,
dst[0],
function,
(mask & SATURATE) ? BRW_MATH_SATURATE_SATURATE : BRW_MATH_SATURATE_NONE,
2,
brw_null_reg(),
BRW_MATH_DATA_VECTOR,
BRW_MATH_PRECISION_FULL);
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_math(p,
offset(dst[0],1),
function,
(mask & SATURATE) ? BRW_MATH_SATURATE_SATURATE : BRW_MATH_SATURATE_NONE,
4,
brw_null_reg(),
BRW_MATH_DATA_VECTOR,
BRW_MATH_PRECISION_FULL);
brw_pop_insn_state(p);
}
static void wm_src_sample_argb(struct brw_compile *p)
{
static const uint32_t fragment[][4] = {
#include "exa_wm_src_affine.g6b"
#include "exa_wm_src_sample_argb.g6b"
#include "exa_wm_write.g6b"
};
int n;
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p,
retype(brw_vec1_grf(0,2), BRW_REGISTER_TYPE_UD),
brw_imm_ud(0));
brw_pop_insn_state(p);
brw_SAMPLE(p,
retype(vec16(brw_vec8_grf(14, 0)), BRW_REGISTER_TYPE_UW),
1,
retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD),
1, 0,
WRITEMASK_XYZW,
GEN5_SAMPLER_MESSAGE_SAMPLE,
8,
5,
true,
BRW_SAMPLER_SIMD_MODE_SIMD16);
for (n = 0; n < p->nr_insn; n++) {
brw_disasm(stdout, &p->store[n], 60);
}
printf("\n\n");
for (n = 0; n < ARRAY_SIZE(fragment); n++) {
brw_disasm(stdout,
(const struct brw_instruction *)&fragment[n][0],
60);
}
}
/**
* Write block of 16 dwords/floats to the data port Render Cache scratch buffer.
* Scratch offset should be a multiple of 64.
* Used for register spilling.
*/
void brw_dp_WRITE_16( struct brw_compile *p,
struct brw_reg src,
GLuint scratch_offset )
{
GLuint msg_reg_nr = 1;
{
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
/* set message header global offset field (reg 0, element 2) */
brw_MOV(p,
retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_D),
brw_imm_d(scratch_offset));
brw_pop_insn_state(p);
}
{
GLuint msg_length = 3;
struct brw_reg dest = retype(brw_null_reg(), BRW_REGISTER_TYPE_UW);
struct brw_instruction *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;
brw_set_dest(insn, dest);
brw_set_src0(insn, src);
brw_set_dp_write_message(p->brw,
insn,
255, /* binding table index (255=stateless) */
BRW_DATAPORT_OWORD_BLOCK_4_OWORDS, /* msg_control */
BRW_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE, /* msg_type */
msg_length,
0, /* pixel scoreboard */
0, /* response_length */
0); /* eot */
}
}
/**
* Read block of 16 dwords/floats from the data port Render Cache scratch buffer.
* Scratch offset should be a multiple of 64.
* Used for register spilling.
*/
void brw_dp_READ_16( struct brw_compile *p,
struct brw_reg dest,
GLuint scratch_offset )
{
GLuint msg_reg_nr = 1;
{
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_mask_control(p, BRW_MASK_DISABLE);
/* set message header global offset field (reg 0, element 2) */
brw_MOV(p,
retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_D),
brw_imm_d(scratch_offset));
brw_pop_insn_state(p);
}
{
struct brw_instruction *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;
brw_set_dest(insn, dest); /* UW? */
brw_set_src0(insn, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW));
brw_set_dp_read_message(p->brw,
insn,
255, /* binding table index (255=stateless) */
3, /* msg_control (3 means 4 Owords) */
BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ, /* msg_type */
1, /* target cache (render/scratch) */
1, /* msg_length */
2, /* response_length */
0); /* eot */
}
}
/**
* Computes the screen-space x,y position of the pixels.
*
* This will be used by emit_delta_xy() or emit_wpos_xy() for
* interpolation of attributes..
*
* Payload R0:
*
* R0.0 -- pixel mask, one bit for each of 4 pixels in 4 tiles,
* corresponding to each of the 16 execution channels.
* R0.1..8 -- ?
* R1.0 -- triangle vertex 0.X
* R1.1 -- triangle vertex 0.Y
* R1.2 -- tile 0 x,y coords (2 packed uwords)
* R1.3 -- tile 1 x,y coords (2 packed uwords)
* R1.4 -- tile 2 x,y coords (2 packed uwords)
* R1.5 -- tile 3 x,y coords (2 packed uwords)
* R1.6 -- ?
* R1.7 -- ?
* R1.8 -- ?
*/
void emit_pixel_xy(struct brw_wm_compile *c,
const struct brw_reg *dst,
GLuint mask)
{
struct brw_compile *p = &c->func;
struct brw_reg r1 = brw_vec1_grf(1, 0);
struct brw_reg r1_uw = retype(r1, BRW_REGISTER_TYPE_UW);
struct brw_reg dst0_uw, dst1_uw;
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
if (c->dispatch_width == 16) {
dst0_uw = vec16(retype(dst[0], BRW_REGISTER_TYPE_UW));
dst1_uw = vec16(retype(dst[1], BRW_REGISTER_TYPE_UW));
} else {
dst0_uw = vec8(retype(dst[0], BRW_REGISTER_TYPE_UW));
dst1_uw = vec8(retype(dst[1], BRW_REGISTER_TYPE_UW));
}
/* Calculate pixel centers by adding 1 or 0 to each of the
* micro-tile coordinates passed in r1.
*/
if (mask & WRITEMASK_X) {
brw_ADD(p,
dst0_uw,
stride(suboffset(r1_uw, 4), 2, 4, 0),
brw_imm_v(0x10101010));
}
if (mask & WRITEMASK_Y) {
brw_ADD(p,
dst1_uw,
stride(suboffset(r1_uw,5), 2, 4, 0),
brw_imm_v(0x11001100));
}
brw_pop_insn_state(p);
}
void brw_compile_init(struct brw_compile *p, int gen, void *store)
{
assert(gen);
p->gen = gen;
p->store = store;
p->nr_insn = 0;
p->current = p->stack;
p->compressed = false;
memset(p->current, 0, sizeof(p->current[0]));
/* Some defaults?
*/
brw_set_mask_control(p, BRW_MASK_ENABLE); /* what does this do? */
brw_set_saturate(p, 0);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_predicate_control_flag_value(p, 0xff);
p->if_stack_depth = 0;
p->if_stack_array_size = 0;
p->if_stack = NULL;
}
/* Post-fragment-program processing. Send the results to the
* framebuffer.
*/
static void emit_fb_write( struct brw_wm_compile *c,
struct brw_reg *arg0,
struct brw_reg *arg1,
struct brw_reg *arg2,
GLuint target,
GLuint eot)
{
struct brw_compile *p = &c->func;
GLuint nr = 2;
GLuint channel;
/* Reserve a space for AA - may not be needed:
*/
if (c->key.aa_dest_stencil_reg)
nr += 1;
/* I don't really understand how this achieves the color interleave
* (ie RGBARGBA) in the result: [Do the saturation here]
*/
{
brw_push_insn_state(p);
for (channel = 0; channel < 4; channel++) {
/* mov (8) m2.0<1>:ud r28.0<8;8,1>:ud { Align1 } */
/* mov (8) m6.0<1>:ud r29.0<8;8,1>:ud { Align1 SecHalf } */
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p,
brw_message_reg(nr + channel),
arg0[channel]);
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_MOV(p,
brw_message_reg(nr + channel + 4),
sechalf(arg0[channel]));
}
/* skip over the regs populated above:
*/
nr += 8;
brw_pop_insn_state(p);
}
if (c->key.source_depth_to_render_target)
{
if (c->key.computes_depth)
brw_MOV(p, brw_message_reg(nr), arg2[2]);
else
brw_MOV(p, brw_message_reg(nr), arg1[1]); /* ? */
nr += 2;
}
if (c->key.dest_depth_reg)
{
GLuint comp = c->key.dest_depth_reg / 2;
GLuint off = c->key.dest_depth_reg % 2;
if (off != 0) {
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p, brw_message_reg(nr), offset(arg1[comp],1));
/* 2nd half? */
brw_MOV(p, brw_message_reg(nr+1), arg1[comp+1]);
brw_pop_insn_state(p);
}
else {
brw_MOV(p, brw_message_reg(nr), arg1[comp]);
}
nr += 2;
}
if (!c->key.runtime_check_aads_emit) {
if (c->key.aa_dest_stencil_reg)
emit_aa(c, arg1, 2);
fire_fb_write(c, 0, nr, target, eot);
}
else {
struct brw_reg v1_null_ud = vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD));
struct brw_reg ip = brw_ip_reg();
struct brw_instruction *jmp;
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_conditionalmod(p, BRW_CONDITIONAL_Z);
brw_AND(p,
v1_null_ud,
get_element_ud(brw_vec8_grf(1,0), 6),
brw_imm_ud(1<<26));
jmp = brw_JMPI(p, ip, ip, brw_imm_w(0));
{
emit_aa(c, arg1, 2);
fire_fb_write(c, 0, nr, target, eot);
/* note - thread killed in subroutine */
}
brw_land_fwd_jump(p, jmp);
/* ELSE: Shuffle up one register to fill in the hole left for AA:
*/
fire_fb_write(c, 1, nr-1, target, eot);
}
}
void emit_math2(struct brw_wm_compile *c,
GLuint function,
const struct brw_reg *dst,
GLuint mask,
const struct brw_reg *arg0,
const struct brw_reg *arg1)
{
struct brw_compile *p = &c->func;
struct intel_context *intel = &p->brw->intel;
int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
if (!(mask & WRITEMASK_XYZW))
return; /* Do not emit dead code */
assert(is_power_of_two(mask & WRITEMASK_XYZW));
brw_push_insn_state(p);
/* math can only operate on up to a vec8 at a time, so in
* dispatch_width==16 we have to do the second half manually.
*/
if (intel->gen >= 6) {
struct brw_reg src0 = arg0[0];
struct brw_reg src1 = arg1[0];
struct brw_reg temp_dst = dst[dst_chan];
if (arg0[0].hstride == BRW_HORIZONTAL_STRIDE_0) {
brw_MOV(p, temp_dst, src0);
src0 = temp_dst;
}
if (arg1[0].hstride == BRW_HORIZONTAL_STRIDE_0) {
/* This is a heinous hack to get a temporary register for use
* in case both arg0 and arg1 are constants. Why you're
* doing exponentiation on constant values in the shader, we
* don't know.
*
* max_wm_grf is almost surely less than the maximum GRF, and
* gen6 doesn't care about the number of GRFs used in a
* shader like pre-gen6 did.
*/
struct brw_reg temp = brw_vec8_grf(c->max_wm_grf, 0);
brw_MOV(p, temp, src1);
src1 = temp;
}
brw_set_saturate(p, (mask & SATURATE) ? 1 : 0);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_math2(p,
temp_dst,
function,
src0,
src1);
if (c->dispatch_width == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_math2(p,
sechalf(temp_dst),
function,
sechalf(src0),
sechalf(src1));
}
} else {
GLuint saturate = ((mask & SATURATE) ?
BRW_MATH_SATURATE_SATURATE :
BRW_MATH_SATURATE_NONE);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p, brw_message_reg(3), arg1[0]);
if (c->dispatch_width == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_MOV(p, brw_message_reg(5), sechalf(arg1[0]));
}
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_math(p,
dst[dst_chan],
function,
saturate,
2,
arg0[0],
BRW_MATH_DATA_VECTOR,
BRW_MATH_PRECISION_FULL);
/* Send two messages to perform all 16 operations:
*/
if (c->dispatch_width == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_math(p,
offset(dst[dst_chan],1),
function,
saturate,
4,
sechalf(arg0[0]),
BRW_MATH_DATA_VECTOR,
BRW_MATH_PRECISION_FULL);
}
}
brw_pop_insn_state(p);
}
/* Emit the fragment program instructions here.
*/
void brw_wm_emit( struct brw_wm_compile *c )
{
struct brw_compile *p = &c->func;
struct intel_context *intel = &p->brw->intel;
GLuint insn;
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
if (intel->gen >= 6)
brw_set_acc_write_control(p, 1);
/* Check if any of the payload regs need to be spilled:
*/
spill_values(c, c->payload.depth, 4);
spill_values(c, c->creg, c->nr_creg);
spill_values(c, c->payload.input_interp, FRAG_ATTRIB_MAX);
for (insn = 0; insn < c->nr_insns; insn++) {
struct brw_wm_instruction *inst = &c->instruction[insn];
struct brw_reg args[3][4], dst[4];
GLuint i, dst_flags;
/* Get argument regs:
*/
for (i = 0; i < 3; i++)
get_argument_regs(c, inst->src[i], args[i]);
/* Get dest regs:
*/
for (i = 0; i < 4; i++)
if (inst->dst[i])
dst[i] = inst->dst[i]->hw_reg;
else
dst[i] = brw_null_reg();
/* Flags
*/
dst_flags = inst->writemask;
if (inst->saturate)
dst_flags |= SATURATE;
switch (inst->opcode) {
/* Generated instructions for calculating triangle interpolants:
*/
case WM_PIXELXY:
emit_pixel_xy(c, dst, dst_flags);
break;
case WM_DELTAXY:
emit_delta_xy(p, dst, dst_flags, args[0]);
break;
case WM_WPOSXY:
emit_wpos_xy(c, dst, dst_flags, args[0]);
break;
case WM_PIXELW:
emit_pixel_w(c, dst, dst_flags, args[0], args[1]);
break;
case WM_LINTERP:
emit_linterp(p, dst, dst_flags, args[0], args[1]);
break;
case WM_PINTERP:
emit_pinterp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case WM_CINTERP:
emit_cinterp(p, dst, dst_flags, args[0]);
break;
case WM_FB_WRITE:
emit_fb_write(c, args[0], args[1], args[2], inst->target, inst->eot);
break;
case WM_FRONTFACING:
emit_frontfacing(p, dst, dst_flags);
break;
/* Straightforward arithmetic:
*/
case OPCODE_ADD:
emit_alu2(p, brw_ADD, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_FRC:
emit_alu1(p, brw_FRC, dst, dst_flags, args[0]);
break;
case OPCODE_FLR:
emit_alu1(p, brw_RNDD, dst, dst_flags, args[0]);
break;
case OPCODE_DDX:
emit_ddxy(p, dst, dst_flags, true, args[0]);
break;
case OPCODE_DDY:
emit_ddxy(p, dst, dst_flags, false, args[0]);
break;
case OPCODE_DP2:
emit_dp2(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DP3:
emit_dp3(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DP4:
emit_dp4(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DPH:
emit_dph(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_TRUNC:
for (i = 0; i < 4; i++) {
if (dst_flags & (1<<i)) {
brw_RNDZ(p, dst[i], args[0][i]);
}
}
break;
case OPCODE_LRP:
emit_lrp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MAD:
emit_mad(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MOV:
case OPCODE_SWZ:
emit_alu1(p, brw_MOV, dst, dst_flags, args[0]);
break;
case OPCODE_MUL:
emit_alu2(p, brw_MUL, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_XPD:
emit_xpd(p, dst, dst_flags, args[0], args[1]);
break;
/* Higher math functions:
*/
case OPCODE_RCP:
emit_math1(c, BRW_MATH_FUNCTION_INV, dst, dst_flags, args[0]);
break;
case OPCODE_RSQ:
emit_math1(c, BRW_MATH_FUNCTION_RSQ, dst, dst_flags, args[0]);
break;
case OPCODE_SIN:
emit_math1(c, BRW_MATH_FUNCTION_SIN, dst, dst_flags, args[0]);
break;
case OPCODE_COS:
emit_math1(c, BRW_MATH_FUNCTION_COS, dst, dst_flags, args[0]);
break;
case OPCODE_EX2:
emit_math1(c, BRW_MATH_FUNCTION_EXP, dst, dst_flags, args[0]);
break;
case OPCODE_LG2:
emit_math1(c, BRW_MATH_FUNCTION_LOG, dst, dst_flags, args[0]);
break;
case OPCODE_SCS:
/* There is an scs math function, but it would need some
* fixup for 16-element execution.
*/
if (dst_flags & WRITEMASK_X)
emit_math1(c, BRW_MATH_FUNCTION_COS, dst, (dst_flags&SATURATE)|WRITEMASK_X, args[0]);
if (dst_flags & WRITEMASK_Y)
emit_math1(c, BRW_MATH_FUNCTION_SIN, dst+1, (dst_flags&SATURATE)|WRITEMASK_X, args[0]);
break;
case OPCODE_POW:
emit_math2(c, BRW_MATH_FUNCTION_POW, dst, dst_flags, args[0], args[1]);
break;
/* Comparisons:
*/
case OPCODE_CMP:
emit_cmp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MAX:
emit_max(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_MIN:
emit_min(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SLT:
emit_slt(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SLE:
emit_sle(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SGT:
emit_sgt(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SGE:
emit_sge(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SEQ:
emit_seq(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SNE:
emit_sne(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SSG:
emit_sign(p, dst, dst_flags, args[0]);
break;
case OPCODE_LIT:
emit_lit(c, dst, dst_flags, args[0]);
break;
/* Texturing operations:
*/
case OPCODE_TEX:
emit_tex(c, dst, dst_flags, args[0], c->payload.depth[0].hw_reg,
inst->tex_idx, inst->tex_unit,
inst->tex_shadow);
break;
case OPCODE_TXB:
emit_txb(c, dst, dst_flags, args[0], c->payload.depth[0].hw_reg,
inst->tex_idx, inst->tex_unit);
break;
case OPCODE_KIL:
emit_kil(c, args[0]);
break;
default:
printf("Unsupported opcode %i (%s) in fragment shader\n",
inst->opcode, inst->opcode < MAX_OPCODE ?
_mesa_opcode_string(inst->opcode) :
"unknown");
}
for (i = 0; i < 4; i++)
if (inst->dst[i] && inst->dst[i]->spill_slot)
emit_spill(c,
inst->dst[i]->hw_reg,
inst->dst[i]->spill_slot);
}
/* Only properly tested on ILK */
if (p->brw->intel.gen == 5) {
brw_remove_duplicate_mrf_moves(p);
if (c->dispatch_width == 16)
brw_remove_grf_to_mrf_moves(p);
}
if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
int i;
printf("wm-native:\n");
for (i = 0; i < p->nr_insn; i++)
brw_disasm(stdout, &p->store[i], p->brw->intel.gen);
printf("\n");
}
}
/* Post-fragment-program processing. Send the results to the
* framebuffer.
* \param arg0 the fragment color
* \param arg1 the pass-through depth value
* \param arg2 the shader-computed depth value
*/
void emit_fb_write(struct brw_wm_compile *c,
struct brw_reg *arg0,
struct brw_reg *arg1,
struct brw_reg *arg2,
GLuint target,
GLuint eot)
{
struct brw_compile *p = &c->func;
struct brw_context *brw = p->brw;
struct intel_context *intel = &brw->intel;
GLuint nr = 2;
GLuint channel;
/* Reserve a space for AA - may not be needed:
*/
if (c->aa_dest_stencil_reg)
nr += 1;
/* I don't really understand how this achieves the color interleave
* (ie RGBARGBA) in the result: [Do the saturation here]
*/
brw_push_insn_state(p);
if (c->key.clamp_fragment_color)
brw_set_saturate(p, 1);
for (channel = 0; channel < 4; channel++) {
if (intel->gen >= 6) {
/* gen6 SIMD16 single source DP write looks like:
* m + 0: r0
* m + 1: r1
* m + 2: g0
* m + 3: g1
* m + 4: b0
* m + 5: b1
* m + 6: a0
* m + 7: a1
*/
if (c->dispatch_width == 16) {
brw_MOV(p, brw_message_reg(nr + channel * 2), arg0[channel]);
} else {
brw_MOV(p, brw_message_reg(nr + channel), arg0[channel]);
}
} else if (c->dispatch_width == 16 && brw->has_compr4) {
/* pre-gen6 SIMD16 single source DP write looks like:
* m + 0: r0
* m + 1: g0
* m + 2: b0
* m + 3: a0
* m + 4: r1
* m + 5: g1
* m + 6: b1
* m + 7: a1
*
* By setting the high bit of the MRF register number, we indicate
* that we want COMPR4 mode - instead of doing the usual destination
* + 1 for the second half we get destination + 4.
*/
brw_MOV(p,
brw_message_reg(nr + channel + BRW_MRF_COMPR4),
arg0[channel]);
} else {
/* mov (8) m2.0<1>:ud r28.0<8;8,1>:ud { Align1 } */
/* mov (8) m6.0<1>:ud r29.0<8;8,1>:ud { Align1 SecHalf } */
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p,
brw_message_reg(nr + channel),
arg0[channel]);
if (c->dispatch_width == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
brw_MOV(p,
brw_message_reg(nr + channel + 4),
sechalf(arg0[channel]));
}
}
}
brw_set_saturate(p, 0);
/* skip over the regs populated above:
*/
if (c->dispatch_width == 16)
nr += 8;
else
nr += 4;
brw_pop_insn_state(p);
if (c->source_depth_to_render_target)
{
if (c->computes_depth)
brw_MOV(p, brw_message_reg(nr), arg2[2]);
else
brw_MOV(p, brw_message_reg(nr), arg1[1]); /* ? */
nr += 2;
}
if (c->dest_depth_reg)
{
GLuint comp = c->dest_depth_reg / 2;
GLuint off = c->dest_depth_reg % 2;
if (off != 0) {
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p, brw_message_reg(nr), offset(arg1[comp],1));
/* 2nd half? */
brw_MOV(p, brw_message_reg(nr+1), arg1[comp+1]);
brw_pop_insn_state(p);
}
else {
brw_MOV(p, brw_message_reg(nr), arg1[comp]);
}
nr += 2;
}
if (intel->gen >= 6) {
/* Load the message header. There's no implied move from src0
* to the base mrf on gen6.
*/
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_MOV(p, retype(brw_message_reg(0), BRW_REGISTER_TYPE_UD),
retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
brw_pop_insn_state(p);
if (target != 0) {
brw_MOV(p, retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE,
0,
2), BRW_REGISTER_TYPE_UD),
brw_imm_ud(target));
}
}
if (!c->runtime_check_aads_emit) {
if (c->aa_dest_stencil_reg)
emit_aa(c, arg1, 2);
fire_fb_write(c, 0, nr, target, eot);
}
else {
struct brw_reg v1_null_ud = vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD));
struct brw_reg ip = brw_ip_reg();
struct brw_instruction *jmp;
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_conditionalmod(p, BRW_CONDITIONAL_Z);
brw_AND(p,
v1_null_ud,
get_element_ud(brw_vec8_grf(1,0), 6),
brw_imm_ud(1<<26));
jmp = brw_JMPI(p, ip, ip, brw_imm_w(0));
{
emit_aa(c, arg1, 2);
fire_fb_write(c, 0, nr, target, eot);
/* note - thread killed in subroutine */
}
brw_land_fwd_jump(p, jmp);
/* ELSE: Shuffle up one register to fill in the hole left for AA:
*/
fire_fb_write(c, 1, nr-1, target, eot);
}
}
/**
* Read float[4] constant(s) from VS constant buffer.
* For relative addressing, two float[4] constants will be read into 'dest'.
* Otherwise, one float[4] constant will be read into the lower half of 'dest'.
*/
void brw_dp_READ_4_vs(struct brw_compile *p,
struct brw_reg dest,
GLuint oword,
GLboolean relAddr,
struct brw_reg addrReg,
GLuint location,
GLuint bind_table_index)
{
GLuint msg_reg_nr = 1;
assert(oword < 2);
/*
printf("vs const read msg, location %u, msg_reg_nr %d\n",
location, msg_reg_nr);
*/
/* Setup MRF[1] with location/offset into const buffer */
{
struct brw_reg b;
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_set_predicate_control(p, BRW_PREDICATE_NONE);
/*brw_set_access_mode(p, BRW_ALIGN_16);*/
/* XXX I think we're setting all the dwords of MRF[1] to 'location'.
* when the docs say only dword[2] should be set. Hmmm. But it works.
*/
b = brw_message_reg(msg_reg_nr);
b = retype(b, BRW_REGISTER_TYPE_UD);
/*b = get_element_ud(b, 2);*/
if (relAddr) {
brw_ADD(p, b, addrReg, brw_imm_ud(location));
}
else {
brw_MOV(p, b, brw_imm_ud(location));
}
brw_pop_insn_state(p);
}
{
struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
insn->header.predicate_control = BRW_PREDICATE_NONE;
insn->header.compression_control = BRW_COMPRESSION_NONE;
insn->header.destreg__conditionalmod = msg_reg_nr;
insn->header.mask_control = BRW_MASK_DISABLE;
/*insn->header.access_mode = BRW_ALIGN_16;*/
brw_set_dest(insn, dest);
brw_set_src0(insn, brw_null_reg());
brw_set_dp_read_message(p->brw,
insn,
bind_table_index,
oword, /* 0 = lower Oword, 1 = upper Oword */
BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ, /* msg_type */
0, /* source cache = data cache */
1, /* msg_length */
1, /* response_length (1 Oword) */
0); /* eot */
}
}
const GLuint *
brw_blorp_const_color_program::compile(struct brw_context *brw,
GLuint *program_size)
{
/* Set up prog_data */
memset(&prog_data, 0, sizeof(prog_data));
prog_data.persample_msaa_dispatch = false;
alloc_regs();
brw_set_compression_control(&func, BRW_COMPRESSION_NONE);
struct brw_reg mrf_rt_write =
retype(vec16(brw_message_reg(base_mrf)), BRW_REGISTER_TYPE_F);
uint32_t mlen, msg_type;
if (key->use_simd16_replicated_data) {
/* The message payload is a single register with the low 4 floats/ints
* filled with the constant clear color.
*/
brw_set_mask_control(&func, BRW_MASK_DISABLE);
brw_MOV(&func, vec4(brw_message_reg(base_mrf)), clear_rgba);
brw_set_mask_control(&func, BRW_MASK_ENABLE);
msg_type = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED;
mlen = 1;
} else {
for (int i = 0; i < 4; i++) {
/* The message payload is pairs of registers for 16 pixels each of r,
* g, b, and a.
*/
brw_set_compression_control(&func, BRW_COMPRESSION_COMPRESSED);
brw_MOV(&func,
brw_message_reg(base_mrf + i * 2),
brw_vec1_grf(clear_rgba.nr, i));
brw_set_compression_control(&func, BRW_COMPRESSION_NONE);
}
msg_type = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE;
mlen = 8;
}
/* Now write to the render target and terminate the thread */
brw_fb_WRITE(&func,
16 /* dispatch_width */,
base_mrf /* msg_reg_nr */,
mrf_rt_write /* src0 */,
msg_type,
BRW_BLORP_RENDERBUFFER_BINDING_TABLE_INDEX,
mlen,
0 /* response_length */,
true /* eot */,
false /* header present */);
if (unlikely(INTEL_DEBUG & DEBUG_BLORP)) {
fprintf(stderr, "Native code for BLORP clear:\n");
brw_dump_compile(&func, stderr, 0, func.next_insn_offset);
fprintf(stderr, "\n");
}
return brw_get_program(&func, program_size);
}
/**
* Texture sample instruction.
* Note: the msg_type plus msg_length values determine exactly what kind
* of sampling operation is performed. See volume 4, page 161 of docs.
*/
void brw_SAMPLE(struct brw_compile *p,
struct brw_reg dest,
GLuint msg_reg_nr,
struct brw_reg src0,
GLuint binding_table_index,
GLuint sampler,
GLuint writemask,
GLuint msg_type,
GLuint response_length,
GLuint msg_length,
GLboolean eot,
GLuint header_present,
GLuint simd_mode)
{
GLboolean need_stall = 0;
if (writemask == 0) {
/*_mesa_printf("%s: zero writemask??\n", __FUNCTION__); */
return;
}
/* Hardware doesn't do destination dependency checking on send
* instructions properly. Add a workaround which generates the
* dependency by other means. In practice it seems like this bug
* only crops up for texture samples, and only where registers are
* written by the send and then written again later without being
* read in between. Luckily for us, we already track that
* information and use it to modify the writemask for the
* instruction, so that is a guide for whether a workaround is
* needed.
*/
if (writemask != WRITEMASK_XYZW) {
GLuint dst_offset = 0;
GLuint i, newmask = 0, len = 0;
for (i = 0; i < 4; i++) {
if (writemask & (1<<i))
break;
dst_offset += 2;
}
for (; i < 4; i++) {
if (!(writemask & (1<<i)))
break;
newmask |= 1<<i;
len++;
}
if (newmask != writemask) {
need_stall = 1;
/* _mesa_printf("need stall %x %x\n", newmask , writemask); */
}
else {
struct brw_reg m1 = brw_message_reg(msg_reg_nr);
newmask = ~newmask & WRITEMASK_XYZW;
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_MOV(p, m1, brw_vec8_grf(0,0));
brw_MOV(p, get_element_ud(m1, 2), brw_imm_ud(newmask << 12));
brw_pop_insn_state(p);
src0 = retype(brw_null_reg(), BRW_REGISTER_TYPE_UW);
dest = offset(dest, dst_offset);
response_length = len * 2;
}
}
{
struct brw_instruction *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;
brw_set_dest(insn, dest);
brw_set_src0(insn, src0);
brw_set_sampler_message(p->brw, insn,
binding_table_index,
sampler,
msg_type,
response_length,
msg_length,
eot,
header_present,
simd_mode);
}
if (need_stall) {
struct brw_reg reg = vec8(offset(dest, response_length-1));
/* mov (8) r9.0<1>:f r9.0<8;8,1>:f { Align1 }
*/
brw_push_insn_state(p);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_MOV(p, reg, reg);
brw_pop_insn_state(p);
}
}
static void brw_wm_projective_st(struct brw_compile *p, int dw,
int channel, int msg)
{
int uv;
if (dw == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
uv = p->gen >= 060 ? 6 : 3;
} else {
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
uv = p->gen >= 060 ? 4 : 3;
}
uv += 2*channel;
msg++;
if (p->gen >= 060) {
/* First compute 1/z */
brw_PLN(p,
brw_vec8_grf(30, 0),
brw_vec1_grf(uv+1, 0),
brw_vec8_grf(2, 0));
if (dw == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_math_invert(p, brw_vec8_grf(30, 0), brw_vec8_grf(30, 0));
brw_math_invert(p, brw_vec8_grf(31, 0), brw_vec8_grf(31, 0));
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
} else
brw_math_invert(p, brw_vec8_grf(30, 0), brw_vec8_grf(30, 0));
brw_PLN(p,
brw_vec8_grf(26, 0),
brw_vec1_grf(uv, 0),
brw_vec8_grf(2, 0));
brw_PLN(p,
brw_vec8_grf(28, 0),
brw_vec1_grf(uv, 4),
brw_vec8_grf(2, 0));
brw_MUL(p,
brw_message_reg(msg),
brw_vec8_grf(26, 0),
brw_vec8_grf(30, 0));
brw_MUL(p,
brw_message_reg(msg + dw/8),
brw_vec8_grf(28, 0),
brw_vec8_grf(30, 0));
} else {
struct brw_reg r = brw_vec1_grf(uv, 0);
/* First compute 1/z */
brw_LINE(p, brw_null_reg(), brw_vec1_grf(uv+1, 0), brw_vec8_grf(X16, 0));
brw_MAC(p, brw_vec8_grf(30, 0), brw_vec1_grf(uv+1, 1), brw_vec8_grf(Y16, 0));
if (dw == 16) {
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
brw_math_invert(p, brw_vec8_grf(30, 0), brw_vec8_grf(30, 0));
brw_math_invert(p, brw_vec8_grf(31, 0), brw_vec8_grf(31, 0));
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
} else
brw_math_invert(p, brw_vec8_grf(30, 0), brw_vec8_grf(30, 0));
/* Now compute the output s,t values */
brw_LINE(p, brw_null_reg(), __suboffset(r, 0), brw_vec8_grf(X16, 0));
brw_MAC(p, brw_vec8_grf(28, 0), __suboffset(r, 1), brw_vec8_grf(Y16, 0));
brw_MUL(p, brw_message_reg(msg), brw_vec8_grf(28, 0), brw_vec8_grf(30, 0));
msg += dw/8;
brw_LINE(p, brw_null_reg(), __suboffset(r, 4), brw_vec8_grf(X16, 0));
brw_MAC(p, brw_vec8_grf(28, 0), __suboffset(r, 5), brw_vec8_grf(Y16, 0));
brw_MUL(p, brw_message_reg(msg), brw_vec8_grf(28, 0), brw_vec8_grf(30, 0));
}
}
/* Emit the fragment program instructions here.
*/
void brw_wm_emit( struct brw_wm_compile *c )
{
struct brw_compile *p = &c->func;
GLuint insn;
brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
/* Check if any of the payload regs need to be spilled:
*/
spill_values(c, c->payload.depth, 4);
spill_values(c, c->creg, c->nr_creg);
spill_values(c, c->payload.input_interp, FRAG_ATTRIB_MAX);
for (insn = 0; insn < c->nr_insns; insn++) {
struct brw_wm_instruction *inst = &c->instruction[insn];
struct brw_reg args[3][4], dst[4];
GLuint i, dst_flags;
/* Get argument regs:
*/
for (i = 0; i < 3; i++)
get_argument_regs(c, inst->src[i], args[i]);
/* Get dest regs:
*/
for (i = 0; i < 4; i++)
if (inst->dst[i])
dst[i] = inst->dst[i]->hw_reg;
else
dst[i] = brw_null_reg();
/* Flags
*/
dst_flags = inst->writemask;
if (inst->saturate)
dst_flags |= SATURATE;
switch (inst->opcode) {
/* Generated instructions for calculating triangle interpolants:
*/
case WM_PIXELXY:
emit_pixel_xy(p, dst, dst_flags, args[0]);
break;
case WM_DELTAXY:
emit_delta_xy(p, dst, dst_flags, args[0], args[1]);
break;
case WM_WPOSXY:
emit_wpos_xy(c, dst, dst_flags, args[0]);
break;
case WM_PIXELW:
emit_pixel_w(p, dst, dst_flags, args[0], args[1]);
break;
case WM_LINTERP:
emit_linterp(p, dst, dst_flags, args[0], args[1]);
break;
case WM_PINTERP:
emit_pinterp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case WM_CINTERP:
emit_cinterp(p, dst, dst_flags, args[0]);
break;
case WM_FB_WRITE:
emit_fb_write(c, args[0], args[1], args[2], inst->target, inst->eot);
break;
/* Straightforward arithmetic:
*/
case OPCODE_ADD:
emit_alu2(p, brw_ADD, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_FRC:
emit_alu1(p, brw_FRC, dst, dst_flags, args[0]);
break;
case OPCODE_FLR:
emit_alu1(p, brw_RNDD, dst, dst_flags, args[0]);
break;
case OPCODE_DP3: /* */
emit_dp3(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DP4:
emit_dp4(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_DPH:
emit_dph(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_LRP: /* */
emit_lrp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MAD:
emit_mad(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MOV:
case OPCODE_SWZ:
emit_alu1(p, brw_MOV, dst, dst_flags, args[0]);
break;
case OPCODE_MUL:
emit_alu2(p, brw_MUL, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_XPD:
emit_xpd(p, dst, dst_flags, args[0], args[1]);
break;
/* Higher math functions:
*/
case OPCODE_RCP:
emit_math1(p, BRW_MATH_FUNCTION_INV, dst, dst_flags, args[0]);
break;
case OPCODE_RSQ:
emit_math1(p, BRW_MATH_FUNCTION_RSQ, dst, dst_flags, args[0]);
break;
case OPCODE_SIN:
emit_math1(p, BRW_MATH_FUNCTION_SIN, dst, dst_flags, args[0]);
break;
case OPCODE_COS:
emit_math1(p, BRW_MATH_FUNCTION_COS, dst, dst_flags, args[0]);
break;
case OPCODE_EX2:
emit_math1(p, BRW_MATH_FUNCTION_EXP, dst, dst_flags, args[0]);
break;
case OPCODE_LG2:
emit_math1(p, BRW_MATH_FUNCTION_LOG, dst, dst_flags, args[0]);
break;
case OPCODE_SCS:
/* There is an scs math function, but it would need some
* fixup for 16-element execution.
*/
if (dst_flags & WRITEMASK_X)
emit_math1(p, BRW_MATH_FUNCTION_COS, dst, (dst_flags&SATURATE)|WRITEMASK_X, args[0]);
if (dst_flags & WRITEMASK_Y)
emit_math1(p, BRW_MATH_FUNCTION_SIN, dst+1, (dst_flags&SATURATE)|WRITEMASK_X, args[0]);
break;
case OPCODE_POW:
emit_math2(p, BRW_MATH_FUNCTION_POW, dst, dst_flags, args[0], args[1]);
break;
/* Comparisons:
*/
case OPCODE_CMP:
emit_cmp(p, dst, dst_flags, args[0], args[1], args[2]);
break;
case OPCODE_MAX:
emit_max(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_MIN:
emit_min(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SLT:
emit_slt(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SLE:
emit_sle(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SGT:
emit_sgt(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SGE:
emit_sge(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SEQ:
emit_seq(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_SNE:
emit_sne(p, dst, dst_flags, args[0], args[1]);
break;
case OPCODE_LIT:
emit_lit(p, dst, dst_flags, args[0]);
break;
/* Texturing operations:
*/
case OPCODE_TEX:
emit_tex(c, inst, dst, dst_flags, args[0]);
break;
case OPCODE_TXB:
emit_txb(c, inst, dst, dst_flags, args[0]);
break;
case OPCODE_KIL:
emit_kil(c, args[0]);
break;
default:
_mesa_printf("unsupport opcode %d in fragment program\n",
inst->opcode);
}
for (i = 0; i < 4; i++)
if (inst->dst[i] && inst->dst[i]->spill_slot)
emit_spill(c,
inst->dst[i]->hw_reg,
inst->dst[i]->spill_slot);
}
}