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);
}
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);
}
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);
}
/* 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);
}
}
/* 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);
}
}
