void
vec4_generator::generate_math2_gen7(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg src0,
struct brw_reg src1)
{
brw_math2(p,
dst,
brw_math_function(inst->opcode),
src0, src1);
}
void
vec4_generator::generate_math2_gen6(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg src0,
struct brw_reg src1)
{
/* Can't do writemask because math can't be align16. */
assert(dst.dw1.bits.writemask == WRITEMASK_XYZW);
/* Source swizzles are ignored. */
check_gen6_math_src_arg(src0);
check_gen6_math_src_arg(src1);
brw_set_access_mode(p, BRW_ALIGN_1);
brw_math2(p,
dst,
brw_math_function(inst->opcode),
src0, src1);
brw_set_access_mode(p, BRW_ALIGN_16);
}
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);
}
