}
} else if (ir->location == FRAG_RESULT_DEPTH) {
this->frag_depth = *reg;
+ } else if (ir->location == FRAG_RESULT_SAMPLE_MASK) {
+ this->sample_mask = *reg;
} else {
/* gl_FragData or a user-defined FS output */
assert(ir->location >= FRAG_RESULT_DATA0 &&
reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
reg->type = brw_type_for_base_type(ir->type);
+
+ } else if (ir->mode == ir_var_system_value) {
+ if (ir->location == SYSTEM_VALUE_SAMPLE_POS) {
+ reg = emit_samplepos_setup(ir);
+ } else if (ir->location == SYSTEM_VALUE_SAMPLE_ID) {
+ reg = emit_sampleid_setup(ir);
+ }
}
if (!reg)
if (brw->gen >= 7 && dispatch_width == 16)
fail("16-wide explicit accumulator operands unsupported\n");
- struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_D);
+ struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
emit(MUL(acc, op[0], op[1]));
emit(MACH(reg_null_d, op[0], op[1]));
emit(MUL(this->result, op[0], op[1]));
}
break;
+ case ir_binop_imul_high: {
+ if (brw->gen >= 7 && dispatch_width == 16)
+ fail("16-wide explicit accumulator operands unsupported\n");
+
+ struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
+
+ emit(MUL(acc, op[0], op[1]));
+ emit(MACH(this->result, op[0], op[1]));
+ break;
+ }
case ir_binop_div:
/* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
assert(ir->type->is_integer());
emit_math(SHADER_OPCODE_INT_QUOTIENT, this->result, op[0], op[1]);
break;
+ case ir_binop_carry: {
+ if (brw->gen >= 7 && dispatch_width == 16)
+ fail("16-wide explicit accumulator operands unsupported\n");
+
+ struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD);
+
+ emit(ADDC(reg_null_ud, op[0], op[1]));
+ emit(MOV(this->result, fs_reg(acc)));
+ break;
+ }
+ case ir_binop_borrow: {
+ if (brw->gen >= 7 && dispatch_width == 16)
+ fail("16-wide explicit accumulator operands unsupported\n");
+
+ struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD);
+
+ emit(SUBB(reg_null_ud, op[0], op[1]));
+ emit(MOV(this->result, fs_reg(acc)));
+ break;
+ }
case ir_binop_mod:
/* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
assert(ir->type->is_integer());
*/
ir_constant *uniform_block = ir->operands[0]->as_constant();
ir_constant *const_offset = ir->operands[1]->as_constant();
- fs_reg surf_index = fs_reg((unsigned)SURF_INDEX_WM_UBO(uniform_block->value.u[0]));
+ fs_reg surf_index = fs_reg(c->prog_data.base.binding_table.ubo_start +
+ uniform_block->value.u[0]);
if (const_offset) {
fs_reg packed_consts = fs_reg(this, glsl_type::float_type);
packed_consts.type = result.type;
break;
case GLSL_TYPE_SAMPLER:
+ case GLSL_TYPE_ATOMIC_UINT:
break;
case GLSL_TYPE_VOID:
const int vector_elements =
ir->coordinate ? ir->coordinate->type->vector_elements : 0;
- if (ir->offset != NULL && ir->op == ir_txf) {
- /* It appears that the ld instruction used for txf does its
- * address bounds check before adding in the offset. To work
- * around this, just add the integer offset to the integer texel
- * coordinate, and don't put the offset in the header.
+ if (ir->offset) {
+ /* The offsets set up by the ir_texture visitor are in the
+ * m1 header, so we can't go headerless.
*/
- ir_constant *offset = ir->offset->as_constant();
- for (int i = 0; i < vector_elements; i++) {
- emit(ADD(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
- coordinate,
- offset->value.i[i]));
- coordinate.reg_offset++;
- }
- } else {
- if (ir->offset) {
- /* The offsets set up by the ir_texture visitor are in the
- * m1 header, so we can't go headerless.
- */
- header_present = true;
- mlen++;
- base_mrf--;
- }
+ header_present = true;
+ mlen++;
+ base_mrf--;
+ }
- for (int i = 0; i < vector_elements; i++) {
- emit(MOV(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
- coordinate));
- coordinate.reg_offset++;
- }
+ for (int i = 0; i < vector_elements; i++) {
+ emit(MOV(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
+ coordinate));
+ coordinate.reg_offset++;
}
mlen += vector_elements * reg_width;
mlen += reg_width;
inst = emit(SHADER_OPCODE_TXS, dst);
break;
+ case ir_query_levels:
+ emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), fs_reg(0u)));
+ mlen += reg_width;
+ inst = emit(SHADER_OPCODE_TXS, dst);
+ break;
case ir_txf:
mlen = header_present + 4 * reg_width;
emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD), lod));
fs_reg shadow_c, fs_reg lod, fs_reg lod2,
fs_reg sample_index)
{
- int mlen = 0;
- int base_mrf = 2;
int reg_width = dispatch_width / 8;
bool header_present = false;
- int offsets[3];
+
+ fs_reg payload = fs_reg(this, glsl_type::float_type);
+ fs_reg next = payload;
if (ir->op == ir_tg4 || (ir->offset && ir->op != ir_txf)) {
- /* * The offsets set up by the ir_texture visitor are in the
- * m1 header, so we can't go headerless.
+ /* For general texture offsets (no txf workaround), we need a header to
+ * put them in. Note that for 16-wide we're making space for two actual
+ * hardware registers here, so the emit will have to fix up for this.
*
* * ir4_tg4 needs to place its channel select in the header,
* for interaction with ARB_texture_swizzle
*/
header_present = true;
- mlen++;
- base_mrf--;
+ next.reg_offset++;
}
if (ir->shadow_comparitor) {
- emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
- mlen += reg_width;
+ emit(MOV(next, shadow_c));
+ next.reg_offset++;
}
+ bool has_nonconstant_offset = ir->offset && !ir->offset->as_constant();
+ bool coordinate_done = false;
+
/* Set up the LOD info */
switch (ir->op) {
case ir_tex:
case ir_lod:
- case ir_tg4:
break;
case ir_txb:
- emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
- mlen += reg_width;
+ emit(MOV(next, lod));
+ next.reg_offset++;
break;
case ir_txl:
- emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
- mlen += reg_width;
+ emit(MOV(next, lod));
+ next.reg_offset++;
break;
case ir_txd: {
if (dispatch_width == 16)
* [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
*/
for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
- emit(MOV(fs_reg(MRF, base_mrf + mlen), coordinate));
+ emit(MOV(next, coordinate));
coordinate.reg_offset++;
- mlen += reg_width;
+ next.reg_offset++;
/* For cube map array, the coordinate is (u,v,r,ai) but there are
* only derivatives for (u, v, r).
*/
if (i < ir->lod_info.grad.dPdx->type->vector_elements) {
- emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
+ emit(MOV(next, lod));
lod.reg_offset++;
- mlen += reg_width;
+ next.reg_offset++;
- emit(MOV(fs_reg(MRF, base_mrf + mlen), lod2));
+ emit(MOV(next, lod2));
lod2.reg_offset++;
- mlen += reg_width;
+ next.reg_offset++;
}
}
+
+ coordinate_done = true;
break;
}
case ir_txs:
- emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
- mlen += reg_width;
+ emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), lod));
+ next.reg_offset++;
+ break;
+ case ir_query_levels:
+ emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), fs_reg(0u)));
+ next.reg_offset++;
break;
case ir_txf:
- /* It appears that the ld instruction used for txf does its
- * address bounds check before adding in the offset. To work
- * around this, just add the integer offset to the integer texel
- * coordinate, and don't put the offset in the header.
- */
- if (ir->offset) {
- ir_constant *offset = ir->offset->as_constant();
- offsets[0] = offset->value.i[0];
- offsets[1] = offset->value.i[1];
- offsets[2] = offset->value.i[2];
- } else {
- memset(offsets, 0, sizeof(offsets));
- }
-
/* Unfortunately, the parameters for LD are intermixed: u, lod, v, r. */
- emit(ADD(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D),
- coordinate, offsets[0]));
+ emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
coordinate.reg_offset++;
- mlen += reg_width;
+ next.reg_offset++;
- emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D), lod));
- mlen += reg_width;
+ emit(MOV(next.retype(BRW_REGISTER_TYPE_D), lod));
+ next.reg_offset++;
for (int i = 1; i < ir->coordinate->type->vector_elements; i++) {
- emit(ADD(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D),
- coordinate, offsets[i]));
+ emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
coordinate.reg_offset++;
- mlen += reg_width;
+ next.reg_offset++;
}
+
+ coordinate_done = true;
break;
case ir_txf_ms:
- emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), sample_index));
- mlen += reg_width;
+ emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), sample_index));
+ next.reg_offset++;
/* constant zero MCS; we arrange to never actually have a compressed
* multisample surface here for now. TODO: issue ld_mcs to get this first,
* if we ever support texturing from compressed multisample surfaces
*/
- emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), fs_reg(0u)));
- mlen += reg_width;
+ emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), fs_reg(0u)));
+ next.reg_offset++;
/* there is no offsetting for this message; just copy in the integer
* texture coordinates
*/
for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
- emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D),
- coordinate));
+ emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
coordinate.reg_offset++;
- mlen += reg_width;
+ next.reg_offset++;
+ }
+
+ coordinate_done = true;
+ break;
+ case ir_tg4:
+ if (has_nonconstant_offset) {
+ if (ir->shadow_comparitor && dispatch_width == 16)
+ fail("Gen7 does not support gather4_po_c in SIMD16 mode.");
+
+ /* More crazy intermixing */
+ ir->offset->accept(this);
+ fs_reg offset_value = this->result;
+
+ for (int i = 0; i < 2; i++) { /* u, v */
+ emit(MOV(next, coordinate));
+ coordinate.reg_offset++;
+ next.reg_offset++;
+ }
+
+ for (int i = 0; i < 2; i++) { /* offu, offv */
+ emit(MOV(next.retype(BRW_REGISTER_TYPE_D), offset_value));
+ offset_value.reg_offset++;
+ next.reg_offset++;
+ }
+
+ if (ir->coordinate->type->vector_elements == 3) { /* r if present */
+ emit(MOV(next, coordinate));
+ coordinate.reg_offset++;
+ next.reg_offset++;
+ }
+
+ coordinate_done = true;
}
break;
}
/* Set up the coordinate (except for cases where it was done above) */
- if (ir->op != ir_txd && ir->op != ir_txs && ir->op != ir_txf && ir->op != ir_txf_ms) {
+ if (ir->coordinate && !coordinate_done) {
for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
- emit(MOV(fs_reg(MRF, base_mrf + mlen), coordinate));
- coordinate.reg_offset++;
- mlen += reg_width;
+ emit(MOV(next, coordinate));
+ coordinate.reg_offset++;
+ next.reg_offset++;
}
}
/* Generate the SEND */
fs_inst *inst = NULL;
switch (ir->op) {
- case ir_tex: inst = emit(SHADER_OPCODE_TEX, dst); break;
- case ir_txb: inst = emit(FS_OPCODE_TXB, dst); break;
- case ir_txl: inst = emit(SHADER_OPCODE_TXL, dst); break;
- case ir_txd: inst = emit(SHADER_OPCODE_TXD, dst); break;
- case ir_txf: inst = emit(SHADER_OPCODE_TXF, dst); break;
- case ir_txf_ms: inst = emit(SHADER_OPCODE_TXF_MS, dst); break;
- case ir_txs: inst = emit(SHADER_OPCODE_TXS, dst); break;
- case ir_lod: inst = emit(SHADER_OPCODE_LOD, dst); break;
- case ir_tg4: inst = emit(SHADER_OPCODE_TG4, dst); break;
+ case ir_tex: inst = emit(SHADER_OPCODE_TEX, dst, payload); break;
+ case ir_txb: inst = emit(FS_OPCODE_TXB, dst, payload); break;
+ case ir_txl: inst = emit(SHADER_OPCODE_TXL, dst, payload); break;
+ case ir_txd: inst = emit(SHADER_OPCODE_TXD, dst, payload); break;
+ case ir_txf: inst = emit(SHADER_OPCODE_TXF, dst, payload); break;
+ case ir_txf_ms: inst = emit(SHADER_OPCODE_TXF_MS, dst, payload); break;
+ case ir_txs: inst = emit(SHADER_OPCODE_TXS, dst, payload); break;
+ case ir_query_levels: inst = emit(SHADER_OPCODE_TXS, dst, payload); break;
+ case ir_lod: inst = emit(SHADER_OPCODE_LOD, dst, payload); break;
+ case ir_tg4:
+ if (has_nonconstant_offset)
+ inst = emit(SHADER_OPCODE_TG4_OFFSET, dst, payload);
+ else
+ inst = emit(SHADER_OPCODE_TG4, dst, payload);
+ break;
}
- inst->base_mrf = base_mrf;
- inst->mlen = mlen;
+ inst->base_mrf = -1;
+ if (reg_width == 2)
+ inst->mlen = next.reg_offset * reg_width - header_present;
+ else
+ inst->mlen = next.reg_offset * reg_width;
inst->header_present = header_present;
inst->regs_written = 4;
- if (mlen > 11) {
+ virtual_grf_sizes[payload.reg] = next.reg_offset;
+ if (inst->mlen > 11) {
fail("Message length >11 disallowed by hardware\n");
}
(brw->gen < 6 ||
(brw->gen >= 6 && (c->key.tex.gl_clamp_mask[0] & (1 << sampler) ||
c->key.tex.gl_clamp_mask[1] & (1 << sampler))))) {
- struct gl_program_parameter_list *params = fp->Base.Parameters;
+ struct gl_program_parameter_list *params = prog->Parameters;
int tokens[STATE_LENGTH] = {
STATE_INTERNAL,
STATE_TEXRECT_SCALE,
GLuint index = _mesa_add_state_reference(params,
(gl_state_index *)tokens);
c->prog_data.param[c->prog_data.nr_params++] =
- &fp->Base.Parameters->ParameterValues[index][0].f;
+ &prog->Parameters->ParameterValues[index][0].f;
c->prog_data.param[c->prog_data.nr_params++] =
- &fp->Base.Parameters->ParameterValues[index][1].f;
+ &prog->Parameters->ParameterValues[index][1].f;
}
/* The 965 requires the EU to do the normalization of GL rectangle
fs_inst *inst = NULL;
int sampler =
- _mesa_get_sampler_uniform_value(ir->sampler, shader_prog, &fp->Base);
+ _mesa_get_sampler_uniform_value(ir->sampler, shader_prog, prog);
/* FINISHME: We're failing to recompile our programs when the sampler is
* updated. This only matters for the texture rectangle scale parameters
* (pre-gen6, or gen6+ with GL_CLAMP).
*/
- int texunit = fp->Base.SamplerUnits[sampler];
+ int texunit = prog->SamplerUnits[sampler];
if (ir->op == ir_tg4) {
/* When tg4 is used with the degenerate ZERO/ONE swizzles, don't bother
* emitting anything other than setting up the constant result.
*/
- int swiz = GET_SWZ(c->key.tex.swizzles[sampler], 0);
+ ir_constant *chan = ir->lod_info.component->as_constant();
+ int swiz = GET_SWZ(c->key.tex.swizzles[sampler], chan->value.i[0]);
if (swiz == SWIZZLE_ZERO || swiz == SWIZZLE_ONE) {
fs_reg res = fs_reg(this, glsl_type::vec4_type);
/* Should be lowered by do_lower_texture_projection */
assert(!ir->projector);
+ /* Should be lowered */
+ assert(!ir->offset || !ir->offset->type->is_array());
+
/* Generate code to compute all the subexpression trees. This has to be
* done before loading any values into MRFs for the sampler message since
* generating these values may involve SEND messages that need the MRFs.
case ir_tex:
case ir_lod:
case ir_tg4:
+ case ir_query_levels:
break;
case ir_txb:
ir->lod_info.bias->accept(this);
lod, lod2);
}
- /* The header is set up by generate_tex() when necessary. */
- inst->src[0] = reg_undef;
-
if (ir->offset != NULL && ir->op != ir_txf)
- inst->texture_offset = brw_texture_offset(ir->offset->as_constant());
+ inst->texture_offset = brw_texture_offset(ctx, ir->offset->as_constant());
if (ir->op == ir_tg4)
inst->texture_offset |= gather_channel(ir, sampler) << 16; // M0.2:16-17
uint32_t
fs_visitor::gather_channel(ir_texture *ir, int sampler)
{
- int swiz = GET_SWZ(c->key.tex.swizzles[sampler], 0 /* red */);
- if (c->key.tex.gather_channel_quirk_mask & (1<<sampler))
- return 2; /* gather4 sampler is broken for green channel on RG32F --
- * we must ask for blue instead.
- */
+ ir_constant *chan = ir->lod_info.component->as_constant();
+ int swiz = GET_SWZ(c->key.tex.swizzles[sampler], chan->value.i[0]);
switch (swiz) {
case SWIZZLE_X: return 0;
- case SWIZZLE_Y: return 1;
+ case SWIZZLE_Y:
+ /* gather4 sampler is broken for green channel on RG32F --
+ * we must ask for blue instead.
+ */
+ if (c->key.tex.gather_channel_quirk_mask & (1<<sampler))
+ return 2;
+ return 1;
case SWIZZLE_Z: return 2;
case SWIZZLE_W: return 3;
default:
void
fs_visitor::swizzle_result(ir_texture *ir, fs_reg orig_val, int sampler)
{
+ if (ir->op == ir_query_levels) {
+ /* # levels is in .w */
+ orig_val.reg_offset += 3;
+ this->result = orig_val;
+ return;
+ }
+
this->result = orig_val;
/* txs,lod don't actually sample the texture, so swizzling the result
pop_force_uncompressed();
}
+ c->prog_data.uses_omask =
+ fp->Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK);
+ if(c->prog_data.uses_omask) {
+ this->current_annotation = "FB write oMask";
+ assert(this->sample_mask.file != BAD_FILE);
+ /* Hand over gl_SampleMask. Only lower 16 bits are relevant. */
+ emit(FS_OPCODE_SET_OMASK, fs_reg(MRF, nr, BRW_REGISTER_TYPE_UW), this->sample_mask);
+ nr += 1;
+ }
+
/* Reserve space for color. It'll be filled in per MRT below. */
int color_mrf = nr;
nr += 4 * reg_width;
fail("Missing support for simd16 depth writes on gen6\n");
}
- if (fp->Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
+ if (prog->OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
/* Hand over gl_FragDepth. */
assert(this->frag_depth.file != BAD_FILE);
emit(MOV(fs_reg(MRF, nr), this->frag_depth));
this->c = c;
this->brw = brw;
this->fp = fp;
+ this->prog = &fp->Base;
this->shader_prog = shader_prog;
+ this->prog = &fp->Base;
+ this->stage_prog_data = &c->prog_data.base;
this->ctx = &brw->ctx;
this->mem_ctx = ralloc_context(NULL);
if (shader_prog)
this->virtual_grf_array_size = 0;
this->virtual_grf_start = NULL;
this->virtual_grf_end = NULL;
- this->live_intervals_valid = false;
+ this->live_intervals = NULL;
this->params_remap = NULL;
this->nr_params_remap = 0;
this->force_uncompressed_stack = 0;
this->force_sechalf_stack = 0;
+ this->spilled_any_registers = false;
+
memset(&this->param_size, 0, sizeof(this->param_size));
}