#include "brw_vec4.h"
#include "brw_cfg.h"
#include "glsl/ir_uniform.h"
-extern "C" {
#include "program/sampler.h"
-}
namespace brw {
-vec4_instruction::vec4_instruction(vec4_visitor *v,
- enum opcode opcode, const dst_reg &dst,
+vec4_instruction::vec4_instruction(enum opcode opcode, const dst_reg &dst,
const src_reg &src0, const src_reg &src1,
const src_reg &src2)
{
this->no_dd_check = false;
this->writes_accumulator = false;
this->conditional_mod = BRW_CONDITIONAL_NONE;
- this->texture_offset = 0;
+ this->predicate = BRW_PREDICATE_NONE;
+ this->predicate_inverse = false;
this->target = 0;
+ this->regs_written = (dst.file == BAD_FILE ? 0 : 1);
this->shadow_compare = false;
- this->ir = v->base_ir;
+ this->ir = NULL;
this->urb_write_flags = BRW_URB_WRITE_NO_FLAGS;
this->header_present = false;
+ this->flag_subreg = 0;
this->mlen = 0;
this->base_mrf = 0;
this->offset = 0;
- this->annotation = v->current_annotation;
+ this->annotation = NULL;
}
vec4_instruction *
vec4_visitor::emit(vec4_instruction *inst)
{
+ inst->ir = this->base_ir;
+ inst->annotation = this->current_annotation;
+
this->instructions.push_tail(inst);
return inst;
}
vec4_instruction *
-vec4_visitor::emit(enum opcode opcode, dst_reg dst,
- src_reg src0, src_reg src1, src_reg src2)
+vec4_visitor::emit(enum opcode opcode, const dst_reg &dst, const src_reg &src0,
+ const src_reg &src1, const src_reg &src2)
{
- return emit(new(mem_ctx) vec4_instruction(this, opcode, dst,
- src0, src1, src2));
+ return emit(new(mem_ctx) vec4_instruction(opcode, dst, src0, src1, src2));
}
vec4_instruction *
-vec4_visitor::emit(enum opcode opcode, dst_reg dst, src_reg src0, src_reg src1)
+vec4_visitor::emit(enum opcode opcode, const dst_reg &dst, const src_reg &src0,
+ const src_reg &src1)
{
- return emit(new(mem_ctx) vec4_instruction(this, opcode, dst, src0, src1));
+ return emit(new(mem_ctx) vec4_instruction(opcode, dst, src0, src1));
}
vec4_instruction *
-vec4_visitor::emit(enum opcode opcode, dst_reg dst, src_reg src0)
+vec4_visitor::emit(enum opcode opcode, const dst_reg &dst, const src_reg &src0)
{
- return emit(new(mem_ctx) vec4_instruction(this, opcode, dst, src0));
+ return emit(new(mem_ctx) vec4_instruction(opcode, dst, src0));
}
vec4_instruction *
-vec4_visitor::emit(enum opcode opcode, dst_reg dst)
+vec4_visitor::emit(enum opcode opcode, const dst_reg &dst)
{
- return emit(new(mem_ctx) vec4_instruction(this, opcode, dst));
+ return emit(new(mem_ctx) vec4_instruction(opcode, dst));
}
vec4_instruction *
vec4_visitor::emit(enum opcode opcode)
{
- return emit(new(mem_ctx) vec4_instruction(this, opcode, dst_reg()));
+ return emit(new(mem_ctx) vec4_instruction(opcode, dst_reg()));
}
#define ALU1(op) \
vec4_instruction * \
vec4_visitor::op(const dst_reg &dst, const src_reg &src0) \
{ \
- return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \
- src0); \
+ return new(mem_ctx) vec4_instruction(BRW_OPCODE_##op, dst, src0); \
}
#define ALU2(op) \
vec4_visitor::op(const dst_reg &dst, const src_reg &src0, \
const src_reg &src1) \
{ \
- return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \
- src0, src1); \
+ return new(mem_ctx) vec4_instruction(BRW_OPCODE_##op, dst, \
+ src0, src1); \
}
#define ALU2_ACC(op) \
vec4_visitor::op(const dst_reg &dst, const src_reg &src0, \
const src_reg &src1) \
{ \
- vec4_instruction *inst = new(mem_ctx) vec4_instruction(this, \
+ vec4_instruction *inst = new(mem_ctx) vec4_instruction( \
BRW_OPCODE_##op, dst, src0, src1); \
- inst->writes_accumulator = true; \
- return inst; \
+ inst->writes_accumulator = true; \
+ return inst; \
}
#define ALU3(op) \
const src_reg &src1, const src_reg &src2) \
{ \
assert(brw->gen >= 6); \
- return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \
+ return new(mem_ctx) vec4_instruction(BRW_OPCODE_##op, dst, \
src0, src1, src2); \
}
{
vec4_instruction *inst;
- inst = new(mem_ctx) vec4_instruction(this, BRW_OPCODE_IF);
+ inst = new(mem_ctx) vec4_instruction(BRW_OPCODE_IF);
inst->predicate = predicate;
return inst;
resolve_ud_negate(&src0);
resolve_ud_negate(&src1);
- inst = new(mem_ctx) vec4_instruction(this, BRW_OPCODE_IF, dst_null_d(),
+ inst = new(mem_ctx) vec4_instruction(BRW_OPCODE_IF, dst_null_d(),
src0, src1);
inst->conditional_mod = condition;
{
vec4_instruction *inst;
- /* original gen4 does type conversion to the destination type
- * before before comparison, producing garbage results for floating
- * point comparisons.
+ /* Take the instruction:
+ *
+ * CMP null<d> src0<f> src1<f>
+ *
+ * Original gen4 does type conversion to the destination type before
+ * comparison, producing garbage results for floating point comparisons.
+ *
+ * The destination type doesn't matter on newer generations, so we set the
+ * type to match src0 so we can compact the instruction.
*/
- if (brw->gen == 4) {
- dst.type = src0.type;
- if (dst.file == HW_REG)
- dst.fixed_hw_reg.type = dst.type;
- }
+ dst.type = src0.type;
+ if (dst.file == HW_REG)
+ dst.fixed_hw_reg.type = dst.type;
resolve_ud_negate(&src0);
resolve_ud_negate(&src1);
- inst = new(mem_ctx) vec4_instruction(this, BRW_OPCODE_CMP, dst, src0, src1);
+ inst = new(mem_ctx) vec4_instruction(BRW_OPCODE_CMP, dst, src0, src1);
inst->conditional_mod = condition;
return inst;
{
vec4_instruction *inst;
- inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_GEN4_SCRATCH_READ,
+ inst = new(mem_ctx) vec4_instruction(SHADER_OPCODE_GEN4_SCRATCH_READ,
dst, index);
inst->base_mrf = 14;
inst->mlen = 2;
{
vec4_instruction *inst;
- inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_GEN4_SCRATCH_WRITE,
+ inst = new(mem_ctx) vec4_instruction(SHADER_OPCODE_GEN4_SCRATCH_WRITE,
dst, src, index);
inst->base_mrf = 13;
inst->mlen = 3;
dst_reg expanded = dst_reg(this, glsl_type::vec4_type);
expanded.type = src.type;
- emit(MOV(expanded, src));
+ emit(VEC4_OPCODE_UNPACK_UNIFORM, expanded, src);
return src_reg(expanded);
}
src_reg
vec4_visitor::fix_math_operand(src_reg src)
{
+ if (brw->gen < 6 || brw->gen >= 8 || src.file == BAD_FILE)
+ return src;
+
/* The gen6 math instruction ignores the source modifiers --
* swizzle, abs, negate, and at least some parts of the register
* region description.
return src_reg(expanded);
}
-void
-vec4_visitor::emit_math1_gen6(enum opcode opcode, dst_reg dst, src_reg src)
-{
- src = fix_math_operand(src);
-
- if (brw->gen == 6 && dst.writemask != WRITEMASK_XYZW) {
- /* The gen6 math instruction must be align1, so we can't do
- * writemasks.
- */
- dst_reg temp_dst = dst_reg(this, glsl_type::vec4_type);
-
- emit(opcode, temp_dst, src);
-
- emit(MOV(dst, src_reg(temp_dst)));
- } else {
- emit(opcode, dst, src);
- }
-}
-
-void
-vec4_visitor::emit_math1_gen4(enum opcode opcode, dst_reg dst, src_reg src)
-{
- vec4_instruction *inst = emit(opcode, dst, src);
- inst->base_mrf = 1;
- inst->mlen = 1;
-}
-
-void
-vec4_visitor::emit_math(opcode opcode, dst_reg dst, src_reg src)
-{
- switch (opcode) {
- case SHADER_OPCODE_RCP:
- case SHADER_OPCODE_RSQ:
- case SHADER_OPCODE_SQRT:
- case SHADER_OPCODE_EXP2:
- case SHADER_OPCODE_LOG2:
- case SHADER_OPCODE_SIN:
- case SHADER_OPCODE_COS:
- break;
- default:
- unreachable("not reached: bad math opcode");
- }
-
- if (brw->gen >= 8) {
- emit(opcode, dst, src);
- } else if (brw->gen >= 6) {
- emit_math1_gen6(opcode, dst, src);
- } else {
- emit_math1_gen4(opcode, dst, src);
- }
-}
-
-void
-vec4_visitor::emit_math2_gen6(enum opcode opcode,
- dst_reg dst, src_reg src0, src_reg src1)
-{
- src0 = fix_math_operand(src0);
- src1 = fix_math_operand(src1);
-
- if (brw->gen == 6 && dst.writemask != WRITEMASK_XYZW) {
- /* The gen6 math instruction must be align1, so we can't do
- * writemasks.
- */
- dst_reg temp_dst = dst_reg(this, glsl_type::vec4_type);
- temp_dst.type = dst.type;
-
- emit(opcode, temp_dst, src0, src1);
-
- emit(MOV(dst, src_reg(temp_dst)));
- } else {
- emit(opcode, dst, src0, src1);
- }
-}
-
-void
-vec4_visitor::emit_math2_gen4(enum opcode opcode,
- dst_reg dst, src_reg src0, src_reg src1)
-{
- vec4_instruction *inst = emit(opcode, dst, src0, src1);
- inst->base_mrf = 1;
- inst->mlen = 2;
-}
-
void
vec4_visitor::emit_math(enum opcode opcode,
- dst_reg dst, src_reg src0, src_reg src1)
+ const dst_reg &dst,
+ const src_reg &src0, const src_reg &src1)
{
- switch (opcode) {
- case SHADER_OPCODE_POW:
- case SHADER_OPCODE_INT_QUOTIENT:
- case SHADER_OPCODE_INT_REMAINDER:
- break;
- default:
- unreachable("not reached: unsupported binary math opcode");
- }
+ vec4_instruction *math =
+ emit(opcode, dst, fix_math_operand(src0), fix_math_operand(src1));
- if (brw->gen >= 8) {
- emit(opcode, dst, src0, src1);
- } else if (brw->gen >= 6) {
- emit_math2_gen6(opcode, dst, src0, src1);
- } else {
- emit_math2_gen4(opcode, dst, src0, src1);
+ if (brw->gen == 6 && dst.writemask != WRITEMASK_XYZW) {
+ /* MATH on Gen6 must be align1, so we can't do writemasks. */
+ math->dst = dst_reg(this, glsl_type::vec4_type);
+ math->dst.type = dst.type;
+ emit(MOV(dst, src_reg(math->dst)));
+ } else if (brw->gen < 6) {
+ math->base_mrf = 1;
+ math->mlen = src1.file == BAD_FILE ? 1 : 2;
}
}
emit(F16TO32(dst, tmp_src));
}
+void
+vec4_visitor::emit_unpack_unorm_4x8(const dst_reg &dst, src_reg src0)
+{
+ /* Instead of splitting the 32-bit integer, shifting, and ORing it back
+ * together, we can shift it by <0, 8, 16, 24>. The packed integer immediate
+ * is not suitable to generate the shift values, but we can use the packed
+ * vector float and a type-converting MOV.
+ */
+ dst_reg shift(this, glsl_type::uvec4_type);
+ emit(MOV(shift, src_reg(0x00, 0x60, 0x70, 0x78)));
+
+ dst_reg shifted(this, glsl_type::uvec4_type);
+ src0.swizzle = BRW_SWIZZLE_XXXX;
+ emit(SHR(shifted, src0, src_reg(shift)));
+
+ shifted.type = BRW_REGISTER_TYPE_UB;
+ dst_reg f(this, glsl_type::vec4_type);
+ emit(VEC4_OPCODE_MOV_BYTES, f, src_reg(shifted));
+
+ emit(MUL(dst, src_reg(f), src_reg(1.0f / 255.0f)));
+}
+
+void
+vec4_visitor::emit_unpack_snorm_4x8(const dst_reg &dst, src_reg src0)
+{
+ /* Instead of splitting the 32-bit integer, shifting, and ORing it back
+ * together, we can shift it by <0, 8, 16, 24>. The packed integer immediate
+ * is not suitable to generate the shift values, but we can use the packed
+ * vector float and a type-converting MOV.
+ */
+ dst_reg shift(this, glsl_type::uvec4_type);
+ emit(MOV(shift, src_reg(0x00, 0x60, 0x70, 0x78)));
+
+ dst_reg shifted(this, glsl_type::uvec4_type);
+ src0.swizzle = BRW_SWIZZLE_XXXX;
+ emit(SHR(shifted, src0, src_reg(shift)));
+
+ shifted.type = BRW_REGISTER_TYPE_B;
+ dst_reg f(this, glsl_type::vec4_type);
+ emit(VEC4_OPCODE_MOV_BYTES, f, src_reg(shifted));
+
+ dst_reg scaled(this, glsl_type::vec4_type);
+ emit(MUL(scaled, src_reg(f), src_reg(1.0f / 127.0f)));
+
+ dst_reg max(this, glsl_type::vec4_type);
+ emit_minmax(BRW_CONDITIONAL_GE, max, src_reg(scaled), src_reg(-1.0f));
+ emit_minmax(BRW_CONDITIONAL_L, dst, src_reg(max), src_reg(1.0f));
+}
+
+void
+vec4_visitor::emit_pack_unorm_4x8(const dst_reg &dst, const src_reg &src0)
+{
+ dst_reg saturated(this, glsl_type::vec4_type);
+ vec4_instruction *inst = emit(MOV(saturated, src0));
+ inst->saturate = true;
+
+ dst_reg scaled(this, glsl_type::vec4_type);
+ emit(MUL(scaled, src_reg(saturated), src_reg(255.0f)));
+
+ dst_reg rounded(this, glsl_type::vec4_type);
+ emit(RNDE(rounded, src_reg(scaled)));
+
+ dst_reg u(this, glsl_type::uvec4_type);
+ emit(MOV(u, src_reg(rounded)));
+
+ src_reg bytes(u);
+ emit(VEC4_OPCODE_PACK_BYTES, dst, bytes);
+}
+
+void
+vec4_visitor::emit_pack_snorm_4x8(const dst_reg &dst, const src_reg &src0)
+{
+ dst_reg max(this, glsl_type::vec4_type);
+ emit_minmax(BRW_CONDITIONAL_GE, max, src0, src_reg(-1.0f));
+
+ dst_reg min(this, glsl_type::vec4_type);
+ emit_minmax(BRW_CONDITIONAL_L, min, src_reg(max), src_reg(1.0f));
+
+ dst_reg scaled(this, glsl_type::vec4_type);
+ emit(MUL(scaled, src_reg(min), src_reg(127.0f)));
+
+ dst_reg rounded(this, glsl_type::vec4_type);
+ emit(RNDE(rounded, src_reg(scaled)));
+
+ dst_reg i(this, glsl_type::ivec4_type);
+ emit(MOV(i, src_reg(rounded)));
+
+ src_reg bytes(i);
+ emit(VEC4_OPCODE_PACK_BYTES, dst, bytes);
+}
+
void
vec4_visitor::visit_instructions(const exec_list *list)
{
return 0;
case GLSL_TYPE_IMAGE:
case GLSL_TYPE_VOID:
+ case GLSL_TYPE_DOUBLE:
case GLSL_TYPE_ERROR:
case GLSL_TYPE_INTERFACE:
unreachable("not reached");
return 0;
}
-int
-vec4_visitor::virtual_grf_alloc(int size)
-{
- if (virtual_grf_array_size <= virtual_grf_count) {
- if (virtual_grf_array_size == 0)
- virtual_grf_array_size = 16;
- else
- virtual_grf_array_size *= 2;
- virtual_grf_sizes = reralloc(mem_ctx, virtual_grf_sizes, int,
- virtual_grf_array_size);
- virtual_grf_reg_map = reralloc(mem_ctx, virtual_grf_reg_map, int,
- virtual_grf_array_size);
- }
- virtual_grf_reg_map[virtual_grf_count] = virtual_grf_reg_count;
- virtual_grf_reg_count += size;
- virtual_grf_sizes[virtual_grf_count] = size;
- return virtual_grf_count++;
-}
-
src_reg::src_reg(class vec4_visitor *v, const struct glsl_type *type)
{
init();
this->file = GRF;
- this->reg = v->virtual_grf_alloc(type_size(type));
+ this->reg = v->alloc.allocate(type_size(type));
if (type->is_array() || type->is_record()) {
this->swizzle = BRW_SWIZZLE_NOOP;
} else {
- this->swizzle = swizzle_for_size(type->vector_elements);
+ this->swizzle = brw_swizzle_for_size(type->vector_elements);
}
this->type = brw_type_for_base_type(type);
init();
this->file = GRF;
- this->reg = v->virtual_grf_alloc(type_size(type) * size);
+ this->reg = v->alloc.allocate(type_size(type) * size);
this->swizzle = BRW_SWIZZLE_NOOP;
init();
this->file = GRF;
- this->reg = v->virtual_grf_alloc(type_size(type));
+ this->reg = v->alloc.allocate(type_size(type));
if (type->is_array() || type->is_record()) {
this->writemask = WRITEMASK_XYZW;
&this->prog->Parameters->ParameterValues[index][0];
assert(this->uniforms < uniform_array_size);
- this->uniform_vector_size[this->uniforms] = 0;
- /* Add each of the unique swizzled channels of the element.
- * This will end up matching the size of the glsl_type of this field.
- */
- int last_swiz = -1;
- for (unsigned int j = 0; j < 4; j++) {
- int swiz = GET_SWZ(slots[i].swizzle, j);
- last_swiz = swiz;
-
- stage_prog_data->param[this->uniforms * 4 + j] = &values[swiz];
- assert(this->uniforms < uniform_array_size);
- if (swiz <= last_swiz)
- this->uniform_vector_size[this->uniforms]++;
- }
+
+ for (unsigned j = 0; j < 4; j++)
+ stage_prog_data->param[this->uniforms * 4 + j] =
+ &values[GET_SWZ(slots[i].swizzle, j)];
+
+ this->uniform_vector_size[this->uniforms] =
+ (ir->type->is_scalar() || ir->type->is_vector() ||
+ ir->type->is_matrix() ? ir->type->vector_elements : 4);
+
this->uniforms++;
}
}
break;
case ir_binop_logic_xor:
- inst = emit(XOR(dst_null_d(), op[0], op[1]));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ if (brw->gen <= 5) {
+ src_reg temp = src_reg(this, ir->type);
+ emit(XOR(dst_reg(temp), op[0], op[1]));
+ inst = emit(AND(dst_null_d(), temp, src_reg(1)));
+ } else {
+ inst = emit(XOR(dst_null_d(), op[0], op[1]));
+ }
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
break;
case ir_binop_logic_or:
- inst = emit(OR(dst_null_d(), op[0], op[1]));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ if (brw->gen <= 5) {
+ src_reg temp = src_reg(this, ir->type);
+ emit(OR(dst_reg(temp), op[0], op[1]));
+ inst = emit(AND(dst_null_d(), temp, src_reg(1)));
+ } else {
+ inst = emit(OR(dst_null_d(), op[0], op[1]));
+ }
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
break;
case ir_binop_logic_and:
- inst = emit(AND(dst_null_d(), op[0], op[1]));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ if (brw->gen <= 5) {
+ src_reg temp = src_reg(this, ir->type);
+ emit(AND(dst_reg(temp), op[0], op[1]));
+ inst = emit(AND(dst_null_d(), temp, src_reg(1)));
+ } else {
+ inst = emit(AND(dst_null_d(), op[0], op[1]));
+ }
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
break;
case ir_unop_f2b:
break;
case ir_binop_all_equal:
+ if (brw->gen <= 5) {
+ resolve_bool_comparison(expr->operands[0], &op[0]);
+ resolve_bool_comparison(expr->operands[1], &op[1]);
+ }
inst = emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z));
*predicate = BRW_PREDICATE_ALIGN16_ALL4H;
break;
case ir_binop_any_nequal:
+ if (brw->gen <= 5) {
+ resolve_bool_comparison(expr->operands[0], &op[0]);
+ resolve_bool_comparison(expr->operands[1], &op[1]);
+ }
inst = emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ));
*predicate = BRW_PREDICATE_ALIGN16_ANY4H;
break;
case ir_unop_any:
+ if (brw->gen <= 5) {
+ resolve_bool_comparison(expr->operands[0], &op[0]);
+ }
inst = emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ));
*predicate = BRW_PREDICATE_ALIGN16_ANY4H;
break;
case ir_binop_lequal:
case ir_binop_equal:
case ir_binop_nequal:
+ if (brw->gen <= 5) {
+ resolve_bool_comparison(expr->operands[0], &op[0]);
+ resolve_bool_comparison(expr->operands[1], &op[1]);
+ }
emit(CMP(dst_null_d(), op[0], op[1],
brw_conditional_for_comparison(expr->operation)));
break;
resolve_ud_negate(&this->result);
- if (brw->gen >= 6) {
- vec4_instruction *inst = emit(AND(dst_null_d(),
- this->result, src_reg(1)));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- } else {
- vec4_instruction *inst = emit(MOV(dst_null_d(), this->result));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
- }
+ vec4_instruction *inst = emit(AND(dst_null_d(), this->result, src_reg(1)));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
}
/**
switch (ir->data.mode) {
case ir_var_shader_in:
+ assert(ir->data.location != -1);
reg = new(mem_ctx) dst_reg(ATTR, ir->data.location);
break;
case ir_var_shader_out:
+ assert(ir->data.location != -1);
reg = new(mem_ctx) dst_reg(this, ir->type);
for (int i = 0; i < type_size(ir->type); i++) {
if (ir->type->base_type != GLSL_TYPE_FLOAT)
return false;
- ir_rvalue *nonmul = ir->operands[1];
- ir_expression *mul = ir->operands[0]->as_expression();
+ ir_rvalue *nonmul;
+ ir_expression *mul;
+ bool mul_negate, mul_abs;
- if (!mul || mul->operation != ir_binop_mul) {
- nonmul = ir->operands[0];
- mul = ir->operands[1]->as_expression();
+ for (int i = 0; i < 2; i++) {
+ mul_negate = false;
+ mul_abs = false;
- if (!mul || mul->operation != ir_binop_mul)
- return false;
+ mul = ir->operands[i]->as_expression();
+ nonmul = ir->operands[1 - i];
+
+ if (mul && mul->operation == ir_unop_abs) {
+ mul = mul->operands[0]->as_expression();
+ mul_abs = true;
+ } else if (mul && mul->operation == ir_unop_neg) {
+ mul = mul->operands[0]->as_expression();
+ mul_negate = true;
+ }
+
+ if (mul && mul->operation == ir_binop_mul)
+ break;
}
+ if (!mul || mul->operation != ir_binop_mul)
+ return false;
+
nonmul->accept(this);
src_reg src0 = fix_3src_operand(this->result);
mul->operands[0]->accept(this);
src_reg src1 = fix_3src_operand(this->result);
+ src1.negate ^= mul_negate;
+ src1.abs = mul_abs;
+ if (mul_abs)
+ src1.negate = false;
mul->operands[1]->accept(this);
src_reg src2 = fix_3src_operand(this->result);
+ src2.abs = mul_abs;
+ if (mul_abs)
+ src2.negate = false;
this->result = src_reg(this, ir->type);
emit(BRW_OPCODE_MAD, dst_reg(this->result), src0, src1, src2);
vec4_visitor::visit(ir_expression *ir)
{
unsigned int operand;
- src_reg op[Elements(ir->operands)];
+ src_reg op[ARRAY_SIZE(ir->operands)];
vec4_instruction *inst;
if (ir->operation == ir_binop_add) {
switch (ir->operation) {
case ir_unop_logic_not:
- if (ctx->Const.UniformBooleanTrue != 1) {
- emit(NOT(result_dst, op[0]));
- } else {
- emit(XOR(result_dst, op[0], src_reg(1u)));
- }
+ emit(NOT(result_dst, op[0]));
break;
case ir_unop_neg:
op[0].negate = !op[0].negate;
break;
}
case ir_binop_mod:
- /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
+ /* Floating point should be lowered by MOD_TO_FLOOR in the compiler. */
assert(ir->type->is_integer());
emit_math(SHADER_OPCODE_INT_REMAINDER, result_dst, op[0], op[1]);
break;
case ir_binop_gequal:
case ir_binop_equal:
case ir_binop_nequal: {
+ if (brw->gen <= 5) {
+ resolve_bool_comparison(ir->operands[0], &op[0]);
+ resolve_bool_comparison(ir->operands[1], &op[1]);
+ }
emit(CMP(result_dst, op[0], op[1],
brw_conditional_for_comparison(ir->operation)));
- if (ctx->Const.UniformBooleanTrue == 1) {
- emit(AND(result_dst, result_src, src_reg(1u)));
- }
break;
}
case ir_binop_all_equal:
+ if (brw->gen <= 5) {
+ resolve_bool_comparison(ir->operands[0], &op[0]);
+ resolve_bool_comparison(ir->operands[1], &op[1]);
+ }
+
/* "==" operator producing a scalar boolean. */
if (ir->operands[0]->type->is_vector() ||
ir->operands[1]->type->is_vector()) {
emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_Z));
emit(MOV(result_dst, src_reg(0)));
- inst = emit(MOV(result_dst, src_reg(ctx->Const.UniformBooleanTrue)));
+ inst = emit(MOV(result_dst, src_reg((int)ctx->Const.UniformBooleanTrue)));
inst->predicate = BRW_PREDICATE_ALIGN16_ALL4H;
} else {
emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_Z));
- if (ctx->Const.UniformBooleanTrue == 1) {
- emit(AND(result_dst, result_src, src_reg(1u)));
- }
}
break;
case ir_binop_any_nequal:
+ if (brw->gen <= 5) {
+ resolve_bool_comparison(ir->operands[0], &op[0]);
+ resolve_bool_comparison(ir->operands[1], &op[1]);
+ }
+
/* "!=" operator producing a scalar boolean. */
if (ir->operands[0]->type->is_vector() ||
ir->operands[1]->type->is_vector()) {
emit(CMP(dst_null_d(), op[0], op[1], BRW_CONDITIONAL_NZ));
emit(MOV(result_dst, src_reg(0)));
- inst = emit(MOV(result_dst, src_reg(ctx->Const.UniformBooleanTrue)));
+ inst = emit(MOV(result_dst, src_reg((int)ctx->Const.UniformBooleanTrue)));
inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H;
} else {
emit(CMP(result_dst, op[0], op[1], BRW_CONDITIONAL_NZ));
- if (ctx->Const.UniformBooleanTrue == 1) {
- emit(AND(result_dst, result_src, src_reg(1u)));
- }
}
break;
case ir_unop_any:
+ if (brw->gen <= 5) {
+ resolve_bool_comparison(ir->operands[0], &op[0]);
+ }
emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ));
emit(MOV(result_dst, src_reg(0)));
- inst = emit(MOV(result_dst, src_reg(ctx->Const.UniformBooleanTrue)));
+ inst = emit(MOV(result_dst, src_reg((int)ctx->Const.UniformBooleanTrue)));
inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H;
break;
emit(MOV(result_dst, op[0]));
break;
case ir_unop_b2i:
- if (ctx->Const.UniformBooleanTrue != 1) {
- emit(AND(result_dst, op[0], src_reg(1u)));
- } else {
- emit(MOV(result_dst, op[0]));
- }
+ emit(AND(result_dst, op[0], src_reg(1)));
break;
case ir_unop_b2f:
- if (ctx->Const.UniformBooleanTrue != 1) {
- op[0].type = BRW_REGISTER_TYPE_UD;
- result_dst.type = BRW_REGISTER_TYPE_UD;
- emit(AND(result_dst, op[0], src_reg(0x3f800000u)));
- result_dst.type = BRW_REGISTER_TYPE_F;
- } else {
- emit(MOV(result_dst, op[0]));
+ if (brw->gen <= 5) {
+ resolve_bool_comparison(ir->operands[0], &op[0]);
}
+ op[0].type = BRW_REGISTER_TYPE_D;
+ result_dst.type = BRW_REGISTER_TYPE_D;
+ emit(AND(result_dst, op[0], src_reg(0x3f800000u)));
+ result_dst.type = BRW_REGISTER_TYPE_F;
break;
case ir_unop_f2b:
- case ir_unop_i2b:
emit(CMP(result_dst, op[0], src_reg(0.0f), BRW_CONDITIONAL_NZ));
- if (ctx->Const.UniformBooleanTrue == 1) {
- emit(AND(result_dst, result_src, src_reg(1u)));
- }
+ break;
+ case ir_unop_i2b:
+ emit(CMP(result_dst, op[0], src_reg(0), BRW_CONDITIONAL_NZ));
break;
case ir_unop_trunc:
emit(RNDZ(result_dst, op[0]));
break;
- case ir_unop_ceil:
- op[0].negate = !op[0].negate;
- inst = emit(RNDD(result_dst, op[0]));
- this->result.negate = true;
+ case ir_unop_ceil: {
+ src_reg tmp = src_reg(this, ir->type);
+ op[0].negate = !op[0].negate;
+ emit(RNDD(dst_reg(tmp), op[0]));
+ tmp.negate = true;
+ emit(MOV(result_dst, tmp));
+ }
break;
case ir_unop_floor:
inst = emit(RNDD(result_dst, op[0]));
emit_minmax(BRW_CONDITIONAL_L, result_dst, op[0], op[1]);
break;
case ir_binop_max:
- emit_minmax(BRW_CONDITIONAL_G, result_dst, op[0], op[1]);
+ emit_minmax(BRW_CONDITIONAL_GE, result_dst, op[0], op[1]);
break;
case ir_binop_pow:
if (brw->gen >= 7) {
dst_reg grf_offset = dst_reg(this, glsl_type::int_type);
+
+ /* We have to use a message header on Skylake to get SIMD4x2 mode.
+ * Reserve space for the register.
+ */
+ if (brw->gen >= 9) {
+ grf_offset.reg_offset++;
+ alloc.sizes[grf_offset.reg] = 2;
+ }
+
grf_offset.type = offset.type;
emit(MOV(grf_offset, offset));
- emit(new(mem_ctx) vec4_instruction(this,
- VS_OPCODE_PULL_CONSTANT_LOAD_GEN7,
- dst_reg(packed_consts),
- surf_index,
- src_reg(grf_offset)));
+ vec4_instruction *pull =
+ emit(new(mem_ctx) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD_GEN7,
+ dst_reg(packed_consts),
+ surf_index,
+ src_reg(grf_offset)));
+ pull->mlen = 1;
} else {
vec4_instruction *pull =
- emit(new(mem_ctx) vec4_instruction(this,
- VS_OPCODE_PULL_CONSTANT_LOAD,
+ emit(new(mem_ctx) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD,
dst_reg(packed_consts),
surf_index,
offset));
pull->mlen = 1;
}
- packed_consts.swizzle = swizzle_for_size(ir->type->vector_elements);
+ packed_consts.swizzle = brw_swizzle_for_size(ir->type->vector_elements);
packed_consts.swizzle += BRW_SWIZZLE4(const_offset % 16 / 4,
const_offset % 16 / 4,
const_offset % 16 / 4,
if (ir->type->base_type == GLSL_TYPE_BOOL) {
emit(CMP(result_dst, packed_consts, src_reg(0u),
BRW_CONDITIONAL_NZ));
- if (ctx->Const.UniformBooleanTrue == 1) {
- emit(AND(result_dst, result, src_reg(1u)));
- }
} else {
emit(MOV(result_dst, packed_consts));
}
case ir_unop_unpack_half_2x16:
emit_unpack_half_2x16(result_dst, op[0]);
break;
- case ir_unop_pack_snorm_2x16:
+ case ir_unop_unpack_unorm_4x8:
+ emit_unpack_unorm_4x8(result_dst, op[0]);
+ break;
+ case ir_unop_unpack_snorm_4x8:
+ emit_unpack_snorm_4x8(result_dst, op[0]);
+ break;
+ case ir_unop_pack_unorm_4x8:
+ emit_pack_unorm_4x8(result_dst, op[0]);
+ break;
case ir_unop_pack_snorm_4x8:
+ emit_pack_snorm_4x8(result_dst, op[0]);
+ break;
+ case ir_unop_pack_snorm_2x16:
case ir_unop_pack_unorm_2x16:
- case ir_unop_pack_unorm_4x8:
case ir_unop_unpack_snorm_2x16:
- case ir_unop_unpack_snorm_4x8:
case ir_unop_unpack_unorm_2x16:
- case ir_unop_unpack_unorm_4x8:
unreachable("not reached: should be handled by lower_packing_builtins");
case ir_unop_unpack_half_2x16_split_x:
case ir_unop_unpack_half_2x16_split_y:
unreachable("not reached: should not occur in vertex shader");
case ir_binop_ldexp:
unreachable("not reached: should be handled by ldexp_to_arith()");
+ case ir_unop_d2f:
+ case ir_unop_f2d:
+ case ir_unop_d2i:
+ case ir_unop_i2d:
+ case ir_unop_d2u:
+ case ir_unop_u2d:
+ case ir_unop_d2b:
+ case ir_unop_pack_double_2x32:
+ case ir_unop_unpack_double_2x32:
+ case ir_unop_frexp_sig:
+ case ir_unop_frexp_exp:
+ unreachable("fp64 todo");
}
}
void
vec4_visitor::visit(ir_swizzle *ir)
{
- src_reg src;
- int i = 0;
- int swizzle[4];
-
/* Note that this is only swizzles in expressions, not those on the left
* hand side of an assignment, which do write masking. See ir_assignment
* for that.
*/
+ const unsigned swz = brw_compose_swizzle(
+ brw_swizzle_for_size(ir->type->vector_elements),
+ BRW_SWIZZLE4(ir->mask.x, ir->mask.y, ir->mask.z, ir->mask.w));
ir->val->accept(this);
- src = this->result;
- assert(src.file != BAD_FILE);
-
- for (i = 0; i < ir->type->vector_elements; i++) {
- switch (i) {
- case 0:
- swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.x);
- break;
- case 1:
- swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.y);
- break;
- case 2:
- swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.z);
- break;
- case 3:
- swizzle[i] = BRW_GET_SWZ(src.swizzle, ir->mask.w);
- break;
- }
- }
- for (; i < 4; i++) {
- /* Replicate the last channel out. */
- swizzle[i] = swizzle[ir->type->vector_elements - 1];
- }
-
- src.swizzle = BRW_SWIZZLE4(swizzle[0], swizzle[1], swizzle[2], swizzle[3]);
-
- this->result = src;
+ this->result = swizzle(this->result, swz);
}
void
return;
if (type->is_scalar() || type->is_vector() || type->is_matrix())
- this->result.swizzle = swizzle_for_size(type->vector_elements);
+ this->result.swizzle = brw_swizzle_for_size(type->vector_elements);
}
/* If the type is smaller than a vec4, replicate the last channel out. */
if (ir->type->is_scalar() || ir->type->is_vector() || ir->type->is_matrix())
- src.swizzle = swizzle_for_size(ir->type->vector_elements);
+ src.swizzle = brw_swizzle_for_size(ir->type->vector_elements);
else
src.swizzle = BRW_SWIZZLE_NOOP;
src.type = brw_type_for_base_type(ir->type);
/* If the type is smaller than a vec4, replicate the last channel out. */
if (ir->type->is_scalar() || ir->type->is_vector() || ir->type->is_matrix())
- this->result.swizzle = swizzle_for_size(ir->type->vector_elements);
+ this->result.swizzle = brw_swizzle_for_size(ir->type->vector_elements);
else
this->result.swizzle = BRW_SWIZZLE_NOOP;
this->result.type = brw_type_for_base_type(ir->type);
dst->writemask = (1 << type->vector_elements) - 1;
- src->swizzle = swizzle_for_size(type->vector_elements);
+ src->swizzle = brw_swizzle_for_size(type->vector_elements);
vec4_instruction *inst = emit(MOV(*dst, *src));
inst->predicate = predicate;
*/
assert(src.swizzle ==
(ir->rhs->type->is_matrix()
- ? swizzle_for_size(ir->rhs->type->vector_elements)
+ ? brw_swizzle_for_size(ir->rhs->type->vector_elements)
: BRW_SWIZZLE_NOOP));
emit_block_move(&dst, &src, ir->rhs->type, predicate);
last_rhs_inst = (vec4_instruction *)this->instructions.get_tail();
- src_reg src = this->result;
-
int swizzles[4];
- int first_enabled_chan = 0;
int src_chan = 0;
assert(ir->lhs->type->is_vector() ||
ir->lhs->type->is_scalar());
dst.writemask = ir->write_mask;
- for (int i = 0; i < 4; i++) {
- if (dst.writemask & (1 << i)) {
- first_enabled_chan = BRW_GET_SWZ(src.swizzle, i);
- break;
- }
- }
-
/* Swizzle a small RHS vector into the channels being written.
*
* glsl ir treats write_mask as dictating how many channels are
* present on the RHS while in our instructions we need to make
* those channels appear in the slots of the vec4 they're written to.
*/
- for (int i = 0; i < 4; i++) {
- if (dst.writemask & (1 << i))
- swizzles[i] = BRW_GET_SWZ(src.swizzle, src_chan++);
- else
- swizzles[i] = first_enabled_chan;
- }
- src.swizzle = BRW_SWIZZLE4(swizzles[0], swizzles[1],
- swizzles[2], swizzles[3]);
+ for (int i = 0; i < 4; i++)
+ swizzles[i] = (ir->write_mask & (1 << i) ? src_chan++ : 0);
+
+ src_reg src = swizzle(this->result,
+ BRW_SWIZZLE4(swizzles[0], swizzles[1],
+ swizzles[2], swizzles[3]));
if (try_rewrite_rhs_to_dst(ir, dst, src, pre_rhs_inst, last_rhs_inst)) {
return;
break;
case GLSL_TYPE_BOOL:
emit(MOV(*dst,
- src_reg(ir->value.b[i] != 0 ? ctx->Const.UniformBooleanTrue
- : 0u)));
+ src_reg(ir->value.b[i] != 0 ? (int)ctx->Const.UniformBooleanTrue
+ : 0)));
break;
default:
unreachable("Non-float/uint/int/bool constant");
src_reg
vec4_visitor::emit_mcs_fetch(ir_texture *ir, src_reg coordinate, src_reg sampler)
{
- vec4_instruction *inst = new(mem_ctx) vec4_instruction(this, SHADER_OPCODE_TXF_MCS);
+ vec4_instruction *inst =
+ new(mem_ctx) vec4_instruction(SHADER_OPCODE_TXF_MCS,
+ dst_reg(this, glsl_type::uvec4_type));
inst->base_mrf = 2;
inst->mlen = 1;
- inst->dst = dst_reg(this, glsl_type::uvec4_type);
- inst->dst.writemask = WRITEMASK_XYZW;
-
inst->src[1] = sampler;
/* parameters are: u, v, r, lod; lod will always be zero due to api restrictions */
unreachable("Unrecognized tex op");
}
- vec4_instruction *inst = new(mem_ctx) vec4_instruction(this, opcode);
+ vec4_instruction *inst = new(mem_ctx) vec4_instruction(
+ opcode, dst_reg(this, ir->type));
if (ir->offset != NULL && !has_nonconstant_offset) {
- inst->texture_offset =
+ inst->offset =
brw_texture_offset(ctx, ir->offset->as_constant()->value.i,
ir->offset->type->vector_elements);
}
/* Stuff the channel select bits in the top of the texture offset */
if (ir->op == ir_tg4)
- inst->texture_offset |= gather_channel(ir, sampler) << 16;
+ inst->offset |= gather_channel(ir, sampler) << 16;
/* The message header is necessary for:
* - Gen4 (always)
+ * - Gen9+ for selecting SIMD4x2
* - Texel offsets
* - Gather channel selection
* - Sampler indices too large to fit in a 4-bit value.
*/
inst->header_present =
- brw->gen < 5 || inst->texture_offset != 0 || ir->op == ir_tg4 ||
+ brw->gen < 5 || brw->gen >= 9 ||
+ inst->offset != 0 || ir->op == ir_tg4 ||
is_high_sampler(brw, sampler_reg);
inst->base_mrf = 2;
inst->mlen = inst->header_present + 1; /* always at least one */
- inst->dst = dst_reg(this, ir->type);
inst->dst.writemask = WRITEMASK_XYZW;
inst->shadow_compare = ir->shadow_comparitor != NULL;
}
}
-void
+vec4_instruction *
vec4_visitor::emit_generic_urb_slot(dst_reg reg, int varying)
{
assert (varying < VARYING_SLOT_MAX);
reg.type = output_reg[varying].type;
current_annotation = output_reg_annotation[varying];
/* Copy the register, saturating if necessary */
- vec4_instruction *inst = emit(MOV(reg,
- src_reg(output_reg[varying])));
- if ((varying == VARYING_SLOT_COL0 ||
- varying == VARYING_SLOT_COL1 ||
- varying == VARYING_SLOT_BFC0 ||
- varying == VARYING_SLOT_BFC1) &&
- key->clamp_vertex_color) {
- inst->saturate = true;
- }
+ return emit(MOV(reg, src_reg(output_reg[varying])));
}
void
case BRW_VARYING_SLOT_PAD:
/* No need to write to this slot */
break;
+ case VARYING_SLOT_COL0:
+ case VARYING_SLOT_COL1:
+ case VARYING_SLOT_BFC0:
+ case VARYING_SLOT_BFC1: {
+ /* These built-in varyings are only supported in compatibility mode,
+ * and we only support GS in core profile. So, this must be a vertex
+ * shader.
+ */
+ assert(stage == MESA_SHADER_VERTEX);
+ vec4_instruction *inst = emit_generic_urb_slot(reg, varying);
+ if (((struct brw_vs_prog_key *) key)->clamp_vertex_color)
+ inst->saturate = true;
+ break;
+ }
+
default:
emit_generic_urb_slot(reg, varying);
break;
* weren't initialized, it will confuse live interval analysis, which will
* make spilling fail to make progress.
*/
- src_reg temp = src_reg(this, glsl_type::vec4_type);
- temp.type = inst->dst.type;
- int first_writemask_chan = ffs(inst->dst.writemask) - 1;
- int swizzles[4];
- for (int i = 0; i < 4; i++)
- if (inst->dst.writemask & (1 << i))
- swizzles[i] = i;
- else
- swizzles[i] = first_writemask_chan;
- temp.swizzle = BRW_SWIZZLE4(swizzles[0], swizzles[1],
- swizzles[2], swizzles[3]);
-
+ const src_reg temp = swizzle(retype(src_reg(this, glsl_type::vec4_type),
+ inst->dst.type),
+ brw_swizzle_for_mask(inst->dst.writemask));
dst_reg dst = dst_reg(brw_writemask(brw_vec8_grf(0, 0),
inst->dst.writemask));
vec4_instruction *write = SCRATCH_WRITE(dst, temp, index);
inst->dst.reladdr = NULL;
}
+/**
+ * Checks if \p src and/or \p src.reladdr require a scratch read, and if so,
+ * adds the scratch read(s) before \p inst. The function also checks for
+ * recursive reladdr scratch accesses, issuing the corresponding scratch
+ * loads and rewriting reladdr references accordingly.
+ *
+ * \return \p src if it did not require a scratch load, otherwise, the
+ * register holding the result of the scratch load that the caller should
+ * use to rewrite src.
+ */
+src_reg
+vec4_visitor::emit_resolve_reladdr(int scratch_loc[], bblock_t *block,
+ vec4_instruction *inst, src_reg src)
+{
+ /* Resolve recursive reladdr scratch access by calling ourselves
+ * with src.reladdr
+ */
+ if (src.reladdr)
+ *src.reladdr = emit_resolve_reladdr(scratch_loc, block, inst,
+ *src.reladdr);
+
+ /* Now handle scratch access on src */
+ if (src.file == GRF && scratch_loc[src.reg] != -1) {
+ dst_reg temp = dst_reg(this, glsl_type::vec4_type);
+ emit_scratch_read(block, inst, temp, src, scratch_loc[src.reg]);
+ src.reg = temp.reg;
+ src.reg_offset = temp.reg_offset;
+ src.reladdr = NULL;
+ }
+
+ return src;
+}
+
/**
* We can't generally support array access in GRF space, because a
* single instruction's destination can only span 2 contiguous
void
vec4_visitor::move_grf_array_access_to_scratch()
{
- int scratch_loc[this->virtual_grf_count];
+ int scratch_loc[this->alloc.count];
memset(scratch_loc, -1, sizeof(scratch_loc));
/* First, calculate the set of virtual GRFs that need to be punted
* scratch.
*/
foreach_block_and_inst(block, vec4_instruction, inst, cfg) {
- if (inst->dst.file == GRF && inst->dst.reladdr &&
- scratch_loc[inst->dst.reg] == -1) {
- scratch_loc[inst->dst.reg] = c->last_scratch;
- c->last_scratch += this->virtual_grf_sizes[inst->dst.reg];
+ if (inst->dst.file == GRF && inst->dst.reladdr) {
+ if (scratch_loc[inst->dst.reg] == -1) {
+ scratch_loc[inst->dst.reg] = c->last_scratch;
+ c->last_scratch += this->alloc.sizes[inst->dst.reg];
+ }
+
+ for (src_reg *iter = inst->dst.reladdr;
+ iter->reladdr;
+ iter = iter->reladdr) {
+ if (iter->file == GRF && scratch_loc[iter->reg] == -1) {
+ scratch_loc[iter->reg] = c->last_scratch;
+ c->last_scratch += this->alloc.sizes[iter->reg];
+ }
+ }
}
for (int i = 0 ; i < 3; i++) {
- src_reg *src = &inst->src[i];
-
- if (src->file == GRF && src->reladdr &&
- scratch_loc[src->reg] == -1) {
- scratch_loc[src->reg] = c->last_scratch;
- c->last_scratch += this->virtual_grf_sizes[src->reg];
- }
+ for (src_reg *iter = &inst->src[i];
+ iter->reladdr;
+ iter = iter->reladdr) {
+ if (iter->file == GRF && scratch_loc[iter->reg] == -1) {
+ scratch_loc[iter->reg] = c->last_scratch;
+ c->last_scratch += this->alloc.sizes[iter->reg];
+ }
+ }
}
}
base_ir = inst->ir;
current_annotation = inst->annotation;
- if (inst->dst.file == GRF && scratch_loc[inst->dst.reg] != -1) {
- emit_scratch_write(block, inst, scratch_loc[inst->dst.reg]);
- }
-
- for (int i = 0 ; i < 3; i++) {
- if (inst->src[i].file != GRF || scratch_loc[inst->src[i].reg] == -1)
- continue;
-
- dst_reg temp = dst_reg(this, glsl_type::vec4_type);
+ /* First handle scratch access on the dst. Notice we have to handle
+ * the case where the dst's reladdr also points to scratch space.
+ */
+ if (inst->dst.reladdr)
+ *inst->dst.reladdr = emit_resolve_reladdr(scratch_loc, block, inst,
+ *inst->dst.reladdr);
- emit_scratch_read(block, inst, temp, inst->src[i],
- scratch_loc[inst->src[i].reg]);
+ /* Now that we have handled any (possibly recursive) reladdr scratch
+ * accesses for dst we can safely do the scratch write for dst itself
+ */
+ if (inst->dst.file == GRF && scratch_loc[inst->dst.reg] != -1)
+ emit_scratch_write(block, inst, scratch_loc[inst->dst.reg]);
- inst->src[i].file = temp.file;
- inst->src[i].reg = temp.reg;
- inst->src[i].reg_offset = temp.reg_offset;
- inst->src[i].reladdr = NULL;
+ /* Now handle scratch access on any src. In this case, since inst->src[i]
+ * already is a src_reg, we can just call emit_resolve_reladdr with
+ * inst->src[i] and it will take care of handling scratch loads for
+ * both src and src.reladdr (recursively).
+ */
+ for (int i = 0 ; i < 3; i++) {
+ inst->src[i] = emit_resolve_reladdr(scratch_loc, block, inst,
+ inst->src[i]);
}
}
}
if (brw->gen >= 7) {
dst_reg grf_offset = dst_reg(this, glsl_type::int_type);
+
+ /* We have to use a message header on Skylake to get SIMD4x2 mode.
+ * Reserve space for the register.
+ */
+ if (brw->gen >= 9) {
+ grf_offset.reg_offset++;
+ alloc.sizes[grf_offset.reg] = 2;
+ }
+
grf_offset.type = offset.type;
emit_before(block, inst, MOV(grf_offset, offset));
- load = new(mem_ctx) vec4_instruction(this,
- VS_OPCODE_PULL_CONSTANT_LOAD_GEN7,
+ load = new(mem_ctx) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD_GEN7,
temp, index, src_reg(grf_offset));
+ load->mlen = 1;
} else {
- load = new(mem_ctx) vec4_instruction(this, VS_OPCODE_PULL_CONSTANT_LOAD,
+ load = new(mem_ctx) vec4_instruction(VS_OPCODE_PULL_CONSTANT_LOAD,
temp, index, offset);
load->base_mrf = 14;
load->mlen = 1;
{
int pull_constant_loc[this->uniforms];
memset(pull_constant_loc, -1, sizeof(pull_constant_loc));
+ bool nested_reladdr;
/* Walk through and find array access of uniforms. Put a copy of that
* uniform in the pull constant buffer.
* Note that we don't move constant-indexed accesses to arrays. No
* testing has been done of the performance impact of this choice.
*/
- foreach_block_and_inst_safe(block, vec4_instruction, inst, cfg) {
- for (int i = 0 ; i < 3; i++) {
- if (inst->src[i].file != UNIFORM || !inst->src[i].reladdr)
- continue;
+ do {
+ nested_reladdr = false;
- int uniform = inst->src[i].reg;
+ foreach_block_and_inst_safe(block, vec4_instruction, inst, cfg) {
+ for (int i = 0 ; i < 3; i++) {
+ if (inst->src[i].file != UNIFORM || !inst->src[i].reladdr)
+ continue;
- /* If this array isn't already present in the pull constant buffer,
- * add it.
- */
- if (pull_constant_loc[uniform] == -1) {
- const gl_constant_value **values =
- &stage_prog_data->param[uniform * 4];
+ int uniform = inst->src[i].reg;
- pull_constant_loc[uniform] = stage_prog_data->nr_pull_params / 4;
+ if (inst->src[i].reladdr->reladdr)
+ nested_reladdr = true; /* will need another pass */
- assert(uniform < uniform_array_size);
- for (int j = 0; j < uniform_size[uniform] * 4; j++) {
- stage_prog_data->pull_param[stage_prog_data->nr_pull_params++]
- = values[j];
- }
- }
+ /* If this array isn't already present in the pull constant buffer,
+ * add it.
+ */
+ if (pull_constant_loc[uniform] == -1) {
+ const gl_constant_value **values =
+ &stage_prog_data->param[uniform * 4];
+
+ pull_constant_loc[uniform] = stage_prog_data->nr_pull_params / 4;
+
+ assert(uniform < uniform_array_size);
+ for (int j = 0; j < uniform_size[uniform] * 4; j++) {
+ stage_prog_data->pull_param[stage_prog_data->nr_pull_params++]
+ = values[j];
+ }
+ }
- /* Set up the annotation tracking for new generated instructions. */
- base_ir = inst->ir;
- current_annotation = inst->annotation;
+ /* Set up the annotation tracking for new generated instructions. */
+ base_ir = inst->ir;
+ current_annotation = inst->annotation;
- dst_reg temp = dst_reg(this, glsl_type::vec4_type);
+ dst_reg temp = dst_reg(this, glsl_type::vec4_type);
- emit_pull_constant_load(block, inst, temp, inst->src[i],
- pull_constant_loc[uniform]);
+ emit_pull_constant_load(block, inst, temp, inst->src[i],
+ pull_constant_loc[uniform]);
- inst->src[i].file = temp.file;
- inst->src[i].reg = temp.reg;
- inst->src[i].reg_offset = temp.reg_offset;
- inst->src[i].reladdr = NULL;
+ inst->src[i].file = temp.file;
+ inst->src[i].reg = temp.reg;
+ inst->src[i].reg_offset = temp.reg_offset;
+ inst->src[i].reladdr = NULL;
+ }
}
- }
+ } while (nested_reladdr);
/* Now there are no accesses of the UNIFORM file with a reladdr, so
* no need to track them as larger-than-vec4 objects. This will be
*reg = temp;
}
+/**
+ * Resolve the result of a Gen4-5 CMP instruction to a proper boolean.
+ *
+ * CMP on Gen4-5 only sets the LSB of the result; the rest are undefined.
+ * If we need a proper boolean value, we have to fix it up to be 0 or ~0.
+ */
+void
+vec4_visitor::resolve_bool_comparison(ir_rvalue *rvalue, src_reg *reg)
+{
+ assert(brw->gen <= 5);
+
+ if (!rvalue->type->is_boolean())
+ return;
+
+ src_reg and_result = src_reg(this, rvalue->type);
+ src_reg neg_result = src_reg(this, rvalue->type);
+ emit(AND(dst_reg(and_result), *reg, src_reg(1)));
+ emit(MOV(dst_reg(neg_result), negate(and_result)));
+ *reg = neg_result;
+}
+
vec4_visitor::vec4_visitor(struct brw_context *brw,
struct brw_vec4_compile *c,
struct gl_program *prog,
- const struct brw_vec4_prog_key *key,
- struct brw_vec4_prog_data *prog_data,
+ const struct brw_vue_prog_key *key,
+ struct brw_vue_prog_data *prog_data,
struct gl_shader_program *shader_prog,
gl_shader_stage stage,
void *mem_ctx,
- bool debug_flag,
bool no_spills,
shader_time_shader_type st_base,
shader_time_shader_type st_written,
fail_msg(NULL),
first_non_payload_grf(0),
need_all_constants_in_pull_buffer(false),
- debug_flag(debug_flag),
no_spills(no_spills),
st_base(st_base),
st_written(st_written),
this->virtual_grf_start = NULL;
this->virtual_grf_end = NULL;
- this->virtual_grf_sizes = NULL;
- this->virtual_grf_count = 0;
- this->virtual_grf_reg_map = NULL;
- this->virtual_grf_reg_count = 0;
- this->virtual_grf_array_size = 0;
- this->live_intervals_valid = false;
+ this->live_intervals = NULL;
this->max_grf = brw->gen >= 7 ? GEN7_MRF_HACK_START : BRW_MAX_GRF;
*/
this->uniform_array_size = 1;
if (prog_data) {
- this->uniform_array_size = MAX2(stage_prog_data->nr_params, 1);
+ this->uniform_array_size =
+ MAX2(DIV_ROUND_UP(stage_prog_data->nr_params, 4), 1);
}
this->uniform_size = rzalloc_array(mem_ctx, int, this->uniform_array_size);
va_start(va, format);
msg = ralloc_vasprintf(mem_ctx, format, va);
va_end(va);
- msg = ralloc_asprintf(mem_ctx, "vec4 compile failed: %s\n", msg);
+ msg = ralloc_asprintf(mem_ctx, "%s compile failed: %s\n", stage_abbrev, msg);
this->fail_msg = msg;
- if (debug_flag) {
+ if (debug_enabled) {
fprintf(stderr, "%s", msg);
}
}
projects / mesa.git / blobdiff