* reciprocal. By breaking the operation down, constant reciprocals
* can get constant folded.
*
- * FDIV_TO_MUL_RCP only lowers single-precision floating point division;
+ * FDIV_TO_MUL_RCP lowers single-precision and half-precision
+ * floating point division;
* DDIV_TO_MUL_RCP only lowers double-precision floating point division.
* DIV_TO_MUL_RCP is a convenience macro that sets both flags.
* INT_DIV_TO_MUL_RCP handles the integer case, converting to and from floating
#include "ir.h"
#include "ir_builder.h"
#include "ir_optimization.h"
+#include "util/half_float.h"
using namespace ir_builder;
void find_msb_to_float_cast(ir_expression *ir);
void imul_high_to_mul(ir_expression *ir);
void sqrt_to_abs_sqrt(ir_expression *ir);
+ void mul64_to_mul_and_mul_high(ir_expression *ir);
ir_expression *_carry(operand a, operand b);
+
+ static ir_constant *_imm_fp(void *mem_ctx,
+ const glsl_type *type,
+ double f,
+ unsigned vector_elements=1);
};
} /* anonymous namespace */
void
lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
{
- assert(ir->operands[1]->type->is_float() || ir->operands[1]->type->is_double());
+ assert(ir->operands[1]->type->is_float_16_32_64());
/* New expression for the 1.0 / op1 */
ir_rvalue *expr;
void
lower_instructions_visitor::int_div_to_mul_rcp(ir_expression *ir)
{
- assert(ir->operands[1]->type->is_integer());
+ assert(ir->operands[1]->type->is_integer_32());
/* Be careful with integer division -- we need to do it as a
* float and re-truncate, since rcp(n > 1) of an integer would
void
lower_instructions_visitor::exp_to_exp2(ir_expression *ir)
{
- ir_constant *log2_e = new(ir) ir_constant(float(M_LOG2E));
+ ir_constant *log2_e = _imm_fp(ir, ir->type, M_LOG2E);
ir->operation = ir_unop_exp2;
ir->init_num_operands();
ir->init_num_operands();
ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
ir->operands[0], NULL);
- ir->operands[1] = new(ir) ir_constant(float(1.0 / M_LOG2E));
+ ir->operands[1] = _imm_fp(ir, ir->operands[0]->type, 1.0 / M_LOG2E);
this->progress = true;
}
ir_assignment *const assign_x =
new(ir) ir_assignment(new(ir) ir_dereference_variable(x),
- ir->operands[0], NULL);
+ ir->operands[0]);
ir_assignment *const assign_y =
new(ir) ir_assignment(new(ir) ir_dereference_variable(y),
- ir->operands[1], NULL);
+ ir->operands[1]);
this->base_ir->insert_before(assign_x);
this->base_ir->insert_before(assign_y);
/* Don't generate new IR that would need to be lowered in an additional
* pass.
*/
- if ((lowering(FDIV_TO_MUL_RCP) && ir->type->is_float()) ||
+ if ((lowering(FDIV_TO_MUL_RCP) && ir->type->is_float_16_32()) ||
(lowering(DDIV_TO_MUL_RCP) && ir->type->is_double()))
div_to_mul_rcp(div_expr);
* into
*
* extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
- * resulting_biased_exp = extracted_biased_exp + exp;
+ * resulting_biased_exp = min(extracted_biased_exp + exp, 255);
+ *
+ * if (extracted_biased_exp >= 255)
+ * return x; // +/-inf, NaN
+ *
+ * sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
*
- * if (resulting_biased_exp < 1 || x == 0.0f) {
- * return copysign(0.0, x);
+ * if (min(resulting_biased_exp, extracted_biased_exp) < 1)
+ * resulting_biased_exp = 0;
+ * if (resulting_biased_exp >= 255 ||
+ * min(resulting_biased_exp, extracted_biased_exp) < 1) {
+ * sign_mantissa &= sign_mask;
* }
*
- * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
+ * return bitcast_u2f(sign_mantissa |
* lshift(i2u(resulting_biased_exp), exp_shift));
*
* which we can't actually implement as such, since the GLSL IR doesn't
* using conditional-select:
*
* extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
- * resulting_biased_exp = extracted_biased_exp + exp;
+ * resulting_biased_exp = min(extracted_biased_exp + exp, 255);
*
- * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
- * gequal(resulting_biased_exp, 1);
- * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
- * resulting_biased_exp = csel(is_not_zero_or_underflow,
- * resulting_biased_exp, 0);
+ * sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
*
- * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
- * lshift(i2u(resulting_biased_exp), exp_shift));
+ * flush_to_zero = lequal(min(resulting_biased_exp, extracted_biased_exp), 0);
+ * resulting_biased_exp = csel(flush_to_zero, 0, resulting_biased_exp)
+ * zero_mantissa = logic_or(flush_to_zero,
+ * gequal(resulting_biased_exp, 255));
+ * sign_mantissa = csel(zero_mantissa, sign_mantissa & sign_mask, sign_mantissa);
+ *
+ * result = sign_mantissa |
+ * lshift(i2u(resulting_biased_exp), exp_shift));
+ *
+ * return csel(extracted_biased_exp >= 255, x, bitcast_u2f(result));
+ *
+ * The definition of ldexp in the GLSL spec says:
+ *
+ * "If this product is too large to be represented in the
+ * floating-point type, the result is undefined."
+ *
+ * However, the definition of ldexp in the GLSL ES spec does not contain
+ * this sentence, so we do need to handle overflow correctly.
+ *
+ * There is additional language limiting the defined range of exp, but this
+ * is merely to allow implementations that store 2^exp in a temporary
+ * variable.
*/
const unsigned vec_elem = ir->type->vector_elements;
/* Types */
const glsl_type *ivec = glsl_type::get_instance(GLSL_TYPE_INT, vec_elem, 1);
+ const glsl_type *uvec = glsl_type::get_instance(GLSL_TYPE_UINT, vec_elem, 1);
const glsl_type *bvec = glsl_type::get_instance(GLSL_TYPE_BOOL, vec_elem, 1);
- /* Constants */
- ir_constant *zeroi = ir_constant::zero(ir, ivec);
-
- ir_constant *sign_mask = new(ir) ir_constant(0x80000000u, vec_elem);
-
- ir_constant *exp_shift = new(ir) ir_constant(23, vec_elem);
-
/* Temporary variables */
ir_variable *x = new(ir) ir_variable(ir->type, "x", ir_var_temporary);
ir_variable *exp = new(ir) ir_variable(ivec, "exp", ir_var_temporary);
-
- ir_variable *zero_sign_x = new(ir) ir_variable(ir->type, "zero_sign_x",
- ir_var_temporary);
+ ir_variable *result = new(ir) ir_variable(uvec, "result", ir_var_temporary);
ir_variable *extracted_biased_exp =
new(ir) ir_variable(ivec, "extracted_biased_exp", ir_var_temporary);
ir_variable *resulting_biased_exp =
new(ir) ir_variable(ivec, "resulting_biased_exp", ir_var_temporary);
- ir_variable *is_not_zero_or_underflow =
- new(ir) ir_variable(bvec, "is_not_zero_or_underflow", ir_var_temporary);
+ ir_variable *sign_mantissa =
+ new(ir) ir_variable(uvec, "sign_mantissa", ir_var_temporary);
+
+ ir_variable *flush_to_zero =
+ new(ir) ir_variable(bvec, "flush_to_zero", ir_var_temporary);
+ ir_variable *zero_mantissa =
+ new(ir) ir_variable(bvec, "zero_mantissa", ir_var_temporary);
ir_instruction &i = *base_ir;
/* Extract the biased exponent from <x>. */
i.insert_before(extracted_biased_exp);
i.insert_before(assign(extracted_biased_exp,
- rshift(bitcast_f2i(abs(x)), exp_shift)));
+ rshift(bitcast_f2i(abs(x)),
+ new(ir) ir_constant(23, vec_elem))));
+ /* The definition of ldexp in the GLSL 4.60 spec says:
+ *
+ * "If exp is greater than +128 (single-precision) or +1024
+ * (double-precision), the value returned is undefined. If exp is less
+ * than -126 (single-precision) or -1022 (double-precision), the value
+ * returned may be flushed to zero."
+ *
+ * So we do not have to guard against the possibility of addition overflow,
+ * which could happen when exp is close to INT_MAX. Addition underflow
+ * cannot happen (the worst case is 0 + (-INT_MAX)).
+ */
i.insert_before(resulting_biased_exp);
i.insert_before(assign(resulting_biased_exp,
- add(extracted_biased_exp, exp)));
-
- /* Test if result is ±0.0, subnormal, or underflow by checking if the
- * resulting biased exponent would be less than 0x1. If so, the result is
- * 0.0 with the sign of x. (Actually, invert the conditions so that
- * immediate values are the second arguments, which is better for i965)
- */
- i.insert_before(zero_sign_x);
- i.insert_before(assign(zero_sign_x,
- bitcast_u2f(bit_and(bitcast_f2u(x), sign_mask))));
+ min2(add(extracted_biased_exp, exp),
+ new(ir) ir_constant(255, vec_elem))));
- i.insert_before(is_not_zero_or_underflow);
- i.insert_before(assign(is_not_zero_or_underflow,
- logic_and(nequal(x, new(ir) ir_constant(0.0f, vec_elem)),
- gequal(resulting_biased_exp,
- new(ir) ir_constant(0x1, vec_elem)))));
- i.insert_before(assign(x, csel(is_not_zero_or_underflow,
- x, zero_sign_x)));
- i.insert_before(assign(resulting_biased_exp,
- csel(is_not_zero_or_underflow,
- resulting_biased_exp, zeroi)));
+ i.insert_before(sign_mantissa);
+ i.insert_before(assign(sign_mantissa,
+ bit_and(bitcast_f2u(x),
+ new(ir) ir_constant(0x807fffffu, vec_elem))));
- /* We could test for overflows by checking if the resulting biased exponent
- * would be greater than 0xFE. Turns out we don't need to because the GLSL
- * spec says:
+ /* We flush to zero if the original or resulting biased exponent is 0,
+ * indicating a +/-0.0 or subnormal input or output.
*
- * "If this product is too large to be represented in the
- * floating-point type, the result is undefined."
+ * The mantissa is set to 0 if the resulting biased exponent is 255, since
+ * an overflow should produce a +/-inf result.
+ *
+ * Note that NaN inputs are handled separately.
*/
-
- ir_constant *exp_shift_clone = exp_shift->clone(ir, NULL);
+ i.insert_before(flush_to_zero);
+ i.insert_before(assign(flush_to_zero,
+ lequal(min2(resulting_biased_exp,
+ extracted_biased_exp),
+ ir_constant::zero(ir, ivec))));
+ i.insert_before(assign(resulting_biased_exp,
+ csel(flush_to_zero,
+ ir_constant::zero(ir, ivec),
+ resulting_biased_exp)));
+
+ i.insert_before(zero_mantissa);
+ i.insert_before(assign(zero_mantissa,
+ logic_or(flush_to_zero,
+ equal(resulting_biased_exp,
+ new(ir) ir_constant(255, vec_elem)))));
+ i.insert_before(assign(sign_mantissa,
+ csel(zero_mantissa,
+ bit_and(sign_mantissa,
+ new(ir) ir_constant(0x80000000u, vec_elem)),
+ sign_mantissa)));
/* Don't generate new IR that would need to be lowered in an additional
* pass.
*/
+ i.insert_before(result);
if (!lowering(INSERT_TO_SHIFTS)) {
- ir_constant *exp_width = new(ir) ir_constant(8, vec_elem);
- ir->operation = ir_unop_bitcast_i2f;
- ir->init_num_operands();
- ir->operands[0] = bitfield_insert(bitcast_f2i(x), resulting_biased_exp,
- exp_shift_clone, exp_width);
- ir->operands[1] = NULL;
+ i.insert_before(assign(result,
+ bitfield_insert(sign_mantissa,
+ i2u(resulting_biased_exp),
+ new(ir) ir_constant(23u, vec_elem),
+ new(ir) ir_constant(8u, vec_elem))));
} else {
- ir_constant *sign_mantissa_mask = new(ir) ir_constant(0x807fffffu, vec_elem);
- ir->operation = ir_unop_bitcast_u2f;
- ir->init_num_operands();
- ir->operands[0] = bit_or(bit_and(bitcast_f2u(x), sign_mantissa_mask),
- lshift(i2u(resulting_biased_exp), exp_shift_clone));
- ir->operands[1] = NULL;
+ i.insert_before(assign(result,
+ bit_or(sign_mantissa,
+ lshift(i2u(resulting_biased_exp),
+ new(ir) ir_constant(23, vec_elem)))));
}
+ ir->operation = ir_triop_csel;
+ ir->init_num_operands();
+ ir->operands[0] = gequal(extracted_biased_exp,
+ new(ir) ir_constant(255, vec_elem));
+ ir->operands[1] = new(ir) ir_dereference_variable(x);
+ ir->operands[2] = bitcast_u2f(result);
+
this->progress = true;
}
ir->operation = ir_binop_min;
ir->init_num_operands();
+
+ ir_constant *zero = _imm_fp(ir, ir->operands[0]->type, 0.0);
ir->operands[0] = new(ir) ir_expression(ir_binop_max, ir->operands[0]->type,
- ir->operands[0],
- new(ir) ir_constant(0.0f));
- ir->operands[1] = new(ir) ir_constant(1.0f);
+ ir->operands[0], zero);
+ ir->operands[1] = _imm_fp(ir, ir->operands[0]->type, 1.0);
this->progress = true;
}
return carry(a, b);
}
+ir_constant *
+lower_instructions_visitor::_imm_fp(void *mem_ctx,
+ const glsl_type *type,
+ double f,
+ unsigned vector_elements)
+{
+ switch (type->base_type) {
+ case GLSL_TYPE_FLOAT:
+ return new(mem_ctx) ir_constant((float) f, vector_elements);
+ case GLSL_TYPE_DOUBLE:
+ return new(mem_ctx) ir_constant((double) f, vector_elements);
+ case GLSL_TYPE_FLOAT16:
+ return new(mem_ctx) ir_constant(float16_t(f), vector_elements);
+ default:
+ assert(!"unknown float type for immediate");
+ return NULL;
+ }
+}
+
void
lower_instructions_visitor::imul_high_to_mul(ir_expression *ir)
{
this->progress = true;
}
+void
+lower_instructions_visitor::mul64_to_mul_and_mul_high(ir_expression *ir)
+{
+ /* Lower 32x32-> 64 to
+ * msb = imul_high(x_lo, y_lo)
+ * lsb = mul(x_lo, y_lo)
+ */
+ const unsigned elements = ir->operands[0]->type->vector_elements;
+
+ const ir_expression_operation operation =
+ ir->type->base_type == GLSL_TYPE_UINT64 ? ir_unop_pack_uint_2x32
+ : ir_unop_pack_int_2x32;
+
+ const glsl_type *var_type = ir->type->base_type == GLSL_TYPE_UINT64
+ ? glsl_type::uvec(elements)
+ : glsl_type::ivec(elements);
+
+ const glsl_type *ret_type = ir->type->base_type == GLSL_TYPE_UINT64
+ ? glsl_type::uvec2_type
+ : glsl_type::ivec2_type;
+
+ ir_instruction &i = *base_ir;
+
+ ir_variable *msb =
+ new(ir) ir_variable(var_type, "msb", ir_var_temporary);
+ ir_variable *lsb =
+ new(ir) ir_variable(var_type, "lsb", ir_var_temporary);
+ ir_variable *x =
+ new(ir) ir_variable(var_type, "x", ir_var_temporary);
+ ir_variable *y =
+ new(ir) ir_variable(var_type, "y", ir_var_temporary);
+
+ i.insert_before(x);
+ i.insert_before(assign(x, ir->operands[0]));
+ i.insert_before(y);
+ i.insert_before(assign(y, ir->operands[1]));
+ i.insert_before(msb);
+ i.insert_before(lsb);
+
+ i.insert_before(assign(msb, imul_high(x, y)));
+ i.insert_before(assign(lsb, mul(x, y)));
+
+ ir_rvalue *result[4] = {NULL};
+ for (unsigned elem = 0; elem < elements; elem++) {
+ ir_rvalue *val = new(ir) ir_expression(ir_quadop_vector, ret_type,
+ swizzle(lsb, elem, 1),
+ swizzle(msb, elem, 1), NULL, NULL);
+ result[elem] = expr(operation, val);
+ }
+
+ ir->operation = ir_quadop_vector;
+ ir->init_num_operands();
+ ir->operands[0] = result[0];
+ ir->operands[1] = result[1];
+ ir->operands[2] = result[2];
+ ir->operands[3] = result[3];
+
+ this->progress = true;
+}
+
ir_visitor_status
lower_instructions_visitor::visit_leave(ir_expression *ir)
{
break;
case ir_binop_div:
- if (ir->operands[1]->type->is_integer() && lowering(INT_DIV_TO_MUL_RCP))
+ if (ir->operands[1]->type->is_integer_32() && lowering(INT_DIV_TO_MUL_RCP))
int_div_to_mul_rcp(ir);
- else if ((ir->operands[1]->type->is_float() && lowering(FDIV_TO_MUL_RCP)) ||
+ else if ((ir->operands[1]->type->is_float_16_32() && lowering(FDIV_TO_MUL_RCP)) ||
(ir->operands[1]->type->is_double() && lowering(DDIV_TO_MUL_RCP)))
div_to_mul_rcp(ir);
break;
break;
case ir_binop_mod:
- if (lowering(MOD_TO_FLOOR) && (ir->type->is_float() || ir->type->is_double()))
+ if (lowering(MOD_TO_FLOOR) && ir->type->is_float_16_32_64())
mod_to_floor(ir);
break;
imul_high_to_mul(ir);
break;
+ case ir_binop_mul:
+ if (lowering(MUL64_TO_MUL_AND_MUL_HIGH) &&
+ (ir->type->base_type == GLSL_TYPE_INT64 ||
+ ir->type->base_type == GLSL_TYPE_UINT64) &&
+ (ir->operands[0]->type->base_type == GLSL_TYPE_INT ||
+ ir->operands[1]->type->base_type == GLSL_TYPE_UINT))
+ mul64_to_mul_and_mul_high(ir);
+ break;
+
case ir_unop_rsq:
case ir_unop_sqrt:
if (lowering(SQRT_TO_ABS_SQRT))
projects / mesa.git / blobdiff