/* * Copyright © 2010 Intel Corporation * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER * DEALINGS IN THE SOFTWARE. */ /** * \file lower_instructions.cpp * * Many GPUs lack native instructions for certain expression operations, and * must replace them with some other expression tree. This pass lowers some * of the most common cases, allowing the lowering code to be implemented once * rather than in each driver backend. */ #include "program/prog_instruction.h" /* for swizzle */ #include "compiler/glsl_types.h" #include "ir.h" #include "ir_builder.h" #include "ir_optimization.h" #include "util/half_float.h" #include /* Operations for lower_instructions() */ #define FIND_LSB_TO_FLOAT_CAST 0x20000 #define FIND_MSB_TO_FLOAT_CAST 0x40000 #define IMUL_HIGH_TO_MUL 0x80000 using namespace ir_builder; namespace { class lower_instructions_visitor : public ir_hierarchical_visitor { public: lower_instructions_visitor(unsigned lower) : progress(false), lower(lower) { } ir_visitor_status visit_leave(ir_expression *); bool progress; private: unsigned lower; /** Bitfield of which operations to lower */ void double_dot_to_fma(ir_expression *); void double_lrp(ir_expression *); void find_lsb_to_float_cast(ir_expression *ir); void find_msb_to_float_cast(ir_expression *ir); void imul_high_to_mul(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 */ /** * Determine if a particular type of lowering should occur */ #define lowering(x) (this->lower & x) bool lower_instructions(exec_list *instructions,bool have_gpu_shader5) { unsigned what_to_lower = /* Assume that if ARB_gpu_shader5 is not supported then all of the * extended integer functions need lowering. It may be necessary to add * some caps for individual instructions. */ (!have_gpu_shader5 ? FIND_LSB_TO_FLOAT_CAST | FIND_MSB_TO_FLOAT_CAST | IMUL_HIGH_TO_MUL : 0); lower_instructions_visitor v(what_to_lower); visit_list_elements(&v, instructions); return v.progress; } void lower_instructions_visitor::double_dot_to_fma(ir_expression *ir) { ir_variable *temp = new(ir) ir_variable(glsl_get_base_glsl_type(ir->operands[0]->type), "dot_res", ir_var_temporary); this->base_ir->insert_before(temp); int nc = glsl_get_components(ir->operands[0]->type); for (int i = nc - 1; i >= 1; i--) { ir_assignment *assig; if (i == (nc - 1)) { assig = assign(temp, mul(swizzle(ir->operands[0]->clone(ir, NULL), i, 1), swizzle(ir->operands[1]->clone(ir, NULL), i, 1))); } else { assig = assign(temp, fma(swizzle(ir->operands[0]->clone(ir, NULL), i, 1), swizzle(ir->operands[1]->clone(ir, NULL), i, 1), temp)); } this->base_ir->insert_before(assig); } ir->operation = ir_triop_fma; ir->init_num_operands(); ir->operands[0] = swizzle(ir->operands[0], 0, 1); ir->operands[1] = swizzle(ir->operands[1], 0, 1); ir->operands[2] = new(ir) ir_dereference_variable(temp); this->progress = true; } void lower_instructions_visitor::double_lrp(ir_expression *ir) { int swizval; ir_rvalue *op0 = ir->operands[0], *op2 = ir->operands[2]; ir_constant *one = new(ir) ir_constant(1.0, op2->type->vector_elements); switch (op2->type->vector_elements) { case 1: swizval = SWIZZLE_XXXX; break; default: assert(op0->type->vector_elements == op2->type->vector_elements); swizval = SWIZZLE_XYZW; break; } ir->operation = ir_triop_fma; ir->init_num_operands(); ir->operands[0] = swizzle(op2, swizval, op0->type->vector_elements); ir->operands[2] = mul(sub(one, op2->clone(ir, NULL)), op0); this->progress = true; } void lower_instructions_visitor::find_lsb_to_float_cast(ir_expression *ir) { /* For more details, see: * * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast */ const unsigned elements = ir->operands[0]->type->vector_elements; ir_constant *c0 = new(ir) ir_constant(unsigned(0), elements); ir_constant *cminus1 = new(ir) ir_constant(int(-1), elements); ir_constant *c23 = new(ir) ir_constant(int(23), elements); ir_constant *c7F = new(ir) ir_constant(int(0x7F), elements); ir_variable *temp = new(ir) ir_variable(glsl_ivec_type(elements), "temp", ir_var_temporary); ir_variable *lsb_only = new(ir) ir_variable(glsl_uvec_type(elements), "lsb_only", ir_var_temporary); ir_variable *as_float = new(ir) ir_variable(glsl_vec_type(elements), "as_float", ir_var_temporary); ir_variable *lsb = new(ir) ir_variable(glsl_ivec_type(elements), "lsb", ir_var_temporary); ir_instruction &i = *base_ir; i.insert_before(temp); if (ir->operands[0]->type->base_type == GLSL_TYPE_INT) { i.insert_before(assign(temp, ir->operands[0])); } else { assert(ir->operands[0]->type->base_type == GLSL_TYPE_UINT); i.insert_before(assign(temp, u2i(ir->operands[0]))); } /* The int-to-float conversion is lossless because (value & -value) is * either a power of two or zero. We don't use the result in the zero * case. The uint() cast is necessary so that 0x80000000 does not * generate a negative value. * * uint lsb_only = uint(value & -value); * float as_float = float(lsb_only); */ i.insert_before(lsb_only); i.insert_before(assign(lsb_only, i2u(bit_and(temp, neg(temp))))); i.insert_before(as_float); i.insert_before(assign(as_float, u2f(lsb_only))); /* This is basically an open-coded frexp. Implementations that have a * native frexp instruction would be better served by that. This is * optimized versus a full-featured open-coded implementation in two ways: * * - We don't care about a correct result from subnormal numbers (including * 0.0), so the raw exponent can always be safely unbiased. * * - The value cannot be negative, so it does not need to be masked off to * extract the exponent. * * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f; */ i.insert_before(lsb); i.insert_before(assign(lsb, sub(rshift(bitcast_f2i(as_float), c23), c7F))); /* Use lsb_only in the comparison instead of temp so that the & (far above) * can possibly generate the result without an explicit comparison. * * (lsb_only == 0) ? -1 : lsb; * * Since our input values are all integers, the unbiased exponent must not * be negative. It will only be negative (-0x7f, in fact) if lsb_only is * 0. Instead of using (lsb_only == 0), we could use (lsb >= 0). Which is * better is likely GPU dependent. Either way, the difference should be * small. */ ir->operation = ir_triop_csel; ir->init_num_operands(); ir->operands[0] = equal(lsb_only, c0); ir->operands[1] = cminus1; ir->operands[2] = new(ir) ir_dereference_variable(lsb); this->progress = true; } void lower_instructions_visitor::find_msb_to_float_cast(ir_expression *ir) { /* For more details, see: * * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast */ const unsigned elements = ir->operands[0]->type->vector_elements; ir_constant *c0 = new(ir) ir_constant(int(0), elements); ir_constant *cminus1 = new(ir) ir_constant(int(-1), elements); ir_constant *c23 = new(ir) ir_constant(int(23), elements); ir_constant *c7F = new(ir) ir_constant(int(0x7F), elements); ir_constant *c000000FF = new(ir) ir_constant(0x000000FFu, elements); ir_constant *cFFFFFF00 = new(ir) ir_constant(0xFFFFFF00u, elements); ir_variable *temp = new(ir) ir_variable(glsl_uvec_type(elements), "temp", ir_var_temporary); ir_variable *as_float = new(ir) ir_variable(glsl_vec_type(elements), "as_float", ir_var_temporary); ir_variable *msb = new(ir) ir_variable(glsl_ivec_type(elements), "msb", ir_var_temporary); ir_instruction &i = *base_ir; i.insert_before(temp); if (ir->operands[0]->type->base_type == GLSL_TYPE_UINT) { i.insert_before(assign(temp, ir->operands[0])); } else { assert(ir->operands[0]->type->base_type == GLSL_TYPE_INT); /* findMSB(uint(abs(some_int))) almost always does the right thing. * There are two problem values: * * * 0x80000000. Since abs(0x80000000) == 0x80000000, findMSB returns * 31. However, findMSB(int(0x80000000)) == 30. * * * 0xffffffff. Since abs(0xffffffff) == 1, findMSB returns * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says: * * For a value of zero or negative one, -1 will be returned. * * For all negative number cases, including 0x80000000 and 0xffffffff, * the correct value is obtained from findMSB if instead of negating the * (already negative) value the logical-not is used. A conditonal * logical-not can be achieved in two instructions. */ ir_variable *as_int = new(ir) ir_variable(glsl_ivec_type(elements), "as_int", ir_var_temporary); ir_constant *c31 = new(ir) ir_constant(int(31), elements); i.insert_before(as_int); i.insert_before(assign(as_int, ir->operands[0])); i.insert_before(assign(temp, i2u(expr(ir_binop_bit_xor, as_int, rshift(as_int, c31))))); } /* The int-to-float conversion is lossless because bits are conditionally * masked off the bottom of temp to ensure the value has at most 24 bits of * data or is zero. We don't use the result in the zero case. The uint() * cast is necessary so that 0x80000000 does not generate a negative value. * * float as_float = float(temp > 255 ? temp & ~255 : temp); */ i.insert_before(as_float); i.insert_before(assign(as_float, u2f(csel(greater(temp, c000000FF), bit_and(temp, cFFFFFF00), temp)))); /* This is basically an open-coded frexp. Implementations that have a * native frexp instruction would be better served by that. This is * optimized versus a full-featured open-coded implementation in two ways: * * - We don't care about a correct result from subnormal numbers (including * 0.0), so the raw exponent can always be safely unbiased. * * - The value cannot be negative, so it does not need to be masked off to * extract the exponent. * * int msb = (floatBitsToInt(as_float) >> 23) - 0x7f; */ i.insert_before(msb); i.insert_before(assign(msb, sub(rshift(bitcast_f2i(as_float), c23), c7F))); /* Use msb in the comparison instead of temp so that the subtract can * possibly generate the result without an explicit comparison. * * (msb < 0) ? -1 : msb; * * Since our input values are all integers, the unbiased exponent must not * be negative. It will only be negative (-0x7f, in fact) if temp is 0. */ ir->operation = ir_triop_csel; ir->init_num_operands(); ir->operands[0] = less(msb, c0); ir->operands[1] = cminus1; ir->operands[2] = new(ir) ir_dereference_variable(msb); this->progress = true; } ir_expression * lower_instructions_visitor::_carry(operand a, operand b) { return i2u(b2i(less(add(a, b), a.val->clone(ralloc_parent(a.val), NULL)))); } void lower_instructions_visitor::imul_high_to_mul(ir_expression *ir) { /* ABCD * * EFGH * ====== * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32 * * In GLSL, (a * b) becomes * * uint m1 = (a & 0x0000ffffu) * (b & 0x0000ffffu); * uint m2 = (a & 0x0000ffffu) * (b >> 16); * uint m3 = (a >> 16) * (b & 0x0000ffffu); * uint m4 = (a >> 16) * (b >> 16); * * uint c1; * uint c2; * uint lo_result; * uint hi_result; * * lo_result = uaddCarry(m1, m2 << 16, c1); * hi_result = m4 + c1; * lo_result = uaddCarry(lo_result, m3 << 16, c2); * hi_result = hi_result + c2; * hi_result = hi_result + (m2 >> 16) + (m3 >> 16); */ const unsigned elements = ir->operands[0]->type->vector_elements; ir_variable *src1 = new(ir) ir_variable(glsl_uvec_type(elements), "src1", ir_var_temporary); ir_variable *src1h = new(ir) ir_variable(glsl_uvec_type(elements), "src1h", ir_var_temporary); ir_variable *src1l = new(ir) ir_variable(glsl_uvec_type(elements), "src1l", ir_var_temporary); ir_variable *src2 = new(ir) ir_variable(glsl_uvec_type(elements), "src2", ir_var_temporary); ir_variable *src2h = new(ir) ir_variable(glsl_uvec_type(elements), "src2h", ir_var_temporary); ir_variable *src2l = new(ir) ir_variable(glsl_uvec_type(elements), "src2l", ir_var_temporary); ir_variable *t1 = new(ir) ir_variable(glsl_uvec_type(elements), "t1", ir_var_temporary); ir_variable *t2 = new(ir) ir_variable(glsl_uvec_type(elements), "t2", ir_var_temporary); ir_variable *lo = new(ir) ir_variable(glsl_uvec_type(elements), "lo", ir_var_temporary); ir_variable *hi = new(ir) ir_variable(glsl_uvec_type(elements), "hi", ir_var_temporary); ir_variable *different_signs = NULL; ir_constant *c0000FFFF = new(ir) ir_constant(0x0000FFFFu, elements); ir_constant *c16 = new(ir) ir_constant(16u, elements); ir_instruction &i = *base_ir; i.insert_before(src1); i.insert_before(src2); i.insert_before(src1h); i.insert_before(src2h); i.insert_before(src1l); i.insert_before(src2l); if (ir->operands[0]->type->base_type == GLSL_TYPE_UINT) { i.insert_before(assign(src1, ir->operands[0])); i.insert_before(assign(src2, ir->operands[1])); } else { assert(ir->operands[0]->type->base_type == GLSL_TYPE_INT); ir_variable *itmp1 = new(ir) ir_variable(glsl_ivec_type(elements), "itmp1", ir_var_temporary); ir_variable *itmp2 = new(ir) ir_variable(glsl_ivec_type(elements), "itmp2", ir_var_temporary); ir_constant *c0 = new(ir) ir_constant(int(0), elements); i.insert_before(itmp1); i.insert_before(itmp2); i.insert_before(assign(itmp1, ir->operands[0])); i.insert_before(assign(itmp2, ir->operands[1])); different_signs = new(ir) ir_variable(glsl_bvec_type(elements), "different_signs", ir_var_temporary); i.insert_before(different_signs); i.insert_before(assign(different_signs, expr(ir_binop_logic_xor, less(itmp1, c0), less(itmp2, c0->clone(ir, NULL))))); i.insert_before(assign(src1, i2u(abs(itmp1)))); i.insert_before(assign(src2, i2u(abs(itmp2)))); } i.insert_before(assign(src1l, bit_and(src1, c0000FFFF))); i.insert_before(assign(src2l, bit_and(src2, c0000FFFF->clone(ir, NULL)))); i.insert_before(assign(src1h, rshift(src1, c16))); i.insert_before(assign(src2h, rshift(src2, c16->clone(ir, NULL)))); i.insert_before(lo); i.insert_before(hi); i.insert_before(t1); i.insert_before(t2); i.insert_before(assign(lo, mul(src1l, src2l))); i.insert_before(assign(t1, mul(src1l, src2h))); i.insert_before(assign(t2, mul(src1h, src2l))); i.insert_before(assign(hi, mul(src1h, src2h))); i.insert_before(assign(hi, add(hi, _carry(lo, lshift(t1, c16->clone(ir, NULL)))))); i.insert_before(assign(lo, add(lo, lshift(t1, c16->clone(ir, NULL))))); i.insert_before(assign(hi, add(hi, _carry(lo, lshift(t2, c16->clone(ir, NULL)))))); i.insert_before(assign(lo, add(lo, lshift(t2, c16->clone(ir, NULL))))); if (different_signs == NULL) { assert(ir->operands[0]->type->base_type == GLSL_TYPE_UINT); ir->operation = ir_binop_add; ir->init_num_operands(); ir->operands[0] = add(hi, rshift(t1, c16->clone(ir, NULL))); ir->operands[1] = rshift(t2, c16->clone(ir, NULL)); } else { assert(ir->operands[0]->type->base_type == GLSL_TYPE_INT); i.insert_before(assign(hi, add(add(hi, rshift(t1, c16->clone(ir, NULL))), rshift(t2, c16->clone(ir, NULL))))); /* For channels where different_signs is set we have to perform a 64-bit * negation. This is *not* the same as just negating the high 32-bits. * Consider -3 * 2. The high 32-bits is 0, but the desired result is * -1, not -0! Recall -x == ~x + 1. */ ir_variable *neg_hi = new(ir) ir_variable(glsl_ivec_type(elements), "neg_hi", ir_var_temporary); ir_constant *c1 = new(ir) ir_constant(1u, elements); i.insert_before(neg_hi); i.insert_before(assign(neg_hi, add(bit_not(u2i(hi)), u2i(_carry(bit_not(lo), c1))))); ir->operation = ir_triop_csel; ir->init_num_operands(); ir->operands[0] = new(ir) ir_dereference_variable(different_signs); ir->operands[1] = new(ir) ir_dereference_variable(neg_hi); ir->operands[2] = u2i(hi); } } ir_visitor_status lower_instructions_visitor::visit_leave(ir_expression *ir) { switch (ir->operation) { case ir_binop_dot: if (glsl_type_is_double(ir->operands[0]->type)) double_dot_to_fma(ir); break; case ir_triop_lrp: if (glsl_type_is_double(ir->operands[0]->type)) double_lrp(ir); break; case ir_unop_find_lsb: if (lowering(FIND_LSB_TO_FLOAT_CAST)) find_lsb_to_float_cast(ir); break; case ir_unop_find_msb: if (lowering(FIND_MSB_TO_FLOAT_CAST)) find_msb_to_float_cast(ir); break; case ir_binop_imul_high: if (lowering(IMUL_HIGH_TO_MUL)) imul_high_to_mul(ir); break; default: return visit_continue; } return visit_continue; }