From b6235651b9c3460cbd63adb3f692c4ed5cecb719 Mon Sep 17 00:00:00 2001 From: =?utf8?q?Timur=20Krist=C3=B3f?= Date: Thu, 26 Sep 2019 17:46:05 +0200 Subject: [PATCH] aco: Support GFX10 SMEM in aco_assembler. MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Signed-off-by: Timur Kristóf Reviewed-by: Daniel Schürmann --- src/amd/compiler/aco_assembler.cpp | 73 ++++++++++++++++++++++++------ 1 file changed, 60 insertions(+), 13 deletions(-) diff --git a/src/amd/compiler/aco_assembler.cpp b/src/amd/compiler/aco_assembler.cpp index 9423a26997d..a9fefb832a9 100644 --- a/src/amd/compiler/aco_assembler.cpp +++ b/src/amd/compiler/aco_assembler.cpp @@ -122,23 +122,70 @@ void emit_instruction(asm_context& ctx, std::vector& out, Instruction* } case Format::SMEM: { SMEM_instruction* smem = static_cast(instr); - uint32_t encoding = (0b110000 << 26); - encoding |= opcode << 18; - if (instr->operands.size() >= 2) - encoding |= instr->operands[1].isConstant() ? 1 << 17 : 0; bool soe = instr->operands.size() >= (!instr->definitions.empty() ? 3 : 4); - assert(!soe || ctx.chip_class >= GFX9); - encoding |= soe ? 1 << 14 : 0; + bool is_load = !instr->definitions.empty(); + + uint32_t encoding = 0; + + if (ctx.chip_class <= GFX9) { + encoding = (0b110000 << 26); + assert(!smem->dlc); /* Device-level coherent is not supported on GFX9 and lower */ + encoding |= smem->nv ? 1 << 15 : 0; + } else { + encoding = (0b111101 << 26); + assert(!smem->nv); /* Non-volatile is not supported on GFX10 */ + encoding |= smem->dlc ? 1 << 14 : 0; + } + + encoding |= opcode << 18; encoding |= smem->glc ? 1 << 16 : 0; - if (!instr->definitions.empty() || instr->operands.size() >= 3) - encoding |= (!instr->definitions.empty() ? instr->definitions[0].physReg() : instr->operands[2].physReg().reg) << 6; - if (instr->operands.size() >= 1) - encoding |= instr->operands[0].physReg() >> 1; + + if (ctx.chip_class <= GFX9) { + if (instr->operands.size() >= 2) + encoding |= instr->operands[1].isConstant() ? 1 << 17 : 0; /* IMM - immediate enable */ + } + if (ctx.chip_class == GFX9) { + encoding |= soe ? 1 << 14 : 0; + } + + if (is_load || instr->operands.size() >= 3) { /* SDATA */ + encoding |= (is_load ? instr->definitions[0].physReg().reg : instr->operands[2].physReg().reg) << 6; + } + if (instr->operands.size() >= 1) { /* SBASE */ + encoding |= instr->operands[0].physReg().reg >> 1; + } + out.push_back(encoding); encoding = 0; - if (instr->operands.size() >= 2) - encoding |= instr->operands[1].isConstant() ? instr->operands[1].constantValue() : instr->operands[1].physReg().reg; - encoding |= soe ? instr->operands.back().physReg() << 25 : 0; + + int32_t offset = 0; + uint32_t soffset = ctx.chip_class >= GFX10 + ? sgpr_null /* On GFX10 this is disabled by specifying SGPR_NULL */ + : 0; /* On GFX9, it is disabled by the SOE bit (and it's not present on GFX8 and below) */ + if (instr->operands.size() >= 2) { + const Operand &op_off1 = instr->operands[1]; + if (ctx.chip_class <= GFX9) { + offset = op_off1.isConstant() ? op_off1.constantValue() : op_off1.physReg(); + } else { + /* GFX10 only supports constants in OFFSET, so put the operand in SOFFSET if it's an SGPR */ + if (op_off1.isConstant()) { + offset = op_off1.constantValue(); + } else { + soffset = op_off1.physReg(); + assert(!soe); /* There is no place to put the other SGPR offset, if any */ + } + } + + if (soe) { + const Operand &op_off2 = instr->operands.back(); + assert(ctx.chip_class >= GFX9); /* GFX8 and below don't support specifying a constant and an SGPR at the same time */ + assert(!op_off2.isConstant()); + soffset = op_off2.physReg(); + } + } + encoding |= offset; + encoding |= soffset << 25; + out.push_back(encoding); return; } -- 2.30.2