wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

merged fpu_port branch

Browse Source
This commit is contained in:
Blaise Tine
2020-07-31 17:13:22 -04:00
parent 4bdab8903e 836a735555
commit 31ee824862
508 changed files with 45407 additions and 41832 deletions
Download Patch File Download Diff File Expand all files Collapse all files

311
hw/rtl/VX_define.vh
View File

@@ -1,63 +1,17 @@
`ifndef VX_DEFINE
`define VX_DEFINE
`include "VX_platform.vh"
`include "VX_config.vh"
`include "VX_scope.vh"
`define QUEUE_FORCE_MLAB 1
///////////////////////////////////////////////////////////////////////////////
// `define SYNTHESIS 1
// `define ASIC 1
///////////////////////////////////////////////////////////////////////////////
`ifndef NDEBUG
`define DEBUG_BLOCK(x) /* verilator lint_off UNUSED */ \
x \
/* verilator lint_on UNUSED */
`else
`define DEBUG_BLOCK(x)
`endif
`define DEBUG_BEGIN /* verilator lint_off UNUSED */
`define DEBUG_END /* verilator lint_on UNUSED */
`define IGNORE_WARNINGS_BEGIN /* verilator lint_off UNUSED */ \
/* verilator lint_off PINCONNECTEMPTY */ \
/* verilator lint_off DECLFILENAME */
`define IGNORE_WARNINGS_END /* verilator lint_on UNUSED */ \
/* verilator lint_on PINCONNECTEMPTY */ \
/* verilator lint_on DECLFILENAME */
`define UNUSED_VAR(x) /* verilator lint_off UNUSED */ \
wire [$bits(x)-1:0] __``x``__ = x; \
/* verilator lint_on UNUSED */
`define UNUSED_PIN(x) /* verilator lint_off PINCONNECTEMPTY */ \
. x () \
/* verilator lint_on PINCONNECTEMPTY */
`define STRINGIFY(x) `"x`"
`define STATIC_ASSERT(cond, msg) \
generate \
if (!(cond)) $error(msg); \
endgenerate
`define CLOG2(x) $clog2(x)
`define FLOG2(x) ($clog2(x) - (((1 << $clog2(x)) > (x)) ? 1 : 0))
`define LOG2UP(x) (((x) > 1) ? $clog2(x) : 1)
`define ISPOW2(x) (((x) != 0) && (0 == ((x) & ((x) - 1))))
`define MIN(x, y) ((x < y) ? (x) : (y))
`define MAX(x, y) ((x > y) ? (x) : (y))
`define UP(x) (((x) > 0) ? x : 1)
///////////////////////////////////////////////////////////////////////////////
`define NW_BITS `LOG2UP(`NUM_WARPS)
`define NT_BITS `LOG2UP(`NUM_THREADS)
@@ -68,17 +22,29 @@
`define REQS_BITS `LOG2UP(NUM_REQUESTS)
`ifdef EXT_F_ENABLE
`define NUM_REGS 64
`else
`define NUM_REGS 32
`endif
`define NR_BITS `LOG2UP(`NUM_REGS)
`define CSR_ADDR_SIZE 12
`define CSR_ADDR_BITS 12
`define CSR_WIDTH 12
`define DIV_LATENCY 2
`define ISTAG_BITS `LOG2UP(`ISSUEQ_SIZE)
`define MUL_LATENCY 2
///////////////////////////////////////////////////////////////////////////////
`define LATENCY_IDIV 24
`define LATENCY_IMUL 2
`define LATENCY_FMULADD 2
`define LATENCY_FDIVSQRT 2
`define LATENCY_FCONV 2
`define LATENCY_FNONCOMP 1
///////////////////////////////////////////////////////////////////////////////
@@ -86,13 +52,22 @@
`define INST_AUIPC 7'b0010111
`define INST_JAL 7'b1101111
`define INST_JALR 7'b1100111
`define INST_B 7'b1100011
`define INST_L 7'b0000011
`define INST_S 7'b0100011
`define INST_I 7'b0010011
`define INST_R 7'b0110011
`define INST_F 7'b0001111
`define INST_SYS 7'b1110011
`define INST_B 7'b1100011 // branch instructions
`define INST_L 7'b0000011 // load instructions
`define INST_S 7'b0100011 // store instructions
`define INST_I 7'b0010011 // immediate instructions
`define INST_R 7'b0110011 // register instructions
`define INST_F 7'b0001111 // Fence instructions
`define INST_SYS 7'b1110011 // system instructions
`define INST_FL 7'b0000111 // float load instruction
`define INST_FS 7'b0100111 // float store instruction
`define INST_FMADD 7'b1000011
`define INST_FMSUB 7'b1000111
`define INST_FNMSUB 7'b1001011
`define INST_FNMADD 7'b1001111
`define INST_FCI 7'b1010011 // float common instructions
`define INST_GPU 7'b1101011
`define BYTEEN_SB 3'h0
@@ -101,6 +76,7 @@
`define BYTEEN_UB 3'h4
`define BYTEEN_UH 3'h5
`define BYTEEN_BITS 3
`define BYTEEN_TYPE(x) x[1:0]
`define BR_EQ 4'h0
`define BR_NE 4'h1
@@ -115,6 +91,7 @@
`define BR_MRET 4'hA
`define BR_SRET 4'hB
`define BR_DRET 4'hC
`define BR_NO 4'hF
`define BR_BITS 4
`define OP_BITS 5
@@ -150,18 +127,6 @@
`define BR_OP(x) x[`BR_BITS-1:0]
`define IS_BR_OP(x) x[4]
`define MUL_MUL 3'h0
`define MUL_MULH 3'h1
`define MUL_MULHSU 3'h2
`define MUL_MULHU 3'h3
`define MUL_DIV 3'h4
`define MUL_DIVU 3'h5
`define MUL_REM 3'h6
`define MUL_REMU 3'h7
`define MUL_BITS 3
`define MUL_OP(x) x[`MUL_BITS-1:0]
`define IS_DIV_OP(x) x[2]
`define LSU_LB {1'b0, `BYTEEN_SB}
`define LSU_LH {1'b0, `BYTEEN_SH}
`define LSU_LW {1'b0, `BYTEEN_SW}
@@ -183,6 +148,53 @@
`define CSR_BITS 2
`define CSR_OP(x) x[`CSR_BITS-1:0]
`define MUL_MUL 3'h0
`define MUL_MULH 3'h1
`define MUL_MULHSU 3'h2
`define MUL_MULHU 3'h3
`define MUL_DIV 3'h4
`define MUL_DIVU 3'h5
`define MUL_REM 3'h6
`define MUL_REMU 3'h7
`define MUL_BITS 3
`define MUL_OP(x) x[`MUL_BITS-1:0]
`define IS_DIV_OP(x) x[2]
`define FPU_ADD 5'h00
`define FPU_SUB 5'h01
`define FPU_MUL 5'h02
`define FPU_DIV 5'h03
`define FPU_SQRT 5'h04
`define FPU_MADD 5'h05
`define FPU_MSUB 5'h06
`define FPU_NMSUB 5'h07
`define FPU_NMADD 5'h08
`define FPU_SGNJ 5'h09 // FSGNJ
`define FPU_SGNJN 5'h0A // FSGNJN
`define FPU_SGNJX 5'h0B // FSGNJX
`define FPU_MIN 5'h0C // FMIN.S
`define FPU_MAX 5'h0D // FMAX.S
`define FPU_CVTWS 5'h0E // FCVT.W.S
`define FPU_CVTWUS 5'h0F // FCVT.WU.S
`define FPU_CVTSW 5'h10 // FCVT.S.W
`define FPU_CVTSWU 5'h11 // FCVT.S.WU
`define FPU_MVXW 5'h12 // MOV FP from fpReg to integer reg
`define FPU_MVWX 5'h13 // MOV FP from integer reg to fpReg
`define FPU_CLASS 5'h14
`define FPU_CMP 5'h15
`define FPU_OTHER 5'h1f
`define FPU_BITS 5
`define FPU_OP(x) x[`FPU_BITS-1:0]
`define FRM_RNE 3'b000 // round to nearest even
`define FRM_RTZ 3'b001 // round to zero
`define FRM_RDN 3'b010 // round to -inf
`define FRM_RUP 3'b011 // round to +inf
`define FRM_RMM 3'b100 // round to nearest max magnitude
`define FRM_DYN 3'b111 // dynamic mode
`define FRM_BITS 3
`define FFG_BITS 5
`define GPU_TMC 3'h0
`define GPU_WSPAWN 3'h1
`define GPU_SPLIT 3'h2
@@ -194,36 +206,43 @@
`define EX_NOP 3'h0
`define EX_ALU 3'h1
`define EX_MUL 3'h2
`define EX_LSU 3'h3
`define EX_CSR 3'h4
`define EX_GPU 3'h5
`define EX_LSU 3'h2
`define EX_CSR 3'h3
`define EX_MUL 3'h4
`define EX_FPU 3'h5
`define EX_GPU 3'h6
`define EX_BITS 3
`define NUM_EXS 5
`define NUM_EXS 6
`define NE_BITS `LOG2UP(`NUM_EXS)
`define WB_NO 2'h0
`define WB_ALU 2'h1
`define WB_MEM 2'h2
`define WB_JAL 2'h3
`define WB_BITS 2
///////////////////////////////////////////////////////////////////////////////
`ifdef EXT_M_ENABLE
`define ISA_EXT_M (1 << 12)
`else
`define ISA_EXT_M 0
`endif
`ifdef EXT_F_ENABLE
`define ISA_EXT_F (1 << 5)
`else
`define ISA_EXT_F 0
`endif
`define ISA_CODE (0 << 0) // A - Atomic Instructions extension \
| (0 << 1) // B - Tentatively reserved for Bit operations extension \
| (0 << 2) // C - Compressed extension \
| (0 << 3) // D - Double precsision floating-point extension \
| (0 << 4) // E - RV32E base ISA \
| (0 << 5) // F - Single precsision floating-point extension \
|`ISA_EXT_F // F - Single precsision floating-point extension \
| (0 << 6) // G - Additional standard extensions present \
| (0 << 7) // H - Hypervisor mode implemented \
| (1 << 8) // I - RV32I/64I/128I base ISA \
| (0 << 9) // J - Reserved \
| (0 << 10) // K - Reserved \
| (0 << 11) // L - Tentatively reserved for Bit operations extension \
| (1 << 12) // M - Integer Multiply/Divide extension \
|`ISA_EXT_M // M - Integer Multiply/Divide extension \
| (0 << 13) // N - User level interrupts supported \
| (0 << 14) // O - Reserved \
| (0 << 15) // P - Tentatively reserved for Packed-SIMD extension \
@@ -240,8 +259,8 @@
///////////////////////////////////////////////////////////////////////////////
`ifdef DBG_CORE_REQ_INFO // pc, wb, rd, warp_num
`define DEBUG_CORE_REQ_MDATA_WIDTH (32 + `WB_BITS + `NR_BITS + `NW_BITS)
`ifdef DBG_CORE_REQ_INFO // pc, wb, rd, warp_num
`define DEBUG_CORE_REQ_MDATA_WIDTH (32 + 1 + `NR_BITS + `NW_BITS)
`else
`define DEBUG_CORE_REQ_MDATA_WIDTH 0
`endif
@@ -252,7 +271,7 @@
`define DCACHE_ID (((`L3_ENABLE && `L2_ENABLE) ? 2 : `L2_ENABLE ? 1 : 0) + (CORE_ID * 3) + 0)
// TAG sharing enable
`define DCORE_TAG_ID_BITS `LOG2UP(`DCREQ_SIZE)
`define DCORE_TAG_ID_BITS `ISTAG_BITS
// Core request tag bits
`define DCORE_TAG_WIDTH (`DEBUG_CORE_REQ_MDATA_WIDTH + `DCORE_TAG_ID_BITS)
@@ -287,7 +306,7 @@
`define ICORE_BYTEEN_WIDTH `DWORD_SIZE
// TAG sharing enable
`define ICORE_TAG_ID_BITS `LOG2UP(`ICREQ_SIZE)
`define ICORE_TAG_ID_BITS `NW_BITS
// Core request tag bits
`define ICORE_TAG_WIDTH (`DEBUG_CORE_REQ_MDATA_WIDTH + `ICORE_TAG_ID_BITS)
@@ -390,116 +409,12 @@
///////////////////////////////////////////////////////////////////////////////
task print_ex_type;
input [`EX_BITS-1:0] ex;
begin
case (ex)
`EX_ALU: $write("ALU");
`EX_LSU: $write("LSU");
`EX_CSR: $write("CSR");
`EX_MUL: $write("MUL");
`EX_GPU: $write("GPU");
default: $write("NOP");
endcase
end
endtask
task print_instr_op;
input [`EX_BITS-1:0] ex;
input [`OP_BITS-1:0] op;
begin
case (ex)
`EX_ALU: begin
case (`ALU_BITS'(op))
`ALU_ADD: $write("ADD");
`ALU_SUB: $write("SUB");
`ALU_SLL: $write("SLL");
`ALU_SRL: $write("SRL");
`ALU_SRA: $write("SRA");
`ALU_SLT: $write("SLT");
`ALU_SLTU: $write("SLTU");
`ALU_XOR: $write("XOR");
`ALU_OR: $write("OR");
`ALU_AND: $write("AND");
`ALU_LUI: $write("LUI");
`ALU_AUIPC: $write("AUIPC");
`ALU_BEQ: $write("EQ");
`ALU_BNE: $write("NE");
`ALU_BLT: $write("LT");
`ALU_BGE: $write("GE");
`ALU_BLTU: $write("LTU");
`ALU_BGEU: $write("GEU");
`ALU_JAL: $write("JAL");
`ALU_JALR: $write("JALR");
`ALU_ECALL: $write("ECALL");
`ALU_EBREAK:$write("EBREAK");
`ALU_MRET: $write("MRET");
`ALU_SRET: $write("SRET");
`ALU_DRET: $write("DRET");
default: $write("?");
endcase
end
`EX_MUL: begin
case (`MUL_BITS'(op))
`MUL_MUL: $write("MUL");
`MUL_MULH: $write("MULH");
`MUL_MULHSU:$write("MULHSU");
`MUL_MULHU: $write("MULHU");
`MUL_DIV: $write("DIV");
`MUL_DIVU: $write("DIVU");
`MUL_REM: $write("REM");
`MUL_REMU: $write("REMU");
default: $write("?");
endcase
end
`EX_LSU: begin
case (`LSU_BITS'(op))
`LSU_LB: $write("LB");
`LSU_LH: $write("LH");
`LSU_LW: $write("LW");
`LSU_LBU: $write("LBU");
`LSU_LHU: $write("LHU");
`LSU_SB: $write("SB");
`LSU_SH: $write("SH");
`LSU_SW: $write("SW");
`LSU_SBU: $write("SBU");
`LSU_SHU: $write("SHU");
default: $write("?");
endcase
end
`EX_CSR: begin
case (`CSR_BITS'(op))
`CSR_RW: $write("CSRW");
`CSR_RS: $write("CSRS");
`CSR_RC: $write("CSRC");
default: $write("?");
endcase
end
`EX_GPU: begin
case (`GPU_BITS'(op))
`GPU_TMC: $write("TMC");
`GPU_WSPAWN:$write("WSPAWN");
`GPU_SPLIT: $write("SPLIT");
`GPU_JOIN: $write("JOIN");
`GPU_BAR: $write("BAR");
default: $write("?");
endcase
end
default:;
endcase
end
endtask
task print_wb;
input [`WB_BITS-1:0] wb;
begin
case (wb)
`WB_ALU: $write("ALU");
`WB_MEM: $write("MEM");
`WB_JAL: $write("JAL");
default: $write("NO");
endcase
end
endtask
typedef struct packed {
logic [`NW_BITS-1:0] warp_num;
logic [`NUM_THREADS-1:0] thread_mask;
logic [31:0] curr_PC;
logic [`NR_BITS-1:0] rd;
logic wb;
} is_data_t;
`endif