Download to read offline
W1M2_Introduction_HLS from under CBased VLSI.pdf
- 1. C-Based VLSI Design-An Overview Dr. Chandan Karfa Department of Computer Science and Engineering IIT Guwahati 1
- 2. 2 VLSI Design Flow SystemSpecification Architectural Design High-level Synthesis Logic Synthesis Physical Design Fabrication Packaging & Testing IIT Guwahati
- 3. High-Level Synthesis (HLS)/C-BasedVLSI Design • C Gates • Design time • Design complexity (10x) • Verification effort • Hardware/software co-design
- 4. 4 Importance of DesignAutomation • Shorter design cycle • Design space exploration • Fewer errors in the design • Less Verification efforts • Specification driven optimization at the higher abstraction level IIT Guwahati
- 5. C-Based VLSI Design •Enables designs at higher abstraction level (e.g., C, C++, Java) • 14 out of the top-20 semiconductor companies use HLS tools • Communications, signal processing, computation, crypto, healthcare, etc. • Tailor • Tailor implementation to match characteristics of target technology (e.g., speed, resources, area budget) – Video components in Tegra X1 chip designed using Catapult HLS. – NVIDIA 4K processing was designed with C-based HLS. – Qualcomm designing parts of Snapdragon with Catapult HLS. – Vivado HLS is part of xilinx design flow – intel HLS is part of quartus design flow 5
- 6. 6 HLS High-level Behaviour High-level Synthesis RegisterTransfer Level Description IIT Guwahati Example: 2nd order differential equation solver Diffeq: (x, dx, u, a, clock, y) input: x, dx, u, a, clock; output: y while(x < a) u1 = u-(3*x*u*dx)-(3*y*dx); y1 = y+(u*dx); x1 = x+dx; x = x1, y = y1, u = u1; end always @(posedge ap_clk) begin if(1'b1 == ap_CS_fsm_state5) begin j_reg_126 <= j_4_reg_293; end else if((1'b1 == ap_CS_fsm_state1) & (ap_start == 1'b1)) begin j_reg_126 <= 3'd0; end end assign tmp_108_fu_235_p1_temp_6 = tmp_108_fu_235_p1 & 63'd12; assign statemt_addr_28_reg_324_temp_7 = statemt_addr_28_reg_324 & 4'd19; assign tmp_108_fu_235_p1_temp_6_temp_8 = tmp_108_fu_235_p1_temp_6 | statemt_addr_28_reg_324_temp_7; ap_ST_fsm_state2: begin if((exitcond_fu_175_p2 == 1'd1) & (1'b1 == ap_CS_fsm_state2)) begin ap_NS_fsm = ap_ST_fsm_state1; end else begin ap_NS_fsm = ap_ST_fsm_state3; end end
- 7. 7 HLS Data-path Controller High-level Behaviour High-level Synthesis RegisterTransfer Level Description IIT Guwahati Example: 2nd order differential equation solver Diffeq: (x, dx, u, a, clock, y) input: x, dx, u, a, clock; output: y while(x < a) u1 = u-(3*x*u*dx)-(3*y*dx); y1 = y+(u*dx); x1 = x+dx; x = x1, y = y1, u = u1; end
- 8. 8 • Preprocessing: Intermediaterepresentation (CDFG) construction, data-dependency, live variable analysis, compiler optimization. • Scheduling: Assigns control step to the operations of the input behaviour. • Allocation: Computes minimum number of functional units and registers. • Binding: Variables are mapped to registers, operation to functional units, data transfers to the interconnection units. • Data path & Controller design: controller is designed based on inter connections among the data path elements, data transfer required in different control steps. High-level Synthesis Steps IIT Guwahati
- 9. 9 High-level Synthesis Steps |* | <6 *> <7 *> 5. <3 *> | * | 4. <5 - > | * | | * | 6. <8 - > <9 +> 7. | * | <4 * > 3. <0 * > <2 + > 2. < 1 *> 1. Input behaviour R1 : 3, v1 R2 : x u, v5 R3 : v0, v6 R4 : v3 FU1: op1, on3. .. FU2: op2, op5, … FU3: … scheduling Data-path generation Allocation & binding FU1: Controller generation Data-path Controller Control signal status signal RTL behaviour IIT Guwahati pre- processing
- 10. Working with anexample Example: 2nd order differential equation solver Diffeq: (x, dx, u, a, clock, y) input: x, dx, u, a, clock; output: y while(x < a) u1 = u-(3*x*u*dx)-(3*y*dx); y1 = y+(u*dx); x1 = x+dx; x = x1, y = y1, u = u1; end CDFG
- 11. Preprocessing I Read(p1, dx) Read(p2, x) Read(p3,a) Read(p1,y) Read(p2, u) c = x < a B1 V1 : t1 = u * dx V2 : t2 = 3 * x V3 : t3 = 3 * y V4 : t4 = u * dx V5 : t5 = t1 * t2 V6 : t6 = t3 * dx V7 : t7 = u – t5 V8 : u = t7 – t6 V9 : y = y + t4 V10 : x = x + dx V11 : c = x < a B2 Write(p1, y) Basic Blocks with 3-address codes Control and Dataflow graph (CDFG) Example: 2nd order differential equation solver Diffeq: (x, dx, u, a, clock, y) input: x, dx, u, a, clock; output: y while(x < a) u1 = u-(3*x*u*dx)-(3*y*dx); y1 = y+(u*dx); x1 = x+dx; x = x1, y = y1, u = u1; end
- 12. B1 V1 : t1= u * dx V2 : t2 = 3 * x V3 : t3 = 3 * y V4 : t4 = u * dx V5 : t5 = t1 * t2 V6 : t6 = t3 * dx V7 : t7 = u – t5 V8 : u = t7 – t6 V9 : y = y + t4 V10 : x = x + dx V11 : c = x < a Preprocessing
- 13. Preprocessing IIT Guwahati 13 B1 V1: t1 = u * dx V2 : t2 = 3 * x V3 : t3 = 3 * y V4 : t4 = u * dx V5 : t5 = t1 * t2 V6 : t6 = t3 * dx V7 : t7 = u – t5 V8 : u = t7 – t6 V9 : y = y + t4 V10 : x = x + dx V11 : c = x < a u dx x 3 y dx * - * * - * * * < + + c V1 V2 V5 V8 V10 V11 V6 V9 V3 V4 V7 t1 t3 t5 t8 t6 t7 t2 t4 Date dependency graph a
- 14. Scheduling S1 S2 S4 S3 u dx x 3y dx * - * * - * * * < + + c I V1 V2 V5 V8 V10 V11 V6 V9 V3 V4 V7 t1 t3 t5 t8 t6 t7 t2 t4 IIT Guwahati 14
- 15. Register Allocation andBinding IIT Guwahati 15 R1: t1, t3, t6 R2: t2, t5, t7 R3: t4 R4: t8 R5: u R6: x R7: dx R8: y R9: c R10: 3 R11: a S1 S2 S4 S3 u dx x 3 y dx * - * * - * * * < + + c I V1 V2 V5 V8 V10 V11 V6 V9 V3 V4 V7 t1 t3 t5 t8 t6 t7 t2 t4 Interval graph Var t1 t8 t7 t6 t5 t4 t3 t2 u x dx y c 3 a S1 S4 S3 S2 R2 R1 R3 R2 R1 R2 R1 R11 R10 R9 R8 R7 R6 R5 R4
- 16. FU Allocation andBinding: Multiplier V1 V6 V5 V4 V3 V2 M1 M2 M3 M2 M1 M3 S1 S4 S3 S2 IIT Guwahati 16 MULT: M1: V1, V5 MULT: M2: V2, V3 MULT: M3: V4, V6 Mult operations with non-overlapping schedule can be mapped to the same Multiplier FU S1 S2 S4 S3 u dx x 3 y dx * - * * - * * * < + + c I V1 V2 V5 V8 V10 V11 V6 V9 V3 V4 V7 t1 t3 t5 t8 t6 t7 t2 t4
- 17. FU Allocation andBinding: Adder IIT Guwahati 17 Var V10 V8 V7 V9 S1 S4 S3 S2 A1 A1 A1 A1 S1 S2 S4 S3 u dx x 3 y dx * - * * - * * * < + + c I V1 V2 V5 V8 V10 V11 V6 V9 V3 V4 V7 t1 t3 t5 t8 t6 t7 t2 t4 FU allocation and Binding A1: v10, v9, v7, v6 Add/Sub operations with non-overlapping schedule can be mapped to the same adder FU
- 18. Functional Unit Allocationand Binding IIT Guwahati 18 FU alloc and bind: MULT: M1: V1, V5 MULT: M2: V2, V3 MULT: M3: V4, V6 ADD: A1: V7, V8, V9, V10 COMP: C1: V11
- 19. Register Transfer Level(RTL) Behaviour IIT Guwahati 19 S1: V1 : t1 = u * dx V2 : t2 = 3 * x V4 : t4 = u * dx V10 : x = x + dx S2: V5 : t5 = t1 * t2 V3 : t3 = 3 * y V9 : y = y + t4 S1: V1 R1 = R5 <M1> R7 V2 R2 = R10 <M2> R6 V4 R3 = R5 <M3> R7 V10 R4 = R6 <A> R7 S2: V5 R2 = R1 <M1> R2 V3 R1 = R10 <M2> R8 V9 R8 = R8 <A> R3 R1: t1, t3, t6 R2: t2, t5, t7 R3: t4 R4: t8 R5: u R6: x R7: dx R8: y R9: c R10: 3 R11: a Original behaviour Register mapping RTL behaviour FU alloc and bind: MULT: M1: V1, V5 MULT: M1: V2, V3 MULT: M1: V4, V6 ADD: A1: V7, V8, V9, V10 COMP: C1: V11 FU mapping
- 20. Datapath Synthesis IIT Guwahati20 FU R1, R2, R1, R5, R4 R6, R1, R5, R6, R2 R1, R1, R2, R7, R4
- 21. Data path Synthesis IITGuwahati 21 S1: V1 R1 = R5 <M1> R7 V2 R2 = R10 <M2> R6 V4 R3 = R5 <M3> R7 V10 R4 = R6 <A> R7 S2: V5 R2 = R1 <M1> R2 V3 R1 = R10 <M2> R8 V9 R8 = R8 <A> R3
- 22. Data path Generation IITGuwahati 22
- 23. Controller Synthesis IIT Guwahati23 * ALU DATA-PATH CONTROL-UNIT r2 r1 u y x dx 3 a REGISTERS enable Mux control ALU control (+,-,<) c
- 24. Control Signals IIT Guwahati24 Control Assertion Pattern: <FU, FU_MUX_in, Reg-en, Reg_Mux_in> FU: 1 bit FU_MUX_in: 7 bits Reg_en: 11 bits Reg_MUX_in: 2 bits Total: 21 bits S1: <1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0 0, 0, 1> S2: <1, 1, 1, 1, 0, 1, 0, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0 0, 1, 0> S3: <…..> S4: <…..> S1: V1 R1 = R5 <M1> R7 V2 R2 = R10 <M2> R6 V4 R3 = R5 <M3> R7 V10 R4 = R6 <A> R7 S2: V5 R2 = R1 <M1> R2 V3 R1 = R10 <M2> R8 V9 R8 = R8 <A> R3
- 25. Final RTL IIT Guwahati25 <1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0 0, 0, 1>
- 26. 26 High-level Synthesis Steps |* | <6 *> <7 *> 5. <3 *> | * | 4. <5 - > | * | | * | 6. <8 - > <9 +> 7. | * | <4 * > 3. <0 * > <2 + > 2. < 1 *> 1. Input behaviour R1 : 3, v1 R2 : x u, v5 R3 : v0, v6 R4 : v3 FU1: op1, on3. .. FU2: op2, op5, … FU3: … scheduling Data-path generation Allocation & binding FU1: Controller generation Data-path Controller Control signal status signal RTL behaviour IIT Guwahati pre- processing
- 27. Topics to becovered • Scheduling Possibilities • Register and FU allocation and binding • Datapath and Controller Synthesis IIT Guwahati 27
- 28. Scheduling Possibilities IIT Guwahati28 S1 S2 S4 S3 u dx x 3 y dx * - * * - * * * < + + c I V1 V2 V5 V8 V10 V11 V6 V9 V3 V4 V7 t1 t3 t5 t8 t6 t7 t2 t4 At least 3 Multipliers required At least 4 Multipliers required S1 S2 S4 S3 u dx x 3 y dx * - * * - * * * < + + c I V1 V2 V5 V8 V10 V11 V6 V9 V3 V4 V7 t1 t3 t5 t8 t6 t7 t2 t4
- 29. Scheduling Possibilities IIT Guwahati29 S1 S2 S4 S3 u dx x 3 y dx * - * * - * * * < + + c I V1 V2 V5 V8 V10 V11 V6 V9 V3 V4 V7 t1 t3 t5 t8 t6 t7 t2 t4 S1 S2 S4 S3 u dx x 3 y dx * - * * - * * * < + + c I V1 V2 V5 V8 V10 V11 V6 V9 V3 V4 V7 t1 t3 t5 t8 t6 t7 t2 t4 At least 3 Multipliers required At least 2 Multipliers required
- 30. Automation of registerand FU allocation and binding • How to automate register allocation and binding? • How to automate FU allocation and binding • Map the problem to Graph colouring problem or clique partitioning problem and solve. IIT Guwahati 30 Data path and Controller Synthesis: -Mux based and Bus based architecture -Various way to optimize interconnections
- 31.