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

MWMT tested + minor opt

Browse Source
This commit is contained in:
felsabbagh3
2019-02-14 13:41:59 -05:00
parent 6c493cc4de
commit 6935d52c39
10 changed files with 315 additions and 295 deletions
Download Patch File Download Diff File Expand all files Collapse all files

427
src/riscv_gpgpu/gpgpu_test.dump
View File

@@ -9,7 +9,7 @@ Disassembly of section .text:
80000004: 00112623 sw ra,12(sp)
80000008: 00812423 sw s0,8(sp)
8000000c: 01010413 addi s0,sp,16
80000010: 354000ef jal ra,80000364 <initiate_stack>
80000010: 348000ef jal ra,80000358 <initiate_stack>
80000014: 810007b7 lui a5,0x81000
80000018: 00078793 mv a5,a5
8000001c: 81000737 lui a4,0x81000
@@ -18,9 +18,9 @@ Disassembly of section .text:
80000028: 04068693 addi a3,a3,64 # 81000040 <y+0xffffffc0>
8000002c: 80000637 lui a2,0x80000
80000030: 05860613 addi a2,a2,88 # 80000058 <y+0xfeffffd8>
80000034: 00800593 li a1,8
80000038: 00200513 li a0,2
8000003c: 1e0000ef jal ra,8000021c <createWarps>
80000034: 00400593 li a1,4
80000038: 00400513 li a0,4
8000003c: 1d4000ef jal ra,80000210 <createWarps>
80000040: 00000793 li a5,0
80000044: 00078513 mv a0,a5
80000048: 00c12083 lw ra,12(sp)
@@ -29,225 +29,222 @@ Disassembly of section .text:
80000054: 00008067 ret
80000058 <matAddition>:
80000058: fc010113 addi sp,sp,-64
8000005c: 02112e23 sw ra,60(sp)
80000060: 02812c23 sw s0,56(sp)
80000064: 04010413 addi s0,sp,64
80000068: fca42623 sw a0,-52(s0)
8000006c: 258000ef jal ra,800002c4 <get_wid>
80000070: fea42623 sw a0,-20(s0)
80000074: 278000ef jal ra,800002ec <get_1st_arg>
80000078: fea42423 sw a0,-24(s0)
8000007c: 298000ef jal ra,80000314 <get_2nd_arg>
80000080: fea42223 sw a0,-28(s0)
80000084: 2b8000ef jal ra,8000033c <get_3rd_arg>
80000088: fea42023 sw a0,-32(s0)
8000008c: fec42783 lw a5,-20(s0)
80000090: 00379793 slli a5,a5,0x3
80000094: fcc42703 lw a4,-52(s0)
80000098: 00f707b3 add a5,a4,a5
8000009c: fcf42e23 sw a5,-36(s0)
800000a0: fdc42783 lw a5,-36(s0)
800000a4: 00279793 slli a5,a5,0x2
800000a8: fe842703 lw a4,-24(s0)
800000ac: 00f707b3 add a5,a4,a5
800000b0: 0007a683 lw a3,0(a5) # 81000000 <y+0xffffff80>
800000b4: fdc42783 lw a5,-36(s0)
800000b8: 00279793 slli a5,a5,0x2
800000bc: fe442703 lw a4,-28(s0)
800000c0: 00f707b3 add a5,a4,a5
800000c4: 0007a703 lw a4,0(a5)
800000c8: fdc42783 lw a5,-36(s0)
800000cc: 00279793 slli a5,a5,0x2
800000d0: fe042603 lw a2,-32(s0)
800000d4: 00f607b3 add a5,a2,a5
800000d8: 00e68733 add a4,a3,a4
800000dc: 00e7a023 sw a4,0(a5)
800000e0: 00000013 nop
800000e4: 03c12083 lw ra,60(sp)
800000e8: 03812403 lw s0,56(sp)
800000ec: 04010113 addi sp,sp,64
800000f0: 00008067 ret
80000058: fd010113 addi sp,sp,-48
8000005c: 02112623 sw ra,44(sp)
80000060: 02812423 sw s0,40(sp)
80000064: 03010413 addi s0,sp,48
80000068: fca42e23 sw a0,-36(s0)
8000006c: fcb42c23 sw a1,-40(s0)
80000070: 270000ef jal ra,800002e0 <get_1st_arg>
80000074: fea42623 sw a0,-20(s0)
80000078: 290000ef jal ra,80000308 <get_2nd_arg>
8000007c: fea42423 sw a0,-24(s0)
80000080: 2b0000ef jal ra,80000330 <get_3rd_arg>
80000084: fea42223 sw a0,-28(s0)
80000088: fd842783 lw a5,-40(s0)
8000008c: 00279793 slli a5,a5,0x2
80000090: fdc42703 lw a4,-36(s0)
80000094: 00f707b3 add a5,a4,a5
80000098: fef42023 sw a5,-32(s0)
8000009c: fe042783 lw a5,-32(s0)
800000a0: 00279793 slli a5,a5,0x2
800000a4: fec42703 lw a4,-20(s0)
800000a8: 00f707b3 add a5,a4,a5
800000ac: 0007a683 lw a3,0(a5) # 81000000 <y+0xffffff80>
800000b0: fe042783 lw a5,-32(s0)
800000b4: 00279793 slli a5,a5,0x2
800000b8: fe842703 lw a4,-24(s0)
800000bc: 00f707b3 add a5,a4,a5
800000c0: 0007a703 lw a4,0(a5)
800000c4: fe042783 lw a5,-32(s0)
800000c8: 00279793 slli a5,a5,0x2
800000cc: fe442603 lw a2,-28(s0)
800000d0: 00f607b3 add a5,a2,a5
800000d4: 00e68733 add a4,a3,a4
800000d8: 00e7a023 sw a4,0(a5)
800000dc: 00000013 nop
800000e0: 02c12083 lw ra,44(sp)
800000e4: 02812403 lw s0,40(sp)
800000e8: 03010113 addi sp,sp,48
800000ec: 00008067 ret
800000f4 <createThreads>:
800000f4: fb010113 addi sp,sp,-80
800000f8: 04812623 sw s0,76(sp)
800000fc: 04912423 sw s1,72(sp)
80000100: 05212223 sw s2,68(sp)
80000104: 05312023 sw s3,64(sp)
80000108: 03412e23 sw s4,60(sp)
8000010c: 03a12c23 sw s10,56(sp)
80000110: 03b12a23 sw s11,52(sp)
80000114: 05010413 addi s0,sp,80
80000118: fca42623 sw a0,-52(s0)
8000011c: fcb42423 sw a1,-56(s0)
80000120: fcc42223 sw a2,-60(s0)
80000124: fcd42023 sw a3,-64(s0)
80000128: fae42e23 sw a4,-68(s0)
8000012c: faf42c23 sw a5,-72(s0)
80000130: fc042903 lw s2,-64(s0)
80000134: fbc42983 lw s3,-68(s0)
80000138: fb842a03 lw s4,-72(s0)
8000013c: fc842483 lw s1,-56(s0)
80000140: 00010f13 mv t5,sp
80000144: 00100793 li a5,1
80000148: fcf42e23 sw a5,-36(s0)
8000014c: 0200006f j 8000016c <createThreads+0x78>
80000150: fdc42503 lw a0,-36(s0)
80000154: fdc42303 lw t1,-36(s0)
80000158: f0010113 addi sp,sp,-256
8000015c: 0003506b 0x3506b
80000160: fdc42783 lw a5,-36(s0)
80000164: 00178793 addi a5,a5,1
80000168: fcf42e23 sw a5,-36(s0)
8000016c: fdc42703 lw a4,-36(s0)
80000170: fcc42783 lw a5,-52(s0)
80000174: fcf76ee3 bltu a4,a5,80000150 <createThreads+0x5c>
80000178: 000f0113 mv sp,t5
8000017c: 00000513 li a0,0
80000180: fc442f83 lw t6,-60(s0)
80000184: fcc42d83 lw s11,-52(s0)
80000188: 01bfe0eb 0x1bfe0eb
8000018c: 00000073 ecall
80000190: 00000013 nop
80000194: 04c12403 lw s0,76(sp)
80000198: 04812483 lw s1,72(sp)
8000019c: 04412903 lw s2,68(sp)
800001a0: 04012983 lw s3,64(sp)
800001a4: 03c12a03 lw s4,60(sp)
800001a8: 03812d03 lw s10,56(sp)
800001ac: 03412d83 lw s11,52(sp)
800001b0: 05010113 addi sp,sp,80
800001b4: 00008067 ret
800000f0 <createThreads>:
800000f0: fb010113 addi sp,sp,-80
800000f4: 04812623 sw s0,76(sp)
800000f8: 05212423 sw s2,72(sp)
800000fc: 05312223 sw s3,68(sp)
80000100: 05412023 sw s4,64(sp)
80000104: 03a12e23 sw s10,60(sp)
80000108: 03b12c23 sw s11,56(sp)
8000010c: 05010413 addi s0,sp,80
80000110: fca42623 sw a0,-52(s0)
80000114: fcb42423 sw a1,-56(s0)
80000118: fcc42223 sw a2,-60(s0)
8000011c: fcd42023 sw a3,-64(s0)
80000120: fae42e23 sw a4,-68(s0)
80000124: faf42c23 sw a5,-72(s0)
80000128: fc042903 lw s2,-64(s0)
8000012c: fbc42983 lw s3,-68(s0)
80000130: fb842a03 lw s4,-72(s0)
80000134: fc842583 lw a1,-56(s0)
80000138: 00010f13 mv t5,sp
8000013c: 00100793 li a5,1
80000140: fcf42e23 sw a5,-36(s0)
80000144: 0200006f j 80000164 <createThreads+0x74>
80000148: fdc42503 lw a0,-36(s0)
8000014c: fdc42303 lw t1,-36(s0)
80000150: f0010113 addi sp,sp,-256
80000154: 0003506b 0x3506b
80000158: fdc42783 lw a5,-36(s0)
8000015c: 00178793 addi a5,a5,1
80000160: fcf42e23 sw a5,-36(s0)
80000164: fdc42703 lw a4,-36(s0)
80000168: fcc42783 lw a5,-52(s0)
8000016c: fcf76ee3 bltu a4,a5,80000148 <createThreads+0x58>
80000170: 000f0113 mv sp,t5
80000174: 00000513 li a0,0
80000178: fc442f83 lw t6,-60(s0)
8000017c: fcc42d83 lw s11,-52(s0)
80000180: 01bfe0eb 0x1bfe0eb
80000184: 00000073 ecall
80000188: 00000013 nop
8000018c: 04c12403 lw s0,76(sp)
80000190: 04812903 lw s2,72(sp)
80000194: 04412983 lw s3,68(sp)
80000198: 04012a03 lw s4,64(sp)
8000019c: 03c12d03 lw s10,60(sp)
800001a0: 03812d83 lw s11,56(sp)
800001a4: 05010113 addi sp,sp,80
800001a8: 00008067 ret
800001b8 <wspawn>:
800001b8: fd010113 addi sp,sp,-48
800001bc: 02812623 sw s0,44(sp)
800001c0: 03010413 addi s0,sp,48
800001c4: fea42623 sw a0,-20(s0)
800001c8: feb42423 sw a1,-24(s0)
800001cc: fec42223 sw a2,-28(s0)
800001d0: fed42023 sw a3,-32(s0)
800001d4: fce42e23 sw a4,-36(s0)
800001d8: fcf42c23 sw a5,-40(s0)
800001dc: fd842383 lw t2,-40(s0)
800001e0: 800007b7 lui a5,0x80000
800001e4: 0f478793 addi a5,a5,244 # 800000f4 <y+0xff000074>
800001e8: 00078313 mv t1,a5
800001ec: fec42503 lw a0,-20(s0)
800001f0: fe842583 lw a1,-24(s0)
800001f4: fe442783 lw a5,-28(s0)
800001f8: 00078613 mv a2,a5
800001fc: fe042683 lw a3,-32(s0)
80000200: fdc42703 lw a4,-36(s0)
80000204: 00038793 mv a5,t2
80000208: 0003006b 0x3006b
8000020c: 00000013 nop
80000210: 02c12403 lw s0,44(sp)
80000214: 03010113 addi sp,sp,48
80000218: 00008067 ret
800001ac <wspawn>:
800001ac: fd010113 addi sp,sp,-48
800001b0: 02812623 sw s0,44(sp)
800001b4: 03010413 addi s0,sp,48
800001b8: fea42623 sw a0,-20(s0)
800001bc: feb42423 sw a1,-24(s0)
800001c0: fec42223 sw a2,-28(s0)
800001c4: fed42023 sw a3,-32(s0)
800001c8: fce42e23 sw a4,-36(s0)
800001cc: fcf42c23 sw a5,-40(s0)
800001d0: fd842383 lw t2,-40(s0)
800001d4: 800007b7 lui a5,0x80000
800001d8: 0f078793 addi a5,a5,240 # 800000f0 <y+0xff000070>
800001dc: 00078313 mv t1,a5
800001e0: fec42503 lw a0,-20(s0)
800001e4: fe842583 lw a1,-24(s0)
800001e8: fe442783 lw a5,-28(s0)
800001ec: 00078613 mv a2,a5
800001f0: fe042683 lw a3,-32(s0)
800001f4: fdc42703 lw a4,-36(s0)
800001f8: 00038793 mv a5,t2
800001fc: 0003006b 0x3006b
80000200: 00000013 nop
80000204: 02c12403 lw s0,44(sp)
80000208: 03010113 addi sp,sp,48
8000020c: 00008067 ret
8000021c <createWarps>:
8000021c: fc010113 addi sp,sp,-64
80000220: 02112e23 sw ra,60(sp)
80000224: 02812c23 sw s0,56(sp)
80000228: 04010413 addi s0,sp,64
8000022c: fca42e23 sw a0,-36(s0)
80000230: fcb42c23 sw a1,-40(s0)
80000234: fcc42a23 sw a2,-44(s0)
80000238: fcd42823 sw a3,-48(s0)
8000023c: fce42623 sw a4,-52(s0)
80000240: fcf42423 sw a5,-56(s0)
80000244: 00010f13 mv t5,sp
80000248: 00100793 li a5,1
8000024c: fef42623 sw a5,-20(s0)
80000250: 0300006f j 80000280 <createWarps+0x64>
80000254: 80010113 addi sp,sp,-2048
80000258: fc842783 lw a5,-56(s0)
8000025c: fcc42703 lw a4,-52(s0)
80000260: fd042683 lw a3,-48(s0)
80000264: fd442603 lw a2,-44(s0)
80000268: fec42583 lw a1,-20(s0)
8000026c: fd842503 lw a0,-40(s0)
80000270: f49ff0ef jal ra,800001b8 <wspawn>
80000274: fec42783 lw a5,-20(s0)
80000278: 00178793 addi a5,a5,1
8000027c: fef42623 sw a5,-20(s0)
80000280: fec42703 lw a4,-20(s0)
80000284: fdc42783 lw a5,-36(s0)
80000288: fcf766e3 bltu a4,a5,80000254 <createWarps+0x38>
8000028c: 000f0113 mv sp,t5
80000290: fd442603 lw a2,-44(s0)
80000294: fc842783 lw a5,-56(s0)
80000298: fcc42703 lw a4,-52(s0)
8000029c: fd042683 lw a3,-48(s0)
800002a0: 00000593 li a1,0
800002a4: fd842503 lw a0,-40(s0)
800002a8: e4dff0ef jal ra,800000f4 <createThreads>
800002ac: 00000073 ecall
800002b0: 00000013 nop
800002b4: 03c12083 lw ra,60(sp)
800002b8: 03812403 lw s0,56(sp)
800002bc: 04010113 addi sp,sp,64
800002c0: 00008067 ret
80000210 <createWarps>:
80000210: fc010113 addi sp,sp,-64
80000214: 02112e23 sw ra,60(sp)
80000218: 02812c23 sw s0,56(sp)
8000021c: 04010413 addi s0,sp,64
80000220: fca42e23 sw a0,-36(s0)
80000224: fcb42c23 sw a1,-40(s0)
80000228: fcc42a23 sw a2,-44(s0)
8000022c: fcd42823 sw a3,-48(s0)
80000230: fce42623 sw a4,-52(s0)
80000234: fcf42423 sw a5,-56(s0)
80000238: 00010f13 mv t5,sp
8000023c: 00100793 li a5,1
80000240: fef42623 sw a5,-20(s0)
80000244: 0300006f j 80000274 <createWarps+0x64>
80000248: 80010113 addi sp,sp,-2048
8000024c: fc842783 lw a5,-56(s0)
80000250: fcc42703 lw a4,-52(s0)
80000254: fd042683 lw a3,-48(s0)
80000258: fd442603 lw a2,-44(s0)
8000025c: fec42583 lw a1,-20(s0)
80000260: fd842503 lw a0,-40(s0)
80000264: f49ff0ef jal ra,800001ac <wspawn>
80000268: fec42783 lw a5,-20(s0)
8000026c: 00178793 addi a5,a5,1
80000270: fef42623 sw a5,-20(s0)
80000274: fec42703 lw a4,-20(s0)
80000278: fdc42783 lw a5,-36(s0)
8000027c: fcf766e3 bltu a4,a5,80000248 <createWarps+0x38>
80000280: 000f0113 mv sp,t5
80000284: fd442603 lw a2,-44(s0)
80000288: fc842783 lw a5,-56(s0)
8000028c: fcc42703 lw a4,-52(s0)
80000290: fd042683 lw a3,-48(s0)
80000294: 00000593 li a1,0
80000298: fd842503 lw a0,-40(s0)
8000029c: e55ff0ef jal ra,800000f0 <createThreads>
800002a0: 00000073 ecall
800002a4: 00000013 nop
800002a8: 03c12083 lw ra,60(sp)
800002ac: 03812403 lw s0,56(sp)
800002b0: 04010113 addi sp,sp,64
800002b4: 00008067 ret
800002c4 <get_wid>:
800002c4: ff010113 addi sp,sp,-16
800002c8: 00812623 sw s0,12(sp)
800002cc: 00912423 sw s1,8(sp)
800002d0: 01010413 addi s0,sp,16
800002d4: 00048793 mv a5,s1
800002d8: 00078513 mv a0,a5
800002dc: 00c12403 lw s0,12(sp)
800002e0: 00812483 lw s1,8(sp)
800002e4: 01010113 addi sp,sp,16
800002e8: 00008067 ret
800002b8 <get_wid>:
800002b8: ff010113 addi sp,sp,-16
800002bc: 00812623 sw s0,12(sp)
800002c0: 00912423 sw s1,8(sp)
800002c4: 01010413 addi s0,sp,16
800002c8: 00048793 mv a5,s1
800002cc: 00078513 mv a0,a5
800002d0: 00c12403 lw s0,12(sp)
800002d4: 00812483 lw s1,8(sp)
800002d8: 01010113 addi sp,sp,16
800002dc: 00008067 ret
800002ec <get_1st_arg>:
800002ec: ff010113 addi sp,sp,-16
800002f0: 00812623 sw s0,12(sp)
800002f4: 01212423 sw s2,8(sp)
800002f8: 01010413 addi s0,sp,16
800002fc: 00090793 mv a5,s2
80000300: 00078513 mv a0,a5
80000304: 00c12403 lw s0,12(sp)
80000308: 00812903 lw s2,8(sp)
8000030c: 01010113 addi sp,sp,16
80000310: 00008067 ret
800002e0 <get_1st_arg>:
800002e0: ff010113 addi sp,sp,-16
800002e4: 00812623 sw s0,12(sp)
800002e8: 01212423 sw s2,8(sp)
800002ec: 01010413 addi s0,sp,16
800002f0: 00090793 mv a5,s2
800002f4: 00078513 mv a0,a5
800002f8: 00c12403 lw s0,12(sp)
800002fc: 00812903 lw s2,8(sp)
80000300: 01010113 addi sp,sp,16
80000304: 00008067 ret
80000314 <get_2nd_arg>:
80000314: ff010113 addi sp,sp,-16
80000318: 00812623 sw s0,12(sp)
8000031c: 01312423 sw s3,8(sp)
80000320: 01010413 addi s0,sp,16
80000324: 00098793 mv a5,s3
80000328: 00078513 mv a0,a5
8000032c: 00c12403 lw s0,12(sp)
80000330: 00812983 lw s3,8(sp)
80000334: 01010113 addi sp,sp,16
80000338: 00008067 ret
80000308 <get_2nd_arg>:
80000308: ff010113 addi sp,sp,-16
8000030c: 00812623 sw s0,12(sp)
80000310: 01312423 sw s3,8(sp)
80000314: 01010413 addi s0,sp,16
80000318: 00098793 mv a5,s3
8000031c: 00078513 mv a0,a5
80000320: 00c12403 lw s0,12(sp)
80000324: 00812983 lw s3,8(sp)
80000328: 01010113 addi sp,sp,16
8000032c: 00008067 ret
8000033c <get_3rd_arg>:
8000033c: ff010113 addi sp,sp,-16
80000340: 00812623 sw s0,12(sp)
80000344: 01412423 sw s4,8(sp)
80000348: 01010413 addi s0,sp,16
8000034c: 000a0793 mv a5,s4
80000350: 00078513 mv a0,a5
80000354: 00c12403 lw s0,12(sp)
80000358: 00812a03 lw s4,8(sp)
8000035c: 01010113 addi sp,sp,16
80000360: 00008067 ret
80000330 <get_3rd_arg>:
80000330: ff010113 addi sp,sp,-16
80000334: 00812623 sw s0,12(sp)
80000338: 01412423 sw s4,8(sp)
8000033c: 01010413 addi s0,sp,16
80000340: 000a0793 mv a5,s4
80000344: 00078513 mv a0,a5
80000348: 00c12403 lw s0,12(sp)
8000034c: 00812a03 lw s4,8(sp)
80000350: 01010113 addi sp,sp,16
80000354: 00008067 ret
80000364 <initiate_stack>:
80000364: ff010113 addi sp,sp,-16
80000368: 00812623 sw s0,12(sp)
8000036c: 01010413 addi s0,sp,16
80000370: 7ffff137 lui sp,0x7ffff
80000374: 00000013 nop
80000378: 00c12403 lw s0,12(sp) # 7ffff00c <main-0xff4>
8000037c: 01010113 addi sp,sp,16
80000380: 00008067 ret
80000358 <initiate_stack>:
80000358: ff010113 addi sp,sp,-16
8000035c: 00812623 sw s0,12(sp)
80000360: 01010413 addi s0,sp,16
80000364: 7ffff137 lui sp,0x7ffff
80000368: 00000013 nop
8000036c: 00c12403 lw s0,12(sp) # 7ffff00c <main-0xff4>
80000370: 01010113 addi sp,sp,16
80000374: 00008067 ret
Disassembly of section .bss: