diff --git a/benchmarks/opencl/guassian/Fan1.dump b/benchmarks/opencl/guassian/Fan1.dump index c23e12d5..bff54585 100644 --- a/benchmarks/opencl/guassian/Fan1.dump +++ b/benchmarks/opencl/guassian/Fan1.dump @@ -1,5 +1,5 @@ -/tmp/pocl_vortex_kernel-db-03-14-35-2b.elf: file format ELF32-riscv +/tmp/pocl_vortex_kernel-00-08-9c-7c-dc.elf: file format ELF32-riscv Disassembly of section .init: @@ -774,7 +774,7 @@ Disassembly of section .comment: c: 6e 20 e: 31 30 10: 2e 30 - 12: 2e 30 + 12: 2e 31 14: 20 28 16: 68 74 18: 74 70 @@ -792,25 +792,25 @@ Disassembly of section .comment: 36: 6a 65 38: 63 74 2e 67 bgeu t3, s2, 1640 3c: 69 74 - 3e: 20 37 - 40: 33 63 35 33 - 44: 65 36 - 46: 31 32 - 48: 61 61 - 4a: 31 32 - 4c: 35 61 - 4e: 32 34 - 50: 65 62 - 52: 32 63 - 54: 37 35 32 34 lui a0, 213795 - 58: 31 62 - 5a: 32 65 - 5c: 62 62 - 5e: 64 35 - 60: 36 30 - 62: 62 38 - 64: 35 63 - 66: 37 29 00 47 lui s2, 290818 + 3e: 20 65 + 40: 66 33 + 42: 32 63 + 44: 36 31 + 46: 31 61 + 48: 61 32 + 4a: 31 34 + 4c: 64 65 + 4e: 61 38 + 50: 35 35 + 52: 33 36 34 65 + 56: 66 64 + 58: 37 62 61 34 lui tp, 214550 + 5c: 35 31 + 5e: 65 63 + 60: 35 65 + 62: 63 33 66 37 + 66: 34 29 + 68: 00 47 6a: 43 43 3a 20 fmadd.s ft6, fs4, ft3, ft4, rmm 6e: 28 47 70: 4e 55 @@ -1258,13 +1258,13 @@ Disassembly of section .strtab: 3e: 5f 6b 65 72 42: 6e 65 44: 6c 2d - 46: 32 34 - 48: 2d 37 - 4a: 61 2d - 4c: 61 31 - 4e: 2d 63 - 50: 30 2d - 52: 33 31 2e 63 + 46: 31 38 + 48: 2d 63 + 4a: 33 2d 38 61 + 4e: 2d 37 + 50: 35 2d + 52: 31 33 + 54: 2e 63 56: 00 70 58: 61 72 5a: 61 6c diff --git a/benchmarks/opencl/guassian/Fan2.dump b/benchmarks/opencl/guassian/Fan2.dump index 4384c0e4..8c52171f 100644 --- a/benchmarks/opencl/guassian/Fan2.dump +++ b/benchmarks/opencl/guassian/Fan2.dump @@ -1,5 +1,5 @@ -/tmp/pocl_vortex_kernel-6b-12-be-02-10.elf: file format ELF32-riscv +/tmp/pocl_vortex_kernel-1b-12-83-97-df.elf: file format ELF32-riscv Disassembly of section .init: @@ -989,7 +989,7 @@ Disassembly of section .comment: c: 6e 20 e: 31 30 10: 2e 30 - 12: 2e 30 + 12: 2e 31 14: 20 28 16: 68 74 18: 74 70 @@ -1007,25 +1007,25 @@ Disassembly of section .comment: 36: 6a 65 38: 63 74 2e 67 bgeu t3, s2, 1640 3c: 69 74 - 3e: 20 37 - 40: 33 63 35 33 - 44: 65 36 - 46: 31 32 - 48: 61 61 - 4a: 31 32 - 4c: 35 61 - 4e: 32 34 - 50: 65 62 - 52: 32 63 - 54: 37 35 32 34 lui a0, 213795 - 58: 31 62 - 5a: 32 65 - 5c: 62 62 - 5e: 64 35 - 60: 36 30 - 62: 62 38 - 64: 35 63 - 66: 37 29 00 47 lui s2, 290818 + 3e: 20 65 + 40: 66 33 + 42: 32 63 + 44: 36 31 + 46: 31 61 + 48: 61 32 + 4a: 31 34 + 4c: 64 65 + 4e: 61 38 + 50: 35 35 + 52: 33 36 34 65 + 56: 66 64 + 58: 37 62 61 34 lui tp, 214550 + 5c: 35 31 + 5e: 65 63 + 60: 35 65 + 62: 63 33 66 37 + 66: 34 29 + 68: 00 47 6a: 43 43 3a 20 fmadd.s ft6, fs4, ft3, ft4, rmm 6e: 28 47 70: 4e 55 @@ -1473,12 +1473,13 @@ Disassembly of section .strtab: 3e: 5f 6b 65 72 42: 6e 65 44: 6c 2d - 46: 61 34 - 48: 2d 61 - 4a: 61 2d - 4c: 37 62 2d 65 lui tp, 414422 - 50: 61 2d - 52: 31 31 + 46: 34 35 + 48: 2d 30 + 4a: 34 2d + 4c: 64 61 + 4e: 2d 31 + 50: 35 2d + 52: 36 65 54: 2e 63 56: 00 70 58: 61 72 diff --git a/benchmarks/opencl/guassian/kernel.pocl b/benchmarks/opencl/guassian/kernel.pocl index 3b14c4b7..23ce271b 100644 Binary files a/benchmarks/opencl/guassian/kernel.pocl and b/benchmarks/opencl/guassian/kernel.pocl differ diff --git a/benchmarks/opencl/nearn/NearestNeighbor.dump b/benchmarks/opencl/nearn/NearestNeighbor.dump index 2c461925..c63db6b8 100644 --- a/benchmarks/opencl/nearn/NearestNeighbor.dump +++ b/benchmarks/opencl/nearn/NearestNeighbor.dump @@ -1,5 +1,5 @@ -/tmp/pocl_vortex_kernel-7f-1f-8c-52-07.elf: file format ELF32-riscv +/tmp/pocl_vortex_kernel-c7-de-a7-37-23.elf: file format ELF32-riscv Disassembly of section .init: @@ -987,7 +987,7 @@ Disassembly of section .comment: c: 6e 20 e: 31 30 10: 2e 30 - 12: 2e 30 + 12: 2e 31 14: 20 28 16: 68 74 18: 74 70 @@ -1005,25 +1005,25 @@ Disassembly of section .comment: 36: 6a 65 38: 63 74 2e 67 bgeu t3, s2, 1640 3c: 69 74 - 3e: 20 37 - 40: 33 63 35 33 - 44: 65 36 - 46: 31 32 - 48: 61 61 - 4a: 31 32 - 4c: 35 61 - 4e: 32 34 - 50: 65 62 - 52: 32 63 - 54: 37 35 32 34 lui a0, 213795 - 58: 31 62 - 5a: 32 65 - 5c: 62 62 - 5e: 64 35 - 60: 36 30 - 62: 62 38 - 64: 35 63 - 66: 37 29 00 47 lui s2, 290818 + 3e: 20 65 + 40: 66 33 + 42: 32 63 + 44: 36 31 + 46: 31 61 + 48: 61 32 + 4a: 31 34 + 4c: 64 65 + 4e: 61 38 + 50: 35 35 + 52: 33 36 34 65 + 56: 66 64 + 58: 37 62 61 34 lui tp, 214550 + 5c: 35 31 + 5e: 65 63 + 60: 35 65 + 62: 63 33 66 37 + 66: 34 29 + 68: 00 47 6a: 43 43 3a 20 fmadd.s ft6, fs4, ft3, ft4, rmm 6e: 28 47 70: 4e 55 @@ -1530,12 +1530,13 @@ Disassembly of section .strtab: 3e: 5f 6b 65 72 42: 6e 65 44: 6c 2d - 46: 61 62 - 48: 2d 63 - 4a: 37 2d 35 39 lui s10, 234322 - 4e: 2d 32 - 50: 32 2d - 52: 32 35 + 46: 35 35 + 48: 2d 30 + 4a: 39 2d + 4c: 39 61 + 4e: 2d 61 + 50: 36 2d + 52: 38 38 54: 2e 63 56: 00 70 58: 61 72 diff --git a/benchmarks/opencl/nearn/kernel.pocl b/benchmarks/opencl/nearn/kernel.pocl index 97b4afb8..0cd2e947 100644 Binary files a/benchmarks/opencl/nearn/kernel.pocl and b/benchmarks/opencl/nearn/kernel.pocl differ diff --git a/benchmarks/opencl/saxpy/kernel.pocl b/benchmarks/opencl/saxpy/kernel.pocl index 0eb739e7..c114817b 100644 Binary files a/benchmarks/opencl/saxpy/kernel.pocl and b/benchmarks/opencl/saxpy/kernel.pocl differ diff --git a/benchmarks/opencl/saxpy/saxpy.dump b/benchmarks/opencl/saxpy/saxpy.dump index 0ed173a8..8fab3750 100644 --- a/benchmarks/opencl/saxpy/saxpy.dump +++ b/benchmarks/opencl/saxpy/saxpy.dump @@ -1,5 +1,5 @@ -/tmp/pocl_vortex_kernel-5f-33-e1-2a-a5.elf: file format ELF32-riscv +/tmp/pocl_vortex_kernel-1b-e7-8a-9c-af.elf: file format ELF32-riscv Disassembly of section .init: @@ -720,7 +720,7 @@ Disassembly of section .comment: c: 6e 20 e: 31 30 10: 2e 30 - 12: 2e 30 + 12: 2e 31 14: 20 28 16: 68 74 18: 74 70 @@ -738,25 +738,25 @@ Disassembly of section .comment: 36: 6a 65 38: 63 74 2e 67 bgeu t3, s2, 1640 3c: 69 74 - 3e: 20 37 - 40: 33 63 35 33 - 44: 65 36 - 46: 31 32 - 48: 61 61 - 4a: 31 32 - 4c: 35 61 - 4e: 32 34 - 50: 65 62 - 52: 32 63 - 54: 37 35 32 34 lui a0, 213795 - 58: 31 62 - 5a: 32 65 - 5c: 62 62 - 5e: 64 35 - 60: 36 30 - 62: 62 38 - 64: 35 63 - 66: 37 29 00 47 lui s2, 290818 + 3e: 20 65 + 40: 66 33 + 42: 32 63 + 44: 36 31 + 46: 31 61 + 48: 61 32 + 4a: 31 34 + 4c: 64 65 + 4e: 61 38 + 50: 35 35 + 52: 33 36 34 65 + 56: 66 64 + 58: 37 62 61 34 lui tp, 214550 + 5c: 35 31 + 5e: 65 63 + 60: 35 65 + 62: 63 33 66 37 + 66: 34 29 + 68: 00 47 6a: 43 43 3a 20 fmadd.s ft6, fs4, ft3, ft4, rmm 6e: 28 47 70: 4e 55 @@ -1199,12 +1199,13 @@ Disassembly of section .strtab: 3e: 5f 6b 65 72 42: 6e 65 44: 6c 2d - 46: 36 35 - 48: 2d 66 - 4a: 33 2d 64 35 - 4e: 2d 36 + 46: 35 36 + 48: 2d 63 + 4a: 30 2d + 4c: 66 37 + 4e: 2d 38 50: 34 2d - 52: 31 37 + 52: 36 39 54: 2e 63 56: 00 70 58: 61 72 diff --git a/benchmarks/opencl/sfilter/kernel.pocl b/benchmarks/opencl/sfilter/kernel.pocl index bae293fa..e1726666 100644 Binary files a/benchmarks/opencl/sfilter/kernel.pocl and b/benchmarks/opencl/sfilter/kernel.pocl differ diff --git a/benchmarks/opencl/sfilter/sfilter.dump b/benchmarks/opencl/sfilter/sfilter.dump index 244b713a..163eb04a 100644 --- a/benchmarks/opencl/sfilter/sfilter.dump +++ b/benchmarks/opencl/sfilter/sfilter.dump @@ -1,5 +1,5 @@ -/tmp/pocl_vortex_kernel-9f-b2-9e-1e-b6.elf: file format ELF32-riscv +/tmp/pocl_vortex_kernel-a9-15-33-2e-25.elf: file format ELF32-riscv Disassembly of section .init: @@ -1076,7 +1076,7 @@ Disassembly of section .comment: c: 6e 20 e: 31 30 10: 2e 30 - 12: 2e 30 + 12: 2e 31 14: 20 28 16: 68 74 18: 74 70 @@ -1094,25 +1094,25 @@ Disassembly of section .comment: 36: 6a 65 38: 63 74 2e 67 bgeu t3, s2, 1640 3c: 69 74 - 3e: 20 37 - 40: 33 63 35 33 - 44: 65 36 - 46: 31 32 - 48: 61 61 - 4a: 31 32 - 4c: 35 61 - 4e: 32 34 - 50: 65 62 - 52: 32 63 - 54: 37 35 32 34 lui a0, 213795 - 58: 31 62 - 5a: 32 65 - 5c: 62 62 - 5e: 64 35 - 60: 36 30 - 62: 62 38 - 64: 35 63 - 66: 37 29 00 47 lui s2, 290818 + 3e: 20 65 + 40: 66 33 + 42: 32 63 + 44: 36 31 + 46: 31 61 + 48: 61 32 + 4a: 31 34 + 4c: 64 65 + 4e: 61 38 + 50: 35 35 + 52: 33 36 34 65 + 56: 66 64 + 58: 37 62 61 34 lui tp, 214550 + 5c: 35 31 + 5e: 65 63 + 60: 35 65 + 62: 63 33 66 37 + 66: 34 29 + 68: 00 47 6a: 43 43 3a 20 fmadd.s ft6, fs4, ft3, ft4, rmm 6e: 28 47 70: 4e 55 @@ -1558,12 +1558,13 @@ Disassembly of section .strtab: 3e: 5f 6b 65 72 42: 6e 65 44: 6c 2d - 46: 37 35 2d 38 lui a0, 230099 - 4a: 62 2d - 4c: 32 61 - 4e: 2d 35 - 50: 34 2d - 52: 35 38 + 46: 30 63 + 48: 2d 32 + 4a: 61 2d + 4c: 64 30 + 4e: 2d 33 + 50: 39 2d + 52: 36 61 54: 2e 63 56: 00 70 58: 61 72 diff --git a/benchmarks/opencl/sgemm/kernel.pocl b/benchmarks/opencl/sgemm/kernel.pocl index 7944ce98..689300e2 100644 Binary files a/benchmarks/opencl/sgemm/kernel.pocl and b/benchmarks/opencl/sgemm/kernel.pocl differ diff --git a/benchmarks/opencl/sgemm/sgemm.dump b/benchmarks/opencl/sgemm/sgemm.dump index caa5cd1c..7c1fe6f4 100644 --- a/benchmarks/opencl/sgemm/sgemm.dump +++ b/benchmarks/opencl/sgemm/sgemm.dump @@ -1,5 +1,5 @@ -/tmp/pocl_vortex_kernel-c0-b8-f5-98-dd.elf: file format ELF32-riscv +/tmp/pocl_vortex_kernel-31-f9-9a-78-75.elf: file format ELF32-riscv Disassembly of section .init: @@ -887,7 +887,7 @@ Disassembly of section .comment: c: 6e 20 e: 31 30 10: 2e 30 - 12: 2e 30 + 12: 2e 31 14: 20 28 16: 68 74 18: 74 70 @@ -905,25 +905,25 @@ Disassembly of section .comment: 36: 6a 65 38: 63 74 2e 67 bgeu t3, s2, 1640 3c: 69 74 - 3e: 20 37 - 40: 33 63 35 33 - 44: 65 36 - 46: 31 32 - 48: 61 61 - 4a: 31 32 - 4c: 35 61 - 4e: 32 34 - 50: 65 62 - 52: 32 63 - 54: 37 35 32 34 lui a0, 213795 - 58: 31 62 - 5a: 32 65 - 5c: 62 62 - 5e: 64 35 - 60: 36 30 - 62: 62 38 - 64: 35 63 - 66: 37 29 00 47 lui s2, 290818 + 3e: 20 65 + 40: 66 33 + 42: 32 63 + 44: 36 31 + 46: 31 61 + 48: 61 32 + 4a: 31 34 + 4c: 64 65 + 4e: 61 38 + 50: 35 35 + 52: 33 36 34 65 + 56: 66 64 + 58: 37 62 61 34 lui tp, 214550 + 5c: 35 31 + 5e: 65 63 + 60: 35 65 + 62: 63 33 66 37 + 66: 34 29 + 68: 00 47 6a: 43 43 3a 20 fmadd.s ft6, fs4, ft3, ft4, rmm 6e: 28 47 70: 4e 55 @@ -1370,12 +1370,11 @@ Disassembly of section .strtab: 3e: 5f 6b 65 72 42: 6e 65 44: 6c 2d - 46: 63 34 2d 66 + 46: 63 65 2d 36 bltu s10, sp, 874 4a: 30 2d - 4c: 35 34 - 4e: 2d 38 - 50: 30 2d - 52: 32 36 + 4c: 37 34 2d 32 lui s0, 205523 + 50: 62 2d + 52: 38 35 54: 2e 63 56: 00 70 58: 61 72 diff --git a/benchmarks/opencl/vecadd/kernel.pocl b/benchmarks/opencl/vecadd/kernel.pocl index 9fd25c4f..a2aef8b2 100644 Binary files a/benchmarks/opencl/vecadd/kernel.pocl and b/benchmarks/opencl/vecadd/kernel.pocl differ diff --git a/benchmarks/opencl/vecadd/vecadd.dump b/benchmarks/opencl/vecadd/vecadd.dump index 90b78049..8a10d60d 100644 --- a/benchmarks/opencl/vecadd/vecadd.dump +++ b/benchmarks/opencl/vecadd/vecadd.dump @@ -1,5 +1,5 @@ -/tmp/pocl_vortex_kernel-d7-2c-61-d5-00.elf: file format ELF32-riscv +/tmp/pocl_vortex_kernel-69-5f-25-33-78.elf: file format ELF32-riscv Disassembly of section .init: @@ -725,7 +725,7 @@ Disassembly of section .comment: c: 6e 20 e: 31 30 10: 2e 30 - 12: 2e 30 + 12: 2e 31 14: 20 28 16: 68 74 18: 74 70 @@ -743,25 +743,25 @@ Disassembly of section .comment: 36: 6a 65 38: 63 74 2e 67 bgeu t3, s2, 1640 3c: 69 74 - 3e: 20 37 - 40: 33 63 35 33 - 44: 65 36 - 46: 31 32 - 48: 61 61 - 4a: 31 32 - 4c: 35 61 - 4e: 32 34 - 50: 65 62 - 52: 32 63 - 54: 37 35 32 34 lui a0, 213795 - 58: 31 62 - 5a: 32 65 - 5c: 62 62 - 5e: 64 35 - 60: 36 30 - 62: 62 38 - 64: 35 63 - 66: 37 29 00 47 lui s2, 290818 + 3e: 20 65 + 40: 66 33 + 42: 32 63 + 44: 36 31 + 46: 31 61 + 48: 61 32 + 4a: 31 34 + 4c: 64 65 + 4e: 61 38 + 50: 35 35 + 52: 33 36 34 65 + 56: 66 64 + 58: 37 62 61 34 lui tp, 214550 + 5c: 35 31 + 5e: 65 63 + 60: 35 65 + 62: 63 33 66 37 + 66: 34 29 + 68: 00 47 6a: 43 43 3a 20 fmadd.s ft6, fs4, ft3, ft4, rmm 6e: 28 47 70: 4e 55 @@ -1207,12 +1207,12 @@ Disassembly of section .strtab: 3e: 5f 6b 65 72 42: 6e 65 44: 6c 2d - 46: 65 39 - 48: 2d 38 - 4a: 63 2d 32 66 + 46: 65 31 + 48: 2d 37 + 4a: 63 2d 39 39 4e: 2d 38 - 50: 30 2d - 52: 64 35 + 50: 66 2d + 52: 32 31 54: 2e 63 56: 00 70 58: 61 72