p = [3, 1]:数值检查通过 p = [5, 0]:数值检查通过 p = [7, 1]:数值检查通过 # function jit(select)/select/mpmd_map [symbol 6, select.1]; image bundles [505, 509) { misc: vtrace 0x80000000 } # entry bundle: %select.1 = custom-call(%x_hbm.1, %p_hbm.1) { s1: sld s9, [smem:0x1] } { s0: sne.s32 p0, s9, 0 } { s0: @p0 sbr.rel L_0211 } # end function jit(select)/select/mpmd_map [symbol 6, select.1] {} # function jit(select)/select/mpmd_map [symbol 6, select.1]; image bundles [510, 524) { s0: simm.s32 s10, 62 ; # materialized constant .encoding { s0.y = pair(imm1, imm0) } } { s0: dma.simple [smem:s10], [hbm:s1], length=1, dst_flag=[sflag:52] } { misc: vwait.ge [sflag:52], 1 ; # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:37:12-37:55 [dma_wait; LLO 69] .encoding { misc.value = imm2 } } { misc: vsyncadd.s32 [sflag:52], -1 ; # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:37:12-37:55 [dma_wait; LLO 70] .encoding { misc.value = imm2 } } { s0: sfence } # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:37:12-37:55 [dma_wait; LLO 26] { s0: simm.s32 s18, 0 ; # materialized constant s1: sld s13, [smem:0x3e] ; .encoding { s0.y = lo(imm2) ; s1.address = pair(imm1, imm0) } } { s0: sshll.u32 s14, s13, 0x3 } # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:39:20-39:33 [mul; LLO 64] { s0: sadd.s32 s17, s14, s0 } # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:40:12-40:68 [dma_start; LLO 29] | chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:40:12-40:75 [dma_wait; LLO 29] { s0: dma.simple [vmem:s18], [hbm:s17], length=8, dst_flag=[sflag:52] } { misc: vwait.ge [sflag:52], 8 ; # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:40:12-40:75 [dma_wait; LLO 71] .encoding { misc.value = imm2 } } { s1: sld s19, [smem:0x3f] ; misc: vsyncadd.s32 [sflag:52], -8 ; # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:40:12-40:75 [dma_wait; LLO 72] .encoding { misc.value = imm2 ; s1.address = pair(imm1, imm0) } } { s0: sle.s32 p1, s19, 0 } # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:42:21-42:34 [gt; LLO 66] | chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:42:13-42:35 { s0: @p1 sbr.rel L_020e ; # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:42:13-42:35 vld: @!p1 vld.8x128 v0, [vmem:0x0] } # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:44:30-44:41 <- 42:13-42:35 [get; LLO 47] { va0: @!p1 vmul.8x128.f32 v1, 2.0, v0 } # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:44:30-44:47 <- 42:13-42:35 [mul; LLO 48] # end function jit(select)/select/mpmd_map [symbol 6, select.1] {} # function jit(select)/select/mpmd_map [symbol 6, select.1]; image bundles [525, 529) { vst: vst.8x128 [vmem:0x0], v1 ; # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:44:16-44:27 <- 42:13-42:35 [swap; LLO 49] .encoding { vst.opcode = s0 } } L_020e: { s0: dma.simple [hbm:s2], [vmem:s18], length=8, dst_flag=[sflag:52] } { misc: vwait.ge [sflag:52], 8 ; # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:46:12-46:55 [dma_wait; LLO 73] .encoding { misc.value = imm2 } } { misc: vsyncadd.s32 [sflag:52], -8 ; # chapter01/08_scalar_control_flow/03_pallas_dynamic_window_when.py:46:12-46:55 [dma_wait; LLO 74] .encoding { misc.value = imm2 } } # end function jit(select)/select/mpmd_map [symbol 6, select.1] # exit bundle: %select.1 = custom-call(%x_hbm.1, %p_hbm.1) # ... 省略 runtime 代码 1 个 bundle # function jit(select)/select/mpmd_map [symbol 6, select.1]; image bundles [530, 546) { s1: sld s20, [smem:0x0] } { s1: sld s21, [smem:0x1] } { s0: sand.u32 s24, 0xfff, s20 } { s0: sadd.s32 s22, 1, s21 ; s1: sshll.u32 s27, s24, 0x12 } { s0: sand.u32 s23, 0x1, s22 } # smod.u32 w/div 2 { s0: sand.u32 s25, 0x3, s23 } { s0: sshll.u32 s26, s25, 0x10 } { s0: sadd.s32 s28, 131072, s26 } { s0: sshrl.u32 s29, s28, 0x2 } { s0: sor.u32 s30, 0x2d, s29 } { s0: sor.u32 s17, s30, s27 } { misc: vsyncadd.remote.s32 [sflag:s17], 1 } { misc: vwait.ge [sflag:45], 1 ; .encoding { misc.value = imm2 } } { misc: vsyncadd.s32 [sflag:45], -1 ; .encoding { misc.value = imm2 } } { misc: cfence } { misc: vtrace 0x90000000 } # end function jit(select)/select/mpmd_map [symbol 6, select.1]