206 lines · plain
1; RUN: llc -mtriple=r600 -mcpu=juniper < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s2 3; === WIDTH ==================================================================4; 9 implicit args = 9 dwords to first image argument.5; First width at dword index 9+1 -> KC0[2].Z6 7; FUNC-LABEL: {{^}}width_2d:8; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]9; EG: MOV * [[VAL]], KC0[2].Z10define amdgpu_kernel void @width_2d (ptr addrspace(1) %in,11 ptr addrspace(1) %out) {12entry:13 %0 = call [3 x i32] @llvm.OpenCL.image.get.size.2d(14 ptr addrspace(1) %in) #015 %1 = extractvalue [3 x i32] %0, 016 store i32 %1, ptr addrspace(1) %out17 ret void18}19 20; FUNC-LABEL: {{^}}width_3d:21; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]22; EG: MOV * [[VAL]], KC0[2].Z23define amdgpu_kernel void @width_3d (ptr addrspace(1) %in,24 ptr addrspace(1) %out) {25entry:26 %0 = call [3 x i32] @llvm.OpenCL.image.get.size.3d(27 ptr addrspace(1) %in) #028 %1 = extractvalue [3 x i32] %0, 029 store i32 %1, ptr addrspace(1) %out30 ret void31}32 33 34; === HEIGHT =================================================================35; First height at dword index 9+2 -> KC0[2].W36 37; FUNC-LABEL: {{^}}height_2d:38; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]39; EG: MOV * [[VAL]], KC0[2].W40define amdgpu_kernel void @height_2d (ptr addrspace(1) %in,41 ptr addrspace(1) %out) {42entry:43 %0 = call [3 x i32] @llvm.OpenCL.image.get.size.2d(44 ptr addrspace(1) %in) #045 %1 = extractvalue [3 x i32] %0, 146 store i32 %1, ptr addrspace(1) %out47 ret void48}49 50; FUNC-LABEL: {{^}}height_3d:51; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]52; EG: MOV * [[VAL]], KC0[2].W53define amdgpu_kernel void @height_3d (ptr addrspace(1) %in,54 ptr addrspace(1) %out) {55entry:56 %0 = call [3 x i32] @llvm.OpenCL.image.get.size.3d(57 ptr addrspace(1) %in) #058 %1 = extractvalue [3 x i32] %0, 159 store i32 %1, ptr addrspace(1) %out60 ret void61}62 63 64; === DEPTH ==================================================================65; First depth at dword index 9+3 -> KC0[3].X66 67; FUNC-LABEL: {{^}}depth_3d:68; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]69; EG: MOV * [[VAL]], KC0[3].X70define amdgpu_kernel void @depth_3d (ptr addrspace(1) %in,71 ptr addrspace(1) %out) {72entry:73 %0 = call [3 x i32] @llvm.OpenCL.image.get.size.3d(74 ptr addrspace(1) %in) #075 %1 = extractvalue [3 x i32] %0, 276 store i32 %1, ptr addrspace(1) %out77 ret void78}79 80 81; === CHANNEL DATA TYPE ======================================================82; First channel data type at dword index 9+4 -> KC0[3].Y83 84; FUNC-LABEL: {{^}}data_type_2d:85; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]86; EG: MOV * [[VAL]], KC0[3].Y87define amdgpu_kernel void @data_type_2d (ptr addrspace(1) %in,88 ptr addrspace(1) %out) {89entry:90 %0 = call [2 x i32] @llvm.OpenCL.image.get.format.2d(91 ptr addrspace(1) %in) #092 %1 = extractvalue [2 x i32] %0, 093 store i32 %1, ptr addrspace(1) %out94 ret void95}96 97; FUNC-LABEL: {{^}}data_type_3d:98; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]99; EG: MOV * [[VAL]], KC0[3].Y100define amdgpu_kernel void @data_type_3d (ptr addrspace(1) %in,101 ptr addrspace(1) %out) {102entry:103 %0 = call [2 x i32] @llvm.OpenCL.image.get.format.3d(104 ptr addrspace(1) %in) #0105 %1 = extractvalue [2 x i32] %0, 0106 store i32 %1, ptr addrspace(1) %out107 ret void108}109 110 111; === CHANNEL ORDER ==========================================================112; First channel order at dword index 9+5 -> KC0[3].Z113 114; FUNC-LABEL: {{^}}channel_order_2d:115; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]116; EG: MOV * [[VAL]], KC0[3].Z117define amdgpu_kernel void @channel_order_2d (ptr addrspace(1) %in,118 ptr addrspace(1) %out) {119entry:120 %0 = call [2 x i32] @llvm.OpenCL.image.get.format.2d(121 ptr addrspace(1) %in) #0122 %1 = extractvalue [2 x i32] %0, 1123 store i32 %1, ptr addrspace(1) %out124 ret void125}126 127; FUNC-LABEL: {{^}}channel_order_3d:128; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]129; EG: MOV * [[VAL]], KC0[3].Z130define amdgpu_kernel void @channel_order_3d (ptr addrspace(1) %in,131 ptr addrspace(1) %out) {132entry:133 %0 = call [2 x i32] @llvm.OpenCL.image.get.format.3d(134 ptr addrspace(1) %in) #0135 %1 = extractvalue [2 x i32] %0, 1136 store i32 %1, ptr addrspace(1) %out137 ret void138}139 140 141; === 2ND IMAGE ==============================================================142; 9 implicit args + 2 explicit args + 5 implicit args for 1st image argument143; = 16 dwords to 2nd image argument.144; Height of the second image is at 16+2 -> KC0[4].Z145;146; FUNC-LABEL: {{^}}image_arg_2nd:147; EG: MEM_RAT_CACHELESS STORE_RAW [[VAL:T[0-9]+\.X]]148; EG: MOV * [[VAL]], KC0[4].Z149define amdgpu_kernel void @image_arg_2nd (ptr addrspace(1) %in1,150 i32 %x,151 ptr addrspace(1) %in2,152 ptr addrspace(1) %out) {153entry:154 %0 = call [3 x i32] @llvm.OpenCL.image.get.size.2d(155 ptr addrspace(1) %in2) #0156 %1 = extractvalue [3 x i32] %0, 1157 store i32 %1, ptr addrspace(1) %out158 ret void159}160 161%opencl.image2d_t = type opaque162%opencl.image3d_t = type opaque163 164declare [3 x i32] @llvm.OpenCL.image.get.size.2d(ptr addrspace(1)) #0165declare [3 x i32] @llvm.OpenCL.image.get.size.3d(ptr addrspace(1)) #0166declare [2 x i32] @llvm.OpenCL.image.get.format.2d(ptr addrspace(1)) #0167declare [2 x i32] @llvm.OpenCL.image.get.format.3d(ptr addrspace(1)) #0168 169attributes #0 = { readnone }170 171!opencl.kernels = !{!0, !1, !2, !3, !4, !5, !6, !7, !8, !9}172!0 = !{ptr @width_2d,173 !10, !20, !30, !40, !50}174!1 = !{ptr @width_3d,175 !10, !21, !31, !41, !50}176!2 = !{ptr @height_2d,177 !10, !20, !30, !40, !50}178!3 = !{ptr @height_3d,179 !10, !21, !31, !41, !50}180!4 = !{ptr @depth_3d,181 !10, !21, !31, !41, !50}182!5 = !{ptr @data_type_2d,183 !10, !20, !30, !40, !50}184!6 = !{ptr @data_type_3d,185 !10, !21, !31, !41, !50}186!7 = !{ptr @channel_order_2d,187 !10, !20, !30, !40, !50}188!8 = !{ptr @channel_order_3d,189 !10, !21, !31, !41, !50}190!9 = !{ptr @image_arg_2nd, !12, !22, !32, !42, !52}191 192!10 = !{!"kernel_arg_addr_space", i32 1, i32 1}193!20 = !{!"kernel_arg_access_qual", !"read_only", !"none"}194!21 = !{!"kernel_arg_access_qual", !"read_only", !"none"}195!30 = !{!"kernel_arg_type", !"image2d_t", !"int*"}196!31 = !{!"kernel_arg_type", !"image3d_t", !"int*"}197!40 = !{!"kernel_arg_base_type", !"image2d_t", !"int*"}198!41 = !{!"kernel_arg_base_type", !"image3d_t", !"int*"}199!50 = !{!"kernel_arg_type_qual", !"", !""}200 201!12 = !{!"kernel_arg_addr_space", i32 1, i32 0, i32 1, i32 1}202!22 = !{!"kernel_arg_access_qual", !"read_only", !"none", !"write_only", !"none"}203!32 = !{!"kernel_arg_type", !"image3d_t", !"sampler_t", !"image2d_t", !"int*"}204!42 = !{!"kernel_arg_base_type", !"image3d_t", !"sampler_t", !"image2d_t", !"int*"}205!52 = !{!"kernel_arg_type_qual", !"", !"", !"", !""}206