| ... | ... | @@ -25,7 +25,7 @@ pub fn addCases(ctx: *TestContext) !void { |
| 25 | 25 | case.compiles( |
| 26 | 26 | \\fn threadIdX() u32 { |
| 27 | 27 | \\ return asm ("mov.u32 \t%[r], %tid.x;" |
| 28 | | \\ : [r] "=r" (-> utid), |
| 28 | \\ : [r] "=r" (-> u32), |
| 29 | 29 | \\ ); |
| 30 | 30 | \\} |
| 31 | 31 | \\ |
| ... | ... | @@ -54,12 +54,12 @@ pub fn addCases(ctx: *TestContext) !void { |
| 54 | 54 | case.compiles( |
| 55 | 55 | \\fn threadIdX() u32 { |
| 56 | 56 | \\ return asm ("mov.u32 \t%[r], %tid.x;" |
| 57 | | \\ : [r] "=r" (-> utid), |
| 57 | \\ : [r] "=r" (-> u32), |
| 58 | 58 | \\ ); |
| 59 | 59 | \\} |
| 60 | 60 | \\ |
| 61 | 61 | \\ var _sdata: [1024]f32 addrspace(.shared) = undefined; |
| 62 | | \\ pub export fn reduceSum(d_x: []const f32, out: *f32) callconv(ptx.Kernel) void { |
| 62 | \\ pub export fn reduceSum(d_x: []const f32, out: *f32) callconv(.PtxKernel) void { |
| 63 | 63 | \\ var sdata = @addrSpaceCast(.generic, &_sdata); |
| 64 | 64 | \\ const tid: u32 = threadIdX(); |
| 65 | 65 | \\ var sum = d_x[tid]; |
| ... | ... | @@ -99,6 +99,8 @@ pub fn addPtx( |
| 99 | 99 | .files = std.ArrayList(TestContext.File).init(ctx.cases.allocator), |
| 100 | 100 | .link_libc = false, |
| 101 | 101 | .backend = .llvm, |
| 102 | // Bug in Debug mode |
| 103 | .optimize_mode = .ReleaseSafe, |
| 102 | 104 | }) catch @panic("out of memory"); |
| 103 | 105 | return &ctx.cases.items[ctx.cases.items.len - 1]; |
| 104 | 106 | } |