diff --git a/CHANGELOG.md b/CHANGELOG.md index 60c8a01204..314c68a851 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -4,6 +4,8 @@ Each release's page on GitHub has its notes and measurements. See [the 1.0.0 rel ## Unreleased +- HIP gate-5 launch records are re-derived with ROCm 10.1's hipcc, now the one reference toolchain; record values only, no kernel or serving change (#1979). Thanks to @laudney. + - GLM on CUDA runs a prompt chunk's MTP pass on its last row past the caches, about 2.7% faster prompts at TP=2 with identical replies (#1577). Thanks to @gilby. - Flash Next on CUDA runs the MTP head's experts on a call's last row only, the one row that reaches the draft; the head's cache is still written for every row, and replies are unchanged (#1925). Thanks to @jschmied. diff --git a/docs/kernel-contract.md b/docs/kernel-contract.md index cd66bcbf6d..1b3f80d2b0 100644 --- a/docs/kernel-contract.md +++ b/docs/kernel-contract.md @@ -88,7 +88,8 @@ On an AMD machine, run `zig build hip-kernel-contract -Dhip-arch=ARCH -Dhipcc=.. controls are built with the kernels). For the limits of every served chip, `zig build hip-kernel-contract-limits -Dhipcc=...` compiles every kernel group and the controls for gfx1030, gfx1100, gfx1151 and gfx1201 and checks each record against them, no GPU needed; `hip-kernel-contract-objects` installs those code objects into `zig-out/hsaco/`. -Its `RECORD` lines are the values `hip_records.zig` holds for that hipcc. +Its `RECORD` lines are the values `hip_records.zig` holds for that hipcc. The records hold ROCm 10.1's hipcc (HIP 7.16), +which ROCm 10.1 installs under `core/bin` of its root. Every run ends with a `SUMMARY` line: kernels passing gates 1 to 4, known findings, controls red, and limit problems. The program exits 1 on any failed gate, any control that stayed green, or any limit problem. diff --git a/zig/src/core/contract/hip_records.zig b/zig/src/core/contract/hip_records.zig index 0b339e28ca..b587c54d84 100644 --- a/zig/src/core/contract/hip_records.zig +++ b/zig/src/core/contract/hip_records.zig @@ -65,10 +65,10 @@ const own_records = [_]Record{ .{ .launch = .{ .name = "mamba tf_mamba_conv", .symbol = "tf_mamba_conv", .image = "mamba", .threads = 256 }, .usage = .{ - .{ .vgprs = 27, .sgprs = 26, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 26, .sgprs = 26, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 26, .sgprs = 26, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 26, .sgprs = 26, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 27, .sgprs = 27, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 26, .sgprs = 27, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 26, .sgprs = 27, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 26, .sgprs = 30, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, }, }, .{ @@ -83,18 +83,18 @@ const own_records = [_]Record{ .{ .launch = .{ .name = "mamba tf_mamba_scan", .symbol = "tf_mamba_scan", .image = "mamba", .threads = 128 }, .usage = .{ - .{ .vgprs = 127, .sgprs = 41, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 126, .sgprs = 41, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 125, .sgprs = 42, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 121, .sgprs = 42, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 127, .sgprs = 42, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 126, .sgprs = 42, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 125, .sgprs = 43, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 121, .sgprs = 43, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = .{ .name = "mamba tf_mamba_gnorm", .symbol = "tf_mamba_gnorm", .image = "mamba", .threads = 128 }, .usage = .{ - .{ .vgprs = 23, .sgprs = 23, .lds = 16, .private = 0, .bound = 128, .wave = 32 }, .{ .vgprs = 23, .sgprs = 24, .lds = 16, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 23, .sgprs = 23, .lds = 16, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 23, .sgprs = 24, .lds = 16, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 23, .sgprs = 24, .lds = 16, .private = 0, .bound = 128, .wave = 32 }, .{ .vgprs = 23, .sgprs = 24, .lds = 16, .private = 0, .bound = 128, .wave = 32 }, }, }, @@ -205,10 +205,10 @@ const own_records = [_]Record{ .threads = 256, }, .usage = .{ - .{ .vgprs = 57, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 57, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 57, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 57, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 42, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 42, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 42, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 42, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, }, }, .{ @@ -219,10 +219,10 @@ const own_records = [_]Record{ .threads = 256, }, .usage = .{ - .{ .vgprs = 57, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 57, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 57, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 57, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 42, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 42, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 42, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 42, .sgprs = 44, .lds = 8192, .private = 0, .bound = 256, .wave = 32 }, }, }, .{ @@ -256,19 +256,19 @@ const own_records = [_]Record{ .{ .launch = .{ .name = "add rms tf_add_rms_f16", .symbol = "tf_add_rms_f16", .image = "act", .threads = 1024 }, .usage = .{ - .{ .vgprs = 43, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 45, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 45, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 44, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 50, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 51, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 51, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 51, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, }, }, .{ .launch = .{ .name = "add rms tf_add_rms_bf16", .symbol = "tf_add_rms_bf16", .image = "act", .threads = 1024 }, .usage = .{ - .{ .vgprs = 44, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 45, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 45, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 44, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 51, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 51, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 51, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 51, .sgprs = 40, .lds = 32772, .private = 0, .bound = 1024, .wave = 32 }, }, }, .{ @@ -279,10 +279,10 @@ const own_records = [_]Record{ .threads = 256, }, .usage = .{ - .{ .vgprs = 32, .sgprs = 42, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 38, .sgprs = 42, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 37, .sgprs = 42, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 37, .sgprs = 44, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 32, .sgprs = 44, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 32, .sgprs = 44, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 31, .sgprs = 44, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 31, .sgprs = 45, .lds = 0, .private = 0, .bound = 256, .wave = 32 }, }, }, .{ @@ -293,82 +293,82 @@ const own_records = [_]Record{ .threads = 128, }, .usage = .{ - .{ .vgprs = 137, .sgprs = 50, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 136, .sgprs = 50, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 135, .sgprs = 51, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 137, .sgprs = 51, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 51, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 135, .sgprs = 52, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, .{ .vgprs = 131, .sgprs = 52, .lds = 20544, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = relu2Tile(1, 2), .usage = .{ - .{ .vgprs = 63, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 65, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 65, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 65, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = relu2Tile(1, 4), .usage = .{ - .{ .vgprs = 101, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = relu2Tile(1, 8), .usage = .{ - .{ .vgprs = 181, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 178, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 178, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 178, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 180, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = relu2Tile(2, 2), .usage = .{ - .{ .vgprs = 75, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 91, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 66, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 66, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 66, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = relu2Tile(2, 4), .usage = .{ - .{ .vgprs = 93, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 92, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = relu2Tile(2, 8), .usage = .{ - .{ .vgprs = 157, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 158, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 158, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 158, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = relu2Tile(4, 2), .usage = .{ - .{ .vgprs = 113, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 114, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 114, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 114, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = relu2Tile(4, 4), .usage = .{ - .{ .vgprs = 147, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 146, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ @@ -473,10 +473,10 @@ const own_records = [_]Record{ .threads = 1024, }, .usage = .{ - .{ .vgprs = 38, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 38, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 38, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 38, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 37, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 37, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 37, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 37, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, }, }, .{ @@ -488,7 +488,7 @@ const own_records = [_]Record{ }, .usage = .{ .{ .vgprs = 36, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, - .{ .vgprs = 37, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, + .{ .vgprs = 36, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, .{ .vgprs = 36, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, .{ .vgprs = 36, .sgprs = 32, .lds = 50688, .private = 0, .bound = 1024, .wave = 32 }, }, diff --git a/zig/src/core/contract/hip_records_pair.zig b/zig/src/core/contract/hip_records_pair.zig index 117d572b5d..1ad358ce80 100644 --- a/zig/src/core/contract/hip_records_pair.zig +++ b/zig/src/core/contract/hip_records_pair.zig @@ -21,73 +21,73 @@ pub const records = [_]Record{ .{ .launch = pairTile(1, 2), .usage = .{ - .{ .vgprs = 101, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 117, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 117, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 117, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 129, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 129, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 129, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = pairTile(1, 4), .usage = .{ - .{ .vgprs = 180, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 172, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 172, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 172, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 179, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 210, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 210, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 210, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = pairTile(1, 8), .usage = .{ - .{ .vgprs = 181, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 178, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 178, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 178, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 180, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = pairTile(2, 2), .usage = .{ - .{ .vgprs = 123, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 152, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 152, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 152, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 95, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 143, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 143, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 143, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = pairTile(2, 4), .usage = .{ - .{ .vgprs = 177, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 236, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 236, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 236, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 235, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 235, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 235, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = pairTile(2, 8), .usage = .{ - .{ .vgprs = 157, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 159, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 159, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 159, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 157, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 157, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 157, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 157, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = pairTile(4, 2), .usage = .{ - .{ .vgprs = 194, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 185, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 185, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 185, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 193, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 183, .sgprs = 77, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 183, .sgprs = 77, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 183, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = pairTile(4, 4), .usage = .{ - .{ .vgprs = 148, .sgprs = 77, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 80, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 147, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 82, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 82, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 80, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, }; diff --git a/zig/src/core/contract/hip_records_prompt.zig b/zig/src/core/contract/hip_records_prompt.zig index dbf91f4871..e737e37650 100644 --- a/zig/src/core/contract/hip_records_prompt.zig +++ b/zig/src/core/contract/hip_records_prompt.zig @@ -12,27 +12,27 @@ fn promptTile(comptime symbol: []const u8, usage: [limits.rdna.len]U) Record { pub const records = [_]Record{ promptTile("tf_fa_wide_f16", .{ - .{ .vgprs = 231, .sgprs = 45, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 231, .sgprs = 45, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 231, .sgprs = 45, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 231, .sgprs = 47, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 230, .sgprs = 47, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 230, .sgprs = 47, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 230, .sgprs = 47, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 230, .sgprs = 49, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, }), promptTile("tf_fa_wide_bf16", .{ - .{ .vgprs = 247, .sgprs = 45, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 231, .sgprs = 45, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 231, .sgprs = 45, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 231, .sgprs = 47, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 246, .sgprs = 46, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 230, .sgprs = 47, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 230, .sgprs = 47, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 230, .sgprs = 49, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, }), promptTile("tf_fa_paged_f16", .{ - .{ .vgprs = 232, .sgprs = 51, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 232, .sgprs = 51, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 232, .sgprs = 51, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 232, .sgprs = 51, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 231, .sgprs = 53, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 231, .sgprs = 53, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 231, .sgprs = 53, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 231, .sgprs = 53, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, }), promptTile("tf_fa_paged_bf16", .{ - .{ .vgprs = 250, .sgprs = 50, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 232, .sgprs = 51, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 232, .sgprs = 51, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, - .{ .vgprs = 232, .sgprs = 51, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 249, .sgprs = 52, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 231, .sgprs = 53, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 231, .sgprs = 53, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, + .{ .vgprs = 231, .sgprs = 53, .lds = 40960, .private = 0, .bound = 256, .wave = 32 }, }), }; diff --git a/zig/src/core/contract/hip_records_stream_plain.zig b/zig/src/core/contract/hip_records_stream_plain.zig index 795d84751a..2891aa9a25 100644 --- a/zig/src/core/contract/hip_records_stream_plain.zig +++ b/zig/src/core/contract/hip_records_stream_plain.zig @@ -6,433 +6,433 @@ pub const records = [_]Record{ .{ .launch = tile(2, 1, 2, false), .usage = .{ - .{ .vgprs = 53, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 55, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 55, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 55, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 52, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 53, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 53, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 53, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 1, 4, false), .usage = .{ - .{ .vgprs = 93, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 94, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 94, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 94, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 92, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 92, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 92, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 92, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 1, 8, false), .usage = .{ - .{ .vgprs = 153, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 2, 2, false), .usage = .{ - .{ .vgprs = 62, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 64, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 64, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 64, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 61, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 2, 4, false), .usage = .{ - .{ .vgprs = 97, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 98, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 98, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 98, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 96, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 96, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 96, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 96, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 2, 8, false), .usage = .{ - .{ .vgprs = 141, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 142, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 142, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 142, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 140, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 140, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 140, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 140, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 4, 2, false), .usage = .{ - .{ .vgprs = 113, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 111, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 111, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 111, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 118, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 118, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 118, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 4, 4, false), .usage = .{ - .{ .vgprs = 141, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 142, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 142, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 142, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 138, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 138, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 138, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 138, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 1, 2, false), .usage = .{ - .{ .vgprs = 128, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 113, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 113, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 113, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 127, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 124, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 124, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 124, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 1, 4, false), .usage = .{ - .{ .vgprs = 92, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 93, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 93, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 93, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 91, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 91, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 91, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 91, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 1, 8, false), .usage = .{ - .{ .vgprs = 170, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 161, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 161, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 161, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 159, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 159, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 159, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 159, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 2, 2, false), .usage = .{ - .{ .vgprs = 102, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 122, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 122, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 122, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 101, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 105, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 106, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 106, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 2, 4, false), .usage = .{ - .{ .vgprs = 96, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 90, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 90, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 90, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 95, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 88, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 88, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 88, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 2, 8, false), .usage = .{ - .{ .vgprs = 137, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 138, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 138, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 138, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 4, 2, false), .usage = .{ - .{ .vgprs = 89, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 88, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 4, 4, false), .usage = .{ - .{ .vgprs = 131, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 134, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 134, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 134, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 130, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 133, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 133, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 133, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 1, 2, false), .usage = .{ - .{ .vgprs = 63, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 65, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 65, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 65, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 1, 4, false), .usage = .{ - .{ .vgprs = 101, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 1, 8, false), .usage = .{ - .{ .vgprs = 181, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 178, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 178, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 178, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 180, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 176, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 2, 2, false), .usage = .{ - .{ .vgprs = 75, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 91, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 66, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 66, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 66, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 2, 4, false), .usage = .{ - .{ .vgprs = 93, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 92, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 2, 8, false), .usage = .{ - .{ .vgprs = 157, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 158, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 158, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 158, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 4, 2, false), .usage = .{ - .{ .vgprs = 113, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 114, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 114, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 114, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 4, 4, false), .usage = .{ - .{ .vgprs = 147, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 150, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 146, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 1, 2, false), .usage = .{ - .{ .vgprs = 130, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 117, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 117, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 117, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 129, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 128, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 128, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 128, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 1, 4, false), .usage = .{ - .{ .vgprs = 108, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 109, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 109, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 109, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 107, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 107, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 107, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 107, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 1, 8, false), .usage = .{ - .{ .vgprs = 201, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 193, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 193, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 193, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 191, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 191, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 191, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 191, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 2, 2, false), .usage = .{ - .{ .vgprs = 90, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 101, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 101, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 101, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 80, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 2, 4, false), .usage = .{ - .{ .vgprs = 88, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 98, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 98, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 98, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 87, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 96, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 96, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 96, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 2, 8, false), .usage = .{ - .{ .vgprs = 153, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 162, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 162, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 162, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 161, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 161, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 161, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 4, 2, false), .usage = .{ - .{ .vgprs = 92, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 91, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 103, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 103, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 103, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 4, 4, false), .usage = .{ - .{ .vgprs = 153, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 142, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 142, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 142, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 140, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 140, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 140, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 1, 2, false), .usage = .{ - .{ .vgprs = 79, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 69, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 69, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 69, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 63, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 1, 4, false), .usage = .{ - .{ .vgprs = 116, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 117, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 117, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 117, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 115, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 115, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 115, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 115, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 1, 8, false), .usage = .{ - .{ .vgprs = 220, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 209, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 209, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 209, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 219, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 214, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 214, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 214, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 2, 2, false), .usage = .{ - .{ .vgprs = 71, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 72, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 72, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 72, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 70, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 110, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 109, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 109, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 2, 4, false), .usage = .{ - .{ .vgprs = 101, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 102, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 100, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 2, 8, false), .usage = .{ - .{ .vgprs = 173, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 162, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 162, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 162, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 172, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 160, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 160, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 160, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 4, 2, false), .usage = .{ - .{ .vgprs = 106, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 112, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 112, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 112, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 105, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 110, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 110, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 110, .sgprs = 77, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 4, 4, false), .usage = .{ - .{ .vgprs = 162, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 170, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 170, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 170, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 161, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 162, .sgprs = 77, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 162, .sgprs = 77, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 162, .sgprs = 79, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 1, 2, false), .usage = .{ - .{ .vgprs = 78, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 77, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 77, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 77, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 74, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 71, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 71, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 71, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 1, 4, false), .usage = .{ - .{ .vgprs = 133, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 134, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 134, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 134, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 132, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 132, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 132, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 132, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 1, 8, false), .usage = .{ - .{ .vgprs = 226, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 256, .sgprs = 62, .lds = 0, .private = 124, .bound = 128, .wave = 32 }, - .{ .vgprs = 256, .sgprs = 62, .lds = 0, .private = 124, .bound = 128, .wave = 32 }, - .{ .vgprs = 256, .sgprs = 64, .lds = 0, .private = 124, .bound = 128, .wave = 32 }, + .{ .vgprs = 225, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 238, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 238, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 238, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 2, 2, false), .usage = .{ - .{ .vgprs = 83, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 75, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 75, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 75, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 90, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 2, 4, false), .usage = .{ - .{ .vgprs = 113, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 114, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 114, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 114, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 112, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 2, 8, false), .usage = .{ - .{ .vgprs = 189, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 190, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 190, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 190, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 188, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 188, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 188, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 188, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 4, 2, false), .usage = .{ - .{ .vgprs = 128, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 125, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 125, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 125, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 127, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 124, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 124, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 124, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 4, 4, false), .usage = .{ - .{ .vgprs = 163, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 164, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 164, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 164, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 162, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 162, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 162, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 162, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, }; diff --git a/zig/src/core/contract/hip_records_stream_swiglu.zig b/zig/src/core/contract/hip_records_stream_swiglu.zig index 44da49994d..033390e3ae 100644 --- a/zig/src/core/contract/hip_records_stream_swiglu.zig +++ b/zig/src/core/contract/hip_records_stream_swiglu.zig @@ -6,433 +6,433 @@ pub const records = [_]Record{ .{ .launch = tile(2, 1, 2, true), .usage = .{ - .{ .vgprs = 52, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 54, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 54, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 54, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 51, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 52, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 52, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 52, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 1, 4, true), .usage = .{ - .{ .vgprs = 96, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 88, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 88, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 88, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 95, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 86, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 86, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 86, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 1, 8, true), .usage = .{ - .{ .vgprs = 157, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 154, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 154, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 154, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 2, 2, true), .usage = .{ - .{ .vgprs = 62, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 64, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 64, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 64, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 61, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 2, 4, true), .usage = .{ - .{ .vgprs = 96, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 92, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 92, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 92, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 95, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 94, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 94, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 94, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 2, 8, true), .usage = .{ - .{ .vgprs = 137, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 138, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 138, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 138, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 4, 2, true), .usage = .{ - .{ .vgprs = 112, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 111, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 111, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 111, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 111, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 118, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 118, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 118, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(2, 4, 4, true), .usage = .{ - .{ .vgprs = 137, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 140, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 140, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 140, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 136, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 1, 2, true), .usage = .{ - .{ .vgprs = 127, .sgprs = 49, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 112, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 112, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 112, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 126, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 123, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 123, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 123, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 1, 4, true), .usage = .{ - .{ .vgprs = 90, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 91, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 91, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 91, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 89, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 89, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 89, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 89, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 1, 8, true), .usage = .{ - .{ .vgprs = 174, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 165, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 165, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 165, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 173, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 163, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 163, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 163, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 2, 2, true), .usage = .{ - .{ .vgprs = 102, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 122, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 122, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 122, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 101, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 105, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 106, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 106, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 2, 4, true), .usage = .{ - .{ .vgprs = 94, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 88, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 88, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 88, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 93, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 86, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 86, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 86, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 2, 8, true), .usage = .{ - .{ .vgprs = 133, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 134, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 134, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 134, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 132, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 132, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 132, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 132, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 4, 2, true), .usage = .{ - .{ .vgprs = 89, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 88, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(3, 4, 4, true), .usage = .{ - .{ .vgprs = 110, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 132, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 132, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 132, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 109, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 131, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 131, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 131, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 1, 2, true), .usage = .{ - .{ .vgprs = 58, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 60, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 60, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 60, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 54, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 54, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 54, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 54, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 1, 4, true), .usage = .{ - .{ .vgprs = 99, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 100, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 100, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 100, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 98, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 98, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 98, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 98, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 1, 8, true), .usage = .{ - .{ .vgprs = 185, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 182, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 182, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 182, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 184, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 180, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 180, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 180, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 2, 2, true), .usage = .{ - .{ .vgprs = 91, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 90, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 66, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 66, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 66, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 2, 4, true), .usage = .{ - .{ .vgprs = 91, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 100, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 100, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 100, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 90, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 98, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 98, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 98, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 2, 8, true), .usage = .{ - .{ .vgprs = 153, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 154, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 154, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 154, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 152, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 4, 2, true), .usage = .{ - .{ .vgprs = 112, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 113, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 113, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 113, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 111, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 111, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 111, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 111, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(4, 4, 4, true), .usage = .{ - .{ .vgprs = 145, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 148, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 148, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 148, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 144, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 146, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 146, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 146, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 1, 2, true), .usage = .{ - .{ .vgprs = 129, .sgprs = 49, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 116, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 116, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 116, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 128, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 127, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 127, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 127, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 1, 4, true), .usage = .{ - .{ .vgprs = 106, .sgprs = 50, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 105, .sgprs = 51, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 105, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 105, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 105, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 1, 8, true), .usage = .{ - .{ .vgprs = 206, .sgprs = 52, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 197, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 197, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 197, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 205, .sgprs = 53, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 202, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 202, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 202, .sgprs = 54, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 2, 2, true), .usage = .{ - .{ .vgprs = 90, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 101, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 101, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 101, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 80, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 2, 4, true), .usage = .{ - .{ .vgprs = 86, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 96, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 96, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 96, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 85, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 94, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 94, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 94, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 2, 8, true), .usage = .{ - .{ .vgprs = 149, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 158, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 158, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 158, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 148, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 157, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 157, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 157, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 4, 2, true), .usage = .{ - .{ .vgprs = 92, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 107, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 91, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 103, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 103, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 103, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(5, 4, 4, true), .usage = .{ - .{ .vgprs = 151, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 140, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 140, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 140, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 150, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 138, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 138, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 138, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 1, 2, true), .usage = .{ - .{ .vgprs = 78, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 68, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 62, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 1, 4, true), .usage = .{ - .{ .vgprs = 114, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 115, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 115, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 115, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 113, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 113, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 113, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 113, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 1, 8, true), .usage = .{ - .{ .vgprs = 224, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 213, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 213, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 213, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 223, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 214, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 214, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 214, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 2, 2, true), .usage = .{ - .{ .vgprs = 71, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 72, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 72, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 72, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 70, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 108, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 108, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 108, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 2, 4, true), .usage = .{ - .{ .vgprs = 99, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 100, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 100, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 100, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 98, .sgprs = 67, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 116, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 116, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 116, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 2, 8, true), .usage = .{ - .{ .vgprs = 157, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 170, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 170, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 170, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 156, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 169, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 169, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 169, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 4, 2, true), .usage = .{ - .{ .vgprs = 105, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 111, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 111, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 111, .sgprs = 77, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 104, .sgprs = 73, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 109, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 109, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 109, .sgprs = 79, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(6, 4, 4, true), .usage = .{ - .{ .vgprs = 150, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 164, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 164, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 164, .sgprs = 79, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 149, .sgprs = 75, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 142, .sgprs = 80, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 142, .sgprs = 80, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 142, .sgprs = 81, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 1, 2, true), .usage = .{ - .{ .vgprs = 77, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 76, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 76, .sgprs = 56, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 76, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 73, .sgprs = 55, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 70, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 70, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 70, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 1, 4, true), .usage = .{ - .{ .vgprs = 131, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 132, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 132, .sgprs = 58, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 132, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 130, .sgprs = 57, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 130, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 130, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 130, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 1, 8, true), .usage = .{ - .{ .vgprs = 230, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 231, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 231, .sgprs = 60, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 231, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 229, .sgprs = 59, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 229, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 229, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 229, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 2, 2, true), .usage = .{ - .{ .vgprs = 90, .sgprs = 61, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 75, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 75, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 75, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 89, .sgprs = 62, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 102, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 2, 4, true), .usage = .{ - .{ .vgprs = 111, .sgprs = 63, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 112, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 112, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 112, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 110, .sgprs = 64, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 110, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 110, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 110, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 2, 8, true), .usage = .{ - .{ .vgprs = 207, .sgprs = 65, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 186, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 186, .sgprs = 68, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 186, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 206, .sgprs = 66, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 198, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 198, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 198, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 4, 2, true), .usage = .{ - .{ .vgprs = 128, .sgprs = 69, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 124, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 124, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 124, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 127, .sgprs = 70, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 123, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 123, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 123, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, .{ .launch = tile(8, 4, 4, true), .usage = .{ - .{ .vgprs = 161, .sgprs = 71, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 162, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, - .{ .vgprs = 162, .sgprs = 74, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 160, .sgprs = 72, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 162, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, .{ .vgprs = 162, .sgprs = 76, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, + .{ .vgprs = 162, .sgprs = 78, .lds = 0, .private = 0, .bound = 128, .wave = 32 }, }, }, };