From e6f1a9ad340282645998ff57c287d306bb088ad9 Mon Sep 17 00:00:00 2001 From: Philippe Tillet Date: Thu, 5 Jan 2023 23:25:41 -0800 Subject: [PATCH] commenting dq but not load/store --- python/bwd.ttgir | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/python/bwd.ttgir b/python/bwd.ttgir index 1a8fd9d6e..f7e9b8bf4 100644 --- a/python/bwd.ttgir +++ b/python/bwd.ttgir @@ -143,10 +143,10 @@ module attributes {"triton_gpu.num-warps" = 8 : i32} { %129 = triton_gpu.convert_layout %90 : (tensor<128x64xf16, #shared1>) -> tensor<128x64xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma1}>> %130 = tt.dot %128, %129, %arg28 {allowTF32 = true} : tensor<128x128xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma1}>> * tensor<128x64xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma1}>> -> tensor<128x64xf32, #mma1> %132 = tt.load %arg29 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<128x64xf32, #mma1> - %133 = triton_gpu.convert_layout %126 : (tensor<128x128xf16, #shared1>) -> tensor<128x128xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma1}>> - %134 = triton_gpu.convert_layout %66 : (tensor<128x64xf16, #shared1>) -> tensor<128x64xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma1}>> - %135 = tt.dot %133, %134, %132 {allowTF32 = true} : tensor<128x128xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma1}>> * tensor<128x64xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma1}>> -> tensor<128x64xf32, #mma1> - tt.store %arg29, %135 : tensor<128x64xf32, #mma1> + //%133 = triton_gpu.convert_layout %126 : (tensor<128x128xf16, #shared1>) -> tensor<128x128xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma1}>> + //%134 = triton_gpu.convert_layout %66 : (tensor<128x64xf16, #shared1>) -> tensor<128x64xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma1}>> + //%135 = tt.dot %133, %134, %132 {allowTF32 = true} : tensor<128x128xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma1}>> * tensor<128x64xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma1}>> -> tensor<128x64xf32, #mma1> + tt.store %arg29, %132 : tensor<128x64xf32, #mma1> %137 = tt.addptr %arg29, %43 : tensor<128x64x!tt.ptr, #mma1>, tensor<128x64xi32, #mma1> %138 = tt.addptr %arg30, %42 : tensor<128x64x!tt.ptr, #blocked1>, tensor<128x64xi32, #blocked1> %139 = tt.addptr %arg31, %42 : tensor<128x64x!tt.ptr, #blocked1>, tensor<128x64xi32, #blocked1>