Closed ahsan-ca closed 10 months ago
Ok, so, we'll want a MLIR module input here, ideally (hence my notes about updating MIGRAPHX_TRACE_MLIR to work correctly with quick tuning enabled).
And we'll want to work out if this needs to be backported to 6.0 or not
With MIGRAPHX_TRACE_MLIR=1: 1/3 gemm_2args_mm_5
[ RUN ] gemm_2args_mm_5
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
mlir_dot12:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot12:y1 = @param:y1 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot12:@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot12:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot12:@3 = contiguous(mlir_dot12:@2) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot12:@4 = dot(mlir_dot12:@3,mlir_dot12:y1) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot12:@5 = @return(mlir_dot12:@4), target_id=0
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%1 = migraphx.dot(%0, %arg1) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %1 : tensor<2x3x2x4xf32>
}
}
Benchmarking gpu::mlir_op: 21 configs
Fastest solution: 64,32,32,4,2,2
FAILED: gpu
RMS Error: 0.504296
Max diff: 1.61328
Mismatch at 24: 0.121094 != -0.171875
module: "main"
2 = @param:2 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
1 = @param:1 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](1) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
@3 = dot(@2,2) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
ref:
module: "main"
2 = @param:2 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
1 = @param:1 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](1) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
@3 = dot(@2,2) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
gpu:
module: "main"
@0 = check_context::migraphx::gpu::context -> float_type, {}, {}, target_id=0
2 = @param:2 -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
1 = @param:1 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
output = @param:output -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
@4 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](1) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
@5 = gpu::code_object[code_object=3624,symbol_name=mlir_dot,global=384,local=64,](@4,2,output) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
void run_verify::verify(const std::string &, const migraphx::program &, const migraphx::compile_options &) const
/code/AMDMIGraphX/AMDMIGraphX/test/verify/run_verify.cpp:264:
FAILED: passed [ 0 ]
[ FAILED ] gemm_2args_mm_5: Test failure
With MIGRAPHX_TRACE_MLIR=1: 2/3: gemm_2args_mm_6
[ RUN ] gemm_2args_mm_6
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
mlir_dot13:y1.0 = @param:y1.0 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:y0.0 = @param:y0.0 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
mlir_dot13:@2 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](mlir_dot13:y1.0) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
mlir_dot13:@3 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](mlir_dot13:y0.0) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
mlir_dot13:@4 = contiguous(mlir_dot13:@3) -> float_type, {2, 3, 2, 3}, {18, 6, 3, 1}, target_id=0
mlir_dot13:@5 = contiguous(mlir_dot13:@2) -> float_type, {2, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
mlir_dot13:@6 = dot(mlir_dot13:@4,mlir_dot13:@5) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
mlir_dot13:@7 = @return(mlir_dot13:@6), target_id=0
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1101", kernel = "mixr", num_cu = 27 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
Benchmarking gpu::mlir_op: 21 configs
Fastest solution: 64,64,32,4,4,2
FAILED: gpu
RMS Error: 0.292345
Max diff: 1.59375
Mismatch at 24: 0.582031 != 0
module: "main"
2 = @param:2 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
1 = @param:1 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](1) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
@3 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](2) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
@4 = dot(@2,@3) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
ref:
module: "main"
2 = @param:2 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
1 = @param:1 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
@2 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](1) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
@3 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](2) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
@4 = dot(@2,@3) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
gpu:
module: "main"
@0 = check_context::migraphx::gpu::context -> float_type, {}, {}, target_id=0
2 = @param:2 -> float_type, {1, 3, 3, 4}, {36, 12, 4, 1}, target_id=0
1 = @param:1 -> float_type, {2, 1, 2, 3}, {6, 6, 3, 1}, target_id=0
@3 = multibroadcast[out_lens={2, 3, 3, 4},out_dyn_dims={}](2) -> float_type, {2, 3, 3, 4}, {0, 12, 4, 1}, target_id=0
@4 = multibroadcast[out_lens={2, 3, 2, 3},out_dyn_dims={}](1) -> float_type, {2, 3, 2, 3}, {6, 0, 3, 1}, target_id=0
output = @param:output -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
@6 = gpu::code_object[code_object=3752,symbol_name=mlir_dot,global=384,local=64,](@4,@3,output) -> float_type, {2, 3, 2, 4}, {24, 8, 4, 1}, target_id=0
void run_verify::verify(const std::string &, const migraphx::program &, const migraphx::compile_options &) const
/code/AMDMIGraphX/AMDMIGraphX/test/verify/run_verify.cpp:264:
FAILED: passed [ 0 ]
[ FAILED ] gemm_2args_mm_6: Test failure
With MIGRAPHX_TRACE_MLIR=1: 3/3 test_gemm_transposea_ex
[ RUN ] test_gemm_transposea_ex
test_verify: /usr/local/cget/build/tmp-192f2a4d08984d25a093b9de6fe237ad/rocMLIR-507bb94ce7873786486d296ec81d2eadaab49003/external/llvm-project/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp:1394: static RankedTensorType mlir::tensor::CollapseShapeOp::inferCollapsedType(RankedTensorType, ArrayRef<AffineMap>): Assertion `isReassociationValid(reassociation) && "invalid reassociation"' failed.
Note: Seems like I do not get any trace for this test case. I get an Aborted (core dumped)
message.
Yeah, because we haven't fixed MIGRAPHX_TRACE_MLIR post-quick tuning (@pfultz2 for an example of what problems we're trying to fix) you'll need to patch src/targets/gpu/mlir.cpp
to add a MLIR print into get_tuning_config
before the call to run_highlevel_pipeline
Update, have built on my machine and have (by adding a strategically placed call to dump()
) found an input to tosa-to-rock
that causes one of the crashes.
func.func @mlir_transpose_dot(%arg0: tensor<1x1x5x4xf32>, %arg1: tensor<1x1x5x3xf32>) -> tensor<1x1x4x3xf32> attributes {arch = "gfx1100", kernel = "mixr", num_cu = 48 : i64} { %cst = arith.constant dense<[0, 1, 3, 2]> : tensor<4xi64>
%0 = "tosa.transpose"(%arg0, %cst) : (tensor<1x1x5x4xf32>, tensor<4xi64>) -> tensor<1x1x4x5xf32>
%collapsed = tensor.collapse_shape %0 [[0, 1], [2], [3]] : tensor<1x1x4x5xf32> into tensor<1x4x5xf32>
%collapsed_0 = tensor.collapse_shape %arg1 [[0, 1], [2], [3]] : tensor<1x1x5x3xf32> into tensor<1x5x3xf32>
%1 = "tosa.matmul"(%collapsed, %collapsed_0) : (tensor<1x4x5xf32>, tensor<1x5x3xf32>) -> tensor<1x4x3xf32>
%expanded = tensor.expand_shape %1 [[0, 1], [2], [3]] : tensor<1x4x3xf32> into tensor<1x1x4x3xf32>
return %expanded : tensor<1x1x4x3xf32>
}
This has been further isolated to the collapse_shape handling in TosaToRock . @manupak in case I can't figure out what's going on in there
Update: I have a fix for the crash, but the numeric failures are a seperate issue Here's one of the tests as an example
module { func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1100", kernel = "mixr", num_cu = 48 : i64} { %0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
which lowers (after being sent down the high-level pipeline) to
#map = affine_map<(d0, d1, d2, d3) -> (0, d1, d2, d3)>
#map1 = affine_map<(d0, d1, d2, d3) -> (d0, 0, d2, d3)>
#map2 = affine_map<(d0, d1, d2) -> (d0 floordiv 3, d0 mod 3, d1, d2)>
#map3 = affine_map<(d0, d1, d2, d3) -> (d0 * 3 + d1, d2, d3)>
#transform_map = #rock.transform_map<#map by [<Broadcast{1} ["dim0"] at [0] -> ["dim0"] at [0]>, <PassThrough ["dim1"] at [1] -> ["dim1"] at [1]>, <PassThrough ["dim2"] at [2]
-> ["dim2"] at [2]>, <PassThrough ["dim3"] at [3] -> ["dim3"] at [3]>] bounds = [2, 3, 3
, 4] -> [1, 3, 3, 4]>
#transform_map1 = #rock.transform_map<#map1 by [<PassThrough ["dim0"] at [0] -> ["dim0"] at [0]>, <Broadcast{1} ["dim1"] at [1] -> ["dim1"] at [1]>, <PassThrough ["dim2"] at [2] -> ["dim2"] at [2]>, <PassThrough ["dim3"] at [3] -> ["dim3"] at [3]>] bounds = [2, 3, 2, 3] -> [2, 1, 2, 3]>
#transform_map2 = #rock.transform_map<#map2 by [<Merge{2, 3} ["dim0"] at [0] -> ["col0", "col1"] at [0, 1]>, <PassThrough ["dim1"] at [1] -> ["dim1"] at [2]>, <PassThrough ["dim2"] at [2] -> ["dim2"] at [3]>] bounds = [6, 2, 3] -> [2, 3, 2, 3]> #transform_map3 = #rock.transform_map<#map2 by [<Merge{2, 3} ["dim0"] at [0] -> ["col0", "col1"] at [0, 1]>, <PassThrough ["dim1"] at [1] -> ["dim1"] at [2]>, <PassThrough ["dim2"] at [2] -> ["dim2"] at [3]>] bounds = [6, 3, 4] -> [2, 3, 3, 4]>
#transform_map4 = #rock.transform_map<#map3 by [<Unmerge{2, 3} ["exp0", "exp1"] at [0, 1] -> ["dim0"] at [0]>, <PassThrough ["dim1"] at [2] -> ["dim1"] at [1]>, <PassThrough ["dim2"] at [3] -> ["dim2"] at [2]>] bounds = [2, 3, 2, 4] -> [6, 2, 4]>
module {
func.func @mlir_dot(%arg0: memref<2x1x2x3xf32>, %arg1: memref<1x3x3x4xf32>, %arg2: memref<2x3x2x4xf32>) attributes {arch = "gfx1100", kernel = "mixr", num_cu = 48 : i64} {
%0 = rock.transform %arg1 by #transform_map : memref<1x3x3x4xf32> to memref<2x3x3x4xf32>
%1 = rock.transform %arg0 by #transform_map1 : memref<2x1x2x3xf32> to memref<2x3x2x3xf32>
%2 = rock.transform %1 by #transform_map2 : memref<2x3x2x3xf32> to memref<6x2x3xf32> %3 = rock.transform %0 by #transform_map3 : memref<2x3x3x4xf32> to memref<6x3x4xf32> %alloc = memref.alloc() {alignment = 64 : i64} : memref<6x2x4xf32>
rock.gemm %alloc = %2 * %3 features = dot|atomic_add|atomic_fmax_f32 storeMethod = set {arch = "gfx1100", numCU = 48 : i32} : memref<6x2x4xf32> = memref<6x2x3xf32> * memref<6x3x4xf32>
%4 = rock.transform %alloc by #transform_map4 : memref<6x2x4xf32> to memref<2x3x2x4xf32>
memref.copy %4, %arg2 : memref<2x3x2x4xf32> to memref<2x3x2x4xf32>
return
}
}
which looks ... entirely fine to me? @giuseros since your name got mentioned in the same breath as some sort of numerical issue on Navi
The following module (not lack of "kernel" attribute)
module {
func.func @mlir_dot(%arg0: tensor<2x1x2x3xf32>, %arg1: tensor<1x3x3x4xf32>) -> tensor<2x3x2x4xf32> attributes {arch = "gfx1100", num_cu = 48 : i64} {
%0 = migraphx.multibroadcast(%arg1) {out_dyn_dims = [], out_lens = [2, 3, 3, 4]} : (tensor<1x3x3x4xf32>) -> tensor<2x3x3x4xf32>
%1 = migraphx.multibroadcast(%arg0) {out_dyn_dims = [], out_lens = [2, 3, 2, 3]} : (tensor<2x1x2x3xf32>) -> tensor<2x3x2x3xf32>
%2 = migraphx.dot(%1, %0) : (tensor<2x3x2x3xf32>, tensor<2x3x3x4xf32>) -> tensor<2x3x2x4xf32>
return %2 : tensor<2x3x2x4xf32>
}
}
when run through (and yes, I checked there's actually a GPU kernel in there that's actually compared to an equivalent host kernel)
./bin/rocmlir-driver -kernel-pipeline=migraphx ./broadcast-batch-no-kern.mlir | ./bin/rocmlir-driver -host-pipeline=partition,highlevel --targets gfx1100 | ./bin/rocmlir-gen --arch gfx1100 -pv -fut mlir_dot --verifier clone - | ./bin/rocmlir-driver -kernel-pipeline full -host-pipeline mhal,runner | rocm-run
(replacing the filename and switching gfx1100 for gfx1101 on Navi32 if needed) prints the
[1 1 1]
that shows the kernel operates correctly as far as we're concerned.
Therefore, with regards to the numerical accuracy bugs, please manually examine the GPU and CPU inputs and outputs and confirm that our kernel's output is, in fact, incorrect, ideally with an example.
test_verify test seems to be failing on Navi32. However, on an MI100 system, it passes.
Following errors are seen:
Env details on the failing system: Hostname: ctr-smc-s29-040 OS: "Ubuntu" VERSION="20.04.6 LTS (Focal Fossa)" rocm version on docker (running the test inside docker): 5.7.0-63 rocm version on bare metal machine: 5.6.1-93
Commit MIGraphX: commit 52203fa4bcaac96e390153fb73a0a140034e30c0 (origin/develop, origin/HEAD, develop) Author: Attila Dusnoki 126579622+attila-dusnoki-htec@users.noreply.github.com Date: Wed Oct 25 16:03:08 2023 +0200
MLIR Commit: rocMLIR@a997d5f51314b45d7a4c04f1599966dcf53f9b4d