Skip to content

Instantly share code, notes, and snippets.

View chelini's full-sized avatar
:octocat:

lorenzo chelini chelini

:octocat:
29 followers · 47 following
View GitHub Profile
@chelini
chelini / mlp-with-transpose
Created January 29, 2024 15:47
// BEFORE
#map = affine_map<(d0, d1) -> (d1)>
#map1 = affine_map<(d0, d1) -> (d0, d1)>
module attributes {torch.debug_module_name = "MLP"} {
func.func @MLP(%arg0: tensor<128x262144xf32>, %arg1: tensor<128xf32>, %arg2: tensor<64x512x512xf32>) -> tensor<64x128xf32> {
%cst = arith.constant 0.000000e+00 : f32
%collapsed = tensor.collapse_shape %arg2 [[0], [1, 2]] : tensor<64x512x512xf32> into tensor<64x262144xf32>
%0 = tensor.empty() : tensor<262144x128xf32>
%transposed = linalg.transpose ins(%arg0 : tensor<128x262144xf32>) outs(%0 : tensor<262144x128xf32>) permutation = [1, 0]
@chelini
chelini / gist:75c0c7995e68228f24d0f5eb1233a97a
Last active September 28, 2023 09:50
Clone.mlir
// -one-shot-bufferize="bufferize-function-boundaries function-boundary-type-conversion=identity-layout-map" -expand-realloc -canonicalize -ownership-based-buffer-deallocation -canonicalize
module {
func.func @unpack2(%arg0: tensor<1x2x2x2x2xf32>, %arg1: tensor<1x2x2x4xf32>) -> tensor<1x2x2x4xf32> {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c2 = arith.constant 2 : index
%0 = scf.for %arg2 = %c0 to %c1 step %c1 iter_args(%arg3 = %arg1) -> (tensor<1x2x2x4xf32>) {
@chelini
chelini / unpack_non_dps.md
Last active July 4, 2023 13:43
Unpack non-DPS lowering 
module {
  func.func @unpack(%arg0: tensor<2x2x32x32xf32>, %arg1: tensor<64x64xf32>) -> tensor<64x64xf32> {
    %unpack = tensor.unpack %arg0 inner_dims_pos = [0, 1] inner_tiles = [32, 32] into %arg1 : tensor<2x2x32x32xf32> -> tensor<64x64xf32>
    return %unpack : tensor<64x64xf32>
  }
  transform.sequence  failures(propagate) {
  ^bb0(%arg0: !transform.any_op):
    %0 = transform.structured.match ops{["tensor.unpack"]} in %arg0 : (!transform.any_op) -> !transform.op<"tensor.unpack">
    %empty_op, %transpose_op, %collapse_shape_op, %extract_slice_op = transform.structured.lower_unpack %0 : (!transform.op<"tensor.unpack">) -> (!transform.op<"tensor.empty">, !transform.op<"linalg.transpose">, !transform.op<"tensor.collapse_shape">, !transform.op<"tensor.extract_slice">)
@chelini
chelini / gist:909dd5536d02eafe83746451bd938340
Last active July 2, 2023 14:26
Polygeist.md 
build/bin# cat hello.c                                                    

#include <stdio.h>
int main() {
   printf("Hello, World!");
   return 0;
}
@chelini
chelini / self-attention-full.mlir
Created June 29, 2023 15:56
This file has been truncated, but you can view the full file.
#map = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2)>
#map1 = affine_map<(d0, d1, d2, d3, d4) -> (d2, d3, d4)>
#map2 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d3, d4)>
#map3 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map4 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3)>
#map5 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d4, d2, d3)>
#map6 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d2, d4, d1)>
#map7 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2)>
#map8 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>
#map9 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d3, d1, d4)>
@chelini
chelini / tf.mlir
Created June 27, 2023 07:15
Self attention from TF
This file has been truncated, but you can view the full file.
// tf-opt self-attention-stable-hlo.mlir -tf-lower-to-mlprogram-and-hlo -stablehlo-legalize-to-hlo -hlo-legalize-to-linalg -canonicalize -cse
#map = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2)>
#map1 = affine_map<(d0, d1, d2, d3, d4) -> (d2, d3, d4)>
#map2 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d3, d4)>
#map3 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map4 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3)>
#map5 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d4, d2, d3)>
#map6 = affine_map<(d0, d1, d2, d3, d4) -> (d0, d2, d4, d1)>
#map7 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2)>
@chelini
chelini / self_attention_stablehlo.mlir
Created June 14, 2023 12:44
This file has been truncated, but you can view the full file.
module {
func.func @main(%arg0: tensor<64x32x512xf32>, %arg1: tensor<64x32x512xf32>, %arg2: tensor<64x32x512xf32>) -> tensor<64x32x512xf32> {
%0 = stablehlo.constant dense<1.250000e-01> : tensor<64x32x8x64xf32>
%1 = stablehlo.constant dense<-0.000000e+00> : tensor<f32>
%2 = stablehlo.constant dense<0xFF800000> : tensor<f32>
%3 = stablehlo.constant dense<"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
@chelini
chelini / self_attention_tf_ir.mlir
Created June 14, 2023 12:38
TF IR extracted from a MHA self-attention python layer.
This file has been truncated, but you can view the full file.
module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 1395 : i32}} {
func.func @main(%arg0: tensor<64x32x512xf32>, %arg1: tensor<64x32x512xf32>, %arg2: tensor<64x32x512xf32>) -> tensor<64x32x512xf32> attributes {tf.entry_function = {control_outputs = "", inputs = "query,value,key", outputs = "Identity:0"}} {
%cst = "tf.Const"() {device = "", value = dense<"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
@chelini
chelini / self_attention.md
Last active June 12, 2023 09:14
Self_attention 
// Self attention key = value = query
// value_dim = key_dim


%0 = "tf.Identity"(%cst_3) {_has_manual_control_dependencies = true, device = ""} : (tensor<8x64x512xf32>) -> tensor<8x64x512xf32>
%1 = "tf.Identity"(%cst_1) {_has_manual_control_dependencies = true, device = ""} : (tensor<512x8x64xf32>) -> tensor<512x8x64xf32>
%2 = "tf.Identity"(%cst_0) {_has_manual_control_dependencies = true, device = ""} : (tensor<512x8x64xf32>) -> tensor<512x8x64xf32>
%3 = "tf.Identity"(%cst) {_has_manual_control_dependencies = true, device = ""} : (tensor<512x8x64xf32>) -> tensor<512x8x64xf32>
@chelini
chelini / TransformerEncoder_layer.md
Last active June 8, 2023 07:07
TransformerEncoder layer

Transformer encoder layer in keras:

keras_nlp.layers.TransformerEncoder(
    intermediate_dim,
    num_heads,
    dropout=0,
    activation="relu",
    layer_norm_epsilon=1e-05,
    kernel_initializer="glorot_uniform",