PR #17440: Inject desired pattern for handling Transpose for fp8 gemm…

… rewrite Imported from GitHub PR #17440 Related to #17276 and #16975. This PR updates the GemmRewriter to handle the transpose of non-descending layouts directly, eliminating the need for the layout_normalization pass to correct this error-prone pattern post-rewrite. The desired transformation is now injected into GemmRewriter, ensuring the problematic layout is handled internally. This PR transforms the following error-prone pattern, where the transpose of a non-descending layout is the issue: ``` a = f8e4m3fn[x,y]{0,1} xxx transpose.0 = f8e4m3fn[y,x]{0,1} transpose(a), dimensions=(1,0) custom-call(a,...) ``` to ``` a = f8e4m3fn[x,y]{0,1} xxx bt = f8e4m3fn[y,x]{1,0} bitcast(a) transpose.1 = f8e4m3fn[x,y]{1,0} transpose(bt), dimensions=(1,0) bt.1= f8e4m3fn[y,x]{0,1} bitcast(transpose.1) custom-call(bt.1,...) ``` Copybara import of the project: -- 237c032 by shuw <[email protected]>: Improve TransposeMatrix -- 508cd69 by Shu Wang <[email protected]>: Fix bug of permutation. -- c55e8a9 by shuw <[email protected]>: clang format -- ad0a4ba by Shu Wang <[email protected]>: Add unittest. -- 1d45b4d by Shu Wang <[email protected]>: Remove uncessary space. -- 7837845 by Shu Wang <[email protected]>: Update unittest. -- b479c21 by shuw <[email protected]>: Improve TransposeMatrix Merging this change closes #17440 FUTURE_COPYBARA_INTEGRATE_REVIEW=#17440 from wenscarl:fp8_regulate_transpose b479c21 PiperOrigin-RevId: 680886834
openxla · Oct 1, 2024 · d6f3a5e · d6f3a5e
1 parent e5fe397
commit d6f3a5e
Show file tree

Hide file tree

Showing 2 changed files with 95 additions and 8 deletions.
diff --git a/xla/service/gpu/transforms/gemm_rewriter.cc b/xla/service/gpu/transforms/gemm_rewriter.cc
@@ -48,6 +48,7 @@ limitations under the License.
 #include "xla/layout.h"
 #include "xla/literal.h"
 #include "xla/literal_util.h"
+#include "xla/permutation_util.h"
 #include "xla/primitive_util.h"
 #include "xla/service/algorithm_util.h"
 #include "xla/service/gpu/backend_configs.pb.h"
@@ -362,27 +363,61 @@ std::optional<MatchedFp8Param> MatchFp8Param(HloInstruction *instr) {
 // dimension. Keeps the layout the same.
 HloInstruction *TransposeMatrix(HloInstruction *instr, int64_t contracting_dim,
                                 absl::Span<const int64_t> batch_dims) {
+  auto input_shape = instr->shape();
   // Identify the dimensional order which describes a transpose of the
   // contracting and non-contracting dimensions of the GEMM.
-  std::vector<int64_t> permutation(instr->shape().dimensions_size(), -1);
+  std::vector<int64_t> permutation(input_shape.dimensions_size(), -1);
   // Discard the batch dimensions.
   for (int64_t batch_dim : batch_dims) {
     permutation[batch_dim] = batch_dim;
   }
   // Identify the non-contracting dimension.
   int non_contracting_dim;
-  for (int i = 0; i < instr->shape().dimensions_size(); ++i) {
+  for (int i = 0; i < input_shape.dimensions_size(); ++i) {
     if (permutation[i] == -1 && contracting_dim != i) {
       non_contracting_dim = i;
     }
   }
-  permutation[non_contracting_dim] = contracting_dim;
-  permutation[contracting_dim] = non_contracting_dim;
 
-  Shape new_shape = ShapeUtil::PermuteDimensions(permutation, instr->shape());
-  *new_shape.mutable_layout() = instr->shape().layout();
-  return instr->AddInstruction(
-      HloInstruction::CreateTranspose(new_shape, instr, permutation));
+  if (Layout::Equal()(input_shape.layout(),
+                      LayoutUtil::GetDefaultLayoutForShape(input_shape))) {
+    permutation[non_contracting_dim] = contracting_dim;
+    permutation[contracting_dim] = non_contracting_dim;
+
+    Shape new_shape = ShapeUtil::PermuteDimensions(permutation, input_shape);
+    *new_shape.mutable_layout() = input_shape.layout();
+
+    return instr->AddInstruction(
+        HloInstruction::CreateTranspose(new_shape, instr, permutation));
+  }
+
+  Shape normalized_input_shape =
+      ShapeUtil::MakeShapeWithDescendingLayoutAndSamePhysicalLayout(
+          input_shape);
+  auto a0 = MakeBitcastHlo(instr, normalized_input_shape);
+
+  std::vector<int64_t> layout_permuation(
+      input_shape.layout().minor_to_major().begin(),
+      input_shape.layout().minor_to_major().end());
+  absl::c_reverse(layout_permuation);
+  auto inv_perm = InversePermutation(layout_permuation);
+
+  int new_contracting_dim = inv_perm[contracting_dim];
+  int new_non_contracting_dim = inv_perm[non_contracting_dim];
+  absl::c_iota(permutation, 0);
+  std::swap(permutation[new_contracting_dim],
+            permutation[new_non_contracting_dim]);
+
+  Shape transpose_shape =
+      ShapeUtil::PermuteDimensions(permutation, a0->shape());
+  *transpose_shape.mutable_layout() = a0->shape().layout();
+
+  HloInstruction *normalized_transpose = instr->AddInstruction(
+      HloInstruction::CreateTranspose(transpose_shape, a0, permutation));
+
+  Shape final_shape = ShapeUtil::PermuteDimensions(inv_perm, transpose_shape);
+  *final_shape.mutable_layout() = input_shape.layout();
+  return MakeBitcastHlo(normalized_transpose, final_shape);
 }
 
 // If the bias is a sequence of ops that depend only on broadcasts of

diff --git a/xla/service/gpu/transforms/gemm_rewriter_test.cc b/xla/service/gpu/transforms/gemm_rewriter_test.cc
@@ -5032,6 +5032,58 @@ TEST_P(ParameterizedFp8GemmRewriteTest, UnscaledABUnscaledDMatrixBiasF8) {
       )");
 }
 
+TEST_P(ParameterizedFp8GemmRewriteTest, ScaledABUnscaledDColMajorLhsF8) {
+  const char* hlo_text = R"(
+HloModule test
+    ENTRY test {
+      x = <<F8E4M3>>[2,16,32]{1,0,2} parameter(0)
+      y = <<F8E4M3>>[2,32,16]{2,1,0} parameter(1)
+      x_scale = f32[] parameter(2)
+      y_scale = f32[] parameter(3)
+      dq_scale = f32[] multiply(x_scale, y_scale)
+      dq_scale_bcast = f32[2,16,16] broadcast(dq_scale), dimensions={}
+      out.0 = f32[2,16,16] dot(x, y), lhs_contracting_dims={2}, rhs_contracting_dims={1}, lhs_batch_dims={0}, rhs_batch_dims={0}
+      ROOT out = f32[2,16,16] multiply(out.0, dq_scale_bcast)
+          }
+)";
+
+  CheckFp8IfSupported(hlo_text);
+  RunAndFilecheckHloRewrite(
+      hlo_text,
+      GemmRewriter(CudaHopperOrRocmMI300(), GetToolkitVersion(),
+                   GemmRewriterOptions{GemmRewriterOptions::DType::kFp8Only}),
+      R"(
+; CHECK-LABEL: ENTRY %test ({{.*}}: <<F8E4M3>>[2,16,32], {{.*}}: <<F8E4M3>>[2,32,16], {{.*}}: f32[], {{.*}}: f32[]) -> f32[2,16,16] {
+; CHECK-NEXT:    [[P0:%[^ ]+]] = <<F8E4M3>>[2,16,32]{1,0,2} parameter(0)
+; CHECK-NEXT:    [[P0_BT:%[^ ]+]] = <<F8E4M3>>[32,2,16]{2,1,0} bitcast([[P0]])
+; CHECK-NEXT:    [[P0_TR:%[^ ]+]] = <<F8E4M3>>[16,2,32]{2,1,0} transpose([[P0_BT]]), dimensions={2,1,0}
+; CHECK-NEXT:    [[P0_BT1:%[^ ]+]] = <<F8E4M3>>[2,32,16]{1,0,2} bitcast([[P0_TR]])
+; CHECK-NEXT:    [[P1:%[^ ]+]] = <<F8E4M3>>[2,32,16]{2,1,0} parameter(1)
+; CHECK-NEXT:    [[P1_TRANSPOSE:%[^ ]+]] = <<F8E4M3>>[2,16,32]{2,1,0} transpose([[P1]]), dimensions={0,2,1}
+; CHECK-NEXT:    [[P2:%[^ ]+]] = f32[] parameter(2)
+; CHECK-NEXT:    [[P3:%[^ ]+]] = f32[] parameter(3)
+; CHECK-NEXT:    [[DQ:%[^ ]+]] = f32[] multiply([[P2]], [[P3]])
+; CHECK-NEXT:    [[C1:%[^ ]+]] = f32[] constant(1)
+; CHECK-NEXT:    [[OUT:%[^ ]+]] = (f32[2,16,16]{2,1,0}, s8[{{[0-9]+}}]{0}) custom-call([[P0_BT1]], [[P1_TRANSPOSE]], [[DQ]], [[C1]], [[C1]], /*index=5*/[[C1]]),
+; CHECK:           custom_call_target="__cublas$lt$matmul$f8",
+; CHECK:           backend_config={
+; CHECK-DAG:         "alpha_real":1
+; CHECK-DAG:         "alpha_imag":0
+; CHECK-DAG:         "beta":0
+; CHECK-DAG:         "dot_dimension_numbers":{
+; CHECK-DAG:           "lhs_contracting_dimensions":["1"]
+; CHECK-DAG:           "rhs_contracting_dimensions":["2"]
+; CHECK-DAG:           "lhs_batch_dimensions":["0"]
+; CHECK-DAG:           "rhs_batch_dimensions":["0"]
+; CHECK-DAG:         }
+; CHECK-DAG:         "precision_config":{
+; CHECK-DAG:           "operand_precision":["DEFAULT","DEFAULT"]
+; CHECK-DAG:         }
+; CHECK-DAG:         "epilogue":"DEFAULT"
+; CHECK:           }
+      )");
+}
+
 TEST_P(ParameterizedFp8GemmRewriteTest, ScaledABUnscaledDF8) {
   const char* hlo_text = R"(
     HloModule test