github/ROCm
1
1
Fork 0
mirror of https://github.com/ROCm/ROCm.git synced 2026-04-05 03:01:17 -04:00
Code Issues Packages Projects Releases Wiki Activity

[MFMA] Remove CTA related code from layout (#429)

Browse Source 
This PR removes CTALayout attribute from MFMA layout, because it is NV specific.
This commit is contained in:
Alexander Efimov
2023-12-27 20:01:28 +03:00
committed by GitHub
parent 1e2fd0dd1a
commit 98589ac013
8 changed files with 52 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
test/Conversion/AMDGPU/mfma_variants.mlir
View File

@@ -5,7 +5,7 @@
!c_ty = f32
#k_width = 8
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -25,7 +25,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 8
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -45,7 +45,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 8
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -65,7 +65,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 8
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -85,7 +85,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 4
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -105,7 +105,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 2
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -125,7 +125,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 4
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -145,7 +145,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 1
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -165,7 +165,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = i32
#k_width = 4
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -185,7 +185,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = i32
#k_width = 8
#non_k_dim = 32
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -205,7 +205,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 8
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -225,7 +225,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 8
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -245,7 +245,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 8
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -265,7 +265,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 8
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -285,7 +285,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 4
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -305,7 +305,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 2
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -325,7 +325,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 4
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -345,7 +345,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 1
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -365,7 +365,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = i32
#k_width = 4
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -385,7 +385,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = i32
#k_width = 8
#non_k_dim = 16
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -405,7 +405,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 4
#non_k_dim = 4
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -425,7 +425,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 2
#non_k_dim = 4
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -445,7 +445,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 4
#non_k_dim = 4
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -465,7 +465,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = f32
#k_width = 1
#non_k_dim = 4
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -485,7 +485,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
!c_ty = i32
#k_width = 4
#non_k_dim = 4
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = #non_k_dim, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = #k_width}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = #k_width}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {

2
test/Conversion/minimize_alloc.mlir
View File

@@ -5,7 +5,7 @@
#blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 8], threadsPerWarp = [8, 8], warpsPerCTA = [4, 1], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#shared = #triton_gpu.shared<{vec = 4, perPhase = 2, maxPhase = 8, order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#shared1 = #triton_gpu.shared<{vec = 1, perPhase = 1, maxPhase = 1, order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = 32, warpsPerCTA = [2, 2], isTransposed=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = 32, warpsPerCTA = [2, 2], isTransposed=false}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 : i32, "triton_gpu.threads-per-warp" = 64 : i32} {
tt.func public @matmul_kernel_0d1d2d3d4d5d6d7c8d9c10d11c(%arg0: !tt.ptr<f16> {tt.divisibility = 16 : i32}, %arg1: !tt.ptr<f16> {tt.divisibility = 16 : i32}, %arg2: !tt.ptr<f16> {tt.divisibility = 16 : i32}, %arg3: i32 {tt.divisibility = 16 : i32}, %arg4: i32 {tt.divisibility = 16 : i32}, %arg5: i32 {tt.divisibility = 16 : i32}, %arg6: i32 {tt.divisibility = 16 : i32}, %arg7: i32 {tt.divisibility = 16 : i32}, %arg8: i32 {tt.divisibility = 16 : i32}) attributes {noinline = false} {
%cst = arith.constant dense<0.000000e+00> : tensor<64x64xf32, #mfma>

6
test/Conversion/tritongpu_to_llvm.mlir
View File

@@ -1355,7 +1355,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [16, 4], warpsPerCTA = [1, 1], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#shared0 = #triton_gpu.shared<{vec = 1, perPhase=1, maxPhase=1, order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma0 = #triton_gpu.mfma<{nonKDim = 32, warpsPerCTA=[1,1], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma0 = #triton_gpu.mfma<{nonKDim = 32, warpsPerCTA=[1,1], isTranspose=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma0, kWidth = 4}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma0, kWidth = 4}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
@@ -1379,7 +1379,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 :
// -----
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [64, 1], warpsPerCTA = [1, 4], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = 32, warpsPerCTA = [2, 2], isTranspose=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = 32, warpsPerCTA = [2, 2], isTranspose=false}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : i32} {
// CHECK: llvm.mlir.global external @global_smem() {addr_space = 3 : i32} : !llvm.array<0 x i8>
// CHECK-LABEL: convert_layout_mfma_block
@@ -1529,7 +1529,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 :
#blocked = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [2, 32], warpsPerCTA = [1, 4], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#shared = #triton_gpu.shared<{vec = 1, perPhase = 1, maxPhase = 1, order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = 32, warpsPerCTA = [2, 2], isTransposed=false, CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
#mfma = #triton_gpu.mfma<{nonKDim = 32, warpsPerCTA = [2, 2], isTransposed=false}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mfma, kWidth = 4}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mfma, kWidth = 4}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 : i32} {