nverke commented on code in PR #13301:
URL: https://github.com/apache/tvm/pull/13301#discussion_r1019893402
##########
tests/python/unittest/test_tir_transform_lower_cross_thread_reduction.py:
##########
@@ -487,6 +487,149 @@ def lowered_single_reduction_loop_with_block_predicate(
)
[email protected]_func
+def single_reduction_loop_with_tensorize(
+ input_A: T.Buffer[(1, 64, 7, 7, 32), "uint8"],
+ input_B: T.Buffer[(16, 64, 1, 1, 8, 32, 4), "int8"],
+ output: T.Buffer[(1, 16, 7, 7, 32), "int32"],
+) -> None:
+ # body
+ # with T.block("root")
+ for i1, i2, i3, i4, i5 in T.grid(16, 4, 98, 2, 32):
+ with T.block("compute_o"):
+ n = T.axis.spatial(1, 0)
+ oc_chunk = T.axis.spatial(16, i1)
+ oh = T.axis.spatial(7, (i2 * 6272 + i3 * 64 + i4 * 32 + i5) //
3584)
+ ow = T.axis.spatial(7, (i2 * 6272 + i3 * 64 + i4 * 32 + i5) % 3584
// 512)
+ kh = T.axis.reduce(1, 0)
+ kw = T.axis.reduce(1, 0)
+ ic_outer = T.axis.reduce(64, (i2 * 6272 + i3 * 64 + i4 * 32 + i5)
% 512 // 8)
+ ic_f_inner = T.axis.reduce(8, (i2 * 6272 + i3 * 64 + i4 * 32 + i5)
% 8)
+ T.reads(
+ input_A[n, ic_outer, oh + kh, ow + kw, ic_f_inner * 4 :
ic_f_inner * 4 + 4],
+ input_B[oc_chunk, ic_outer, kh, kw, ic_f_inner, 0:32, 0:4],
+ )
+ T.writes(output[n, oc_chunk, oh, ow, 0:32])
+ with T.init():
+ for x in T.serial(32):
+ with T.block("compute_init"):
+ oc_block_i_init = T.axis.spatial(32, x)
+ T.reads()
+ T.writes(output[n, oc_chunk, oh, ow, oc_block_i_init])
+ output[n, oc_chunk, oh, ow, oc_block_i_init] = 0
+ with T.block("compute_o"):
+ T.reads(
+ output[n, oc_chunk, oh, ow, 0:32],
+ input_A[n, ic_outer, oh + kh, ow + kw, ic_f_inner * 4 :
ic_f_inner * 4 + 4],
+ input_B[oc_chunk, ic_outer, kh, kw, ic_f_inner, 0:32, 0:4],
+ )
+ T.writes(output[n, oc_chunk, oh, ow, 0:32])
+ A = T.match_buffer(
+ input_A[n, ic_outer, oh + kh, ow + kw, ic_f_inner * 4 :
ic_f_inner * 4 + 4],
+ [4],
+ dtype="uint8",
+ offset_factor=1,
+ )
+ B = T.match_buffer(
+ input_B[oc_chunk, ic_outer, kh, kw, ic_f_inner, 0:32, 0:4],
+ [32, 4],
+ dtype="int8",
+ offset_factor=1,
+ )
+ C = T.match_buffer(
+ output[n, oc_chunk, oh, ow, 0:32], [32], dtype="int32",
offset_factor=1
+ )
+ A_u8x4: T.uint8x4 = A[0:4]
+ A_i32: T.int32 = T.reinterpret(A_u8x4, dtype="int32")
+ B_i8x128 = B[0, 0:128]
+ B_i32x32: T.int32x32 = T.reinterpret(B_i8x128,
dtype="int32x32")
+ C[0:32] = T.call_llvm_pure_intrin(
+ 4217, T.uint32(3), C[0:32], T.broadcast(A_i32, 32),
B_i32x32, dtype="int32x32"
+ )
+
+
[email protected]_func
+def nested_reduction_loop_with_inner_match_buffers(
+ in0: T.Buffer[(4, 16), "int8"],
+ in1: T.Buffer[(4, 16), "int8"],
+ out: T.Buffer[(4, 4), "int32"],
+) -> None:
+ # body
+ # with T.block("root")
+ for y in T.serial(4):
+ with T.block("C"):
+ yi = T.axis.spatial(4, y)
+ T.reads(in0[yi, 0:16], in1[yi, 0:16])
+ T.writes(out[yi, 0:4])
+ for x in T.serial(4):
+ xr = T.axis.reduce(4, x)
+ with T.init():
+ for i in T.serial(4):
+ with T.block("C_init"):
+ ii = T.axis.spatial(4, i)
+ T.reads()
+ T.writes(out[yi, ii])
+ out[yi, ii] = 0
+ with T.block("C"):
+ T.reads(
+ out[yi, xr],
+ in0[yi, yi * 4 + xr : yi * 4 + xr + 4],
+ in1[yi, yi * 4 + xr : yi * 4 + xr + 4],
+ )
+ T.writes(out[yi, xr])
+ A = T.match_buffer(
+ in0[yi, yi * 4 + xr : yi * 4 + xr + 4], [4],
dtype="int8", offset_factor=1
+ )
+ B = T.match_buffer(
+ in1[yi, yi * 4 + xr : yi * 4 + xr + 4], [4],
dtype="int8", offset_factor=1
+ )
+ C = T.match_buffer(out[yi, xr], [1], dtype="int32",
offset_factor=1)
+ A_i8x4: T.int8x4 = A[0:4]
+ A_i32: T.int32 = T.reinterpret(A_i8x4, dtype="int32")
+ B_i8x4: T.int8x4 = B[0:4]
+ B_i32: T.int32 = T.reinterpret(B_i8x4, dtype="int32")
+ C[0] = A_i32 + B_i32 + C[0]
+
+
[email protected]_func
+def nested_reduction_loop_with_outer_match_buffers(
+ in0: T.Buffer[(4, 16), "int8"],
+ in1: T.Buffer[(4, 16), "int8"],
+ out: T.Buffer[(4, 4), "int32"],
+) -> None:
+ # body
+ # with T.block("root")
+ for y in T.serial(4):
+ with T.block("C"):
+ yi = T.axis.spatial(4, y)
+ T.reads(in0[yi, 0:16], in1[yi, 0:16])
+ T.writes(out[yi, 0:4])
+ A = T.match_buffer(in0[yi, 0:16], [16], dtype="int8",
offset_factor=1)
+ B = T.match_buffer(in1[yi, 0:16], [16], dtype="int8",
offset_factor=1)
+ C = T.match_buffer(out[yi, 0:4], [4], dtype="int32",
offset_factor=1)
+ for x in T.serial(4):
+ xr = T.axis.reduce(4, x)
+ with T.init():
+ for i in T.serial(4):
+ with T.block("C_init"):
+ ii = T.axis.spatial(4, i)
+ T.reads()
+ T.writes(out[yi, ii])
+ out[yi, ii] = 0
+ with T.block("C"):
Review Comment:
Oh interesting! I see so the reduction axis is tied to the block despite
being within the loop. I can follow up with a commit in a few days or you can
remove the test if its causing issues.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
