Skip to content

GitLab

Commit ee6c4ff7 authored by Anthony Chang's avatar Anthony Chang
Browse files

accumulators are zeroed in gemm_pipeline

parent 7f177f07
Show whitespace changes
Inline Side-by-side
Showing with 2 additions and 15 deletions
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
View file @ ee6c4ff7
...@@ -1762,21 +1762,8 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle ...@@ -1762,21 +1762,8 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
vgrad_grid_buf); vgrad_grid_buf);
// gemm dP // gemm dP
pgrad_acc_thread_buf.Clear();
#if 0
if (hipBlockIdx_x == 0 && hipThreadIdx_x % 32 < 4)
{
printf("j loop idx %d, tid %zd, clear dP[0:3] = %f, %f, %f, %f\n",
gemm1_k_block_outer_index,
hipThreadIdx_x,
pgrad_acc_thread_buf[I0],
pgrad_acc_thread_buf[I1],
pgrad_acc_thread_buf[I2],
pgrad_acc_thread_buf[I3]);
}
#endif
block_sync_lds();
// assume size K == size O so has main block loop // assume size K == size O so has main block loop
block_sync_lds();
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>( gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(
ygrad_grid_desc_o0_m_o1, ygrad_grid_desc_o0_m_o1,
a_block_desc_ak0_m_ak1, // reuse a_block_desc_ak0_m_ak1, // reuse
...@@ -1793,7 +1780,7 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle ...@@ -1793,7 +1780,7 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
pgrad_blockwise_gemm, pgrad_blockwise_gemm,
pgrad_acc_thread_buf, pgrad_acc_thread_buf,
num_o_block_main_loop); num_o_block_main_loop);
#if 1 #if 0
if (hipBlockIdx_x == 0 && hipThreadIdx_x % 32 < 4) if (hipBlockIdx_x == 0 && hipThreadIdx_x % 32 < 4)
{ {
printf("j loop idx %d, tid %zd, dP[0:3] = %f, %f, %f, %f\n", printf("j loop idx %d, tid %zd, dP[0:3] = %f, %f, %f, %f\n",
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment