From 86e8df60f1598760511b059b42a9e4f9dddfa150 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Fri, 12 May 2017 17:43:56 -0700 Subject: Fixed a bug in the TRSM routine; tests now pass --- src/routines/level3/xtrsm.cpp | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) (limited to 'src') diff --git a/src/routines/level3/xtrsm.cpp b/src/routines/level3/xtrsm.cpp index 4378ca94..c1c07d10 100644 --- a/src/routines/level3/xtrsm.cpp +++ b/src/routines/level3/xtrsm.cpp @@ -138,7 +138,7 @@ void Xtrsm::TrsmColMajor(const Side side, const Triangle triangle, DoGemm(Layout::kColMajor, a_transpose, Transpose::kNo, m - i - block_size, n, block_size, ConstantNegOne(), a_buffer, this_a_offset, a_ld, - x_buffer, x_offset + i, x_ld, ConstantOne(), + x_buffer, x_offset + i, x_ld, gemm_alpha, b_buffer, b_offset + i + block_size, b_ld); } } @@ -159,7 +159,7 @@ void Xtrsm::TrsmColMajor(const Side side, const Triangle triangle, DoGemm(Layout::kColMajor, a_transpose, Transpose::kNo, i, n, block_size, ConstantNegOne(), a_buffer, this_a_offset, a_ld, - x_buffer, x_offset + i, x_ld, ConstantOne(), + x_buffer, x_offset + i, x_ld, gemm_alpha, b_buffer, b_offset, b_ld); } } @@ -182,9 +182,9 @@ void Xtrsm::TrsmColMajor(const Side side, const Triangle triangle, if (i - static_cast(block_size) < 0) { break; } const auto this_a_offset = (a_transpose == Transpose::kNo) ? i : i * a_ld; DoGemm(Layout::kColMajor, Transpose::kNo, a_transpose, - m, i, current_block_size, ConstantNegOne(), + m, i, block_size, ConstantNegOne(), x_buffer, x_offset + i * x_ld, x_ld, - a_buffer, this_a_offset, a_ld, ConstantOne(), + a_buffer, this_a_offset, a_ld, gemm_alpha, b_buffer, b_offset, b_ld); } } @@ -204,7 +204,7 @@ void Xtrsm::TrsmColMajor(const Side side, const Triangle triangle, DoGemm(Layout::kColMajor, Transpose::kNo, a_transpose, m, n - i - block_size, block_size, ConstantNegOne(), x_buffer, x_offset + i * x_ld, x_ld, - a_buffer, this_a_offset, a_ld, ConstantOne(), + a_buffer, this_a_offset, a_ld, gemm_alpha, b_buffer, b_offset + (i + block_size) * b_ld, b_ld); } } -- cgit v1.2.3