Add support for mixed precision in contraction scale and bilinear

c84d6f43 · Bartlomiej Wroblewski · a8747955 · c84d6f43 · c84d6f43 · c84d6f43
Commit c84d6f43 authored Sep 19, 2023 by Bartlomiej Wroblewski
16 changed files
--- a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
--- a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance.cpp
--- a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance.cpp
--- a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance.cpp
--- a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance.cpp
--- a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
--- a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
--- a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
--- a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp
--- a/profiler/README.md
+++ b/profiler/README.md
@@ -50,21 +50,23 @@ Best Perf: 1.42509 ms, 102.988 TFlops, 234.086 GB/s
 ## Profile contraction kernels
 ```bash
 #arg1: tensor operation (contraction_bilinear=CONTRACTION+Bilinear)
-#arg2: data type (0: fp32; 1: f64)\n"
+#arg2: data type (0: fp32; 1: f64; 2: f16; 3: bf16)
-#arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
+#arg3: compute data type (0: fp32; 1: f64; 2: f16; 3: bf16)
+#arg4: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
 #                     1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
 #                     2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
 #                     3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1])
-#arg4: verification (0: no; 1: yes)
+#arg5: verification (0: no; 1: yes)
-#arg5: initialization (0: no init; 1: integer value; 2: decimal value)
+#arg6: initialization (0: no init; 1: integer value; 2: decimal value)
-#arg6: print tensor value (0: no; 1: yes)
+#arg7: print tensor value (0: no; 1: yes)
-#arg7: time kernel (0: no, 1: yes)
+#arg8: time kernel (0: no, 1: yes)
-#arg8 and arg9: alpha and beta
+#arg9: alpha
-#arg10 to 15: M0, M1, N0, N1, K0, K1
+#arg10: beta
-#arg16 to 31: Strides for A, B, D and E (skip for default)
+#arg11 to 16: M0, M1, N0, N1, K0, K1
+#arg17 to 32: Strides for A, B, D and E (skip for default)
-################                   op  datatype  layout  verify  init  log  time  alpha  beta  M0  M1  N0  N1  K0  K1
-./bin/ckProfiler contraction_bilinear         0       1       0     0    0     1    1.0   1.0 128 128 128 128 128 128
+################                   op  datatype  compute_datatype  layout  verify  init  log  time  alpha  beta  M0  M1  N0  N1  K0  K1
+./bin/ckProfiler contraction_bilinear         0                 0       1       0     0    0     1    1.0   1.0 128 128 128 128 128 128
 ```
 Result (MI100)

--- a/profiler/include/profiler/profile_contraction_impl.hpp
+++ b/profiler/include/profiler/profile_contraction_impl.hpp
--- a/profiler/include/profiler/profile_contraction_utils.hpp
+++ b/profiler/include/profiler/profile_contraction_utils.hpp
--- a/profiler/src/profile_contraction_bilinear.cpp
+++ b/profiler/src/profile_contraction_bilinear.cpp
--- a/profiler/src/profile_contraction_scale.cpp
+++ b/profiler/src/profile_contraction_scale.cpp
--- a/test/contraction/test_contraction.cpp
+++ b/test/contraction/test_contraction.cpp
--- a/test/contraction/test_contraction_interface.cpp
+++ b/test/contraction/test_contraction_interface.cpp