Merge branch 'ae' into ae-volta

kernels/sgemm_tcore/kernel.cpp

@@ -95,8 +95,6 @@ void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
   constexpr uint32_t quartile = (128 << 10) >> 2; // 128KB / 4
   static_assert((quartile * 4) == SMEM_SIZE, "wrong quartile constant");
 
-  MARK_BEG();
-
   constexpr uint32_t smem_a_offset = 0;
   constexpr uint32_t smem_a_dbuf_offset = 1 * quartile;
   constexpr uint32_t smem_b_offset =
@@ -121,8 +119,6 @@ void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
                       threadblocks_per_cluster, threadblock_id_in_cluster,
                       sharedmem_per_threadblock);
 
-  MARK_END();
-
   float *gmem_tmp_d0 = reinterpret_cast<float *>(0xd0000000UL);
   float *gmem_tmp_d1 = reinterpret_cast<float *>(0xd1000000UL);
   float *gmem_tmp_d2 = reinterpret_cast<float *>(0xd2000000UL);

kernels/sgemm_tcore/sgemm_impl.hpp

@@ -19,7 +19,7 @@ using float_type = float16_t;
 
 // Generate kernel for the Hopper-style SMEM-decoupled tensor core.  This uses
 // asynchronous HGMMA and HGMMA_WAIT instructions.
-#define TENSOR_HOPPER 1
+#define TENSOR_HOPPER 0
 
 // Constraints on parameters:
 // * Memory:
@@ -110,7 +110,7 @@ static_assert(WMITER * WNITER * TCM * TCN * NUM_WARPS * CORES_PER_CLUSTER ==
 // result matrix will be stored in a swizzled form in the global memory.
 #define WMMA_STORE_FAST 0
 
-#define GEMMINI_DMA 1
+#define GEMMINI_DMA 0
 #define GEMMINI_DMA_FAST 1
 #if SMEM_SIZE == 0x4000
 #define SMEM_ADDR_Q0 ((float * const) 0xff000000)

lib/include/VX_config.h

@@ -84,7 +84,7 @@
 #endif
 
 #ifndef NUM_CORES
-#define NUM_CORES 4
+#define NUM_CORES 8
 #endif
 
 #ifndef NUM_WARPS

lib/include/vx_spawn.h

@@ -18,7 +18,7 @@
 #include <stdio.h>
 
 #ifndef CORES_PER_CLUSTER
-#define CORES_PER_CLUSTER 4
+#define CORES_PER_CLUSTER 8
 #endif
 
 #ifdef __cplusplus