Add support for Gaudi2

Add the support for running TPC kernels on Gaudi2 device.
Gaudi2 support was first added to the Linux kernel in kernel 5.7.

Signed-off-by: Guy Malinovitch <[email protected]>
Reviewed-by: Oded Gabbay <[email protected]>
Signed-off-by: Oded Gabbay <[email protected]>

.gitignore

@@ -1,4 +1,13 @@
-build
+*.so
+*.cmake
+Makefile
+CMakeFiles
+CMakeCache.txt
+tests/*_test
+build/
 .cproject
 .project
-.settings
+.settings/
+.vscode/
+.gdbinit
+.cache

CMakeLists.txt

@@ -7,6 +7,7 @@ project(synapse_core)
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -Wall -Werror -Wno-sign-compare -fno-strict-aliasing")
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pipe")
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 if (CMAKE_CXX_COMPILER_VERSION VERSION_LESS 8.0)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-attributes")

build.sh

@@ -23,8 +23,19 @@ set -e
 # then you need to define EXTRA_CMAKE_FLAGS as:
 # EXTRA_CMAKE_FLAGS=-DDRIVER_INCLUDE_PATH=$HOME/habanalabs/include/uapi
 
+SECONDS=0;
 SRCDIR=`dirname $0`
 BUILDDIR="$SRCDIR/build"
+build_type="Release"
+
+while [ -n "$1" ]; do
+    case $1 in 
+        -debug)
+            build_type="Debug"
+            ;;
+    esac;
+    shift;
+done
 
 if [ -d $BUILDDIR ]; then
 	rm -rf $BUILDDIR
@@ -41,5 +52,8 @@ fi
 
 cd "$BUILDDIR"
 
-$CMAKE -DCMAKE_BUILD_TYPE="Release" ${EXTRA_CMAKE_FLAGS:-} ..
+$CMAKE -DCMAKE_BUILD_TYPE=${build_type} ${EXTRA_CMAKE_FLAGS:-} ..
 make -j8
+
+printf "\nElapsed time: %02u:%02u:%02u \n\n" $(($SECONDS / 3600)) $((($SECONDS / 60) % 60)) $(($SECONDS % 60));
+

external_includes/SpecialFuncCoefficients.h

@@ -15,3 +15,10 @@ namespace tpc_gaudi {
 	uint8_t *getLUTStartAddressFromFuncID(uint32_t funcID, uint8_t* baseAddr);
 	uint32_t getCoefAddrFromOffset(uint8_t offset, bool isLOOKUP, uint32_t elementSize);
 }
+
+namespace tpc_gaudi2
+{
+void     buildSpecialFunctionCoefficients(std::vector<std::vector<uint8_t>>& specialFunctionCoefficients);
+uint8_t* getLUTStartAddressFromFuncID(uint32_t funcID, uint8_t* baseAddr);
+uint32_t getCoefAddrFromOffset(uint8_t offset, bool isLOOKUP, uint32_t elementSize);
+} // namespace tpc_gaudi2

external_includes/TensorDescriptor.h

@@ -7,6 +7,7 @@
 
 #pragma once
 
+#include "synapse_common_types.h"
 #include <stdint.h>
 #include <cassert>
 
@@ -42,7 +43,7 @@ DEF_PACK(
 
 typedef struct _TensorDescriptorBase  TensorDescriptorBase;
 
-struct TensorDescriptorGaudi : public TensorDescriptorBase
+struct TensorDescriptor : public TensorDescriptorBase
 {
 	/* for Gaudi:
 	TENSOR_0_BASE_ADDR_LOW		+0x00	32	0		Read/Write	bits 31 to 0 of the base address		no	no	no
@@ -84,20 +85,40 @@ typedef enum TensorDataType_
 uint32_t gen2_TensorDescriptorElementSizeType(uint32_t configuration);
 uint32_t gen2_TensorDescriptorValidDimMask(uint32_t configuration);
 uint32_t gen2_TensorDescriptorLastDim(uint32_t configuration);
+uint32_t gen6_TensorDescriptorElementSizeType(uint32_t configuration);
+uint32_t gen6_TensorDescriptorValidDimMask(uint32_t configuration);
+uint32_t gen6_TensorDescriptorLastDim(uint32_t configuration);
 
 namespace tpc_gaudi {
     inline uint32_t get_TensorDescriptorLastDim(uint32_t configuration)
     {
         return gen2_TensorDescriptorLastDim(configuration);
     }
-    inline uint32_t get_TensorDescriptorElementSizeType(uint32_t configuration) 
-    { 
-        return gen2_TensorDescriptorElementSizeType(configuration); 
+    inline uint32_t get_TensorDescriptorElementSizeType(uint32_t configuration)
+    {
+        return gen2_TensorDescriptorElementSizeType(configuration);
     }
     inline uint32_t get_TensorDescriptorValidDimMask(uint32_t configuration)
-    { 
+    {
         return gen2_TensorDescriptorValidDimMask(configuration);
     }
 	uint32_t TensorDescriptorConfiguration(uint32_t dataType, uint32_t validDimMask, uint32_t lastDim,  uint32_t rmwSel, uint32_t rmwOp);
     uint32_t get_ElementSizeInBytesFromDataType(TensorDataType DataType);
 }
+
+namespace tpc_gaudi2 {
+    inline uint32_t get_TensorDescriptorLastDim(uint32_t configuration)
+    {
+        return gen6_TensorDescriptorLastDim(configuration);
+    }
+    inline uint32_t get_TensorDescriptorElementSizeType(uint32_t configuration)
+    {
+        return gen6_TensorDescriptorElementSizeType(configuration);
+    }
+    inline uint32_t get_TensorDescriptorValidDimMask(uint32_t configuration)
+    {
+        return gen6_TensorDescriptorValidDimMask(configuration);
+    }
+	uint32_t TensorDescriptorConfiguration(uint32_t dataType, uint32_t validDimMask, uint32_t lastDim,  uint32_t rmwSel, uint32_t rmwOp);
+    uint32_t get_ElementSizeInBytesFromDataType(TensorDataType DataType);
+}

external_includes/asic_reg_structs/dma_core_regs.h → external_includes/gaudi/asic_reg_structs/dma_core_regs.h

@@ -9,7 +9,8 @@
 #include <stdint.h>
 
 #pragma pack(push, 1)
-
+namespace gaudi
+{
 namespace dma_core {
 /*
  CFG_0 
@@ -1039,6 +1040,6 @@ const offsetVal block_dma_core_defaults[]
 	{ 0x130 , 0x1                 , 1 }, // wr_awuser_31_11
 	{ 0x230 , 0x178               , 1 }, // dbg_sts
 };
-
+} // namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_DMA_CORE_H_ */

external_includes/asic_reg_structs/gaudi_types.h → external_includes/gaudi/asic_reg_structs/gaudi_types.h

@@ -10,12 +10,15 @@
 
 #pragma pack(push, 1)
 
+namespace gaudi
+{
 struct offsetVal
 {
 	uint32_t offset;
 	uint32_t val;
 	int copies;
 };
 
+} // namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_GAUDI_TYPES_H_ */

external_includes/asic_reg_structs/qman_regs.h → external_includes/gaudi/asic_reg_structs/qman_regs.h

@@ -9,7 +9,8 @@
 #include <stdint.h>
 
 #pragma pack(push, 1)
-
+namespace gaudi
+{
 namespace qman {
 /*
  GLBL_CFG0 
@@ -1954,7 +1955,10 @@ struct block_qman {
 	uint32_t _pad3292[9];
 	struct qman::reg_glbl_mem_init_busy glbl_mem_init_busy;
 };
+} // namespace gaudi
 #include "gaudi_types.h"
+namespace gaudi
+{
 const offsetVal block_qman_defaults[]
 {
 	// offset	// value
@@ -1986,5 +1990,6 @@ const offsetVal block_qman_defaults[]
 	{ 0xc84 , 0x2000              , 1 }, // local_range_size
 };
 
+} //namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_QMAN_H_ */

external_includes/asic_reg_structs/sob_glbl_regs.h → external_includes/gaudi/asic_reg_structs/sob_glbl_regs.h

@@ -9,7 +9,8 @@
 #include <stdint.h>
 
 #pragma pack(push, 1)
-
+namespace gaudi
+{
 namespace sob_glbl {
 /*
  SM_SEI_MASK 
@@ -50,5 +51,6 @@ struct block_sob_glbl {
 	struct sob_glbl::reg_sm_sei_cause sm_sei_cause;
 };
 
+} // namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_SOB_GLBL_H_ */

external_includes/asic_reg_structs/sob_objs_regs.h → external_includes/gaudi/asic_reg_structs/sob_objs_regs.h

@@ -9,7 +9,8 @@
 #include <stdint.h>
 
 #pragma pack(push, 1)
-
+namespace gaudi
+{
 namespace sob_objs {
 /*
  SM_SEC 
@@ -127,5 +128,6 @@ struct block_sob_objs {
 	struct sob_objs::reg_sm_sec sm_sec[256];
 };
 
+} // namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_SOB_OBJS_H_ */

external_includes/asic_reg_structs/sync_mngr_regs.h → external_includes/gaudi/asic_reg_structs/sync_mngr_regs.h

@@ -18,11 +18,14 @@ namespace sync_mngr {
 /*
  SYNC_MNGR block
 */
+namespace gaudi
+{
 struct block_sync_mngr {
 	struct block_sob_glbl sync_mngr_glbl;
 	uint32_t _pad8[1022];
 	struct block_sob_objs sync_mngr_objs;
 };
 
+} // namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_SYNC_MNGR_H_ */

external_includes/asic_reg_structs/sync_object_regs.h → external_includes/gaudi/asic_reg_structs/sync_object_regs.h

@@ -9,7 +9,8 @@
 #include <stdint.h>
 
 #pragma pack(push, 1)
-
+namespace gaudi
+{
 namespace sync_object {
 /*
  MESSAGE 
@@ -47,5 +48,6 @@ struct block_sync_object {
 	struct sync_object::reg_addr addr;
 };
 
+} //namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_SYNC_OBJECT_H_ */

external_includes/asic_reg_structs/tpc_non_tensor_descriptor_regs.h → external_includes/gaudi/asic_reg_structs/tpc_non_tensor_descriptor_regs.h

@@ -9,7 +9,8 @@
 #include <stdint.h>
 
 #pragma pack(push, 1)
-
+namespace gaudi
+{
 namespace tpc_non_tensor_descriptor {
 /*
  KERNEL_BASE_ADDRESS_LOW 
@@ -222,12 +223,16 @@ struct block_tpc_non_tensor_descriptor {
 	struct tpc_non_tensor_descriptor::reg_kernel_id kernel_id;
 	struct tpc_non_tensor_descriptor::reg_srf srf[32];
 };
+} // namespace gaudi
 #include "gaudi_types.h"
+namespace gaudi
+{
 const offsetVal block_tpc_non_tensor_descriptor_defaults[]
 {
 	// offset	// value
 	{ 0x30  , 0x60882             , 1 }, // kernel_config
 };
 
+} //namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_TPC_NON_TENSOR_DESCRIPTOR_H_ */

external_includes/asic_reg_structs/tpc_regs.h → external_includes/gaudi/asic_reg_structs/tpc_regs.h

@@ -9,7 +9,8 @@
 #include <stdint.h>
 
 #pragma pack(push, 1)
-
+namespace gaudi
+{
 namespace tpc {
 /*
  CFG_BASE_ADDRESS_HIGH 
@@ -705,13 +706,15 @@ struct reg_irq_occoupy_cntr {
 };
 static_assert((sizeof(struct reg_irq_occoupy_cntr) == 4), "reg_irq_occoupy_cntr size is not 32-bit");
 } /* tpc namespace */
-
+} //namespace gaudi
 #include "sync_object_regs.h"
 #include "tpc_non_tensor_descriptor_regs.h"
 #include "tpc_tensor_regs.h"
 /*
  TPC block
 */
+namespace gaudi
+{
 struct block_tpc {
 	uint32_t _pad0[256];
 	struct block_tpc_tensor kernel_tensor_0;
@@ -807,7 +810,10 @@ struct block_tpc {
 	struct block_sync_object qm_sync_object;
 	struct block_tpc_non_tensor_descriptor qm;
 };
+} // namespace gaudi
 #include "gaudi_types.h"
+namespace gaudi
+{
 const offsetVal block_tpc_defaults[]
 {
 	// offset	// value
@@ -863,6 +869,6 @@ const offsetVal block_tpc_defaults[]
 	{ 0xd54 , 0x50000             , 1 }, // tensor_config
 	{ 0xdb8 , 0x60882             , 1 }, // kernel_config
 };
-
+} // namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_TPC_H_ */

external_includes/asic_reg_structs/tpc_tensor_regs.h → external_includes/gaudi/asic_reg_structs/tpc_tensor_regs.h

@@ -9,7 +9,8 @@
 #include <stdint.h>
 
 #pragma pack(push, 1)
-
+namespace gaudi
+{
 namespace tpc_tensor {
 /*
  BASE_ADDR_LOW 
@@ -222,12 +223,16 @@ struct block_tpc_tensor {
 	struct tpc_tensor::reg_dim_4_size dim_4_size;
 	struct tpc_tensor::reg_dim_4_stride dim_4_stride;
 };
+} // namespace gaudi
 #include "gaudi_types.h"
+namespace gaudi
+{
 const offsetVal block_tpc_tensor_defaults[]
 {
 	// offset	// value
 	{ 0xc   , 0x50000             , 1 }, // tensor_config
 };
 
+} // namespace gaudi
 #pragma pack(pop)
 #endif /* ASIC_REG_STRUCTS_TPC_TENSOR_H_ */

external_includes/gaudi_packets.h → external_includes/gaudi/gaudi_packets.h

@@ -12,6 +12,8 @@
 
 #define PACKET_HEADER_PACKET_ID_SHIFT		56
 #define PACKET_HEADER_PACKET_ID_MASK		0x1F00000000000000ull
+namespace gaudi
+{
 
 enum packet_id {
 	PACKET_WREG_32 = 0x1,
@@ -325,4 +327,5 @@ struct packet_cp_dma {
 	__u64 src_addr;
 };
 
+} // namespace gaudi
 #endif /* GAUDI_PACKETS_H */

external_includes/gaudi_tpc_descriptor.h → external_includes/gaudi/gaudi_tpc_descriptor.h

@@ -15,6 +15,8 @@
 
 #pragma pack(push, 4)
 
+namespace gaudi
+{
 struct GaudiTpcDesc
 {
     static const int c_max_tensor_dims = 5;
@@ -25,5 +27,6 @@ struct GaudiTpcDesc
     block_tpc_non_tensor_descriptor m_desc;
 };
 
+} // namespace gaudi
 #pragma pack(pop)
 #endif // _GAUDI_TPC_DESC_H_