Merge pull request hardkernel#28 from yan-wyb/khadas-vims-4.9.y

update npu drivers to 6.4.2.1

drivers/amlogic/npu/Kbuild

@@ -134,7 +134,8 @@ OBJS += $(ARCH_KERNEL_DIR)/gc_hal_kernel_context.o \
  $(ARCH_KERNEL_DIR)/gc_hal_kernel_hardware.o \
  $(ARCH_KERNEL_DIR)/gc_hal_kernel_hardware_async_fe.o \
  $(ARCH_KERNEL_DIR)/gc_hal_kernel_hardware_mc_fe.o \
- $(ARCH_KERNEL_DIR)/gc_hal_kernel_hardware_waitlink_fe.o
+ $(ARCH_KERNEL_DIR)/gc_hal_kernel_hardware_waitlink_fe.o \
+ $(ARCH_KERNEL_DIR)/gc_hal_kernel_hardware_func.o
 
 ifeq ($(VIVANTE_ENABLE_3D),1)
 OBJS += $(ARCH_KERNEL_DIR)/gc_hal_kernel_recorder.o

drivers/amlogic/npu/inc/drvi/gc_vsc_drvi_interface.h

@@ -1,6 +1,6 @@
 /****************************************************************************
 *
-* Copyright (c) 2005 - 2019 by Vivante Corp. All rights reserved.
+* Copyright (c) 2005 - 2020 by Vivante Corp. All rights reserved.
 *
 * The material in this file is confidential and contains trade secrets
 * of Vivante Corporation. This is proprietary information owned by
@@ -55,8 +55,11 @@
 /* 0.0.1.32 Add extension flag in VIR_Symbol 05/27/2019 */
 /* 0.0.1.33 Add a new opcode LOGICAL_RSHIFT 05/28/2019 */
 /* 0.0.1.34 Add WorkGroupSizeFactor in VIR_ComputeLayout on 07/18/2019 */
-#define gcdVIR_SHADER_BINARY_FILE_VERSION gcmCC(SHADER_64BITMODE, 0, 1, 34)
-#define gcdVIR_PROGRAM_BINARY_FILE_VERSION gcmCC(SHADER_64BITMODE, 0, 1, 34)
+/* 0.0.1.35 Add skhp flag in VIR_Instruction */
+/* 0.0.1.36 Add a function to patch clipDistance in GL VIR lib shader */
+/* 0.0.1.37 Add the sampled image information on 03/11/2020 */
+#define gcdVIR_SHADER_BINARY_FILE_VERSION gcmCC(SHADER_64BITMODE, 0, 1, 37)
+#define gcdVIR_PROGRAM_BINARY_FILE_VERSION gcmCC(SHADER_64BITMODE, 0, 1, 37)
 
 #if !defined(gcdTARGETHOST_BIGENDIAN)
 #define gcdTARGETHOST_BIGENDIAN 0 /* default host little endian, to change the
@@ -491,8 +494,8 @@ typedef VSC_TyQualifier VIR_TyQualifier;
  * to achieve best performance */
 typedef enum {
  VSC_OCLImgLibKind_UseLoadStore = 0, /* for v54x GC chips, use LOAD/STORE/TEXLD */
- VSC_OCLImgLibKind_UseImgLoadTexldU = 1, /* for v55 GC chips, use IMG_LOAD/IMG_STORE/TEXLD_U */
- VSC_OCLImgLibKind_UseImgLoadTexldUXY = 2, /* for v60 GC and v620 GC chips */
+ VSC_OCLImgLibKind_UseImgLoadTexldU = 1, /* for v55 GC chips, use IMG_LOAD/IMG_STORE/TEXLD_U  -- unused now */
+ VSC_OCLImgLibKind_UseImgLoadTexldUXY = 2, /* for v60 GC and v620 GC chips -- unused now*/
  VSC_OCLImgLibKind_UseImgLoadVIP = 3, /* v60 VIP chip, use IMG_LOAD/IMG_STORE */
  VSC_OCLImgLibKind_Counts, /* count of img libs */
  VSC_OCLImgLibKind_BasedOnHWFeature /* select library based on HW feature */
@@ -732,7 +735,10 @@ typedef struct _VSC_HW_CONFIG
  /* word 3 */
  gctUINT hasPointSizeFix : 1;
  gctUINT supportVectorB0 : 1;
- gctUINT reserved1 : 30;
+ gctUINT hasAtomTimingFix : 1;
+ gctUINT hasUSCAtomicFix2 : 1;
+ gctUINT hasFloatingMadFix : 1;
+ gctUINT reserved1 : 27;
 
  /* Last word */
  /* Followings will be removed after shader programming is removed out of VSC */
@@ -750,16 +756,15 @@ typedef struct _VSC_HW_CONFIG
  gctUINT robustAtomic : 1;
  gctUINT newGPIPE : 1;
  gctUINT FEDrawDirect : 1;
-
- gctUINT hasUSCAtomicFix2 : 1;
- gctUINT reserved2 : 18;
+ gctUINT reserved2 : 19;
  } hwFeatureFlags;
 
  gctUINT chipModel;
  gctUINT chipRevision;
  gctUINT productID;
  gctUINT customerID;
  gctUINT maxCoreCount;
+ gctUINT maxClusterCount;
  gctUINT maxThreadCountPerCore;
  gctUINT maxVaryingCount;
  gctUINT maxAttributeCount;
@@ -841,8 +846,9 @@ typedef gcsGLSLCaps VSC_GL_API_CONFIG, *PVSC_GL_API_CONFIG;
 #define VSC_COMPILER_OPT_FULL_ACTIVE_IO 0x0000000000008000ULL
 #define VSC_COMPILER_OPT_DUAL16 0x0000000000010000ULL
 #define VSC_COMPILER_OPT_ILF_LINK 0x0000000000020000ULL
+#define VSC_COMPILER_OPT_LOOP 0x0000000000040000ULL
 
-#define VSC_COMPILER_OPT_FULL 0x000000000003FFFFULL
+#define VSC_COMPILER_OPT_FULL 0x000000000007FFFFULL
 
 #define VSC_COMPILER_OPT_NO_ALGE_SIMP 0x0000000100000000ULL
 #define VSC_COMPILER_OPT_NO_GCP 0x0000000200000000ULL
@@ -862,8 +868,9 @@ typedef gcsGLSLCaps VSC_GL_API_CONFIG, *PVSC_GL_API_CONFIG;
 #define VSC_COMPILER_OPT_NO_FULL_ACTIVE_IO 0x0000800000000000ULL
 #define VSC_COMPILER_OPT_NO_DUAL16 0x0001000000000000ULL
 #define VSC_COMPILER_OPT_NO_ILF_LINK 0x0002000000000000ULL
+#define VSC_COMPILER_OPT_NO_LOOP 0x0004000000000000ULL
 
-#define VSC_COMPILER_OPT_NO_OPT 0x0003FFFF00000000ULL
+#define VSC_COMPILER_OPT_NO_OPT 0x0007FFFF00000000ULL
 
 /* Compiler flag for special purpose */
 #define VSC_COMPILER_FLAG_COMPILE_TO_HL 0x00000001 /* Compile IR to HL, including doing all opts in HL */

drivers/amlogic/npu/inc/drvi/gc_vsc_drvi_kernel_profile.h

@@ -1,6 +1,6 @@
 /****************************************************************************
 *
-* Copyright (c) 2005 - 2019 by Vivante Corp. All rights reserved.
+* Copyright (c) 2005 - 2020 by Vivante Corp. All rights reserved.
 *
 * The material in this file is confidential and contains trade secrets
 * of Vivante Corporation. This is proprietary information owned by

drivers/amlogic/npu/inc/drvi/gc_vsc_drvi_lib_link.h

@@ -1,6 +1,6 @@
 /****************************************************************************
 *
-* Copyright (c) 2005 - 2019 by Vivante Corp. All rights reserved.
+* Copyright (c) 2005 - 2020 by Vivante Corp. All rights reserved.
 *
 * The material in this file is confidential and contains trade secrets
 * of Vivante Corporation. This is proprietary information owned by
@@ -25,6 +25,10 @@ typedef union _SPECIALIZATION_CONSTANT_VALUE
 
 typedef struct _VSC_LIB_SPECIALIZATION_CONSTANT
 {
+ /*
+ ** Some specialized variables for a function parameter, note that a name of the function parameter may be padding with "#dup%d",
+ ** so we need to use gcoOS_StrNCmp to compare with the function parameter.
+ */
  gctCONST_STRING varName;
  SPECIALIZATION_CONSTANT_VALUE value;
  VSC_SHADER_DATA_TYPE type;
@@ -114,12 +118,14 @@ typedef enum _VSC_RES_ACT_BIT
 
 typedef enum _VSC_LINK_POINT_RESOURCE_SUBTYPE
 {
- VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXLD_EXTRA_LATYER = 1,
- VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXGRAD_EXTRA_LATYER = 2,
- VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXFETCH_REPLACE_WITH_IMGLD = 3,
- VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXGATHER_EXTRA_LAYTER = 4,
- VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXGATHERPCF_D32F = 5,
- VSC_LINK_POINT_RESOURCE_SUBTYPE_NORMALIZE_TEXCOORD = 6,
+ VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXLD_EXTRA_LAYER = 1,
+ VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXLD_EXTRA_LAYER_SPECIFIED_OP = 2,
+ VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXGRAD_EXTRA_LAYER = 3,
+ VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXFETCH_REPLACE_WITH_IMGLD = 4,
+ VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXGATHER_EXTRA_LAYTER = 5,
+ VSC_LINK_POINT_RESOURCE_SUBTYPE_TEXGATHERPCF_D32F = 6,
+ VSC_LINK_POINT_RESOURCE_SUBTYPE_NORMALIZE_TEXCOORD = 7,
+ VSC_LINK_POINT_RESOURCE_SUBTYPE_YCBCR_TEXTURE = 8,
 } VSC_LINK_POINT_RESOURCE_SUBTYPE;
 
 typedef struct _VSC_LIB_LINK_POINT_FUNC_NAME
@@ -140,6 +146,7 @@ typedef struct _VSC_LIB_LINK_POINT_RESOURCE
  gctUINT arrayIndex;
  VSC_RES_OP_BIT opTypeBits;
  VSC_RES_ACT_BIT actBits;
+ void* pPrivData;
  VSC_LINK_POINT_RESOURCE_SUBTYPE subType;
 } VSC_LIB_LINK_POINT_RESOURCE;
 
@@ -165,12 +172,21 @@ typedef struct _VSC_LIB_LINK_IMAGE_READ_WRITE
  VSC_SAMPLER_INFO * samplerInfo;
 } VSC_LIB_LINK_IMAGE_READ_WRITE;
 
+/* Same value with VIR_IMAGE_ACCESS_STRATEGY. */
+typedef enum _VSC_LIB_LINK_IMAGE_ACCESS_STRATEGY
+{
+ VSC_LIB_LINK_IMAGE_ACCESS_STRATEGY_USE_FORMAT = 0,
+ VSC_LIB_LINK_IMAGE_ACCESS_STRATEGY_LOAD_ZERO_STORE_NOP = 1,
+ VSC_LIB_LINK_IMAGE_ACCESS_STRATEGY_LOAD_ZERO_STORE_ZERO = 2,
+} VSC_LIB_LINK_IMAGE_ACCESS_STRATEGY;
+
 typedef struct _VSC_LIB_LINK_IMAGE_FORMAT
 {
- gctUINT set;
- gctUINT binding;
- gctUINT arrayIndex;
- VSC_IMAGE_FORMAT imageFormat;
+ gctUINT set;
+ gctUINT binding;
+ gctUINT arrayIndex;
+ VSC_IMAGE_FORMAT imageFormat;
+ VSC_LIB_LINK_IMAGE_ACCESS_STRATEGY replaceStrategy;
 } VSC_LIB_LINK_IMAGE_FORMAT;
 
 typedef struct _VSC_LIB_LINK_POINT
@@ -194,13 +210,15 @@ typedef struct _VSC_LIB_LINK_POINT
  } u;
 }VSC_LIB_LINK_POINT;
 
+#define LIB_NUM 2
 typedef struct _VSC_SHADER_LIB_LINK_ENTRY
 {
  /* Which level this link entry should be applied. */
  VSC_SH_LEVEL applyLevel;
 
  /* Lib shader */
  SHADER_HANDLE hShaderLib;
+ SHADER_HANDLE hShaderLibs[LIB_NUM];
 
  /* vreg map from libShader to the current shader */
  void* pTempHashTable;

drivers/amlogic/npu/inc/drvi/gc_vsc_drvi_program_profile.h

@@ -1,6 +1,6 @@
 /****************************************************************************
 *
-* Copyright (c) 2005 - 2019 by Vivante Corp. All rights reserved.
+* Copyright (c) 2005 - 2020 by Vivante Corp. All rights reserved.
 *
 * The material in this file is confidential and contains trade secrets
 * of Vivante Corporation. This is proprietary information owned by
@@ -469,6 +469,30 @@ PROG_GL_XFB_OUT_TABLE;
  ********************************** VK program mapping table definitions *********************************
  **********************************************************************************************************************/
 
+typedef struct PROG_VK_IMAGE_FORMAT_INFO
+{
+ VSC_IMAGE_FORMAT imageFormat;
+ gctBOOL bSetInSpriv;
+}
+PROG_VK_IMAGE_FORMAT_INFO;
+
+typedef struct PROG_VK_IMAGE_DERIVED_INFO
+{
+ /* For a image, it might need a image-size attached. As each image in
+ Binding::arraySize array has image-size, so this is the first entry
+ of image-size array. */
+ SHADER_PRIV_CONSTANT_ENTRY* pImageSize;
+
+ /* Extra layer HW mapping. As currently, for images in in tBinding::arraySize
+ array, if one image has extra image, all other images must have extra image, so
+ this is the first entry of extra-image */
+ SHADER_PRIV_UAV_ENTRY* pExtraLayer;
+
+ /* ImageFormat, can be NONE. */
+ PROG_VK_IMAGE_FORMAT_INFO imageFormatInfo;
+}
+PROG_VK_IMAGE_DERIVED_INFO;
+
 typedef struct PROG_VK_SUB_RESOURCE_BINDING
 {
  /* Pointer to original API resource binding */
@@ -528,6 +552,7 @@ typedef struct PROG_VK_PRIV_COMB_TEX_SAMP_HW_MAPPING_LIST
  VSC_SHADER_RESOURCE_TYPE_COMBINED_IMAGE_SAMPLER
 */
 
+#define __YCBCR_PLANE_COUNT__ 3
 typedef struct PROG_VK_COMBINED_TEXTURE_SAMPLER_HW_MAPPING
 {
  /* For the case
@@ -541,6 +566,9 @@ typedef struct PROG_VK_COMBINED_TEXTURE_SAMPLER_HW_MAPPING
  combTsBinding::arraySize */
  SHADER_PRIV_SAMPLER_ENTRY** ppExtraSamplerArray;
 
+ /* For the ycbcr texture recompilation. */
+ SHADER_PRIV_UAV_ENTRY* pYcbcrPlanes[__YCBCR_PLANE_COUNT__];
+
  /* For the case that HW natively supports separated texture, so texture part of API
  combined texture sampler will be directly mapped to HW separated texture */
  SHADER_RESOURCE_SLOT_MAPPING texMapping;
@@ -663,15 +691,7 @@ typedef union PROG_VK_SEPARATED_TEXTURE_HW_MAPPING
  /* For HW does not natively supports separated texture */
  struct
  {
- /* For a separated image, it might need a image-size attached. As each image in
- storageBinding::arraySize array has image-size, so this is the first entry
- of image-size array. */
- SHADER_PRIV_CONSTANT_ENTRY* pImageSize;
-
- /* Extra layer HW mapping. As currently, for images in in texBinding::arraySize
- array, if one image has extra image, all other images must have extra image, so
- this is the first entry of extra-image */
- SHADER_PRIV_UAV_ENTRY* pExtraLayer;
+ PROG_VK_IMAGE_DERIVED_INFO imageDerivedInfo;
 
  /* We still need to allocate a constant image for this separated texture for the imageFetch operation.*/
  SHADER_UAV_SLOT_MAPPING hwMapping;
@@ -749,6 +769,13 @@ typedef struct PROG_VK_UNIFORM_TEXEL_BUFFER_HW_MAPPING
 }
 PROG_VK_UNIFORM_TEXEL_BUFFER_HW_MAPPING;
 
+typedef enum PROG_VK_UNIFORM_TEXEL_BUFFER_ENTRY_FLAG
+{
+ PROG_VK_UTB_ENTRY_FLAG_NONE = 0x0000,
+ /* Treat a texelBuffer as an image, now from a recompilation only. */
+ PROG_VK_UTB_ENTRY_FLAG_TREAT_TEXELBUFFER_AS_IMAGE = 0x0002,
+} PROG_VK_UNIFORM_TEXEL_BUFFER_ENTRY_FLAG;
+
 typedef struct PROG_VK_UNIFORM_TEXEL_BUFFER_TABLE_ENTRY
 {
  /* API resource binding */
@@ -763,14 +790,16 @@ typedef struct PROG_VK_UNIFORM_TEXEL_BUFFER_TABLE_ENTRY
  /* Is this entry really used by shader */
  gctUINT activeStageMask;
 
+ PROG_VK_UNIFORM_TEXEL_BUFFER_ENTRY_FLAG utbEntryFlag;
+
  /*----------------------------------Sampler-related----------------------------------*/
  /* For texel buffer, it might need a texture-size attached. As each texel buffer in
  utbBinding::arraySize array has texture-size, so this is the first entry
  of texture-size array. */
  SHADER_PRIV_CONSTANT_ENTRY* pTextureSize[VSC_MAX_SHADER_STAGE_COUNT][2];
 
  /*----------------------------------Image-related----------------------------------*/
- VSC_IMAGE_FORMAT  imageFormat;
+ PROG_VK_IMAGE_DERIVED_INFO imageDerivedInfo[VSC_MAX_SHADER_STAGE_COUNT];
 
  /* Which kinds of inst operation acting on texture. The count of this
  resOpBit is same as utbBinding::arraySize */
@@ -826,15 +855,7 @@ typedef struct PROG_VK_INPUT_ATTACHMENT_TABLE_ENTRY
  ** it is treated as a sampler, otherwise it is treated as an image.
  */
  /*----------------------------------Image-related----------------------------------*/
- /* For image storage, it might need a image-size attached. As each image in
- iaBinding::arraySize array has image-size, so this is the first entry
- of image-size array. */
- SHADER_PRIV_CONSTANT_ENTRY* pImageSize[VSC_MAX_SHADER_STAGE_COUNT];
-
- /* Extra layer HW mapping. As currently, for images in in iaBinding::arraySize
- array, if one image has extra image, all other images must have extra image, so
- this is the first entry of extra-image */
- SHADER_PRIV_UAV_ENTRY* pExtraLayer[VSC_MAX_SHADER_STAGE_COUNT];
+ PROG_VK_IMAGE_DERIVED_INFO imageDerivedInfo[VSC_MAX_SHADER_STAGE_COUNT];
 
  /*----------------------------------Sampler-related----------------------------------*/
  /* For texel buffer, it might need a texture-size attached. As each texel buffer in
@@ -888,15 +909,12 @@ typedef struct PROG_VK_STORAGE_TABLE_ENTRY
  /* Is this entry really used by shader */
  gctUINT activeStageMask;
 
- /* For image storage, it might need a image-size attached. As each image in
- storageBinding::arraySize array has image-size, so this is the first entry
- of image-size array. */
- SHADER_PRIV_CONSTANT_ENTRY* pImageSize[VSC_MAX_SHADER_STAGE_COUNT];
+ /*----------------------------------Image-related----------------------------------*/
+ PROG_VK_IMAGE_DERIVED_INFO imageDerivedInfo[VSC_MAX_SHADER_STAGE_COUNT];
 
- /* Extra layer HW mapping. As currently, for images in in storageBinding::arraySize
- array, if one image has extra image, all other images must have extra image, so
- this is the first entry of extra-image */
- SHADER_PRIV_UAV_ENTRY* pExtraLayer[VSC_MAX_SHADER_STAGE_COUNT];
+ /* Which kinds of inst operation acting on texture. The count of this
+ resOpBit is same as storageBinding::arraySize */
+ VSC_RES_OP_BIT* pResOpBits;
 
  /* Different shader stage may have different HW mappings. */
  SHADER_UAV_SLOT_MAPPING hwMappings[VSC_MAX_SHADER_STAGE_COUNT];

drivers/amlogic/npu/inc/drvi/gc_vsc_drvi_shader_priv_mapping.h

@@ -1,6 +1,6 @@
 /****************************************************************************
 *
-* Copyright (c) 2005 - 2019 by Vivante Corp. All rights reserved.
+* Copyright (c) 2005 - 2020 by Vivante Corp. All rights reserved.
 *
 * The material in this file is confidential and contains trade secrets
 * of Vivante Corporation. This is proprietary information owned by
@@ -61,13 +61,16 @@ typedef enum SHS_PRIV_CONSTANT_KIND
 /* Shader mem static priv-mapping kind */
 typedef enum SHS_PRIV_MEM_KIND
 {
- SHS_PRIV_MEM_KIND_GPR_SPILLED_MEMORY = 0, /* For gpr register spillage */
- SHS_PRIV_MEM_KIND_CONSTANT_SPILLED_MEMORY = 1, /* For constant register spillage */
- SHS_PRIV_MEM_KIND_STREAMOUT_BY_STORE = 2, /* Stream buffer for SO */
- SHS_PRIV_MEM_KIND_CL_PRIVATE_MEMORY = 3, /* For CL private mem */
- SHS_PRIV_MEM_KIND_SHARED_MEMORY = 4, /* For CL local memory or DirectCompute shared mem */
- SHS_PRIV_MEM_KIND_EXTRA_UAV_LAYER = 5,
- SHS_PRIV_MEM_KIND_COUNT = 6,
+ SHS_PRIV_MEM_KIND_NONE = 0,
+ SHS_PRIV_MEM_KIND_GPR_SPILLED_MEMORY = 1, /* For gpr register spillage */
+ SHS_PRIV_MEM_KIND_CONSTANT_SPILLED_MEMORY = 2, /* For constant register spillage */
+ SHS_PRIV_MEM_KIND_STREAMOUT_BY_STORE = 3, /* Stream buffer for SO */
+ SHS_PRIV_MEM_KIND_CL_PRIVATE_MEMORY = 4, /* For CL private mem */
+ SHS_PRIV_MEM_KIND_SHARED_MEMORY = 5, /* For CL local memory or DirectCompute shared mem */
+ SHS_PRIV_MEM_KIND_EXTRA_UAV_LAYER = 6,
+ SHS_PRIV_MEM_KIND_THREAD_ID_MEM_ADDR = 7, /* The global memory to save the consecutive thread ID. */
+ SHS_PRIV_MEM_KIND_YCBCR_PLANE = 8, /* The YCBCR plane. */
+ SHS_PRIV_MEM_KIND_COUNT = 9,
 }SHS_PRIV_MEM_KIND;
 
 /* !!!!!NOTE: For dynamic (lib-link) patch, the priv-mapping flag will directly use VSC_LIB_LINK_TYPE!!!!! */
@@ -154,6 +157,8 @@ typedef struct SHADER_PRIV_MEM_DATA_MAPPING
 
 typedef struct SHADER_PRIV_UAV_ENTRY
 {
+ gctUINT uavEntryIndex;
+
  SHADER_PRIV_MAPPING_COMMON_ENTRY commonPrivm;
 
  /* The data which will be set to this memory */