Improve copy engine register documentation in nvdebug.h + cleanup

1 files changed, 108 insertions, 54 deletions

diff --git a/nvdebug.h b/nvdebug.h
index 8db07ee..630d40d 100644
--- a/nvdebug.h
+++ b/nvdebug.h
@@ -1,5 +1,12 @@
 /* Copyright 2021 Joshua Bakita
  * SPDX-License-Identifier: MIT
+ *
+ * File outline:
+ * - Runlist, preemption, and channel control (FIFO)
+ * - Basic GPU information (MC)
+ * - Detailed GPU information (PTOP, FUSE, and CE)
+ * - PRAMIN, BAR1/2, and page table status
+ * - Helper functions for nvdebug
  */
 
 // TODO(jbakita): Don't depend on these.
@@ -448,53 +455,6 @@ typedef union {
         } __attribute__((packed));
 } mc_boot_0_t;
 
-enum DEVICE_INFO_TYPE {INFO_TYPE_NOT_VALID = 0, INFO_TYPE_DATA = 1, INFO_TYPE_ENUM = 2, INFO_TYPE_ENGINE_TYPE = 3};
-enum ENGINE_TYPES {
-        ENGINE_GRAPHICS = 0, // GRAPHICS [/compute]
-        ENGINE_COPY0 = 1, // [raw/physical] COPY #0
-        ENGINE_COPY1 = 2, // [raw/physical] COPY #1
-        ENGINE_COPY2 = 3, // [raw/physical] COPY #2
-
-        ENGINE_MSPDEC = 8, // Picture DECoder
-        ENGINE_MSPPP = 9, // [Video] Post Processing
-        ENGINE_MSVLD = 10, // [Video] Variable Length Decoder
-        ENGINE_MSENC = 11, // [Video] ENCoding
-        ENGINE_VIC = 12, // Video Image Compositor
-        ENGINE_SEC = 13, // SEquenCer [?]
-        ENGINE_NVENC0 = 14, // Nvidia Video ENCoder #0
-        ENGINE_NVENC1 = 15, // Nvidia Video ENCoder #1
-        ENGINE_NVDEC = 16, // Nvidia Video DECoder
-
-        ENGINE_IOCTRL = 18, // I/O ConTRoLler [of NVLINK at least]
-        ENGINE_LCE = 19, // Logical Copy Engine
-        ENGINE_GSP = 20, // Gpu System Processor
-        ENGINE_NVJPG = 21, // NVidia JPeG [Decoder] (Ampere+)
-};
-#define ENGINE_TYPES_LEN 22
-static const char* const ENGINE_TYPES_NAMES[ENGINE_TYPES_LEN] = {
-        "Graphics/Compute",
-        "COPY0",
-        "COPY1",
-        "COPY2",
-        "Unknown Engine ID#4",
-        "Unknown Engine ID#5",
-        "Unknown Engine ID#6",
-        "Unknown Engine ID#7",
-        "MSPDEC: Picture Decoder",
-        "MSPPP: Post Processing",
-        "MSVLD: Variable Length Decoder",
-        "MSENC: Encoder",
-        "VIC: Video Image Compositor",
-        "SEC: Sequencer",
-        "NVENC0: NVIDIA Video Encoder #0",
-        "NVENC1: NVIDIA Video Encoder #1",
-        "NVDEC: NVIDIA Video Decoder",
-        "Unknown Engine ID#17",
-        "IOCTRL: I/O Controller",
-        "LCE: Logical Copy Engine",
-        "GSP: GPU System Processor",
-        "NVJPG: NVIDIA JPEG Decoder",
-};
 
 /* GPU engine information and control register offsets
   Each engine is described by one or more entries (terminated by an entry with
@@ -553,6 +513,54 @@ static const char* const ENGINE_TYPES_NAMES[ENGINE_TYPES_LEN] = {
 */
 #define NV_PTOP_DEVICE_INFO(i) (0x00022700+(i)*4)
 #define NV_PTOP_DEVICE_INFO__SIZE_1 64
+enum DEVICE_INFO_TYPE {INFO_TYPE_NOT_VALID = 0, INFO_TYPE_DATA = 1, INFO_TYPE_ENUM = 2, INFO_TYPE_ENGINE_TYPE = 3};
+enum ENGINE_TYPES {
+        ENGINE_GRAPHICS = 0, // GRAPHICS [/compute]
+        ENGINE_COPY0 = 1, // [raw/physical] COPY #0
+        ENGINE_COPY1 = 2, // [raw/physical] COPY #1
+        ENGINE_COPY2 = 3, // [raw/physical] COPY #2
+
+        ENGINE_MSPDEC = 8, // Picture DECoder
+        ENGINE_MSPPP = 9, // [Video] Post Processing
+        ENGINE_MSVLD = 10, // [Video] Variable Length Decoder
+        ENGINE_MSENC = 11, // [Video] ENCoding
+        ENGINE_VIC = 12, // Video Image Compositor
+        ENGINE_SEC = 13, // SEquenCer [?]
+        ENGINE_NVENC0 = 14, // Nvidia Video ENCoder #0
+        ENGINE_NVENC1 = 15, // Nvidia Video ENCoder #1
+        ENGINE_NVDEC = 16, // Nvidia Video DECoder
+
+        ENGINE_IOCTRL = 18, // I/O ConTRoLler [of NVLINK at least]
+        ENGINE_LCE = 19, // Logical Copy Engine
+        ENGINE_GSP = 20, // Gpu System Processor
+        ENGINE_NVJPG = 21, // NVidia JPeG [Decoder] (Ampere+)
+};
+#define ENGINE_TYPES_LEN 22
+static const char* const ENGINE_TYPES_NAMES[ENGINE_TYPES_LEN] = {
+        "Graphics/Compute",
+        "COPY0",
+        "COPY1",
+        "COPY2",
+        "Unknown Engine ID#4",
+        "Unknown Engine ID#5",
+        "Unknown Engine ID#6",
+        "Unknown Engine ID#7",
+        "MSPDEC: Picture Decoder",
+        "MSPPP: Post Processing",
+        "MSVLD: Variable Length Decoder",
+        "MSENC: Encoder",
+        "VIC: Video Image Compositor",
+        "SEC: Sequencer",
+        "NVENC0: NVIDIA Video Encoder #0",
+        "NVENC1: NVIDIA Video Encoder #1",
+        "NVDEC: NVIDIA Video Decoder",
+        "Unknown Engine ID#17",
+        "IOCTRL: I/O Controller",
+        "LCE: Logical Copy Engine",
+        "GSP: GPU System Processor",
+        "NVJPG: NVIDIA JPEG Decoder",
+};
+
 typedef union {
         // DATA type fields
         struct {
@@ -592,16 +600,60 @@ typedef union {
         uint32_t raw;
 } ptop_device_info_t;
 
+/* Graphics Processing Cluster (GPC) information
+  The GPU's Compute/Graphics engine is subdivided into Graphics Processing
+  Clusters (also known as GPU Processing Clusters, starting with Ampere).
+
+  Each GPC is subdivided into Texture Processing Clusters (TPCs) which contain
+  Streaming Multiprocessors (SMs).
+
+
+*/
+// Support: Fermi through Blackwell
+// Get the number of GPCs **on die**
 #define NV_PTOP_SCAL_NUM_GPCS 0x00022430
+// Get the number of TPCs per GPC **on die**
 #define NV_PTOP_SCAL_NUM_TPC_PER_GPC 0x00022434
-#define NV_PTOP_SCAL_NUM_CES 0x00022444
-// PCE_MAP is Volta+ only
-#define NV_CE_PCE_MAP 0x00104028
-
 // GPC and TPC masks
-// Support: Maxwell+
+// Support: Maxwell, Pascal, Volta, Turing
+// Bitmask of which GPC **are enabled** of the max on die
 #define NV_FUSE_GPC 0x00021c1c
+// Bitmask of which TPCs **are enabled** on each GPC
 #define NV_FUSE_TPC_FOR_GPC(i) (0x00021c38+(i)*4)
+// Support: Ampere, Ada, Hopper, Blackwell
+//#define NV_FUSE_GPC 0x00820c1c
+//#define NV_FUSE_TPC_FOR_GPC(i) (0x00820c38+(i)*4)
+
+/* Logical Copy Engine (LCE) Information
+  Every GPU has some number of copy engines which can process transfers to,
+  from, or within a GPU. Up until Maxwell, the hardware engines were directly
+  accessible, and this register exposes how many there are.
+
+  Starting with Pascal, an additional layer of indirection was added---logical
+  copy engines. Only logical copy engines can be directly dispatched to, and
+  there are normally more logical copy engines than there are physical ones. On
+  Pascal+ this register stores the number of logical copy engines.
+
+  SCAL_NUM_CES : Number of externally accessible copy engines
+
+  Support: Kepler through (at least) Blackwell
+  Also see dev_ce.ref.txt of NVIDIA's open-gpu-doc for info.
+*/
+#define NV_PTOP_SCAL_NUM_CES 0x00022444
+
+/* Physical Copy Engine (PCE) information
+  On Pascal GPUs or newer, this register complements the above information by
+  exposing which, and how many, physical copy engines are enabled on the GPU.
+
+  CE_PCE_MAP : A bitmask, where a set bit indicates that the PCE for that index
+               is enabled (not floorswept) on this GPU. Count the number of set
+               bits to get the number of PCEs.
+
+  Support: Kepler through (at least) Blackwell
+  Also see dev_ce.ref.txt of NVIDIA's open-gpu-doc for info.
+*/
+#define NV_CE_PCE_MAP 0x00104028
+
 
 /* Location of the 1Kb instance block with page tables for BAR1 and BAR2.
   Support: Fermi+ (?), Pascal
@@ -665,7 +717,7 @@ typedef union {
 
   Note: Format changed with Pascal (how?)
 
-  Support: Pascal, Volta, Turing, Ampere
+  Support: Pascal, Volta, Turing, Ampere, Ada
 */
 // FIXME: PDE/PTEs are actually 64 bits =S
 // Important: Aperture keys are different with PDEs
@@ -704,7 +756,9 @@ static inline char* pd_target_to_text(enum PD_TARGET t) {
 // Note: As the meaning of target (bits 2:1) changes depending on if the entry
 //       is a PTE or not, this combines them into a single target field to
 //       simplify comparisons.
-// Support: Pascal, Turing, Ampere
+// Support: Pascal, Volta, Turing, Ampere, Ada
+//
+// V3 introduced with Hopper, but Hopper and Blackwell also support V2
 typedef union {
         // Page Directory Entry (PDE)
         struct {