[PATCH i-g-t v3 2/4] lib/gppgu_shader: Add read D32 from ppgtt virtual address
Gwan-gyeong Mun
gwan-gyeong.mun at intel.com
Fri Nov 22 14:34:44 UTC 2024
Create a function that adds the capabilty to read an dword size from a
given ppgtt address. Use an Untyped 2D Block Array Load DataPort
functionality of XE2+ with A64 flat addressing to direct accessing an
entire ppgtt address space.
For the read to succeed, the given ppgtt virtual address has to be bound.
Otherwise a load page fault will be triggered.
v2: Fix the function name to be more clear. (Christoph)
v3: Make block read to1 word (d32) from 4 words(d32 x 4) (Andrzej)
Update the function name (Andrzej)
gpgpu_shader__read_a64_dword() to gpgpu_shader__read_a64_d32()
Drop not needed code (Andrzej)
Signed-off-by: Gwan-gyeong Mun <gwan-gyeong.mun at intel.com>
Reviewed-by: Andrzej Hajda <andrzej.hajda at intel.com>
---
lib/gpgpu_shader.c | 90 +++++++++++++++++++++++++++++++++++++
lib/gpgpu_shader.h | 1 +
lib/iga64_generated_codes.c | 20 ++++++++-
3 files changed, 110 insertions(+), 1 deletion(-)
diff --git a/lib/gpgpu_shader.c b/lib/gpgpu_shader.c
index 973b5ec4c..38d9cc680 100644
--- a/lib/gpgpu_shader.c
+++ b/lib/gpgpu_shader.c
@@ -893,3 +893,93 @@ void gpgpu_shader__write_a64_d32(struct gpgpu_shader *shdr, uint64_t ppgtt_addr,
#endif \n\
", lower_32_bits(addr), upper_32_bits(addr), value);
}
+
+/**
+ * gpgpu_shader__read_a64_d32:
+ * @shdr: shader to be modified
+ * @ppgtt_addr: read target ppgtt virtual address
+ *
+ * Read one D32 data (DW; DoubleWord) directly from the target ppgtt virtual
+ * address (A64 Flat Address model).
+ *
+ * Note: for the read to succeed, the address specified by @ppgtt_addr has
+ * to be bound. Otherwise a load page fault will be triggered.
+ */
+void gpgpu_shader__read_a64_d32(struct gpgpu_shader *shdr, uint64_t ppgtt_addr)
+{
+ uint64_t addr = CANONICAL(ppgtt_addr);
+
+ igt_assert_f((addr & 0x3) == 0, "address must be aligned to DWord!\n");
+
+ emit_iga64_code(shdr, read_a64_d32, " \n\
+#if GEN_VER >= 2000 \n\
+// Unyped 2D Block Array Load \n\
+// Instruction_Load2DBlockArray \n\
+// bspec: 63972 \n\
+// src0 address payload (Untyped2DBLOCKAddressPayload) specifies both \n\
+// the block parameters and the 2D Surface parameters. \n\
+// Untyped2DBLOCKAddressPayload \n\
+// bspec: 63986 \n\
+// [243:240] Array Length: 0 (length is 1) \n\
+// [239:232] Block Height: 0 (height is 1) \n\
+// [231:224] Block Width: 0x3 (width is 4 bytes) \n\
+// [223:192] Block Start Y: 0 \n\
+// [191:160] Block Start X: 0 \n\
+// [159:128] Untyped 2D Surface Pitch: 0x3f (pitch is 64 bytes) \n\
+// [127:96] Untyped 2D Surface Height: 0 (height is 1) \n\
+// [95:64] Untyped 2D Surface Width: 0x3f (width is 64 bytes) \n\
+// [63:0] Untyped 2D Surface Base Address \n\
+// initialize register \n\
+(W) mov (8) r30.0<1>:uq 0x0:uq \n\
+// [31:0] Untyped 2D Surface Base Address low \n\
+(W) mov (1) r30.0<1>:ud ARG(0):ud \n\
+// [63:32] Untyped 2D Surface Base Address high \n\
+(W) mov (1) r30.1<1>:ud ARG(1):ud \n\
+// [95:64] Untyped 2D Surface Width: 0x3f \n\
+// (Width minus 1 (in bytes) of the 2D surface, it represents 64) \n\
+(W) mov (1) r30.2<1>:ud 0x3f:ud \n\
+// [159:128] Untyped 2D Surface Pitch: 0x3f \n\
+// (Pitch minus 1 (in bytes) of the 2D surface, it represents 64) \n\
+(W) mov (1) r30.4<1>:ud 0x3f:ud \n\
+// [231:224] Block Width: 0x3 (4 bytes) \n\
+// (Specifies the width minus 1 (in number of data elements) for this \n\
+// rectangular region, it represents 4) \n\
+// Block width (encoded_value + 1) must be a multiple of DW (4 bytes). \n\
+// [239:232] Block Height: 0 \n\
+// (Specifies the height minus 1 (in number of data elements) for \n\
+// this rectangular region, it represents 1) \n\
+// [243:240] Array Length: 0 \n\
+// (Specifies Array Length minus 1 for Load2DBlockArray messages, \n\
+// must be zero for 2D Block Store messages, it represents 1) \n\
+(W) mov (1) r30.7<1>:ud 0x3:ud \n\
+// \n\
+// dest data payload format is selected by Data Size. \n\
+// Block Height x Block Width x Data size / GRF Register size \n\
+// => 1 x 16 x 32bit / 512bit = 1 \n\
+// data payload format size is 1 GRF Register. \n\
+// \n\
+// send.ugm Untyped 2D Block Array Load \n\
+// Format: send.ugm (1) dst src0 src1 ExtMsg MsgDesc \n\
+// Execution Mask restriction: SIMT1 \n\
+// \n\
+// Extended Message Descriptor (Dataport Extended Descriptor Imm 2D Block) \n\
+// bspec: 67780 \n\
+// 0x0 => \n\
+// [32:22] Global Y_offset: 0 \n\
+// [21:12] Global X_offset: 0 \n\
+// \n\
+// Message Descriptor \n\
+// bspec: 63972 \n\
+// 0x2128403 => \n\
+// [30:29] Address Type: 0 (FLAT) \n\
+// [28:25] Src0 Length: 1 \n\
+// [24:20] Dest Length: 1 \n\
+// [19:16] Cache : 2 (L1UC_L3UC) 10 \n\
+// [15] Transpose Block: 1 \n\
+// [11:9] Data Size: 2 (D32) 10 \n\
+// [7] VNNI Transform: 0 \n\
+// [5:0] Load Operation: 3 (Load 2D Block) 11 \n\
+(W) send.ugm (1) r31 r30 null 0x0 0x2128403 \n\
+#endif \n\
+ ", lower_32_bits(addr), upper_32_bits(addr));
+}
diff --git a/lib/gpgpu_shader.h b/lib/gpgpu_shader.h
index e5eefa0a1..2ad6a7010 100644
--- a/lib/gpgpu_shader.h
+++ b/lib/gpgpu_shader.h
@@ -87,6 +87,7 @@ void gpgpu_shader__write_on_exception(struct gpgpu_shader *shdr, uint32_t dw, ui
uint32_t y_offset, uint32_t mask, uint32_t value);
void gpgpu_shader__write_a64_d32(struct gpgpu_shader *shdr, uint64_t ppgtt_addr,
uint32_t value);
+void gpgpu_shader__read_a64_d32(struct gpgpu_shader *shdr, uint64_t ppgtt_addr);
void gpgpu_shader__label(struct gpgpu_shader *shdr, int label_id);
void gpgpu_shader__jump(struct gpgpu_shader *shdr, int label_id);
void gpgpu_shader__jump_neq(struct gpgpu_shader *shdr, int label_id,
diff --git a/lib/iga64_generated_codes.c b/lib/iga64_generated_codes.c
index e778a4617..9db6bdd95 100644
--- a/lib/iga64_generated_codes.c
+++ b/lib/iga64_generated_codes.c
@@ -3,7 +3,7 @@
#include "gpgpu_shader.h"
-#define MD5_SUM_IGA64_ASMS e9da0ec3ffba2cd02079f91c38b24e8b
+#define MD5_SUM_IGA64_ASMS 85067cf5769bd5f5ab2a2abe5bdc9645
struct iga64_template const iga64_code_gpgpu_fill[] = {
{ .gen_ver = 2000, .size = 44, .code = (const uint32_t []) {
@@ -79,6 +79,24 @@ struct iga64_template const iga64_code_gpgpu_fill[] = {
}}
};
+struct iga64_template const iga64_code_read_a64_d32[] = {
+ { .gen_ver = 2000, .size = 40, .code = (const uint32_t []) {
+ 0x800c0061, 0x1e054330, 0x00000000, 0x00000000,
+ 0x80000061, 0x1e054220, 0x00000000, 0xc0ded000,
+ 0x80000061, 0x1e154220, 0x00000000, 0xc0ded001,
+ 0x80000061, 0x1e254220, 0x00000000, 0x0000003f,
+ 0x80000061, 0x1e454220, 0x00000000, 0x0000003f,
+ 0x80000061, 0x1e754220, 0x00000000, 0x00000003,
+ 0x80032031, 0x1f0c0000, 0xf8061e0c, 0x00a00000,
+ 0x80000001, 0x00010000, 0x20000000, 0x00000000,
+ 0x80000001, 0x00010000, 0x30000000, 0x00000000,
+ 0x80000901, 0x00010000, 0x00000000, 0x00000000,
+ }},
+ { .gen_ver = 0, .size = 0, .code = (const uint32_t []) {
+
+ }}
+};
+
struct iga64_template const iga64_code_write_a64_d32[] = {
{ .gen_ver = 2000, .size = 48, .code = (const uint32_t []) {
0x800c0061, 0x1e054330, 0x00000000, 0x00000000,
--
2.46.1
More information about the igt-dev
mailing list