Re: [PATCH i-g-t v2 2/2] tests/intel/xe_prefetch_fault: add L2 prefetch fault subtests

W dniu 18.08.2026 o 12:35, Varun Gupta pisze:
> Add L2 prefetch fault testing. When L1 cache policy is
> set to uncached (UC), the prefetch bypasses LSC and is sourced from L2
> instead.
> 
> Add a new shader gpgpu_shader__l2_prefetch_fault() that uses message
> descriptor 0x49C00 (L1UC_L2C_L3UC, cache control value 4) instead of
> the existing 0x99C00 (L1C_L2C_L3C, cache control value 9). Only bits
> [19:16] of the descriptor differ.
> 
> Add two new subtests:
>    - l2-prefetch-fault: validates L2 prefetch fault with unmapped address
>      followed by hit-under-miss with mapped page
>    - l2-prefetch-fault-svm: same as above but in SVM mode with HMM-backed
>      CPU page table resolution
> 
> v2:
>    - Include generated iga64 codes for the new shader to fix compilation
>      failures (Kamil)
> 
> Signed-off-by: Varun Gupta <[email protected]>
> ---
>   tests/intel/xe_prefetch_fault.c               | 97 +++++++++++++++++++
>   .../xe_prefetch_fault.c.gen.iga64_codes.c     | 17 +++-
>   2 files changed, 113 insertions(+), 1 deletion(-)
> 
> diff --git a/tests/intel/xe_prefetch_fault.c b/tests/intel/xe_prefetch_fault.c
> index e86cd68d9..4843f56fd 100644
> --- a/tests/intel/xe_prefetch_fault.c
> +++ b/tests/intel/xe_prefetch_fault.c
> @@ -81,6 +81,56 @@ L0:
>   	)", lower_32_bits(addr), upper_32_bits(addr));
>   }
>   
> +/**
> + * gpgpu_shader__l2_prefetch_fault:
> + * @shdr: shader to be modified
> + * @addr: ppgtt virtual address to raise L2 prefetch fault
> + *
> + * This shader can only be used when in efficient 64bit mode.
> + * For a given arbitrary ppgtt virtual address, it raises an L2 prefetch fault
> + * using load instruction with L1 uncached + L2 cached cache policy.
> + * With L1 uncached, the prefetch bypasses LSC and is sourced from L2.
> + */
> +static void gpgpu_shader__l2_prefetch_fault(struct gpgpu_shader *shdr,
> +					    uint64_t addr)

Please use emit_ prefix, gpgpu_shader_ should be reserverd for 
gpggpu_shader library functions.

> +{
> +	igt_assert_f((addr & 0x7) == 0, "address must be aligned to QWord!\n");
> +
> +	emit_iga64_code(shdr, xe_l2_prefetch_fault_prefetch, R"(
> +#define IGA64_FLAGS ""

Why do we need this define? Without it assembler should generate proper 
dependencies without hand crafted ones.

> +#if GFX_VER >= 4000

Please drop above check, in upstream we support up to 3.5.

> +#elif GFX_VER >= 3500
> +L0:
> +// Set base address with scalar register
> +(W)	mov (1)		s0.0<1>:ud		ARG(0):ud
> +(W)	mov (1)		s0.1<1>:ud		ARG(1):ud
> +
> +// A64 offset
> +(W)	mov (8)		r30.0<1>:uq		0x0:uq
> +
> +// efficient 64bit Read with uncached L1, cached L2 and uncached L3
> +// sendg ugm load - L2 prefetch (L1 bypass)
> +// Message Descriptor
> +//      DP_LOAD_STORE_STATELESS_DESC (HAS:2209865465)
> +//      DP_CACHE_LOAD (HAS:2209865465) value 4 = L1UC_L2C_L3UC
> +//      0x49C00 =>
> +//      [45:44] Offset Scaling: 0 (disable)
> +//      [43:22] Global Offset: 0
> +//      [21] Overfetch: 0 (disable)
> +//      [19:16] Cache: 4 (L1 uncached, L2 cached and L3 uncached)
> +//      [15:14] Address Type and Size: 2 (Flat A64 Base, A64 Index)
> +//      [13:11] Data Size: 3 (D64)
> +//      [10:10] Transpose : 1 (enable)
> +//      [9:7] Vector Size: 0 (Vector length 1)
> +//      [5:0] Opcode: 0 (Load)

Whole this message descriptor is too chatty and too general. You can 
drop it in favor of much shorter: load.ugm.d64t.a64.uc.ca.uc, see below.

> +// Prefetch operations are implemented using a NULL destination register.
> +// L1 uncached forces the prefetch to bypass LSC, making L2 the fault source.

Above comment is nice, specific to prefetching.

> +(W)	sendg.ugm (1|M0)	null	r30:1	null:0	s0.0	0x49C00	{A@1,$5}

You can replace it with two liners (after removal of the whole message 
descriptor comment):
+// load.ugm.d64t.a64.uc.ca.uc [src0]
+(W)	sendg.ugm (1)	null	r30:1	null:0	s0.0	0x49C00

> +
> +#endif
> +	)", lower_32_bits(addr), upper_32_bits(addr));
> +}
> +
>   static struct intel_buf *
>   create_buf(int fd, int width, int height, uint32_t color)
>   {
> @@ -130,6 +180,17 @@ static struct gpgpu_shader *get_prefetch_shader(int fd)
>   	return shader;
>   }
>   
> +static struct gpgpu_shader *get_l2_prefetch_shader(int fd)
> +{
> +	struct gpgpu_shader *shader;
> +
> +	shader = gpgpu_shader_create(fd);
> +	gpgpu_shader__l2_prefetch_fault(shader, xe_canonical_va(fd, PREFETCH_ADDR));

You can incorporate body of gpgpu_shader__l2_prefetch_fault here, I see 
no more users of this function.

Regards
Andrzej


> +	gpgpu_shader__eot(shader);
> +
> +	return shader;
> +}
> +
>   /**
>    * SUBTEST: prefetch-fault
>    * Description: Validate L1 prefetch fault and hit-under-miss behavior with
> @@ -140,6 +201,16 @@ static struct gpgpu_shader *get_prefetch_shader(int fd)
>    * Description: Validate L1 prefetch fault and hit-under-miss behavior in SVM
>    *		mode with L1 cached, L2 cached cache policy (fault source: LSC)
>    * Run type: FULL
> + *
> + * SUBTEST: l2-prefetch-fault
> + * Description: Validate L2 prefetch fault and hit-under-miss behavior with
> + *		L1 uncached, L2 cached cache policy (fault source: L2)
> + * Run type: FULL
> + *
> + * SUBTEST: l2-prefetch-fault-svm
> + * Description: Validate L2 prefetch fault and hit-under-miss behavior in SVM
> + *		mode with L1 uncached, L2 cached cache policy (fault source: L2)
> + * Run type: FULL
>    */
>   static void test_prefetch_fault(int fd, struct drm_xe_engine_class_instance *hwe,
>   				bool svm, get_shader_fn get_shader)
> @@ -296,6 +367,32 @@ int igt_main()
>   		}
>   	}
>   
> +	igt_subtest_with_dynamic("l2-prefetch-fault") {
> +		xe_for_each_engine(fd, hwe) {
> +			if (hwe->engine_class == DRM_XE_ENGINE_CLASS_RENDER ||
> +			    hwe->engine_class == DRM_XE_ENGINE_CLASS_COMPUTE) {
> +				igt_dynamic_f("%s%d", xe_engine_class_string(hwe->engine_class),
> +					      hwe->engine_instance)
> +					test_prefetch_fault(fd, hwe, false,
> +							    get_l2_prefetch_shader);
> +			}
> +		}
> +	}
> +
> +	igt_subtest_with_dynamic("l2-prefetch-fault-svm") {
> +		if (!svm_supported)
> +			igt_skip("SVM not supported on this device, skipping.\n");
> +		xe_for_each_engine(fd, hwe) {
> +			if (hwe->engine_class == DRM_XE_ENGINE_CLASS_RENDER ||
> +			    hwe->engine_class == DRM_XE_ENGINE_CLASS_COMPUTE) {
> +				igt_dynamic_f("%s%d", xe_engine_class_string(hwe->engine_class),
> +					      hwe->engine_instance)
> +					test_prefetch_fault(fd, hwe, true,
> +							    get_l2_prefetch_shader);
> +			}
> +		}
> +	}
> +
>   	igt_fixture() {
>   		drm_close_driver(fd);
>   	}
> diff --git a/tests/intel/xe_prefetch_fault.c.gen.iga64_codes.c b/tests/intel/xe_prefetch_fault.c.gen.iga64_codes.c
> index 0df7fe20e..acbfa7a75 100644
> --- a/tests/intel/xe_prefetch_fault.c.gen.iga64_codes.c
> +++ b/tests/intel/xe_prefetch_fault.c.gen.iga64_codes.c
> @@ -3,7 +3,22 @@
>   
>   #include "gpgpu_shader.h"
>   
> -#define MD5_SUM_IGA64_ASMS 153c8c0a5a3ed448d86f86bb51e8d140
> +#define MD5_SUM_IGA64_ASMS 0cd443620cdd553ad899b68d685d15ef
> +
> +struct iga64_template const iga64_code_xe_l2_prefetch_fault_prefetch[] = {
> +	{ .gfx_ver = 3500, .size = 28, .code = (const uint32_t []) {
> +		0x80000061, 0x60014220, 0x00000000, 0xc0ded000,
> +		0x80000061, 0x60114220, 0x00000000, 0xc0ded001,
> +		0x800c0061, 0x1e054330, 0x00000000, 0x00000000,
> +		0x80012033, 0x00000004, 0xf0041e0c, 0x9c000000,
> +		0x80000001, 0x00010000, 0x20000000, 0x00000000,
> +		0x80000001, 0x00010000, 0x30000000, 0x00000000,
> +		0x80000901, 0x00010000, 0x00000000, 0x00000000,
> +	}},
> +	{ .gfx_ver = 0, .size = 0, .code = (const uint32_t []) {
> +
> +	}}
> +};
>   
>   struct iga64_template const iga64_code_xe_prefetch_fault_prefetch[] = {
>   	{ .gfx_ver = 3500, .size = 28, .code = (const uint32_t []) {