[igt-dev] [PATCH i-g-t v2 8/9] lib/intel_compute: Adding pvc compute pipeline implementation
Zbigniew Kempczyński
zbigniew.kempczynski at intel.com
Tue Sep 5 13:33:08 UTC 2023
Add square compute pipeline which works on PVC. Currently limited
to Xe driver.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
Cc: Christoph Manszewski <christoph.manszewski at intel.com>
Cc: Francois Dugast <francois.dugast at intel.com>
Cc: Mauro Carvalho Chehab <mchehab at kernel.org>
---
lib/intel_compute.c | 218 ++++++++++++++++++++++++++++-
lib/intel_compute_square_kernels.c | 39 ++++++
2 files changed, 256 insertions(+), 1 deletion(-)
diff --git a/lib/intel_compute.c b/lib/intel_compute.c
index 29a5ec168f..4a232ce72b 100644
--- a/lib/intel_compute.c
+++ b/lib/intel_compute.c
@@ -71,9 +71,18 @@ static void bo_execenv_create(int fd, struct bo_execenv *execenv)
execenv->driver = get_intel_driver(fd);
if (execenv->driver == INTEL_DRIVER_XE) {
+ uint16_t engine_class;
+ uint32_t devid = intel_get_drm_devid(fd);
+ const struct intel_device_info *info = intel_get_device_info(devid);
+
+ if (info->graphics_ver >= 12 && info->graphics_rel < 60)
+ engine_class = DRM_XE_ENGINE_CLASS_RENDER;
+ else
+ engine_class = DRM_XE_ENGINE_CLASS_COMPUTE;
+
execenv->vm = xe_vm_create(fd, DRM_XE_VM_CREATE_ASYNC_BIND_OPS, 0);
execenv->exec_queue = xe_exec_queue_create_class(fd, execenv->vm,
- DRM_XE_ENGINE_CLASS_RENDER);
+ engine_class);
}
}
@@ -877,6 +886,208 @@ static void xehp_compute_exec(int fd, const unsigned char *kernel,
bo_execenv_destroy(&execenv);
}
+static void xehpc_create_indirect_data(uint32_t *addr_bo_buffer_batch,
+ uint64_t addr_input,
+ uint64_t addr_output)
+{
+ int b = 0;
+
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000400;
+ addr_bo_buffer_batch[b++] = 0x00000001;
+ addr_bo_buffer_batch[b++] = 0x00000001;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = addr_input & 0xffffffff;
+ addr_bo_buffer_batch[b++] = addr_input >> 32;
+ addr_bo_buffer_batch[b++] = addr_output & 0xffffffff;
+ addr_bo_buffer_batch[b++] = addr_output >> 32;
+ addr_bo_buffer_batch[b++] = 0x00000400;
+ addr_bo_buffer_batch[b++] = 0x00000400;
+ addr_bo_buffer_batch[b++] = 0x00000001;
+ addr_bo_buffer_batch[b++] = 0x00000001;
+}
+
+static void xehpc_compute_exec_compute(uint32_t *addr_bo_buffer_batch,
+ uint64_t addr_general_state_base,
+ uint64_t addr_surface_state_base,
+ uint64_t addr_dynamic_state_base,
+ uint64_t addr_instruction_state_base,
+ uint64_t offset_indirect_data_start,
+ uint64_t kernel_start_pointer)
+{
+ int b = 0;
+
+ igt_debug("general state base: %lx\n", addr_general_state_base);
+ igt_debug("surface state base: %lx\n", addr_surface_state_base);
+ igt_debug("dynamic state base: %lx\n", addr_dynamic_state_base);
+ igt_debug("instruct base addr: %lx\n", addr_instruction_state_base);
+ igt_debug("bindless base addr: %lx\n", addr_surface_state_base);
+ igt_debug("offset indirect addr: %lx\n", offset_indirect_data_start);
+ igt_debug("kernel start pointer: %lx\n", kernel_start_pointer);
+
+ addr_bo_buffer_batch[b++] = GEN7_PIPELINE_SELECT | GEN9_PIPELINE_SELECTION_MASK |
+ PIPELINE_SELECT_GPGPU;
+
+ addr_bo_buffer_batch[b++] = XEHP_STATE_COMPUTE_MODE;
+ addr_bo_buffer_batch[b++] = 0xE0186010;
+
+ addr_bo_buffer_batch[b++] = XEHP_CFE_STATE | 0x4;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x10008800;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+
+ addr_bo_buffer_batch[b++] = MI_LOAD_REGISTER_IMM(1);
+ addr_bo_buffer_batch[b++] = 0x00002580;
+ addr_bo_buffer_batch[b++] = 0x00060002;
+
+ addr_bo_buffer_batch[b++] = STATE_BASE_ADDRESS | 0x14;
+ addr_bo_buffer_batch[b++] = (addr_general_state_base & 0xffffffff) | 0x41;
+ addr_bo_buffer_batch[b++] = addr_general_state_base >> 32;
+ addr_bo_buffer_batch[b++] = 0x00044000;
+ addr_bo_buffer_batch[b++] = (addr_surface_state_base & 0xffffffff) | 0x41;
+ addr_bo_buffer_batch[b++] = addr_surface_state_base >> 32;
+ addr_bo_buffer_batch[b++] = (addr_dynamic_state_base & 0xffffffff) | 0x41;
+ addr_bo_buffer_batch[b++] = addr_dynamic_state_base >> 32;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = (addr_instruction_state_base & 0xffffffff) | 0x41;
+ addr_bo_buffer_batch[b++] = addr_instruction_state_base >> 32;
+ addr_bo_buffer_batch[b++] = 0xfffff001;
+ addr_bo_buffer_batch[b++] = 0x00010001;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0xfffff001;
+ addr_bo_buffer_batch[b++] = (addr_surface_state_base & 0xffffffff) | 0x41;
+ addr_bo_buffer_batch[b++] = addr_surface_state_base >> 32;
+ addr_bo_buffer_batch[b++] = 0x00007fbf;
+ addr_bo_buffer_batch[b++] = 0x00000041;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+
+ addr_bo_buffer_batch[b++] = GEN8_3DSTATE_BINDING_TABLE_POOL_ALLOC | 2;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+
+ addr_bo_buffer_batch[b++] = XEHP_COMPUTE_WALKER | 0x25;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000040;
+ addr_bo_buffer_batch[b++] = offset_indirect_data_start;
+ addr_bo_buffer_batch[b++] = 0xbe040000;
+ addr_bo_buffer_batch[b++] = 0xffffffff;
+ addr_bo_buffer_batch[b++] = 0x0000003f;
+ addr_bo_buffer_batch[b++] = 0x00000010;
+
+ addr_bo_buffer_batch[b++] = 0x00000001;
+ addr_bo_buffer_batch[b++] = 0x00000001;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+
+ addr_bo_buffer_batch[b++] = kernel_start_pointer;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00180000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x0c000020;
+
+ addr_bo_buffer_batch[b++] = 0x00000008;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00001047;
+ addr_bo_buffer_batch[b++] = ADDR_BATCH;
+ addr_bo_buffer_batch[b++] = ADDR_BATCH >> 32;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000040;
+ addr_bo_buffer_batch[b++] = 0x00000001;
+ addr_bo_buffer_batch[b++] = 0x00000001;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+ addr_bo_buffer_batch[b++] = 0x00000000;
+
+ addr_bo_buffer_batch[b++] = MI_BATCH_BUFFER_END;
+}
+
+/**
+ * xehpc_compute_exec - run a pipeline compatible with XEHP
+ *
+ * @fd: file descriptor of the opened DRM device
+ * @kernel: GPU Kernel binary to be executed
+ * @size: size of @kernel.
+ */
+static void xehpc_compute_exec(int fd, const unsigned char *kernel,
+ unsigned int size)
+{
+#define XEHPC_BO_DICT_ENTRIES 6
+ struct bo_dict_entry bo_dict[XEHP_BO_DICT_ENTRIES] = {
+ { .addr = XEHP_ADDR_INSTRUCTION_STATE_BASE + OFFSET_KERNEL,
+ .name = "instr state base"},
+ { .addr = XEHP_ADDR_GENERAL_STATE_BASE + OFFSET_INDIRECT_DATA_START,
+ .size = 0x10000,
+ .name = "indirect object base"},
+ { .addr = ADDR_INPUT, .size = SIZE_BUFFER_INPUT,
+ .name = "addr input"},
+ { .addr = ADDR_OUTPUT, .size = SIZE_BUFFER_OUTPUT,
+ .name = "addr output" },
+ { .addr = XEHP_ADDR_GENERAL_STATE_BASE, .size = 0x10000,
+ .name = "general state base" },
+ { .addr = ADDR_BATCH, .size = SIZE_BATCH,
+ .name = "batch" },
+ };
+ struct bo_execenv execenv;
+ float *dinput;
+
+ bo_execenv_create(fd, &execenv);
+
+ /* Sets Kernel size */
+ bo_dict[0].size = ALIGN(size, 0x1000);
+
+ bo_execenv_bind(&execenv, bo_dict, XEHPC_BO_DICT_ENTRIES);
+
+ memcpy(bo_dict[0].data, kernel, size);
+ xehpc_create_indirect_data(bo_dict[1].data, ADDR_INPUT, ADDR_OUTPUT);
+
+ dinput = (float *)bo_dict[2].data;
+ srand(time(NULL));
+ for (int i = 0; i < SIZE_DATA; i++)
+ ((float *)dinput)[i] = rand() / (float)RAND_MAX;
+
+ xehpc_compute_exec_compute(bo_dict[5].data,
+ XEHP_ADDR_GENERAL_STATE_BASE,
+ ADDR_SURFACE_STATE_BASE,
+ ADDR_DYNAMIC_STATE_BASE,
+ XEHP_ADDR_INSTRUCTION_STATE_BASE,
+ OFFSET_INDIRECT_DATA_START,
+ OFFSET_KERNEL);
+
+ bo_execenv_exec(&execenv, ADDR_BATCH);
+
+ for (int i = 0; i < SIZE_DATA; i++) {
+ float f1, f2;
+
+ f1 = ((float *) bo_dict[3].data)[i];
+ f2 = ((float *) bo_dict[2].data)[i];
+ if (f1 != f2 * f2)
+ igt_debug("[%4d] f1: %f != %f\n", i, f1, f2 * f2);
+ igt_assert(f1 == f2 * f2);
+ }
+
+ bo_execenv_unbind(&execenv, bo_dict, XEHPC_BO_DICT_ENTRIES);
+ bo_execenv_destroy(&execenv);
+}
+
/*
* Compatibility flags.
*
@@ -905,6 +1116,11 @@ static const struct {
.compute_exec = xehp_compute_exec,
.compat = COMPAT_I915,
},
+ {
+ .ip_ver = IP_VER(12, 60),
+ .compute_exec = xehpc_compute_exec,
+ .compat = COMPAT_XE,
+ },
};
bool run_compute_kernel(int fd)
diff --git a/lib/intel_compute_square_kernels.c b/lib/intel_compute_square_kernels.c
index da73a3747c..de93a3bdfd 100644
--- a/lib/intel_compute_square_kernels.c
+++ b/lib/intel_compute_square_kernels.c
@@ -112,6 +112,40 @@ static const unsigned char xehp_kernel_square_bin[] = {
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
+static const unsigned char xehpc_kernel_square_bin[] = {
+ 0x65, 0xa1, 0x00, 0x80, 0x20, 0x82, 0x05, 0x7f, 0x04, 0x00, 0x00, 0x02,
+ 0xc0, 0xff, 0xff, 0xff, 0x40, 0x19, 0x00, 0x80, 0x20, 0x82, 0x05, 0x7f,
+ 0x04, 0x7f, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00, 0x31, 0x22, 0x03, 0x00,
+ 0x00, 0x00, 0x0c, 0x04, 0x8f, 0x7f, 0x00, 0xfa, 0x03, 0x00, 0x34, 0xf6,
+ 0x66, 0x09, 0x84, 0xb4, 0x80, 0x80, 0x00, 0x4c, 0x41, 0x22, 0x03, 0x80,
+ 0x60, 0x06, 0x01, 0x20, 0xd4, 0x04, 0x00, 0x01, 0x14, 0x00, 0x00, 0x00,
+ 0x53, 0x80, 0x00, 0x80, 0x60, 0x06, 0x05, 0x02, 0xd4, 0x04, 0x00, 0x06,
+ 0x14, 0x00, 0x00, 0x00, 0x52, 0x19, 0x14, 0x00, 0x60, 0x06, 0x04, 0x05,
+ 0x04, 0x02, 0x0e, 0x01, 0x04, 0x01, 0x04, 0x04, 0x70, 0x19, 0x14, 0x00,
+ 0x20, 0x02, 0x01, 0x00, 0x04, 0x05, 0x10, 0x52, 0xc4, 0x04, 0x00, 0x00,
+ 0x2e, 0x00, 0x14, 0x14, 0x00, 0xc0, 0x00, 0x00, 0x78, 0x00, 0x00, 0x00,
+ 0x78, 0x00, 0x00, 0x00, 0x61, 0x00, 0x00, 0x6c, 0x13, 0x05, 0x00, 0x00,
+ 0x61, 0x00, 0x08, 0x6c, 0x15, 0x06, 0x00, 0x00, 0x69, 0x1a, 0x00, 0xf9,
+ 0x17, 0x13, 0x20, 0x00, 0x69, 0x1a, 0x08, 0xf9, 0x19, 0x15, 0x20, 0x00,
+ 0x40, 0x1a, 0x00, 0x20, 0x07, 0x17, 0x60, 0x04, 0x40, 0x1a, 0x08, 0x20,
+ 0x09, 0x19, 0x60, 0x04, 0x31, 0x23, 0x15, 0x00, 0x00, 0x00, 0x14, 0x0b,
+ 0x24, 0x07, 0x00, 0xfb, 0x00, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x20,
+ 0x0f, 0x17, 0x30, 0x04, 0x40, 0x00, 0x08, 0x20, 0x11, 0x19, 0x30, 0x04,
+ 0x41, 0x83, 0x14, 0x2c, 0x0d, 0x0b, 0x10, 0x0b, 0x31, 0x24, 0x15, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x24, 0x0f, 0x08, 0xfb, 0x14, 0x0d, 0x00, 0x00,
+ 0x2f, 0x00, 0x14, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x10, 0x00, 0x00, 0x00, 0x61, 0x00, 0x1c, 0x34, 0x7f, 0x00, 0x00, 0x00,
+ 0x31, 0x11, 0x0c, 0x80, 0x04, 0x00, 0x00, 0x00, 0x0c, 0x7f, 0x20, 0x30,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+};
+
const struct compute_kernels compute_square_kernels[] = {
{
.ip_ver = IP_VER(12, 0),
@@ -123,5 +157,10 @@ const struct compute_kernels compute_square_kernels[] = {
.size = sizeof(xehp_kernel_square_bin),
.kernel = xehp_kernel_square_bin,
},
+ {
+ .ip_ver = IP_VER(12, 60),
+ .size = sizeof(xehpc_kernel_square_bin),
+ .kernel = xehpc_kernel_square_bin,
+ },
{}
};
--
2.34.1
More information about the igt-dev
mailing list