[PATCH v6 2/5] lib/gpgpu_shader: tooling for preparing and running gpgpu shaders

[Andrzej Hajda <andrzej.hajda@intel.com>]

Implement tooling for building shaders for specific generations.
The library allows you to build and run shader from precompiled blocks
and provides an abstraction layer over gpgpu pipeline.

Signed-off-by: Dominik Grzegorzek <dominik.grzegorzek@intel.com>
Signed-off-by: Christoph Manszewski <christoph.manszewski@intel.com>
Signed-off-by: Andrzej Hajda <andrzej.hajda@intel.com>
---
 lib/gpgpu_shader.c | 211 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 lib/gpgpu_shader.h |  38 ++++++++++
 lib/meson.build    |   1 +
 3 files changed, 250 insertions(+)

diff --git a/lib/gpgpu_shader.c b/lib/gpgpu_shader.c
new file mode 100644
index 000000000000..d14301789421
--- /dev/null
+++ b/lib/gpgpu_shader.c
@@ -0,0 +1,211 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2024 Intel Corporation
+ *
+ * Author: Dominik Grzegorzek <dominik.grzegorzek@intel.com>
+ */
+
+#include <i915_drm.h>
+
+#include "ioctl_wrappers.h"
+#include "gpgpu_shader.h"
+#include "gpu_cmds.h"
+
+#define SUPPORTED_GEN_VER 1200 /* Support TGL and up */
+
+#define PAGE_SIZE 4096
+#define BATCH_STATE_SPLIT 2048
+/* VFE STATE params */
+#define THREADS (1 << 16) /* max value */
+#define GEN8_GPGPU_URB_ENTRIES 1
+#define GPGPU_URB_SIZE 0
+#define GPGPU_CURBE_SIZE 0
+#define GEN7_VFE_STATE_GPGPU_MODE 1
+
+static uint32_t fill_sip(struct intel_bb *ibb,
+			 const uint32_t sip[][4],
+			 const size_t size)
+{
+	uint32_t *sip_dst;
+	uint32_t offset;
+
+	intel_bb_ptr_align(ibb, 16);
+	sip_dst = intel_bb_ptr(ibb);
+	offset = intel_bb_offset(ibb);
+
+	memcpy(sip_dst, sip, size);
+
+	intel_bb_ptr_add(ibb, size);
+
+	return offset;
+}
+
+static void emit_sip(struct intel_bb *ibb, const uint64_t offset)
+{
+	intel_bb_out(ibb, GEN4_STATE_SIP | (3 - 2));
+	intel_bb_out(ibb, lower_32_bits(offset));
+	intel_bb_out(ibb, upper_32_bits(offset));
+}
+
+static void
+__xelp_gpgpu_execfunc(struct intel_bb *ibb,
+		      struct intel_buf *target,
+		      unsigned int x_dim, unsigned int y_dim,
+		      struct gpgpu_shader *shdr,
+		      struct gpgpu_shader *sip,
+		      uint64_t ring, bool explicit_engine)
+{
+	uint32_t interface_descriptor, sip_offset;
+	uint64_t engine;
+
+	intel_bb_add_intel_buf(ibb, target, true);
+
+	intel_bb_ptr_set(ibb, BATCH_STATE_SPLIT);
+
+	interface_descriptor = gen8_fill_interface_descriptor(ibb, target,
+							      shdr->instr,
+							      4 * shdr->size);
+
+	if (sip && sip->size)
+		sip_offset = fill_sip(ibb, sip->instr, 4 * sip->size);
+	else
+		sip_offset = 0;
+
+	intel_bb_ptr_set(ibb, 0);
+
+	/* GPGPU pipeline */
+	intel_bb_out(ibb, GEN7_PIPELINE_SELECT | GEN9_PIPELINE_SELECTION_MASK |
+		     PIPELINE_SELECT_GPGPU);
+
+	gen9_emit_state_base_address(ibb);
+
+	xelp_emit_vfe_state(ibb, THREADS, GEN8_GPGPU_URB_ENTRIES,
+			    GPGPU_URB_SIZE, GPGPU_CURBE_SIZE, true);
+
+	gen7_emit_interface_descriptor_load(ibb, interface_descriptor);
+
+	if (sip_offset)
+		emit_sip(ibb, sip_offset);
+
+	gen8_emit_gpgpu_walk(ibb, 0, 0, x_dim * 16, y_dim);
+
+	intel_bb_out(ibb, MI_BATCH_BUFFER_END);
+	intel_bb_ptr_align(ibb, 32);
+
+	engine = explicit_engine ? ring : I915_EXEC_DEFAULT;
+	intel_bb_exec(ibb, intel_bb_offset(ibb),
+		      engine | I915_EXEC_NO_RELOC, false);
+}
+
+static void
+__xehp_gpgpu_execfunc(struct intel_bb *ibb,
+		      struct intel_buf *target,
+		      unsigned int x_dim, unsigned int y_dim,
+		      struct gpgpu_shader *shdr,
+		      struct gpgpu_shader *sip,
+		      uint64_t ring, bool explicit_engine)
+{
+	struct xehp_interface_descriptor_data idd;
+	uint32_t sip_offset;
+	uint64_t engine;
+
+	intel_bb_add_intel_buf(ibb, target, true);
+
+	intel_bb_ptr_set(ibb, BATCH_STATE_SPLIT);
+
+	xehp_fill_interface_descriptor(ibb, target, shdr->instr,
+				       4 * shdr->size, &idd);
+
+	if (sip && sip->size)
+		sip_offset = fill_sip(ibb, sip->instr, 4 * sip->size);
+	else
+		sip_offset = 0;
+
+	intel_bb_ptr_set(ibb, 0);
+
+	/* GPGPU pipeline */
+	intel_bb_out(ibb, GEN7_PIPELINE_SELECT | GEN9_PIPELINE_SELECTION_MASK |
+		     PIPELINE_SELECT_GPGPU);
+	xehp_emit_state_base_address(ibb);
+	xehp_emit_state_compute_mode(ibb);
+	xehp_emit_state_binding_table_pool_alloc(ibb);
+	xehp_emit_cfe_state(ibb, THREADS);
+
+	if (sip_offset)
+		emit_sip(ibb, sip_offset);
+
+	xehp_emit_compute_walk(ibb, 0, 0, x_dim * 16, y_dim, &idd, 0x0);
+
+	intel_bb_out(ibb, MI_BATCH_BUFFER_END);
+	intel_bb_ptr_align(ibb, 32);
+
+	engine = explicit_engine ? ring : I915_EXEC_DEFAULT;
+	intel_bb_exec(ibb, intel_bb_offset(ibb),
+		      engine | I915_EXEC_NO_RELOC, false);
+
+}
+
+/**
+ * gpgpu_shader_exec:
+ * @ibb: pointer to initialized intel_bb
+ * @target: pointer to initialized intel_buf to be written by shader/sip
+ * @x_dim: gpgpu/compute walker thread group width
+ * @y_dim: gpgpu/compute walker thread group height
+ * @shdr: shader to be executed
+ * @sip: sip to be executed, can be NULL
+ * @ring: engine index
+ * @explicit_engine: whether to use provided engine index
+ *
+ * Execute provided shader in asynchronous fashion. To wait for completion,
+ * caller has to use the provided ibb handle.
+ */
+void gpgpu_shader_exec(struct intel_bb *ibb,
+		       struct intel_buf *target,
+		       unsigned int x_dim, unsigned int y_dim,
+		       struct gpgpu_shader *shdr,
+		       struct gpgpu_shader *sip,
+		       uint64_t ring, bool explicit_engine)
+{
+	igt_require(shdr->gen_ver >= SUPPORTED_GEN_VER);
+	igt_assert(ibb->size >= PAGE_SIZE);
+	igt_assert(ibb->ptr == ibb->batch);
+
+	if (shdr->gen_ver >= 1250)
+		__xehp_gpgpu_execfunc(ibb, target, x_dim, y_dim, shdr, sip,
+				      ring, explicit_engine);
+	else
+		__xelp_gpgpu_execfunc(ibb, target, x_dim, y_dim, shdr, sip,
+				      ring, explicit_engine);
+}
+
+/**
+ * gpgpu_shader_create:
+ * @fd: drm fd - i915 or xe
+ *
+ * Creates empty shader.
+ *
+ * Returns: pointer to empty shader struct.
+ */
+struct gpgpu_shader *gpgpu_shader_create(int fd)
+{
+	struct gpgpu_shader *shdr = calloc(1, sizeof(struct gpgpu_shader));
+	const struct intel_device_info *info;
+
+	info = intel_get_device_info(intel_get_drm_devid(fd));
+	shdr->gen_ver = 100 * info->graphics_ver + info->graphics_rel;
+	shdr->max_size = 16 * 4;
+	shdr->code = malloc(4 * shdr->max_size);
+	return shdr;
+}
+
+/**
+ * gpgpu_shader_destroy:
+ * @shdr: pointer to shader struct created with 'gpgpu_shader_create'
+ *
+ * Frees resources of gpgpu_shader struct.
+ */
+void gpgpu_shader_destroy(struct gpgpu_shader *shdr)
+{
+	free(shdr->code);
+	free(shdr);
+}
diff --git a/lib/gpgpu_shader.h b/lib/gpgpu_shader.h
new file mode 100644
index 000000000000..02f6f1aad1e3
--- /dev/null
+++ b/lib/gpgpu_shader.h
@@ -0,0 +1,38 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2024 Intel Corporation
+ */
+
+#ifndef GPGPU_SHADER_H
+#define GPGPU_SHADER_H
+
+#include <stdbool.h>
+#include <stdint.h>
+#include <stdlib.h>
+
+struct intel_bb;
+struct intel_buf;
+
+struct gpgpu_shader {
+	uint32_t gen_ver;
+	uint32_t size;
+	uint32_t max_size;
+	union {
+		uint32_t *code;
+		uint32_t (*instr)[4];
+	};
+};
+
+struct gpgpu_shader *gpgpu_shader_create(int fd);
+void gpgpu_shader_destroy(struct gpgpu_shader *shdr);
+
+void gpgpu_shader_dump(struct gpgpu_shader *shdr);
+
+void gpgpu_shader_exec(struct intel_bb *ibb,
+		       struct intel_buf *target,
+		       unsigned int x_dim, unsigned int y_dim,
+		       struct gpgpu_shader *shdr,
+		       struct gpgpu_shader *sip,
+		       uint64_t ring, bool explicit_engine);
+
+#endif /* GPGPU_SHADER_H */
diff --git a/lib/meson.build b/lib/meson.build
index e2f740c116f8..0a3084f8aea2 100644
--- a/lib/meson.build
+++ b/lib/meson.build
@@ -72,6 +72,7 @@ lib_sources = [
 	'media_spin.c',
 	'media_fill.c',
 	'gpgpu_fill.c',
+	'gpgpu_shader.c',
 	'gpu_cmds.c',
 	'rendercopy_i915.c',
 	'rendercopy_i830.c',

-- 
2.34.1