[igt-dev] [PATCH i-g-t 4/6] lib/gpu_cmds: Add gpgpu pipeline functions based on intel_bb

["Zbigniew Kempczyński" <zbigniew.kempczynski@intel.com>]

Add no-libdrm functions which will replace libdrm version in the
final version.

Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
---
 lib/gpu_cmds.c | 340 +++++++++++++++++++++++++++++++++++++++++++++++++
 lib/gpu_cmds.h |  33 +++++
 2 files changed, 373 insertions(+)

diff --git a/lib/gpu_cmds.c b/lib/gpu_cmds.c
index dc0ae96c..c5d5e809 100644
--- a/lib/gpu_cmds.c
+++ b/lib/gpu_cmds.c
@@ -768,3 +768,343 @@ gen9_emit_state_base_address(struct intel_batchbuffer *batch)
 	OUT_BATCH(0);
 	OUT_BATCH(0xfffff000);
 }
+
+/*
+ * Here we start version of the gpgpu fill pipeline creation which is based
+ * on intel_bb.
+ */
+uint32_t
+gen7_fill_curbe_buffer_data_v2(struct intel_bb *ibb, uint8_t color)
+{
+	uint32_t *curbe_buffer;
+	uint32_t offset;
+
+	intel_bb_ptr_align(ibb, 64);
+	curbe_buffer = intel_bb_ptr(ibb);
+	offset = intel_bb_offset(ibb);
+
+	*curbe_buffer = color;
+	intel_bb_ptr_add(ibb, 32);
+
+	return offset;
+}
+
+static uint32_t
+gen7_fill_kernel_v2(struct intel_bb *ibb,
+		    const uint32_t kernel[][4],
+		    size_t size)
+{
+	uint32_t *kernel_dst;
+	uint32_t offset;
+
+	intel_bb_ptr_align(ibb, 64);
+	kernel_dst = intel_bb_ptr(ibb);
+	offset = intel_bb_offset(ibb);
+
+	memcpy(kernel_dst, kernel, size);
+
+	intel_bb_ptr_add(ibb, size);
+
+	return offset;
+}
+
+static uint32_t
+gen7_fill_surface_state_v2(struct intel_bb *ibb,
+			   struct intel_buf *buf,
+			   uint32_t format,
+			   int is_dst)
+{
+	struct gen7_surface_state *ss;
+	uint32_t write_domain, read_domain, offset;
+
+	if (is_dst) {
+		write_domain = read_domain = I915_GEM_DOMAIN_RENDER;
+	} else {
+		write_domain = 0;
+		read_domain = I915_GEM_DOMAIN_SAMPLER;
+	}
+
+	intel_bb_ptr_align(ibb, 64);
+	offset = intel_bb_offset(ibb);
+	ss = intel_bb_ptr(ibb);
+	intel_bb_ptr_add(ibb, 64);
+
+	ss->ss0.surface_type = SURFACE_2D;
+	ss->ss0.surface_format = format;
+	ss->ss0.render_cache_read_write = 1;
+
+	if (buf->tiling == I915_TILING_X)
+		ss->ss0.tiled_mode = 2;
+	else if (buf->tiling == I915_TILING_Y)
+		ss->ss0.tiled_mode = 3;
+
+	ss->ss1.base_addr = 0;
+
+	intel_bb_offset_reloc(ibb, buf->handle, offset + 4,
+			      read_domain, write_domain);
+
+	ss->ss2.height = intel_buf_height(buf) - 1;
+	ss->ss2.width  = intel_buf_width(buf) - 1;
+	ss->ss3.pitch  = buf->stride - 1;
+
+	ss->ss7.shader_chanel_select_r = 4;
+	ss->ss7.shader_chanel_select_g = 5;
+	ss->ss7.shader_chanel_select_b = 6;
+	ss->ss7.shader_chanel_select_a = 7;
+
+	return offset;
+}
+
+static uint32_t
+gen8_fill_surface_state_v2(struct intel_bb *ibb,
+			   struct intel_buf *buf,
+			   uint32_t format,
+			   int is_dst)
+{
+	struct gen8_surface_state *ss;
+	uint32_t write_domain, read_domain, offset;
+
+	if (is_dst) {
+		write_domain = read_domain = I915_GEM_DOMAIN_RENDER;
+	} else {
+		write_domain = 0;
+		read_domain = I915_GEM_DOMAIN_SAMPLER;
+	}
+
+	intel_bb_ptr_align(ibb, 64);
+	offset = intel_bb_offset(ibb);
+	ss = intel_bb_ptr(ibb);
+	intel_bb_ptr_add(ibb, 64);
+
+	ss->ss0.surface_type = SURFACE_2D;
+	ss->ss0.surface_format = format;
+	ss->ss0.render_cache_read_write = 1;
+	ss->ss0.vertical_alignment = 1; /* align 4 */
+	ss->ss0.horizontal_alignment = 1; /* align 4 */
+
+	if (buf->tiling == I915_TILING_X)
+		ss->ss0.tiled_mode = 2;
+	else if (buf->tiling == I915_TILING_Y)
+		ss->ss0.tiled_mode = 3;
+
+	ss->ss8.base_addr = 0;
+
+	intel_bb_offset_reloc(ibb, buf->handle, offset + 4,
+			      read_domain, write_domain);
+
+	ss->ss2.height = intel_buf_height(buf) - 1;
+	ss->ss2.width  = intel_buf_width(buf) - 1;
+	ss->ss3.pitch  = buf->stride - 1;
+
+	ss->ss7.shader_chanel_select_r = 4;
+	ss->ss7.shader_chanel_select_g = 5;
+	ss->ss7.shader_chanel_select_b = 6;
+	ss->ss7.shader_chanel_select_a = 7;
+
+	return offset;
+}
+
+static uint32_t
+gen7_fill_binding_table_v2(struct intel_bb *ibb,
+			   struct intel_buf *buf)
+{
+	uint32_t binding_table_offset;
+	uint32_t *binding_table;
+	uint32_t devid = intel_get_drm_devid(ibb->i915);
+
+	intel_bb_ptr_align(ibb, 64);
+	binding_table_offset = intel_bb_offset(ibb);
+	binding_table = intel_bb_ptr(ibb);
+	intel_bb_ptr_add(ibb, 64);
+
+	if (IS_GEN7(devid))
+		binding_table[0] = gen7_fill_surface_state_v2(ibb, buf,
+							      SURFACEFORMAT_R8_UNORM, 1);
+
+	else
+		binding_table[0] = gen8_fill_surface_state_v2(ibb, buf,
+							      SURFACEFORMAT_R8_UNORM, 1);
+
+	return binding_table_offset;
+}
+
+uint32_t
+gen7_fill_interface_descriptor_v2(struct intel_bb *ibb,
+				  struct intel_buf *buf,
+				  const uint32_t kernel[][4],
+				  size_t size)
+{
+	struct gen7_interface_descriptor_data *idd;
+	uint32_t offset;
+	uint32_t binding_table_offset, kernel_offset;
+
+	binding_table_offset = gen7_fill_binding_table_v2(ibb, buf);
+	kernel_offset = gen7_fill_kernel_v2(ibb, kernel, size);
+
+	intel_bb_ptr_align(ibb, 64);
+	idd = intel_bb_ptr(ibb);
+	offset = intel_bb_offset(ibb);
+
+	idd->desc0.kernel_start_pointer = (kernel_offset >> 6);
+
+	idd->desc1.single_program_flow = 1;
+	idd->desc1.floating_point_mode = GEN7_FLOATING_POINT_IEEE_754;
+
+	idd->desc2.sampler_count = 0;      /* 0 samplers used */
+	idd->desc2.sampler_state_pointer = 0;
+
+	idd->desc3.binding_table_entry_count = 0;
+	idd->desc3.binding_table_pointer = (binding_table_offset >> 5);
+
+	idd->desc4.constant_urb_entry_read_offset = 0;
+	idd->desc4.constant_urb_entry_read_length = 1; /* grf 1 */
+
+	intel_bb_ptr_add(ibb, sizeof(*idd));
+
+	return offset;
+}
+
+void
+gen7_emit_state_base_address_v2(struct intel_bb *ibb)
+{
+	intel_bb_out(ibb, GEN7_STATE_BASE_ADDRESS | (10 - 2));
+
+	/* general */
+	intel_bb_out(ibb, 0);
+
+	/* surface */
+	intel_bb_out_reloc(ibb,
+			   I915_GEM_DOMAIN_INSTRUCTION, 0,
+			   BASE_ADDRESS_MODIFY);
+
+	/* dynamic */
+	intel_bb_out_reloc(ibb,
+			   I915_GEM_DOMAIN_INSTRUCTION, 0,
+			   BASE_ADDRESS_MODIFY);
+
+	/* indirect */
+	intel_bb_out(ibb, 0);
+
+	/* instruction */
+	intel_bb_out_reloc(ibb,
+			   I915_GEM_DOMAIN_INSTRUCTION, 0,
+			   BASE_ADDRESS_MODIFY);
+
+	/* general/dynamic/indirect/instruction access Bound */
+	intel_bb_out(ibb, 0);
+	intel_bb_out(ibb, 0 | BASE_ADDRESS_MODIFY);
+	intel_bb_out(ibb, 0);
+	intel_bb_out(ibb, 0 | BASE_ADDRESS_MODIFY);
+}
+
+void
+gen7_emit_vfe_state_v2(struct intel_bb *ibb, uint32_t threads,
+		       uint32_t urb_entries, uint32_t urb_size,
+		       uint32_t curbe_size, uint32_t mode)
+{
+	intel_bb_out(ibb, GEN7_MEDIA_VFE_STATE | (8 - 2));
+
+	/* scratch buffer */
+	intel_bb_out(ibb, 0);
+
+	/* number of threads & urb entries */
+	intel_bb_out(ibb, threads << 16 |
+		     urb_entries << 8 |
+		     mode << 2); /* GPGPU vs media mode */
+
+	intel_bb_out(ibb, 0);
+
+	/* urb entry size & curbe size */
+	intel_bb_out(ibb, urb_size << 16 |	/* in 256 bits unit */
+		     curbe_size);		/* in 256 bits unit */
+
+	/* scoreboard */
+	intel_bb_out(ibb, 0);
+	intel_bb_out(ibb, 0);
+	intel_bb_out(ibb, 0);
+}
+
+void
+gen7_emit_curbe_load_v2(struct intel_bb *ibb, uint32_t curbe_buffer)
+{
+	intel_bb_out(ibb, GEN7_MEDIA_CURBE_LOAD | (4 - 2));
+	intel_bb_out(ibb, 0);
+	/* curbe total data length */
+	intel_bb_out(ibb, 64);
+	/* curbe data start address, is relative to the dynamics base address */
+	intel_bb_out(ibb, curbe_buffer);
+}
+
+void
+gen7_emit_interface_descriptor_load_v2(struct intel_bb *ibb,
+				       uint32_t interface_descriptor)
+{
+	intel_bb_out(ibb, GEN7_MEDIA_INTERFACE_DESCRIPTOR_LOAD | (4 - 2));
+	intel_bb_out(ibb, 0);
+	/* interface descriptor data length */
+	if (ibb->gen == 7)
+		intel_bb_out(ibb, sizeof(struct gen7_interface_descriptor_data));
+	else
+		intel_bb_out(ibb, sizeof(struct gen8_interface_descriptor_data));
+	/* interface descriptor address, is relative to the dynamics base
+	 * address
+	 */
+	intel_bb_out(ibb, interface_descriptor);
+}
+
+void
+gen7_emit_gpgpu_walk_v2(struct intel_bb *ibb,
+			unsigned int x, unsigned int y,
+			unsigned int width, unsigned int height)
+{
+	uint32_t x_dim, y_dim, tmp, right_mask;
+
+	/*
+	 * Simply do SIMD16 based dispatch, so every thread uses
+	 * SIMD16 channels.
+	 *
+	 * Define our own thread group size, e.g 16x1 for every group, then
+	 * will have 1 thread each group in SIMD16 dispatch. So thread
+	 * width/height/depth are all 1.
+	 *
+	 * Then thread group X = width / 16 (aligned to 16)
+	 * thread group Y = height;
+	 */
+	x_dim = (width + 15) / 16;
+	y_dim = height;
+
+	tmp = width & 15;
+	if (tmp == 0)
+		right_mask = (1 << 16) - 1;
+	else
+		right_mask = (1 << tmp) - 1;
+
+	intel_bb_out(ibb, GEN7_GPGPU_WALKER | 9);
+
+	/* interface descriptor offset */
+	intel_bb_out(ibb, 0);
+
+	/* SIMD size, thread w/h/d */
+	intel_bb_out(ibb, 1 << 30 | /* SIMD16 */
+		  0 << 16 | /* depth:1 */
+		  0 << 8 | /* height:1 */
+		  0); /* width:1 */
+
+	/* thread group X */
+	intel_bb_out(ibb, 0);
+	intel_bb_out(ibb, x_dim);
+
+	/* thread group Y */
+	intel_bb_out(ibb, 0);
+	intel_bb_out(ibb, y_dim);
+
+	/* thread group Z */
+	intel_bb_out(ibb, 0);
+	intel_bb_out(ibb, 1);
+
+	/* right mask */
+	intel_bb_out(ibb, right_mask);
+
+	/* bottom mask, height 1, always 0xffffffff */
+	intel_bb_out(ibb, 0xffffffff);
+}
diff --git a/lib/gpu_cmds.h b/lib/gpu_cmds.h
index 1321af44..728b4e68 100644
--- a/lib/gpu_cmds.h
+++ b/lib/gpu_cmds.h
@@ -35,6 +35,7 @@
 #include "drmtest.h"
 #include "intel_batchbuffer.h"
 #include "intel_chipset.h"
+#include "intel_bufops.h"
 #include <assert.h>
 
 void
@@ -148,4 +149,36 @@ gen_emit_media_object(struct intel_batchbuffer *batch, unsigned int xoffset,
 void
 gen9_emit_state_base_address(struct intel_batchbuffer *batch);
 
+
+/* No libdrm */
+uint32_t
+gen7_fill_curbe_buffer_data_v2(struct intel_bb *ibb,
+			       uint8_t color);
+
+uint32_t
+gen7_fill_interface_descriptor_v2(struct intel_bb *ibb,
+				  struct intel_buf *buf,
+				  const uint32_t kernel[][4],
+				  size_t size);
+
+void
+gen7_emit_state_base_address_v2(struct intel_bb *ibb);
+
+void
+gen7_emit_vfe_state_v2(struct intel_bb *ibb, uint32_t threads,
+		       uint32_t urb_entries, uint32_t urb_size,
+		       uint32_t curbe_size, uint32_t mode);
+
+void
+gen7_emit_curbe_load_v2(struct intel_bb *ibb, uint32_t curbe_buffer);
+
+void
+gen7_emit_interface_descriptor_load_v2(struct intel_bb *ibb,
+				       uint32_t interface_descriptor);
+
+void
+gen7_emit_gpgpu_walk_v2(struct intel_bb *ibb,
+			unsigned int x, unsigned int y,
+			unsigned int width, unsigned int height);
+
 #endif /* GPU_CMDS_H */
-- 
2.26.0

_______________________________________________
igt-dev mailing list
igt-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/igt-dev