[igt-dev] [PATCH i-g-t v4 4/8] lib/gpu_cmds: Add gpgpu pipeline functions based on intel_bb
Zbigniew Kempczyński
zbigniew.kempczynski at intel.com
Tue May 19 08:11:30 UTC 2020
Add no-libdrm functions which will replace libdrm version in the
final version.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
Cc: Chris Wilson <chris at chris-wilson.co.uk>
---
lib/gpu_cmds.c | 559 +++++++++++++++++++++++++++++++++++++++++++++++++
lib/gpu_cmds.h | 54 +++++
2 files changed, 613 insertions(+)
diff --git a/lib/gpu_cmds.c b/lib/gpu_cmds.c
index dc0ae96c..db1a5f70 100644
--- a/lib/gpu_cmds.c
+++ b/lib/gpu_cmds.c
@@ -768,3 +768,562 @@ 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,
+ read_domain, write_domain,
+ offset + 4, 0x0);
+
+ 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,
+ read_domain, write_domain,
+ offset + 4, 0x0);
+
+ 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;
+}
+
+uint32_t
+gen8_fill_interface_descriptor_v2(struct intel_bb *ibb,
+ struct intel_buf *buf,
+ const uint32_t kernel[][4],
+ size_t size)
+{
+ struct gen8_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->desc2.single_program_flow = 1;
+ idd->desc2.floating_point_mode = GEN8_FLOATING_POINT_IEEE_754;
+
+ idd->desc3.sampler_count = 0; /* 0 samplers used */
+ idd->desc3.sampler_state_pointer = 0;
+
+ idd->desc4.binding_table_entry_count = 0;
+ idd->desc4.binding_table_pointer = (binding_table_offset >> 5);
+
+ idd->desc5.constant_urb_entry_read_offset = 0;
+ idd->desc5.constant_urb_entry_read_length = 1; /* grf 1 */
+
+ idd->desc6.num_threads_in_tg = 1;
+
+ 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_emit_reloc(ibb, ibb->handle,
+ I915_GEM_DOMAIN_INSTRUCTION, 0,
+ BASE_ADDRESS_MODIFY, 0x0);
+
+ /* dynamic */
+ intel_bb_emit_reloc(ibb, ibb->handle,
+ I915_GEM_DOMAIN_INSTRUCTION, 0,
+ BASE_ADDRESS_MODIFY, 0x0);
+
+ /* indirect */
+ intel_bb_out(ibb, 0);
+
+ /* instruction */
+ intel_bb_emit_reloc(ibb, ibb->handle,
+ I915_GEM_DOMAIN_INSTRUCTION, 0,
+ BASE_ADDRESS_MODIFY, 0x0);
+
+ /* 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
+gen8_emit_state_base_address_v2(struct intel_bb *ibb)
+{
+ intel_bb_out(ibb, GEN8_STATE_BASE_ADDRESS | (16 - 2));
+
+ /* general */
+ intel_bb_out(ibb, 0 | BASE_ADDRESS_MODIFY);
+ intel_bb_out(ibb, 0);
+
+ /* stateless data port */
+ intel_bb_out(ibb, 0 | BASE_ADDRESS_MODIFY);
+
+ /* surface */
+ intel_bb_emit_reloc(ibb, ibb->handle,
+ I915_GEM_DOMAIN_SAMPLER, 0,
+ BASE_ADDRESS_MODIFY, 0x0);
+
+ /* dynamic */
+ intel_bb_emit_reloc(ibb, ibb->handle,
+ I915_GEM_DOMAIN_RENDER | I915_GEM_DOMAIN_INSTRUCTION,
+ 0,
+ BASE_ADDRESS_MODIFY, 0x0);
+
+ /* indirect */
+ intel_bb_out(ibb, 0);
+ intel_bb_out(ibb, 0);
+
+ /* instruction */
+ intel_bb_emit_reloc(ibb, ibb->handle,
+ I915_GEM_DOMAIN_INSTRUCTION,
+ 0,
+ BASE_ADDRESS_MODIFY, 0x0);
+
+
+ /* general state buffer size */
+ intel_bb_out(ibb, 0xfffff000 | 1);
+ /* dynamic state buffer size */
+ intel_bb_out(ibb, 1 << 12 | 1);
+ /* indirect object buffer size */
+ intel_bb_out(ibb, 0xfffff000 | 1);
+ /* instruction buffer size, must set modify enable bit, otherwise it may
+ * result in GPU hang
+ */
+ intel_bb_out(ibb, 1 << 12 | 1);
+}
+
+
+void
+gen9_emit_state_base_address_v2(struct intel_bb *ibb)
+{
+ intel_bb_out(ibb, GEN8_STATE_BASE_ADDRESS | (19 - 2));
+
+ /* general */
+ intel_bb_out(ibb, 0 | BASE_ADDRESS_MODIFY);
+ intel_bb_out(ibb, 0);
+
+ /* stateless data port */
+ intel_bb_out(ibb, 0 | BASE_ADDRESS_MODIFY);
+
+ /* surface */
+ intel_bb_emit_reloc(ibb, ibb->handle,
+ I915_GEM_DOMAIN_SAMPLER, 0,
+ BASE_ADDRESS_MODIFY, 0x0);
+
+ /* dynamic */
+ intel_bb_emit_reloc(ibb, ibb->handle,
+ I915_GEM_DOMAIN_RENDER | I915_GEM_DOMAIN_INSTRUCTION,
+ 0,
+ BASE_ADDRESS_MODIFY, 0x0);
+
+ /* indirect */
+ intel_bb_out(ibb, 0);
+ intel_bb_out(ibb, 0);
+
+ /* instruction */
+ intel_bb_emit_reloc(ibb, ibb->handle,
+ I915_GEM_DOMAIN_INSTRUCTION, 0,
+ BASE_ADDRESS_MODIFY, 0x0);
+
+ /* general state buffer size */
+ intel_bb_out(ibb, 0xfffff000 | 1);
+ /* dynamic state buffer size */
+ intel_bb_out(ibb, 1 << 12 | 1);
+ /* indirect object buffer size */
+ intel_bb_out(ibb, 0xfffff000 | 1);
+ /* intruction buffer size, must set modify enable bit, otherwise it may
+ * result in GPU hang
+ */
+ intel_bb_out(ibb, 1 << 12 | 1);
+
+ /* Bindless surface state base address */
+ intel_bb_out(ibb, 0 | BASE_ADDRESS_MODIFY);
+ intel_bb_out(ibb, 0);
+ intel_bb_out(ibb, 0xfffff000);
+}
+
+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
+gen8_emit_vfe_state_v2(struct intel_bb *ibb, uint32_t threads,
+ uint32_t urb_entries, uint32_t urb_size,
+ uint32_t curbe_size)
+{
+ intel_bb_out(ibb, GEN7_MEDIA_VFE_STATE | (9 - 2));
+
+ /* scratch buffer */
+ intel_bb_out(ibb, 0);
+ intel_bb_out(ibb, 0);
+
+ /* number of threads & urb entries */
+ intel_bb_out(ibb, threads << 16 | urb_entries << 8);
+
+ intel_bb_out(ibb, 0);
+
+ /* urb entry size & curbe size */
+ intel_bb_out(ibb, urb_size << 16 | curbe_size);
+
+ /* 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);
+}
+
+void
+gen8_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 | 13);
+
+ intel_bb_out(ibb, 0); /* kernel offset */
+ intel_bb_out(ibb, 0); /* indirect data length */
+ intel_bb_out(ibb, 0); /* indirect data offset */
+
+ /* 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, 0);
+ intel_bb_out(ibb, x_dim);
+
+ /* thread group Y */
+ intel_bb_out(ibb, 0);
+ 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..61aff153 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 +149,57 @@ 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);
+
+uint32_t
+gen8_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
+gen8_emit_state_base_address_v2(struct intel_bb *ibb);
+
+void
+gen9_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
+gen8_emit_vfe_state_v2(struct intel_bb *ibb, uint32_t threads,
+ uint32_t urb_entries, uint32_t urb_size,
+ uint32_t curbe_size);
+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);
+
+void
+gen8_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
More information about the igt-dev
mailing list