/* * Copyright © 2017 Intel Corporation * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included * in all copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER * DEALINGS IN THE SOFTWARE. */ /** * @file iris_state.c * * ============================= GENXML CODE ============================= * [This file is compiled once per generation.] * ======================================================================= * * This is the main state upload code. * * Gallium uses Constant State Objects, or CSOs, for most state. Large, * complex, or highly reusable state can be created once, and bound and * rebound multiple times. This is modeled with the pipe->create_*_state() * and pipe->bind_*_state() hooks. Highly dynamic or inexpensive state is * streamed out on the fly, via pipe->set_*_state() hooks. * * OpenGL involves frequently mutating context state, which is mirrored in * core Mesa by highly mutable data structures. However, most applications * typically draw the same things over and over - from frame to frame, most * of the same objects are still visible and need to be redrawn. So, rather * than inventing new state all the time, applications usually mutate to swap * between known states that we've seen before. * * Gallium isolates us from this mutation by tracking API state, and * distilling it into a set of Constant State Objects, or CSOs. Large, * complex, or typically reusable state can be created once, then reused * multiple times. Drivers can create and store their own associated data. * This create/bind model corresponds to the pipe->create_*_state() and * pipe->bind_*_state() driver hooks. * * Some state is cheap to create, or expected to be highly dynamic. Rather * than creating and caching piles of CSOs for these, Gallium simply streams * them out, via the pipe->set_*_state() driver hooks. * * To reduce draw time overhead, we try to compute as much state at create * time as possible. Wherever possible, we translate the Gallium pipe state * to 3DSTATE commands, and store those commands in the CSO. At draw time, * we can simply memcpy them into a batch buffer. * * No hardware matches the abstraction perfectly, so some commands require * information from multiple CSOs. In this case, we can store two copies * of the packet (one in each CSO), and simply | together their DWords at * draw time. Sometimes the second set is trivial (one or two fields), so * we simply pack it at draw time. * * There are two main components in the file below. First, the CSO hooks * create/bind/track state. The second are the draw-time upload functions, * iris_upload_render_state() and iris_upload_compute_state(), which read * the context state and emit the commands into the actual batch. */ #include #include #if HAVE_VALGRIND #include #include #define VG(x) x #ifdef DEBUG #define __gen_validate_value(x) VALGRIND_CHECK_MEM_IS_DEFINED(&(x), sizeof(x)) #endif #else #define VG(x) #endif #include "pipe/p_defines.h" #include "pipe/p_state.h" #include "pipe/p_context.h" #include "pipe/p_screen.h" #include "util/u_dual_blend.h" #include "util/u_inlines.h" #include "util/u_format.h" #include "util/u_framebuffer.h" #include "util/u_transfer.h" #include "util/u_upload_mgr.h" #include "util/u_viewport.h" #include "drm-uapi/i915_drm.h" #include "nir.h" #include "intel/compiler/brw_compiler.h" #include "intel/common/gen_l3_config.h" #include "intel/common/gen_sample_positions.h" #include "iris_batch.h" #include "iris_context.h" #include "iris_defines.h" #include "iris_pipe.h" #include "iris_resource.h" #define __gen_address_type struct iris_address #define __gen_user_data struct iris_batch #define ARRAY_BYTES(x) (sizeof(uint32_t) * ARRAY_SIZE(x)) static uint64_t __gen_combine_address(struct iris_batch *batch, void *location, struct iris_address addr, uint32_t delta) { uint64_t result = addr.offset + delta; if (addr.bo) { iris_use_pinned_bo(batch, addr.bo, addr.write); /* Assume this is a general address, not relative to a base. */ result += addr.bo->gtt_offset; } return result; } #define __genxml_cmd_length(cmd) cmd ## _length #define __genxml_cmd_length_bias(cmd) cmd ## _length_bias #define __genxml_cmd_header(cmd) cmd ## _header #define __genxml_cmd_pack(cmd) cmd ## _pack #define _iris_pack_command(batch, cmd, dst, name) \ for (struct cmd name = { __genxml_cmd_header(cmd) }, \ *_dst = (void *)(dst); __builtin_expect(_dst != NULL, 1); \ ({ __genxml_cmd_pack(cmd)(batch, (void *)_dst, &name); \ _dst = NULL; \ })) #define iris_pack_command(cmd, dst, name) \ _iris_pack_command(NULL, cmd, dst, name) #define iris_pack_state(cmd, dst, name) \ for (struct cmd name = {}, \ *_dst = (void *)(dst); __builtin_expect(_dst != NULL, 1); \ __genxml_cmd_pack(cmd)(NULL, (void *)_dst, &name), \ _dst = NULL) #define iris_emit_cmd(batch, cmd, name) \ _iris_pack_command(batch, cmd, iris_get_command_space(batch, 4 * __genxml_cmd_length(cmd)), name) #define iris_emit_merge(batch, dwords0, dwords1, num_dwords) \ do { \ uint32_t *dw = iris_get_command_space(batch, 4 * num_dwords); \ for (uint32_t i = 0; i < num_dwords; i++) \ dw[i] = (dwords0)[i] | (dwords1)[i]; \ VG(VALGRIND_CHECK_MEM_IS_DEFINED(dw, num_dwords)); \ } while (0) #include "genxml/genX_pack.h" #include "genxml/gen_macros.h" #include "genxml/genX_bits.h" #if GEN_GEN == 8 #define MOCS_PTE 0x18 #define MOCS_WB 0x78 #else #define MOCS_PTE (1 << 1) #define MOCS_WB (2 << 1) #endif static uint32_t mocs(const struct iris_bo *bo) { return bo && bo->external ? MOCS_PTE : MOCS_WB; } /** * Statically assert that PIPE_* enums match the hardware packets. * (As long as they match, we don't need to translate them.) */ UNUSED static void pipe_asserts() { #define PIPE_ASSERT(x) STATIC_ASSERT((int)x) /* pipe_logicop happens to match the hardware. */ PIPE_ASSERT(PIPE_LOGICOP_CLEAR == LOGICOP_CLEAR); PIPE_ASSERT(PIPE_LOGICOP_NOR == LOGICOP_NOR); PIPE_ASSERT(PIPE_LOGICOP_AND_INVERTED == LOGICOP_AND_INVERTED); PIPE_ASSERT(PIPE_LOGICOP_COPY_INVERTED == LOGICOP_COPY_INVERTED); PIPE_ASSERT(PIPE_LOGICOP_AND_REVERSE == LOGICOP_AND_REVERSE); PIPE_ASSERT(PIPE_LOGICOP_INVERT == LOGICOP_INVERT); PIPE_ASSERT(PIPE_LOGICOP_XOR == LOGICOP_XOR); PIPE_ASSERT(PIPE_LOGICOP_NAND == LOGICOP_NAND); PIPE_ASSERT(PIPE_LOGICOP_AND == LOGICOP_AND); PIPE_ASSERT(PIPE_LOGICOP_EQUIV == LOGICOP_EQUIV); PIPE_ASSERT(PIPE_LOGICOP_NOOP == LOGICOP_NOOP); PIPE_ASSERT(PIPE_LOGICOP_OR_INVERTED == LOGICOP_OR_INVERTED); PIPE_ASSERT(PIPE_LOGICOP_COPY == LOGICOP_COPY); PIPE_ASSERT(PIPE_LOGICOP_OR_REVERSE == LOGICOP_OR_REVERSE); PIPE_ASSERT(PIPE_LOGICOP_OR == LOGICOP_OR); PIPE_ASSERT(PIPE_LOGICOP_SET == LOGICOP_SET); /* pipe_blend_func happens to match the hardware. */ PIPE_ASSERT(PIPE_BLENDFACTOR_ONE == BLENDFACTOR_ONE); PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_COLOR == BLENDFACTOR_SRC_COLOR); PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA == BLENDFACTOR_SRC_ALPHA); PIPE_ASSERT(PIPE_BLENDFACTOR_DST_ALPHA == BLENDFACTOR_DST_ALPHA); PIPE_ASSERT(PIPE_BLENDFACTOR_DST_COLOR == BLENDFACTOR_DST_COLOR); PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE == BLENDFACTOR_SRC_ALPHA_SATURATE); PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_COLOR == BLENDFACTOR_CONST_COLOR); PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_ALPHA == BLENDFACTOR_CONST_ALPHA); PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_COLOR == BLENDFACTOR_SRC1_COLOR); PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_ALPHA == BLENDFACTOR_SRC1_ALPHA); PIPE_ASSERT(PIPE_BLENDFACTOR_ZERO == BLENDFACTOR_ZERO); PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_COLOR == BLENDFACTOR_INV_SRC_COLOR); PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_ALPHA == BLENDFACTOR_INV_SRC_ALPHA); PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_ALPHA == BLENDFACTOR_INV_DST_ALPHA); PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_COLOR == BLENDFACTOR_INV_DST_COLOR); PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_COLOR == BLENDFACTOR_INV_CONST_COLOR); PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_ALPHA == BLENDFACTOR_INV_CONST_ALPHA); PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_COLOR == BLENDFACTOR_INV_SRC1_COLOR); PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_ALPHA == BLENDFACTOR_INV_SRC1_ALPHA); /* pipe_blend_func happens to match the hardware. */ PIPE_ASSERT(PIPE_BLEND_ADD == BLENDFUNCTION_ADD); PIPE_ASSERT(PIPE_BLEND_SUBTRACT == BLENDFUNCTION_SUBTRACT); PIPE_ASSERT(PIPE_BLEND_REVERSE_SUBTRACT == BLENDFUNCTION_REVERSE_SUBTRACT); PIPE_ASSERT(PIPE_BLEND_MIN == BLENDFUNCTION_MIN); PIPE_ASSERT(PIPE_BLEND_MAX == BLENDFUNCTION_MAX); /* pipe_stencil_op happens to match the hardware. */ PIPE_ASSERT(PIPE_STENCIL_OP_KEEP == STENCILOP_KEEP); PIPE_ASSERT(PIPE_STENCIL_OP_ZERO == STENCILOP_ZERO); PIPE_ASSERT(PIPE_STENCIL_OP_REPLACE == STENCILOP_REPLACE); PIPE_ASSERT(PIPE_STENCIL_OP_INCR == STENCILOP_INCRSAT); PIPE_ASSERT(PIPE_STENCIL_OP_DECR == STENCILOP_DECRSAT); PIPE_ASSERT(PIPE_STENCIL_OP_INCR_WRAP == STENCILOP_INCR); PIPE_ASSERT(PIPE_STENCIL_OP_DECR_WRAP == STENCILOP_DECR); PIPE_ASSERT(PIPE_STENCIL_OP_INVERT == STENCILOP_INVERT); /* pipe_sprite_coord_mode happens to match 3DSTATE_SBE */ PIPE_ASSERT(PIPE_SPRITE_COORD_UPPER_LEFT == UPPERLEFT); PIPE_ASSERT(PIPE_SPRITE_COORD_LOWER_LEFT == LOWERLEFT); #undef PIPE_ASSERT } static unsigned translate_prim_type(enum pipe_prim_type prim, uint8_t verts_per_patch) { static const unsigned map[] = { [PIPE_PRIM_POINTS] = _3DPRIM_POINTLIST, [PIPE_PRIM_LINES] = _3DPRIM_LINELIST, [PIPE_PRIM_LINE_LOOP] = _3DPRIM_LINELOOP, [PIPE_PRIM_LINE_STRIP] = _3DPRIM_LINESTRIP, [PIPE_PRIM_TRIANGLES] = _3DPRIM_TRILIST, [PIPE_PRIM_TRIANGLE_STRIP] = _3DPRIM_TRISTRIP, [PIPE_PRIM_TRIANGLE_FAN] = _3DPRIM_TRIFAN, [PIPE_PRIM_QUADS] = _3DPRIM_QUADLIST, [PIPE_PRIM_QUAD_STRIP] = _3DPRIM_QUADSTRIP, [PIPE_PRIM_POLYGON] = _3DPRIM_POLYGON, [PIPE_PRIM_LINES_ADJACENCY] = _3DPRIM_LINELIST_ADJ, [PIPE_PRIM_LINE_STRIP_ADJACENCY] = _3DPRIM_LINESTRIP_ADJ, [PIPE_PRIM_TRIANGLES_ADJACENCY] = _3DPRIM_TRILIST_ADJ, [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY] = _3DPRIM_TRISTRIP_ADJ, [PIPE_PRIM_PATCHES] = _3DPRIM_PATCHLIST_1 - 1, }; return map[prim] + (prim == PIPE_PRIM_PATCHES ? verts_per_patch : 0); } static unsigned translate_compare_func(enum pipe_compare_func pipe_func) { static const unsigned map[] = { [PIPE_FUNC_NEVER] = COMPAREFUNCTION_NEVER, [PIPE_FUNC_LESS] = COMPAREFUNCTION_LESS, [PIPE_FUNC_EQUAL] = COMPAREFUNCTION_EQUAL, [PIPE_FUNC_LEQUAL] = COMPAREFUNCTION_LEQUAL, [PIPE_FUNC_GREATER] = COMPAREFUNCTION_GREATER, [PIPE_FUNC_NOTEQUAL] = COMPAREFUNCTION_NOTEQUAL, [PIPE_FUNC_GEQUAL] = COMPAREFUNCTION_GEQUAL, [PIPE_FUNC_ALWAYS] = COMPAREFUNCTION_ALWAYS, }; return map[pipe_func]; } static unsigned translate_shadow_func(enum pipe_compare_func pipe_func) { /* Gallium specifies the result of shadow comparisons as: * * 1 if ref texel, * 0 otherwise. * * The hardware does: * * 0 if texel ref, * 1 otherwise. * * So we need to flip the operator and also negate. */ static const unsigned map[] = { [PIPE_FUNC_NEVER] = PREFILTEROPALWAYS, [PIPE_FUNC_LESS] = PREFILTEROPLEQUAL, [PIPE_FUNC_EQUAL] = PREFILTEROPNOTEQUAL, [PIPE_FUNC_LEQUAL] = PREFILTEROPLESS, [PIPE_FUNC_GREATER] = PREFILTEROPGEQUAL, [PIPE_FUNC_NOTEQUAL] = PREFILTEROPEQUAL, [PIPE_FUNC_GEQUAL] = PREFILTEROPGREATER, [PIPE_FUNC_ALWAYS] = PREFILTEROPNEVER, }; return map[pipe_func]; } static unsigned translate_cull_mode(unsigned pipe_face) { static const unsigned map[4] = { [PIPE_FACE_NONE] = CULLMODE_NONE, [PIPE_FACE_FRONT] = CULLMODE_FRONT, [PIPE_FACE_BACK] = CULLMODE_BACK, [PIPE_FACE_FRONT_AND_BACK] = CULLMODE_BOTH, }; return map[pipe_face]; } static unsigned translate_fill_mode(unsigned pipe_polymode) { static const unsigned map[4] = { [PIPE_POLYGON_MODE_FILL] = FILL_MODE_SOLID, [PIPE_POLYGON_MODE_LINE] = FILL_MODE_WIREFRAME, [PIPE_POLYGON_MODE_POINT] = FILL_MODE_POINT, [PIPE_POLYGON_MODE_FILL_RECTANGLE] = FILL_MODE_SOLID, }; return map[pipe_polymode]; } static unsigned translate_mip_filter(enum pipe_tex_mipfilter pipe_mip) { static const unsigned map[] = { [PIPE_TEX_MIPFILTER_NEAREST] = MIPFILTER_NEAREST, [PIPE_TEX_MIPFILTER_LINEAR] = MIPFILTER_LINEAR, [PIPE_TEX_MIPFILTER_NONE] = MIPFILTER_NONE, }; return map[pipe_mip]; } static uint32_t translate_wrap(unsigned pipe_wrap) { static const unsigned map[] = { [PIPE_TEX_WRAP_REPEAT] = TCM_WRAP, [PIPE_TEX_WRAP_CLAMP] = TCM_HALF_BORDER, [PIPE_TEX_WRAP_CLAMP_TO_EDGE] = TCM_CLAMP, [PIPE_TEX_WRAP_CLAMP_TO_BORDER] = TCM_CLAMP_BORDER, [PIPE_TEX_WRAP_MIRROR_REPEAT] = TCM_MIRROR, [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE] = TCM_MIRROR_ONCE, /* These are unsupported. */ [PIPE_TEX_WRAP_MIRROR_CLAMP] = -1, [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER] = -1, }; return map[pipe_wrap]; } static struct iris_address ro_bo(struct iris_bo *bo, uint64_t offset) { /* CSOs must pass NULL for bo! Otherwise it will add the BO to the * validation list at CSO creation time, instead of draw time. */ return (struct iris_address) { .bo = bo, .offset = offset }; } static struct iris_address rw_bo(struct iris_bo *bo, uint64_t offset) { /* CSOs must pass NULL for bo! Otherwise it will add the BO to the * validation list at CSO creation time, instead of draw time. */ return (struct iris_address) { .bo = bo, .offset = offset, .write = true }; } /** * Allocate space for some indirect state. * * Return a pointer to the map (to fill it out) and a state ref (for * referring to the state in GPU commands). */ static void * upload_state(struct u_upload_mgr *uploader, struct iris_state_ref *ref, unsigned size, unsigned alignment) { void *p = NULL; u_upload_alloc(uploader, 0, size, alignment, &ref->offset, &ref->res, &p); return p; } /** * Stream out temporary/short-lived state. * * This allocates space, pins the BO, and includes the BO address in the * returned offset (which works because all state lives in 32-bit memory * zones). */ static uint32_t * stream_state(struct iris_batch *batch, struct u_upload_mgr *uploader, struct pipe_resource **out_res, unsigned size, unsigned alignment, uint32_t *out_offset) { void *ptr = NULL; u_upload_alloc(uploader, 0, size, alignment, out_offset, out_res, &ptr); struct iris_bo *bo = iris_resource_bo(*out_res); iris_use_pinned_bo(batch, bo, false); *out_offset += iris_bo_offset_from_base_address(bo); iris_record_state_size(batch->state_sizes, *out_offset, size); return ptr; } /** * stream_state() + memcpy. */ static uint32_t emit_state(struct iris_batch *batch, struct u_upload_mgr *uploader, struct pipe_resource **out_res, const void *data, unsigned size, unsigned alignment) { unsigned offset = 0; uint32_t *map = stream_state(batch, uploader, out_res, size, alignment, &offset); if (map) memcpy(map, data, size); return offset; } /** * Did field 'x' change between 'old_cso' and 'new_cso'? * * (If so, we may want to set some dirty flags.) */ #define cso_changed(x) (!old_cso || (old_cso->x != new_cso->x)) #define cso_changed_memcmp(x) \ (!old_cso || memcmp(old_cso->x, new_cso->x, sizeof(old_cso->x)) != 0) static void flush_for_state_base_change(struct iris_batch *batch) { /* Flush before emitting STATE_BASE_ADDRESS. * * This isn't documented anywhere in the PRM. However, it seems to be * necessary prior to changing the surface state base adress. We've * seen issues in Vulkan where we get GPU hangs when using multi-level * command buffers which clear depth, reset state base address, and then * go render stuff. * * Normally, in GL, we would trust the kernel to do sufficient stalls * and flushes prior to executing our batch. However, it doesn't seem * as if the kernel's flushing is always sufficient and we don't want to * rely on it. * * We make this an end-of-pipe sync instead of a normal flush because we * do not know the current status of the GPU. On Haswell at least, * having a fast-clear operation in flight at the same time as a normal * rendering operation can cause hangs. Since the kernel's flushing is * insufficient, we need to ensure that any rendering operations from * other processes are definitely complete before we try to do our own * rendering. It's a bit of a big hammer but it appears to work. */ iris_emit_end_of_pipe_sync(batch, PIPE_CONTROL_RENDER_TARGET_FLUSH | PIPE_CONTROL_DEPTH_CACHE_FLUSH | PIPE_CONTROL_DATA_CACHE_FLUSH); } static void _iris_emit_lri(struct iris_batch *batch, uint32_t reg, uint32_t val) { iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_IMM), lri) { lri.RegisterOffset = reg; lri.DataDWord = val; } } #define iris_emit_lri(b, r, v) _iris_emit_lri(b, GENX(r##_num), v) static void _iris_emit_lrr(struct iris_batch *batch, uint32_t dst, uint32_t src) { iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_REG), lrr) { lrr.SourceRegisterAddress = src; lrr.DestinationRegisterAddress = dst; } } static void emit_pipeline_select(struct iris_batch *batch, uint32_t pipeline) { #if GEN_GEN >= 8 && GEN_GEN < 10 /* From the Broadwell PRM, Volume 2a: Instructions, PIPELINE_SELECT: * * Software must clear the COLOR_CALC_STATE Valid field in * 3DSTATE_CC_STATE_POINTERS command prior to send a PIPELINE_SELECT * with Pipeline Select set to GPGPU. * * The internal hardware docs recommend the same workaround for Gen9 * hardware too. */ if (pipeline == GPGPU) iris_emit_cmd(batch, GENX(3DSTATE_CC_STATE_POINTERS), t); #endif /* From "BXML » GT » MI » vol1a GPU Overview » [Instruction] * PIPELINE_SELECT [DevBWR+]": * * "Project: DEVSNB+ * * Software must ensure all the write caches are flushed through a * stalling PIPE_CONTROL command followed by another PIPE_CONTROL * command to invalidate read only caches prior to programming * MI_PIPELINE_SELECT command to change the Pipeline Select Mode." */ iris_emit_pipe_control_flush(batch, PIPE_CONTROL_RENDER_TARGET_FLUSH | PIPE_CONTROL_DEPTH_CACHE_FLUSH | PIPE_CONTROL_DATA_CACHE_FLUSH | PIPE_CONTROL_CS_STALL); iris_emit_pipe_control_flush(batch, PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE | PIPE_CONTROL_CONST_CACHE_INVALIDATE | PIPE_CONTROL_STATE_CACHE_INVALIDATE | PIPE_CONTROL_INSTRUCTION_INVALIDATE); iris_emit_cmd(batch, GENX(PIPELINE_SELECT), sel) { #if GEN_GEN >= 9 sel.MaskBits = 3; #endif sel.PipelineSelection = pipeline; } } UNUSED static void init_glk_barrier_mode(struct iris_batch *batch, uint32_t value) { #if GEN_GEN == 9 /* Project: DevGLK * * "This chicken bit works around a hardware issue with barrier * logic encountered when switching between GPGPU and 3D pipelines. * To workaround the issue, this mode bit should be set after a * pipeline is selected." */ uint32_t reg_val; iris_pack_state(GENX(SLICE_COMMON_ECO_CHICKEN1), ®_val, reg) { reg.GLKBarrierMode = value; reg.GLKBarrierModeMask = 1; } iris_emit_lri(batch, SLICE_COMMON_ECO_CHICKEN1, reg_val); #endif } static void init_state_base_address(struct iris_batch *batch) { flush_for_state_base_change(batch); /* We program most base addresses once at context initialization time. * Each base address points at a 4GB memory zone, and never needs to * change. See iris_bufmgr.h for a description of the memory zones. * * The one exception is Surface State Base Address, which needs to be * updated occasionally. See iris_binder.c for the details there. */ iris_emit_cmd(batch, GENX(STATE_BASE_ADDRESS), sba) { sba.GeneralStateMOCS = MOCS_WB; sba.StatelessDataPortAccessMOCS = MOCS_WB; sba.DynamicStateMOCS = MOCS_WB; sba.IndirectObjectMOCS = MOCS_WB; sba.InstructionMOCS = MOCS_WB; sba.GeneralStateBaseAddressModifyEnable = true; sba.DynamicStateBaseAddressModifyEnable = true; sba.IndirectObjectBaseAddressModifyEnable = true; sba.InstructionBaseAddressModifyEnable = true; sba.GeneralStateBufferSizeModifyEnable = true; sba.DynamicStateBufferSizeModifyEnable = true; #if (GEN_GEN >= 9) sba.BindlessSurfaceStateBaseAddressModifyEnable = true; sba.BindlessSurfaceStateMOCS = MOCS_WB; #endif sba.IndirectObjectBufferSizeModifyEnable = true; sba.InstructionBuffersizeModifyEnable = true; sba.InstructionBaseAddress = ro_bo(NULL, IRIS_MEMZONE_SHADER_START); sba.DynamicStateBaseAddress = ro_bo(NULL, IRIS_MEMZONE_DYNAMIC_START); sba.GeneralStateBufferSize = 0xfffff; sba.IndirectObjectBufferSize = 0xfffff; sba.InstructionBufferSize = 0xfffff; sba.DynamicStateBufferSize = 0xfffff; } } static void iris_emit_l3_config(struct iris_batch *batch, const struct gen_l3_config *cfg, bool has_slm, bool wants_dc_cache) { uint32_t reg_val; iris_pack_state(GENX(L3CNTLREG), ®_val, reg) { reg.SLMEnable = has_slm; #if GEN_GEN == 11 /* WA_1406697149: Bit 9 "Error Detection Behavior Control" must be set * in L3CNTLREG register. The default setting of the bit is not the * desirable behavior. */ reg.ErrorDetectionBehaviorControl = true; reg.UseFullWays = true; #endif reg.URBAllocation = cfg->n[GEN_L3P_URB]; reg.ROAllocation = cfg->n[GEN_L3P_RO]; reg.DCAllocation = cfg->n[GEN_L3P_DC]; reg.AllAllocation = cfg->n[GEN_L3P_ALL]; } iris_emit_lri(batch, L3CNTLREG, reg_val); } static void iris_emit_default_l3_config(struct iris_batch *batch, const struct gen_device_info *devinfo, bool compute) { bool wants_dc_cache = true; bool has_slm = compute; const struct gen_l3_weights w = gen_get_default_l3_weights(devinfo, wants_dc_cache, has_slm); const struct gen_l3_config *cfg = gen_get_l3_config(devinfo, w); iris_emit_l3_config(batch, cfg, has_slm, wants_dc_cache); } #if GEN_GEN == 9 || GEN_GEN == 10 static void iris_enable_obj_preemption(struct iris_batch *batch, bool enable) { uint32_t reg_val; /* A fixed function pipe flush is required before modifying this field */ iris_emit_end_of_pipe_sync(batch, PIPE_CONTROL_RENDER_TARGET_FLUSH); /* enable object level preemption */ iris_pack_state(GENX(CS_CHICKEN1), ®_val, reg) { reg.ReplayMode = enable; reg.ReplayModeMask = true; } iris_emit_lri(batch, CS_CHICKEN1, reg_val); } #endif /** * Upload the initial GPU state for a render context. * * This sets some invariant state that needs to be programmed a particular * way, but we never actually change. */ static void iris_init_render_context(struct iris_screen *screen, struct iris_batch *batch, struct iris_vtable *vtbl, struct pipe_debug_callback *dbg) { UNUSED const struct gen_device_info *devinfo = &screen->devinfo; uint32_t reg_val; emit_pipeline_select(batch, _3D); iris_emit_default_l3_config(batch, devinfo, false); init_state_base_address(batch); #if GEN_GEN >= 9 iris_pack_state(GENX(CS_DEBUG_MODE2), ®_val, reg) { reg.CONSTANT_BUFFERAddressOffsetDisable = true; reg.CONSTANT_BUFFERAddressOffsetDisableMask = true; } iris_emit_lri(batch, CS_DEBUG_MODE2, reg_val); #else iris_pack_state(GENX(INSTPM), ®_val, reg) { reg.CONSTANT_BUFFERAddressOffsetDisable = true; reg.CONSTANT_BUFFERAddressOffsetDisableMask = true; } iris_emit_lri(batch, INSTPM, reg_val); #endif #if GEN_GEN == 9 iris_pack_state(GENX(CACHE_MODE_1), ®_val, reg) { reg.FloatBlendOptimizationEnable = true; reg.FloatBlendOptimizationEnableMask = true; reg.PartialResolveDisableInVC = true; reg.PartialResolveDisableInVCMask = true; } iris_emit_lri(batch, CACHE_MODE_1, reg_val); if (devinfo->is_geminilake) init_glk_barrier_mode(batch, GLK_BARRIER_MODE_3D_HULL); #endif #if GEN_GEN == 11 iris_pack_state(GENX(SAMPLER_MODE), ®_val, reg) { reg.HeaderlessMessageforPreemptableContexts = 1; reg.HeaderlessMessageforPreemptableContextsMask = 1; } iris_emit_lri(batch, SAMPLER_MODE, reg_val); /* Bit 1 must be set in HALF_SLICE_CHICKEN7. */ iris_pack_state(GENX(HALF_SLICE_CHICKEN7), ®_val, reg) { reg.EnabledTexelOffsetPrecisionFix = 1; reg.EnabledTexelOffsetPrecisionFixMask = 1; } iris_emit_lri(batch, HALF_SLICE_CHICKEN7, reg_val); /* WA_2204188704: Pixel Shader Panic dispatch must be disabled. */ iris_pack_state(GENX(COMMON_SLICE_CHICKEN3), ®_val, reg) { reg.PSThreadPanicDispatch = 0x3; reg.PSThreadPanicDispatchMask = 0x3; } iris_emit_lri(batch, COMMON_SLICE_CHICKEN3, reg_val); iris_pack_state(GENX(SLICE_COMMON_ECO_CHICKEN1), ®_val, reg) { reg.StateCacheRedirectToCSSectionEnable = true; reg.StateCacheRedirectToCSSectionEnableMask = true; } iris_emit_lri(batch, SLICE_COMMON_ECO_CHICKEN1, reg_val); // XXX: 3D_MODE? #endif /* 3DSTATE_DRAWING_RECTANGLE is non-pipelined, so we want to avoid * changing it dynamically. We set it to the maximum size here, and * instead include the render target dimensions in the viewport, so * viewport extents clipping takes care of pruning stray geometry. */ iris_emit_cmd(batch, GENX(3DSTATE_DRAWING_RECTANGLE), rect) { rect.ClippedDrawingRectangleXMax = UINT16_MAX; rect.ClippedDrawingRectangleYMax = UINT16_MAX; } /* Set the initial MSAA sample positions. */ iris_emit_cmd(batch, GENX(3DSTATE_SAMPLE_PATTERN), pat) { GEN_SAMPLE_POS_1X(pat._1xSample); GEN_SAMPLE_POS_2X(pat._2xSample); GEN_SAMPLE_POS_4X(pat._4xSample); GEN_SAMPLE_POS_8X(pat._8xSample); #if GEN_GEN >= 9 GEN_SAMPLE_POS_16X(pat._16xSample); #endif } /* Use the legacy AA line coverage computation. */ iris_emit_cmd(batch, GENX(3DSTATE_AA_LINE_PARAMETERS), foo); /* Disable chromakeying (it's for media) */ iris_emit_cmd(batch, GENX(3DSTATE_WM_CHROMAKEY), foo); /* We want regular rendering, not special HiZ operations. */ iris_emit_cmd(batch, GENX(3DSTATE_WM_HZ_OP), foo); /* No polygon stippling offsets are necessary. */ /* TODO: may need to set an offset for origin-UL framebuffers */ iris_emit_cmd(batch, GENX(3DSTATE_POLY_STIPPLE_OFFSET), foo); /* Set a static partitioning of the push constant area. */ /* TODO: this may be a bad idea...could starve the push ringbuffers... */ for (int i = 0; i <= MESA_SHADER_FRAGMENT; i++) { iris_emit_cmd(batch, GENX(3DSTATE_PUSH_CONSTANT_ALLOC_VS), alloc) { alloc._3DCommandSubOpcode = 18 + i; alloc.ConstantBufferOffset = 6 * i; alloc.ConstantBufferSize = i == MESA_SHADER_FRAGMENT ? 8 : 6; } } #if GEN_GEN == 10 /* Gen11+ is enabled for us by the kernel. */ iris_enable_obj_preemption(batch, true); #endif } static void iris_init_compute_context(struct iris_screen *screen, struct iris_batch *batch, struct iris_vtable *vtbl, struct pipe_debug_callback *dbg) { UNUSED const struct gen_device_info *devinfo = &screen->devinfo; emit_pipeline_select(batch, GPGPU); iris_emit_default_l3_config(batch, devinfo, true); init_state_base_address(batch); #if GEN_GEN == 9 if (devinfo->is_geminilake) init_glk_barrier_mode(batch, GLK_BARRIER_MODE_GPGPU); #endif } struct iris_vertex_buffer_state { /** The VERTEX_BUFFER_STATE hardware structure. */ uint32_t state[GENX(VERTEX_BUFFER_STATE_length)]; /** The resource to source vertex data from. */ struct pipe_resource *resource; }; struct iris_depth_buffer_state { /* Depth/HiZ/Stencil related hardware packets. */ uint32_t packets[GENX(3DSTATE_DEPTH_BUFFER_length) + GENX(3DSTATE_STENCIL_BUFFER_length) + GENX(3DSTATE_HIER_DEPTH_BUFFER_length) + GENX(3DSTATE_CLEAR_PARAMS_length)]; }; /** * Generation-specific context state (ice->state.genx->...). * * Most state can go in iris_context directly, but these encode hardware * packets which vary by generation. */ struct iris_genx_state { struct iris_vertex_buffer_state vertex_buffers[33]; struct iris_depth_buffer_state depth_buffer; uint32_t so_buffers[4 * GENX(3DSTATE_SO_BUFFER_length)]; #if GEN_GEN == 9 /* Is object level preemption enabled? */ bool object_preemption; #endif struct { #if GEN_GEN == 8 struct brw_image_param image_param[PIPE_MAX_SHADER_IMAGES]; #endif } shaders[MESA_SHADER_STAGES]; }; /** * The pipe->set_blend_color() driver hook. * * This corresponds to our COLOR_CALC_STATE. */ static void iris_set_blend_color(struct pipe_context *ctx, const struct pipe_blend_color *state) { struct iris_context *ice = (struct iris_context *) ctx; /* Our COLOR_CALC_STATE is exactly pipe_blend_color, so just memcpy */ memcpy(&ice->state.blend_color, state, sizeof(struct pipe_blend_color)); ice->state.dirty |= IRIS_DIRTY_COLOR_CALC_STATE; } /** * Gallium CSO for blend state (see pipe_blend_state). */ struct iris_blend_state { /** Partial 3DSTATE_PS_BLEND */ uint32_t ps_blend[GENX(3DSTATE_PS_BLEND_length)]; /** Partial BLEND_STATE */ uint32_t blend_state[GENX(BLEND_STATE_length) + BRW_MAX_DRAW_BUFFERS * GENX(BLEND_STATE_ENTRY_length)]; bool alpha_to_coverage; /* for shader key */ /** Bitfield of whether blending is enabled for RT[i] - for aux resolves */ uint8_t blend_enables; /** Bitfield of whether color writes are enabled for RT[i] */ uint8_t color_write_enables; /** Does RT[0] use dual color blending? */ bool dual_color_blending; }; static enum pipe_blendfactor fix_blendfactor(enum pipe_blendfactor f, bool alpha_to_one) { if (alpha_to_one) { if (f == PIPE_BLENDFACTOR_SRC1_ALPHA) return PIPE_BLENDFACTOR_ONE; if (f == PIPE_BLENDFACTOR_INV_SRC1_ALPHA) return PIPE_BLENDFACTOR_ZERO; } return f; } /** * The pipe->create_blend_state() driver hook. * * Translates a pipe_blend_state into iris_blend_state. */ static void * iris_create_blend_state(struct pipe_context *ctx, const struct pipe_blend_state *state) { struct iris_blend_state *cso = malloc(sizeof(struct iris_blend_state)); uint32_t *blend_entry = cso->blend_state + GENX(BLEND_STATE_length); cso->blend_enables = 0; cso->color_write_enables = 0; STATIC_ASSERT(BRW_MAX_DRAW_BUFFERS <= 8); cso->alpha_to_coverage = state->alpha_to_coverage; bool indep_alpha_blend = false; for (int i = 0; i < BRW_MAX_DRAW_BUFFERS; i++) { const struct pipe_rt_blend_state *rt = &state->rt[state->independent_blend_enable ? i : 0]; enum pipe_blendfactor src_rgb = fix_blendfactor(rt->rgb_src_factor, state->alpha_to_one); enum pipe_blendfactor src_alpha = fix_blendfactor(rt->alpha_src_factor, state->alpha_to_one); enum pipe_blendfactor dst_rgb = fix_blendfactor(rt->rgb_dst_factor, state->alpha_to_one); enum pipe_blendfactor dst_alpha = fix_blendfactor(rt->alpha_dst_factor, state->alpha_to_one); if (rt->rgb_func != rt->alpha_func || src_rgb != src_alpha || dst_rgb != dst_alpha) indep_alpha_blend = true; if (rt->blend_enable) cso->blend_enables |= 1u << i; if (rt->colormask) cso->color_write_enables |= 1u << i; iris_pack_state(GENX(BLEND_STATE_ENTRY), blend_entry, be) { be.LogicOpEnable = state->logicop_enable; be.LogicOpFunction = state->logicop_func; be.PreBlendSourceOnlyClampEnable = false; be.ColorClampRange = COLORCLAMP_RTFORMAT; be.PreBlendColorClampEnable = true; be.PostBlendColorClampEnable = true; be.ColorBufferBlendEnable = rt->blend_enable; be.ColorBlendFunction = rt->rgb_func; be.AlphaBlendFunction = rt->alpha_func; be.SourceBlendFactor = src_rgb; be.SourceAlphaBlendFactor = src_alpha; be.DestinationBlendFactor = dst_rgb; be.DestinationAlphaBlendFactor = dst_alpha; be.WriteDisableRed = !(rt->colormask & PIPE_MASK_R); be.WriteDisableGreen = !(rt->colormask & PIPE_MASK_G); be.WriteDisableBlue = !(rt->colormask & PIPE_MASK_B); be.WriteDisableAlpha = !(rt->colormask & PIPE_MASK_A); } blend_entry += GENX(BLEND_STATE_ENTRY_length); } iris_pack_command(GENX(3DSTATE_PS_BLEND), cso->ps_blend, pb) { /* pb.HasWriteableRT is filled in at draw time. * pb.AlphaTestEnable is filled in at draw time. * * pb.ColorBufferBlendEnable is filled in at draw time so we can avoid * setting it when dual color blending without an appropriate shader. */ pb.AlphaToCoverageEnable = state->alpha_to_coverage; pb.IndependentAlphaBlendEnable = indep_alpha_blend; pb.SourceBlendFactor = fix_blendfactor(state->rt[0].rgb_src_factor, state->alpha_to_one); pb.SourceAlphaBlendFactor = fix_blendfactor(state->rt[0].alpha_src_factor, state->alpha_to_one); pb.DestinationBlendFactor = fix_blendfactor(state->rt[0].rgb_dst_factor, state->alpha_to_one); pb.DestinationAlphaBlendFactor = fix_blendfactor(state->rt[0].alpha_dst_factor, state->alpha_to_one); } iris_pack_state(GENX(BLEND_STATE), cso->blend_state, bs) { bs.AlphaToCoverageEnable = state->alpha_to_coverage; bs.IndependentAlphaBlendEnable = indep_alpha_blend; bs.AlphaToOneEnable = state->alpha_to_one; bs.AlphaToCoverageDitherEnable = state->alpha_to_coverage; bs.ColorDitherEnable = state->dither; /* bl.AlphaTestEnable and bs.AlphaTestFunction are filled in later. */ } cso->dual_color_blending = util_blend_state_is_dual(state, 0); return cso; } /** * The pipe->bind_blend_state() driver hook. * * Bind a blending CSO and flag related dirty bits. */ static void iris_bind_blend_state(struct pipe_context *ctx, void *state) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_blend_state *cso = state; ice->state.cso_blend = cso; ice->state.blend_enables = cso ? cso->blend_enables : 0; ice->state.dirty |= IRIS_DIRTY_PS_BLEND; ice->state.dirty |= IRIS_DIRTY_BLEND_STATE; ice->state.dirty |= IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES; ice->state.dirty |= ice->state.dirty_for_nos[IRIS_NOS_BLEND]; } /** * Return true if the FS writes to any color outputs which are not disabled * via color masking. */ static bool has_writeable_rt(const struct iris_blend_state *cso_blend, const struct shader_info *fs_info) { if (!fs_info) return false; unsigned rt_outputs = fs_info->outputs_written >> FRAG_RESULT_DATA0; if (fs_info->outputs_written & BITFIELD64_BIT(FRAG_RESULT_COLOR)) rt_outputs = (1 << BRW_MAX_DRAW_BUFFERS) - 1; return cso_blend->color_write_enables & rt_outputs; } /** * Gallium CSO for depth, stencil, and alpha testing state. */ struct iris_depth_stencil_alpha_state { /** Partial 3DSTATE_WM_DEPTH_STENCIL. */ uint32_t wmds[GENX(3DSTATE_WM_DEPTH_STENCIL_length)]; /** Outbound to BLEND_STATE, 3DSTATE_PS_BLEND, COLOR_CALC_STATE. */ struct pipe_alpha_state alpha; /** Outbound to resolve and cache set tracking. */ bool depth_writes_enabled; bool stencil_writes_enabled; }; /** * The pipe->create_depth_stencil_alpha_state() driver hook. * * We encode most of 3DSTATE_WM_DEPTH_STENCIL, and just save off the alpha * testing state since we need pieces of it in a variety of places. */ static void * iris_create_zsa_state(struct pipe_context *ctx, const struct pipe_depth_stencil_alpha_state *state) { struct iris_depth_stencil_alpha_state *cso = malloc(sizeof(struct iris_depth_stencil_alpha_state)); bool two_sided_stencil = state->stencil[1].enabled; cso->alpha = state->alpha; cso->depth_writes_enabled = state->depth.writemask; cso->stencil_writes_enabled = state->stencil[0].writemask != 0 || (two_sided_stencil && state->stencil[1].writemask != 0); /* The state tracker needs to optimize away EQUAL writes for us. */ assert(!(state->depth.func == PIPE_FUNC_EQUAL && state->depth.writemask)); iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL), cso->wmds, wmds) { wmds.StencilFailOp = state->stencil[0].fail_op; wmds.StencilPassDepthFailOp = state->stencil[0].zfail_op; wmds.StencilPassDepthPassOp = state->stencil[0].zpass_op; wmds.StencilTestFunction = translate_compare_func(state->stencil[0].func); wmds.BackfaceStencilFailOp = state->stencil[1].fail_op; wmds.BackfaceStencilPassDepthFailOp = state->stencil[1].zfail_op; wmds.BackfaceStencilPassDepthPassOp = state->stencil[1].zpass_op; wmds.BackfaceStencilTestFunction = translate_compare_func(state->stencil[1].func); wmds.DepthTestFunction = translate_compare_func(state->depth.func); wmds.DoubleSidedStencilEnable = two_sided_stencil; wmds.StencilTestEnable = state->stencil[0].enabled; wmds.StencilBufferWriteEnable = state->stencil[0].writemask != 0 || (two_sided_stencil && state->stencil[1].writemask != 0); wmds.DepthTestEnable = state->depth.enabled; wmds.DepthBufferWriteEnable = state->depth.writemask; wmds.StencilTestMask = state->stencil[0].valuemask; wmds.StencilWriteMask = state->stencil[0].writemask; wmds.BackfaceStencilTestMask = state->stencil[1].valuemask; wmds.BackfaceStencilWriteMask = state->stencil[1].writemask; /* wmds.[Backface]StencilReferenceValue are merged later */ } return cso; } /** * The pipe->bind_depth_stencil_alpha_state() driver hook. * * Bind a depth/stencil/alpha CSO and flag related dirty bits. */ static void iris_bind_zsa_state(struct pipe_context *ctx, void *state) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_depth_stencil_alpha_state *old_cso = ice->state.cso_zsa; struct iris_depth_stencil_alpha_state *new_cso = state; if (new_cso) { if (cso_changed(alpha.ref_value)) ice->state.dirty |= IRIS_DIRTY_COLOR_CALC_STATE; if (cso_changed(alpha.enabled)) ice->state.dirty |= IRIS_DIRTY_PS_BLEND | IRIS_DIRTY_BLEND_STATE; if (cso_changed(alpha.func)) ice->state.dirty |= IRIS_DIRTY_BLEND_STATE; if (cso_changed(depth_writes_enabled)) ice->state.dirty |= IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES; ice->state.depth_writes_enabled = new_cso->depth_writes_enabled; ice->state.stencil_writes_enabled = new_cso->stencil_writes_enabled; } ice->state.cso_zsa = new_cso; ice->state.dirty |= IRIS_DIRTY_CC_VIEWPORT; ice->state.dirty |= IRIS_DIRTY_WM_DEPTH_STENCIL; ice->state.dirty |= ice->state.dirty_for_nos[IRIS_NOS_DEPTH_STENCIL_ALPHA]; } /** * Gallium CSO for rasterizer state. */ struct iris_rasterizer_state { uint32_t sf[GENX(3DSTATE_SF_length)]; uint32_t clip[GENX(3DSTATE_CLIP_length)]; uint32_t raster[GENX(3DSTATE_RASTER_length)]; uint32_t wm[GENX(3DSTATE_WM_length)]; uint32_t line_stipple[GENX(3DSTATE_LINE_STIPPLE_length)]; uint8_t num_clip_plane_consts; bool clip_halfz; /* for CC_VIEWPORT */ bool depth_clip_near; /* for CC_VIEWPORT */ bool depth_clip_far; /* for CC_VIEWPORT */ bool flatshade; /* for shader state */ bool flatshade_first; /* for stream output */ bool clamp_fragment_color; /* for shader state */ bool light_twoside; /* for shader state */ bool rasterizer_discard; /* for 3DSTATE_STREAMOUT and 3DSTATE_CLIP */ bool half_pixel_center; /* for 3DSTATE_MULTISAMPLE */ bool line_stipple_enable; bool poly_stipple_enable; bool multisample; bool force_persample_interp; bool conservative_rasterization; bool fill_mode_point_or_line; enum pipe_sprite_coord_mode sprite_coord_mode; /* PIPE_SPRITE_* */ uint16_t sprite_coord_enable; }; static float get_line_width(const struct pipe_rasterizer_state *state) { float line_width = state->line_width; /* From the OpenGL 4.4 spec: * * "The actual width of non-antialiased lines is determined by rounding * the supplied width to the nearest integer, then clamping it to the * implementation-dependent maximum non-antialiased line width." */ if (!state->multisample && !state->line_smooth) line_width = roundf(state->line_width); if (!state->multisample && state->line_smooth && line_width < 1.5f) { /* For 1 pixel line thickness or less, the general anti-aliasing * algorithm gives up, and a garbage line is generated. Setting a * Line Width of 0.0 specifies the rasterization of the "thinnest" * (one-pixel-wide), non-antialiased lines. * * Lines rendered with zero Line Width are rasterized using the * "Grid Intersection Quantization" rules as specified by the * "Zero-Width (Cosmetic) Line Rasterization" section of the docs. */ line_width = 0.0f; } return line_width; } /** * The pipe->create_rasterizer_state() driver hook. */ static void * iris_create_rasterizer_state(struct pipe_context *ctx, const struct pipe_rasterizer_state *state) { struct iris_rasterizer_state *cso = malloc(sizeof(struct iris_rasterizer_state)); cso->multisample = state->multisample; cso->force_persample_interp = state->force_persample_interp; cso->clip_halfz = state->clip_halfz; cso->depth_clip_near = state->depth_clip_near; cso->depth_clip_far = state->depth_clip_far; cso->flatshade = state->flatshade; cso->flatshade_first = state->flatshade_first; cso->clamp_fragment_color = state->clamp_fragment_color; cso->light_twoside = state->light_twoside; cso->rasterizer_discard = state->rasterizer_discard; cso->half_pixel_center = state->half_pixel_center; cso->sprite_coord_mode = state->sprite_coord_mode; cso->sprite_coord_enable = state->sprite_coord_enable; cso->line_stipple_enable = state->line_stipple_enable; cso->poly_stipple_enable = state->poly_stipple_enable; cso->conservative_rasterization = state->conservative_raster_mode == PIPE_CONSERVATIVE_RASTER_POST_SNAP; cso->fill_mode_point_or_line = state->fill_front == PIPE_POLYGON_MODE_LINE || state->fill_front == PIPE_POLYGON_MODE_POINT || state->fill_back == PIPE_POLYGON_MODE_LINE || state->fill_back == PIPE_POLYGON_MODE_POINT; if (state->clip_plane_enable != 0) cso->num_clip_plane_consts = util_logbase2(state->clip_plane_enable) + 1; else cso->num_clip_plane_consts = 0; float line_width = get_line_width(state); iris_pack_command(GENX(3DSTATE_SF), cso->sf, sf) { sf.StatisticsEnable = true; sf.ViewportTransformEnable = true; sf.AALineDistanceMode = AALINEDISTANCE_TRUE; sf.LineEndCapAntialiasingRegionWidth = state->line_smooth ? _10pixels : _05pixels; sf.LastPixelEnable = state->line_last_pixel; sf.LineWidth = line_width; sf.SmoothPointEnable = (state->point_smooth || state->multisample) && !state->point_quad_rasterization; sf.PointWidthSource = state->point_size_per_vertex ? Vertex : State; sf.PointWidth = state->point_size; if (state->flatshade_first) { sf.TriangleFanProvokingVertexSelect = 1; } else { sf.TriangleStripListProvokingVertexSelect = 2; sf.TriangleFanProvokingVertexSelect = 2; sf.LineStripListProvokingVertexSelect = 1; } } iris_pack_command(GENX(3DSTATE_RASTER), cso->raster, rr) { rr.FrontWinding = state->front_ccw ? CounterClockwise : Clockwise; rr.CullMode = translate_cull_mode(state->cull_face); rr.FrontFaceFillMode = translate_fill_mode(state->fill_front); rr.BackFaceFillMode = translate_fill_mode(state->fill_back); rr.DXMultisampleRasterizationEnable = state->multisample; rr.GlobalDepthOffsetEnableSolid = state->offset_tri; rr.GlobalDepthOffsetEnableWireframe = state->offset_line; rr.GlobalDepthOffsetEnablePoint = state->offset_point; rr.GlobalDepthOffsetConstant = state->offset_units * 2; rr.GlobalDepthOffsetScale = state->offset_scale; rr.GlobalDepthOffsetClamp = state->offset_clamp; rr.SmoothPointEnable = state->point_smooth; rr.AntialiasingEnable = state->line_smooth; rr.ScissorRectangleEnable = state->scissor; #if GEN_GEN >= 9 rr.ViewportZNearClipTestEnable = state->depth_clip_near; rr.ViewportZFarClipTestEnable = state->depth_clip_far; rr.ConservativeRasterizationEnable = cso->conservative_rasterization; #else rr.ViewportZClipTestEnable = (state->depth_clip_near || state->depth_clip_far); #endif } iris_pack_command(GENX(3DSTATE_CLIP), cso->clip, cl) { /* cl.NonPerspectiveBarycentricEnable is filled in at draw time from * the FS program; cl.ForceZeroRTAIndexEnable is filled in from the FB. */ cl.EarlyCullEnable = true; cl.UserClipDistanceClipTestEnableBitmask = state->clip_plane_enable; cl.ForceUserClipDistanceClipTestEnableBitmask = true; cl.APIMode = state->clip_halfz ? APIMODE_D3D : APIMODE_OGL; cl.GuardbandClipTestEnable = true; cl.ClipEnable = true; cl.MinimumPointWidth = 0.125; cl.MaximumPointWidth = 255.875; if (state->flatshade_first) { cl.TriangleFanProvokingVertexSelect = 1; } else { cl.TriangleStripListProvokingVertexSelect = 2; cl.TriangleFanProvokingVertexSelect = 2; cl.LineStripListProvokingVertexSelect = 1; } } iris_pack_command(GENX(3DSTATE_WM), cso->wm, wm) { /* wm.BarycentricInterpolationMode and wm.EarlyDepthStencilControl are * filled in at draw time from the FS program. */ wm.LineAntialiasingRegionWidth = _10pixels; wm.LineEndCapAntialiasingRegionWidth = _05pixels; wm.PointRasterizationRule = RASTRULE_UPPER_RIGHT; wm.LineStippleEnable = state->line_stipple_enable; wm.PolygonStippleEnable = state->poly_stipple_enable; } /* Remap from 0..255 back to 1..256 */ const unsigned line_stipple_factor = state->line_stipple_factor + 1; iris_pack_command(GENX(3DSTATE_LINE_STIPPLE), cso->line_stipple, line) { line.LineStipplePattern = state->line_stipple_pattern; line.LineStippleInverseRepeatCount = 1.0f / line_stipple_factor; line.LineStippleRepeatCount = line_stipple_factor; } return cso; } /** * The pipe->bind_rasterizer_state() driver hook. * * Bind a rasterizer CSO and flag related dirty bits. */ static void iris_bind_rasterizer_state(struct pipe_context *ctx, void *state) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_rasterizer_state *old_cso = ice->state.cso_rast; struct iris_rasterizer_state *new_cso = state; if (new_cso) { /* Try to avoid re-emitting 3DSTATE_LINE_STIPPLE, it's non-pipelined */ if (cso_changed_memcmp(line_stipple)) ice->state.dirty |= IRIS_DIRTY_LINE_STIPPLE; if (cso_changed(half_pixel_center)) ice->state.dirty |= IRIS_DIRTY_MULTISAMPLE; if (cso_changed(line_stipple_enable) || cso_changed(poly_stipple_enable)) ice->state.dirty |= IRIS_DIRTY_WM; if (cso_changed(rasterizer_discard)) ice->state.dirty |= IRIS_DIRTY_STREAMOUT | IRIS_DIRTY_CLIP; if (cso_changed(flatshade_first)) ice->state.dirty |= IRIS_DIRTY_STREAMOUT; if (cso_changed(depth_clip_near) || cso_changed(depth_clip_far) || cso_changed(clip_halfz)) ice->state.dirty |= IRIS_DIRTY_CC_VIEWPORT; if (cso_changed(sprite_coord_enable) || cso_changed(sprite_coord_mode) || cso_changed(light_twoside)) ice->state.dirty |= IRIS_DIRTY_SBE; if (cso_changed(conservative_rasterization)) ice->state.dirty |= IRIS_DIRTY_FS; } ice->state.cso_rast = new_cso; ice->state.dirty |= IRIS_DIRTY_RASTER; ice->state.dirty |= IRIS_DIRTY_CLIP; ice->state.dirty |= ice->state.dirty_for_nos[IRIS_NOS_RASTERIZER]; } /** * Return true if the given wrap mode requires the border color to exist. * * (We can skip uploading it if the sampler isn't going to use it.) */ static bool wrap_mode_needs_border_color(unsigned wrap_mode) { return wrap_mode == TCM_CLAMP_BORDER || wrap_mode == TCM_HALF_BORDER; } /** * Gallium CSO for sampler state. */ struct iris_sampler_state { union pipe_color_union border_color; bool needs_border_color; uint32_t sampler_state[GENX(SAMPLER_STATE_length)]; }; /** * The pipe->create_sampler_state() driver hook. * * We fill out SAMPLER_STATE (except for the border color pointer), and * store that on the CPU. It doesn't make sense to upload it to a GPU * buffer object yet, because 3DSTATE_SAMPLER_STATE_POINTERS requires * all bound sampler states to be in contiguous memor. */ static void * iris_create_sampler_state(struct pipe_context *ctx, const struct pipe_sampler_state *state) { struct iris_sampler_state *cso = CALLOC_STRUCT(iris_sampler_state); if (!cso) return NULL; STATIC_ASSERT(PIPE_TEX_FILTER_NEAREST == MAPFILTER_NEAREST); STATIC_ASSERT(PIPE_TEX_FILTER_LINEAR == MAPFILTER_LINEAR); unsigned wrap_s = translate_wrap(state->wrap_s); unsigned wrap_t = translate_wrap(state->wrap_t); unsigned wrap_r = translate_wrap(state->wrap_r); memcpy(&cso->border_color, &state->border_color, sizeof(cso->border_color)); cso->needs_border_color = wrap_mode_needs_border_color(wrap_s) || wrap_mode_needs_border_color(wrap_t) || wrap_mode_needs_border_color(wrap_r); float min_lod = state->min_lod; unsigned mag_img_filter = state->mag_img_filter; // XXX: explain this code ported from ilo...I don't get it at all... if (state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE && state->min_lod > 0.0f) { min_lod = 0.0f; mag_img_filter = state->min_img_filter; } iris_pack_state(GENX(SAMPLER_STATE), cso->sampler_state, samp) { samp.TCXAddressControlMode = wrap_s; samp.TCYAddressControlMode = wrap_t; samp.TCZAddressControlMode = wrap_r; samp.CubeSurfaceControlMode = state->seamless_cube_map; samp.NonnormalizedCoordinateEnable = !state->normalized_coords; samp.MinModeFilter = state->min_img_filter; samp.MagModeFilter = mag_img_filter; samp.MipModeFilter = translate_mip_filter(state->min_mip_filter); samp.MaximumAnisotropy = RATIO21; if (state->max_anisotropy >= 2) { if (state->min_img_filter == PIPE_TEX_FILTER_LINEAR) { samp.MinModeFilter = MAPFILTER_ANISOTROPIC; samp.AnisotropicAlgorithm = EWAApproximation; } if (state->mag_img_filter == PIPE_TEX_FILTER_LINEAR) samp.MagModeFilter = MAPFILTER_ANISOTROPIC; samp.MaximumAnisotropy = MIN2((state->max_anisotropy - 2) / 2, RATIO161); } /* Set address rounding bits if not using nearest filtering. */ if (state->min_img_filter != PIPE_TEX_FILTER_NEAREST) { samp.UAddressMinFilterRoundingEnable = true; samp.VAddressMinFilterRoundingEnable = true; samp.RAddressMinFilterRoundingEnable = true; } if (state->mag_img_filter != PIPE_TEX_FILTER_NEAREST) { samp.UAddressMagFilterRoundingEnable = true; samp.VAddressMagFilterRoundingEnable = true; samp.RAddressMagFilterRoundingEnable = true; } if (state->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) samp.ShadowFunction = translate_shadow_func(state->compare_func); const float hw_max_lod = GEN_GEN >= 7 ? 14 : 13; samp.LODPreClampMode = CLAMP_MODE_OGL; samp.MinLOD = CLAMP(min_lod, 0, hw_max_lod); samp.MaxLOD = CLAMP(state->max_lod, 0, hw_max_lod); samp.TextureLODBias = CLAMP(state->lod_bias, -16, 15); /* .BorderColorPointer is filled in by iris_bind_sampler_states. */ } return cso; } /** * The pipe->bind_sampler_states() driver hook. */ static void iris_bind_sampler_states(struct pipe_context *ctx, enum pipe_shader_type p_stage, unsigned start, unsigned count, void **states) { struct iris_context *ice = (struct iris_context *) ctx; gl_shader_stage stage = stage_from_pipe(p_stage); struct iris_shader_state *shs = &ice->state.shaders[stage]; assert(start + count <= IRIS_MAX_TEXTURE_SAMPLERS); for (int i = 0; i < count; i++) { shs->samplers[start + i] = states[i]; } ice->state.dirty |= IRIS_DIRTY_SAMPLER_STATES_VS << stage; } /** * Upload the sampler states into a contiguous area of GPU memory, for * for 3DSTATE_SAMPLER_STATE_POINTERS_*. * * Also fill out the border color state pointers. */ static void iris_upload_sampler_states(struct iris_context *ice, gl_shader_stage stage) { struct iris_shader_state *shs = &ice->state.shaders[stage]; const struct shader_info *info = iris_get_shader_info(ice, stage); /* We assume the state tracker will call pipe->bind_sampler_states() * if the program's number of textures changes. */ unsigned count = info ? util_last_bit(info->textures_used) : 0; if (!count) return; /* Assemble the SAMPLER_STATEs into a contiguous table that lives * in the dynamic state memory zone, so we can point to it via the * 3DSTATE_SAMPLER_STATE_POINTERS_* commands. */ unsigned size = count * 4 * GENX(SAMPLER_STATE_length); uint32_t *map = upload_state(ice->state.dynamic_uploader, &shs->sampler_table, size, 32); if (unlikely(!map)) return; struct pipe_resource *res = shs->sampler_table.res; shs->sampler_table.offset += iris_bo_offset_from_base_address(iris_resource_bo(res)); iris_record_state_size(ice->state.sizes, shs->sampler_table.offset, size); /* Make sure all land in the same BO */ iris_border_color_pool_reserve(ice, IRIS_MAX_TEXTURE_SAMPLERS); ice->state.need_border_colors &= ~(1 << stage); for (int i = 0; i < count; i++) { struct iris_sampler_state *state = shs->samplers[i]; struct iris_sampler_view *tex = shs->textures[i]; if (!state) { memset(map, 0, 4 * GENX(SAMPLER_STATE_length)); } else if (!state->needs_border_color) { memcpy(map, state->sampler_state, 4 * GENX(SAMPLER_STATE_length)); } else { ice->state.need_border_colors |= 1 << stage; /* We may need to swizzle the border color for format faking. * A/LA formats are faked as R/RG with 000R or R00G swizzles. * This means we need to move the border color's A channel into * the R or G channels so that those read swizzles will move it * back into A. */ union pipe_color_union *color = &state->border_color; union pipe_color_union tmp; if (tex) { enum pipe_format internal_format = tex->res->internal_format; if (util_format_is_alpha(internal_format)) { unsigned char swz[4] = { PIPE_SWIZZLE_W, PIPE_SWIZZLE_0, PIPE_SWIZZLE_0, PIPE_SWIZZLE_0 }; util_format_apply_color_swizzle(&tmp, color, swz, true); color = &tmp; } else if (util_format_is_luminance_alpha(internal_format) && internal_format != PIPE_FORMAT_L8A8_SRGB) { unsigned char swz[4] = { PIPE_SWIZZLE_X, PIPE_SWIZZLE_W, PIPE_SWIZZLE_0, PIPE_SWIZZLE_0 }; util_format_apply_color_swizzle(&tmp, color, swz, true); color = &tmp; } } /* Stream out the border color and merge the pointer. */ uint32_t offset = iris_upload_border_color(ice, color); uint32_t dynamic[GENX(SAMPLER_STATE_length)]; iris_pack_state(GENX(SAMPLER_STATE), dynamic, dyns) { dyns.BorderColorPointer = offset; } for (uint32_t j = 0; j < GENX(SAMPLER_STATE_length); j++) map[j] = state->sampler_state[j] | dynamic[j]; } map += GENX(SAMPLER_STATE_length); } } static enum isl_channel_select fmt_swizzle(const struct iris_format_info *fmt, enum pipe_swizzle swz) { switch (swz) { case PIPE_SWIZZLE_X: return fmt->swizzle.r; case PIPE_SWIZZLE_Y: return fmt->swizzle.g; case PIPE_SWIZZLE_Z: return fmt->swizzle.b; case PIPE_SWIZZLE_W: return fmt->swizzle.a; case PIPE_SWIZZLE_1: return SCS_ONE; case PIPE_SWIZZLE_0: return SCS_ZERO; default: unreachable("invalid swizzle"); } } static void fill_buffer_surface_state(struct isl_device *isl_dev, struct iris_resource *res, void *map, enum isl_format format, struct isl_swizzle swizzle, unsigned offset, unsigned size) { const struct isl_format_layout *fmtl = isl_format_get_layout(format); const unsigned cpp = format == ISL_FORMAT_RAW ? 1 : fmtl->bpb / 8; /* The ARB_texture_buffer_specification says: * * "The number of texels in the buffer texture's texel array is given by * * floor( / ( * sizeof()), * * where is the size of the buffer object, in basic * machine units and and are the element count * and base data type for elements, as specified in Table X.1. The * number of texels in the texel array is then clamped to the * implementation-dependent limit MAX_TEXTURE_BUFFER_SIZE_ARB." * * We need to clamp the size in bytes to MAX_TEXTURE_BUFFER_SIZE * stride, * so that when ISL divides by stride to obtain the number of texels, that * texel count is clamped to MAX_TEXTURE_BUFFER_SIZE. */ unsigned final_size = MIN3(size, res->bo->size - res->offset - offset, IRIS_MAX_TEXTURE_BUFFER_SIZE * cpp); isl_buffer_fill_state(isl_dev, map, .address = res->bo->gtt_offset + res->offset + offset, .size_B = final_size, .format = format, .swizzle = swizzle, .stride_B = cpp, .mocs = mocs(res->bo)); } #define SURFACE_STATE_ALIGNMENT 64 /** * Allocate several contiguous SURFACE_STATE structures, one for each * supported auxiliary surface mode. */ static void * alloc_surface_states(struct u_upload_mgr *mgr, struct iris_state_ref *ref, unsigned aux_usages) { const unsigned surf_size = 4 * GENX(RENDER_SURFACE_STATE_length); /* If this changes, update this to explicitly align pointers */ STATIC_ASSERT(surf_size == SURFACE_STATE_ALIGNMENT); assert(aux_usages != 0); void *map = upload_state(mgr, ref, util_bitcount(aux_usages) * surf_size, SURFACE_STATE_ALIGNMENT); ref->offset += iris_bo_offset_from_base_address(iris_resource_bo(ref->res)); return map; } static void fill_surface_state(struct isl_device *isl_dev, void *map, struct iris_resource *res, struct isl_view *view, unsigned aux_usage) { struct isl_surf_fill_state_info f = { .surf = &res->surf, .view = view, .mocs = mocs(res->bo), .address = res->bo->gtt_offset + res->offset, }; if (aux_usage != ISL_AUX_USAGE_NONE) { f.aux_surf = &res->aux.surf; f.aux_usage = aux_usage; f.aux_address = res->aux.bo->gtt_offset + res->aux.offset; struct iris_bo *clear_bo = NULL; uint64_t clear_offset = 0; f.clear_color = iris_resource_get_clear_color(res, &clear_bo, &clear_offset); if (clear_bo) { f.clear_address = clear_bo->gtt_offset + clear_offset; f.use_clear_address = isl_dev->info->gen > 9; } } isl_surf_fill_state_s(isl_dev, map, &f); } /** * The pipe->create_sampler_view() driver hook. */ static struct pipe_sampler_view * iris_create_sampler_view(struct pipe_context *ctx, struct pipe_resource *tex, const struct pipe_sampler_view *tmpl) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_screen *screen = (struct iris_screen *)ctx->screen; const struct gen_device_info *devinfo = &screen->devinfo; struct iris_sampler_view *isv = calloc(1, sizeof(struct iris_sampler_view)); if (!isv) return NULL; /* initialize base object */ isv->base = *tmpl; isv->base.context = ctx; isv->base.texture = NULL; pipe_reference_init(&isv->base.reference, 1); pipe_resource_reference(&isv->base.texture, tex); if (util_format_is_depth_or_stencil(tmpl->format)) { struct iris_resource *zres, *sres; const struct util_format_description *desc = util_format_description(tmpl->format); iris_get_depth_stencil_resources(tex, &zres, &sres); tex = util_format_has_depth(desc) ? &zres->base : &sres->base; } isv->res = (struct iris_resource *) tex; void *map = alloc_surface_states(ice->state.surface_uploader, &isv->surface_state, isv->res->aux.sampler_usages); if (!unlikely(map)) return NULL; isl_surf_usage_flags_t usage = ISL_SURF_USAGE_TEXTURE_BIT; if (isv->base.target == PIPE_TEXTURE_CUBE || isv->base.target == PIPE_TEXTURE_CUBE_ARRAY) usage |= ISL_SURF_USAGE_CUBE_BIT; const struct iris_format_info fmt = iris_format_for_usage(devinfo, tmpl->format, usage); isv->clear_color = isv->res->aux.clear_color; isv->view = (struct isl_view) { .format = fmt.fmt, .swizzle = (struct isl_swizzle) { .r = fmt_swizzle(&fmt, tmpl->swizzle_r), .g = fmt_swizzle(&fmt, tmpl->swizzle_g), .b = fmt_swizzle(&fmt, tmpl->swizzle_b), .a = fmt_swizzle(&fmt, tmpl->swizzle_a), }, .usage = usage, }; /* Fill out SURFACE_STATE for this view. */ if (tmpl->target != PIPE_BUFFER) { isv->view.base_level = tmpl->u.tex.first_level; isv->view.levels = tmpl->u.tex.last_level - tmpl->u.tex.first_level + 1; // XXX: do I need to port f9fd0cf4790cb2a530e75d1a2206dbb9d8af7cb2? isv->view.base_array_layer = tmpl->u.tex.first_layer; isv->view.array_len = tmpl->u.tex.last_layer - tmpl->u.tex.first_layer + 1; unsigned aux_modes = isv->res->aux.sampler_usages; while (aux_modes) { enum isl_aux_usage aux_usage = u_bit_scan(&aux_modes); /* If we have a multisampled depth buffer, do not create a sampler * surface state with HiZ. */ fill_surface_state(&screen->isl_dev, map, isv->res, &isv->view, aux_usage); map += SURFACE_STATE_ALIGNMENT; } } else { fill_buffer_surface_state(&screen->isl_dev, isv->res, map, isv->view.format, isv->view.swizzle, tmpl->u.buf.offset, tmpl->u.buf.size); } return &isv->base; } static void iris_sampler_view_destroy(struct pipe_context *ctx, struct pipe_sampler_view *state) { struct iris_sampler_view *isv = (void *) state; pipe_resource_reference(&state->texture, NULL); pipe_resource_reference(&isv->surface_state.res, NULL); free(isv); } /** * The pipe->create_surface() driver hook. * * In Gallium nomenclature, "surfaces" are a view of a resource that * can be bound as a render target or depth/stencil buffer. */ static struct pipe_surface * iris_create_surface(struct pipe_context *ctx, struct pipe_resource *tex, const struct pipe_surface *tmpl) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_screen *screen = (struct iris_screen *)ctx->screen; const struct gen_device_info *devinfo = &screen->devinfo; struct iris_surface *surf = calloc(1, sizeof(struct iris_surface)); struct pipe_surface *psurf = &surf->base; struct iris_resource *res = (struct iris_resource *) tex; if (!surf) return NULL; pipe_reference_init(&psurf->reference, 1); pipe_resource_reference(&psurf->texture, tex); psurf->context = ctx; psurf->format = tmpl->format; psurf->width = tex->width0; psurf->height = tex->height0; psurf->texture = tex; psurf->u.tex.first_layer = tmpl->u.tex.first_layer; psurf->u.tex.last_layer = tmpl->u.tex.last_layer; psurf->u.tex.level = tmpl->u.tex.level; isl_surf_usage_flags_t usage = 0; if (tmpl->writable) usage = ISL_SURF_USAGE_STORAGE_BIT; else if (util_format_is_depth_or_stencil(tmpl->format)) usage = ISL_SURF_USAGE_DEPTH_BIT; else usage = ISL_SURF_USAGE_RENDER_TARGET_BIT; const struct iris_format_info fmt = iris_format_for_usage(devinfo, psurf->format, usage); if ((usage & ISL_SURF_USAGE_RENDER_TARGET_BIT) && !isl_format_supports_rendering(devinfo, fmt.fmt)) { /* Framebuffer validation will reject this invalid case, but it * hasn't had the opportunity yet. In the meantime, we need to * avoid hitting ISL asserts about unsupported formats below. */ free(surf); return NULL; } struct isl_view *view = &surf->view; *view = (struct isl_view) { .format = fmt.fmt, .base_level = tmpl->u.tex.level, .levels = 1, .base_array_layer = tmpl->u.tex.first_layer, .array_len = tmpl->u.tex.last_layer - tmpl->u.tex.first_layer + 1, .swizzle = ISL_SWIZZLE_IDENTITY, .usage = usage, }; surf->clear_color = res->aux.clear_color; /* Bail early for depth/stencil - we don't want SURFACE_STATE for them. */ if (res->surf.usage & (ISL_SURF_USAGE_DEPTH_BIT | ISL_SURF_USAGE_STENCIL_BIT)) return psurf; void *map = alloc_surface_states(ice->state.surface_uploader, &surf->surface_state, res->aux.possible_usages); if (!unlikely(map)) return NULL; if (!isl_format_is_compressed(res->surf.format)) { /* This is a normal surface. Fill out a SURFACE_STATE for each possible * auxiliary surface mode and return the pipe_surface. */ unsigned aux_modes = res->aux.possible_usages; while (aux_modes) { enum isl_aux_usage aux_usage = u_bit_scan(&aux_modes); fill_surface_state(&screen->isl_dev, map, res, view, aux_usage); map += SURFACE_STATE_ALIGNMENT; } return psurf; } /* The resource has a compressed format, which is not renderable, but we * have a renderable view format. We must be attempting to upload blocks * of compressed data via an uncompressed view. * * In this case, we can assume there are no auxiliary buffers, a single * miplevel, and that the resource is single-sampled. Gallium may try * and create an uncompressed view with multiple layers, however. */ assert(!isl_format_is_compressed(fmt.fmt)); assert(res->aux.possible_usages == 1 << ISL_AUX_USAGE_NONE); assert(res->surf.samples == 1); assert(view->levels == 1); struct isl_surf isl_surf; uint32_t offset_B = 0, tile_x_sa = 0, tile_y_sa = 0; if (view->base_level > 0) { /* We can't rely on the hardware's miplevel selection with such * a substantial lie about the format, so we select a single image * using the Tile X/Y Offset fields. In this case, we can't handle * multiple array slices. * * On Broadwell, HALIGN and VALIGN are specified in pixels and are * hard-coded to align to exactly the block size of the compressed * texture. This means that, when reinterpreted as a non-compressed * texture, the tile offsets may be anything and we can't rely on * X/Y Offset. * * Return NULL to force the state tracker to take fallback paths. */ if (view->array_len > 1 || GEN_GEN == 8) return NULL; const bool is_3d = res->surf.dim == ISL_SURF_DIM_3D; isl_surf_get_image_surf(&screen->isl_dev, &res->surf, view->base_level, is_3d ? 0 : view->base_array_layer, is_3d ? view->base_array_layer : 0, &isl_surf, &offset_B, &tile_x_sa, &tile_y_sa); /* We use address and tile offsets to access a single level/layer * as a subimage, so reset level/layer so it doesn't offset again. */ view->base_array_layer = 0; view->base_level = 0; } else { /* Level 0 doesn't require tile offsets, and the hardware can find * array slices using QPitch even with the format override, so we * can allow layers in this case. Copy the original ISL surface. */ memcpy(&isl_surf, &res->surf, sizeof(isl_surf)); } /* Scale down the image dimensions by the block size. */ const struct isl_format_layout *fmtl = isl_format_get_layout(res->surf.format); isl_surf.format = fmt.fmt; isl_surf.logical_level0_px.width = DIV_ROUND_UP(isl_surf.logical_level0_px.width, fmtl->bw); isl_surf.logical_level0_px.height = DIV_ROUND_UP(isl_surf.logical_level0_px.height, fmtl->bh); isl_surf.phys_level0_sa.width /= fmtl->bw; isl_surf.phys_level0_sa.height /= fmtl->bh; tile_x_sa /= fmtl->bw; tile_y_sa /= fmtl->bh; psurf->width = isl_surf.logical_level0_px.width; psurf->height = isl_surf.logical_level0_px.height; struct isl_surf_fill_state_info f = { .surf = &isl_surf, .view = view, .mocs = mocs(res->bo), .address = res->bo->gtt_offset + offset_B, .x_offset_sa = tile_x_sa, .y_offset_sa = tile_y_sa, }; isl_surf_fill_state_s(&screen->isl_dev, map, &f); return psurf; } #if GEN_GEN < 9 static void fill_default_image_param(struct brw_image_param *param) { memset(param, 0, sizeof(*param)); /* Set the swizzling shifts to all-ones to effectively disable swizzling -- * See emit_address_calculation() in brw_fs_surface_builder.cpp for a more * detailed explanation of these parameters. */ param->swizzling[0] = 0xff; param->swizzling[1] = 0xff; } static void fill_buffer_image_param(struct brw_image_param *param, enum pipe_format pfmt, unsigned size) { const unsigned cpp = util_format_get_blocksize(pfmt); fill_default_image_param(param); param->size[0] = size / cpp; param->stride[0] = cpp; } #else #define isl_surf_fill_image_param(x, ...) #define fill_default_image_param(x, ...) #define fill_buffer_image_param(x, ...) #endif /** * The pipe->set_shader_images() driver hook. */ static void iris_set_shader_images(struct pipe_context *ctx, enum pipe_shader_type p_stage, unsigned start_slot, unsigned count, const struct pipe_image_view *p_images) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_screen *screen = (struct iris_screen *)ctx->screen; const struct gen_device_info *devinfo = &screen->devinfo; gl_shader_stage stage = stage_from_pipe(p_stage); struct iris_shader_state *shs = &ice->state.shaders[stage]; #if GEN_GEN == 8 struct iris_genx_state *genx = ice->state.genx; struct brw_image_param *image_params = genx->shaders[stage].image_param; #endif shs->bound_image_views &= ~u_bit_consecutive(start_slot, count); for (unsigned i = 0; i < count; i++) { struct iris_image_view *iv = &shs->image[start_slot + i]; if (p_images && p_images[i].resource) { const struct pipe_image_view *img = &p_images[i]; struct iris_resource *res = (void *) img->resource; // XXX: these are not retained forever, use a separate uploader? void *map = alloc_surface_states(ice->state.surface_uploader, &iv->surface_state, 1 << ISL_AUX_USAGE_NONE); if (!unlikely(map)) return; iv->base = *img; iv->base.resource = NULL; pipe_resource_reference(&iv->base.resource, &res->base); shs->bound_image_views |= 1 << (start_slot + i); res->bind_history |= PIPE_BIND_SHADER_IMAGE; isl_surf_usage_flags_t usage = ISL_SURF_USAGE_STORAGE_BIT; enum isl_format isl_fmt = iris_format_for_usage(devinfo, img->format, usage).fmt; bool untyped_fallback = false; if (img->shader_access & PIPE_IMAGE_ACCESS_READ) { /* On Gen8, try to use typed surfaces reads (which support a * limited number of formats), and if not possible, fall back * to untyped reads. */ untyped_fallback = GEN_GEN == 8 && !isl_has_matching_typed_storage_image_format(devinfo, isl_fmt); if (untyped_fallback) isl_fmt = ISL_FORMAT_RAW; else isl_fmt = isl_lower_storage_image_format(devinfo, isl_fmt); } if (res->base.target != PIPE_BUFFER) { struct isl_view view = { .format = isl_fmt, .base_level = img->u.tex.level, .levels = 1, .base_array_layer = img->u.tex.first_layer, .array_len = img->u.tex.last_layer - img->u.tex.first_layer + 1, .swizzle = ISL_SWIZZLE_IDENTITY, .usage = usage, }; if (untyped_fallback) { fill_buffer_surface_state(&screen->isl_dev, res, map, isl_fmt, ISL_SWIZZLE_IDENTITY, 0, res->bo->size); } else { /* Images don't support compression */ unsigned aux_modes = 1 << ISL_AUX_USAGE_NONE; while (aux_modes) { enum isl_aux_usage usage = u_bit_scan(&aux_modes); fill_surface_state(&screen->isl_dev, map, res, &view, usage); map += SURFACE_STATE_ALIGNMENT; } } isl_surf_fill_image_param(&screen->isl_dev, &image_params[start_slot + i], &res->surf, &view); } else { util_range_add(&res->valid_buffer_range, img->u.buf.offset, img->u.buf.offset + img->u.buf.size); fill_buffer_surface_state(&screen->isl_dev, res, map, isl_fmt, ISL_SWIZZLE_IDENTITY, img->u.buf.offset, img->u.buf.size); fill_buffer_image_param(&image_params[start_slot + i], img->format, img->u.buf.size); } } else { pipe_resource_reference(&iv->base.resource, NULL); pipe_resource_reference(&iv->surface_state.res, NULL); fill_default_image_param(&image_params[start_slot + i]); } } ice->state.dirty |= IRIS_DIRTY_BINDINGS_VS << stage; ice->state.dirty |= stage == MESA_SHADER_COMPUTE ? IRIS_DIRTY_COMPUTE_RESOLVES_AND_FLUSHES : IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES; /* Broadwell also needs brw_image_params re-uploaded */ if (GEN_GEN < 9) { ice->state.dirty |= IRIS_DIRTY_CONSTANTS_VS << stage; shs->cbuf0_needs_upload = true; } } /** * The pipe->set_sampler_views() driver hook. */ static void iris_set_sampler_views(struct pipe_context *ctx, enum pipe_shader_type p_stage, unsigned start, unsigned count, struct pipe_sampler_view **views) { struct iris_context *ice = (struct iris_context *) ctx; gl_shader_stage stage = stage_from_pipe(p_stage); struct iris_shader_state *shs = &ice->state.shaders[stage]; shs->bound_sampler_views &= ~u_bit_consecutive(start, count); for (unsigned i = 0; i < count; i++) { struct pipe_sampler_view *pview = views ? views[i] : NULL; pipe_sampler_view_reference((struct pipe_sampler_view **) &shs->textures[start + i], pview); struct iris_sampler_view *view = (void *) pview; if (view) { view->res->bind_history |= PIPE_BIND_SAMPLER_VIEW; shs->bound_sampler_views |= 1 << (start + i); } } ice->state.dirty |= (IRIS_DIRTY_BINDINGS_VS << stage); ice->state.dirty |= stage == MESA_SHADER_COMPUTE ? IRIS_DIRTY_COMPUTE_RESOLVES_AND_FLUSHES : IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES; } /** * The pipe->set_tess_state() driver hook. */ static void iris_set_tess_state(struct pipe_context *ctx, const float default_outer_level[4], const float default_inner_level[2]) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_shader_state *shs = &ice->state.shaders[MESA_SHADER_TESS_CTRL]; memcpy(&ice->state.default_outer_level[0], &default_outer_level[0], 4 * sizeof(float)); memcpy(&ice->state.default_inner_level[0], &default_inner_level[0], 2 * sizeof(float)); ice->state.dirty |= IRIS_DIRTY_CONSTANTS_TCS; shs->cbuf0_needs_upload = true; } static void iris_surface_destroy(struct pipe_context *ctx, struct pipe_surface *p_surf) { struct iris_surface *surf = (void *) p_surf; pipe_resource_reference(&p_surf->texture, NULL); pipe_resource_reference(&surf->surface_state.res, NULL); free(surf); } static void iris_set_clip_state(struct pipe_context *ctx, const struct pipe_clip_state *state) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_shader_state *shs = &ice->state.shaders[MESA_SHADER_VERTEX]; memcpy(&ice->state.clip_planes, state, sizeof(*state)); ice->state.dirty |= IRIS_DIRTY_CONSTANTS_VS; shs->cbuf0_needs_upload = true; } /** * The pipe->set_polygon_stipple() driver hook. */ static void iris_set_polygon_stipple(struct pipe_context *ctx, const struct pipe_poly_stipple *state) { struct iris_context *ice = (struct iris_context *) ctx; memcpy(&ice->state.poly_stipple, state, sizeof(*state)); ice->state.dirty |= IRIS_DIRTY_POLYGON_STIPPLE; } /** * The pipe->set_sample_mask() driver hook. */ static void iris_set_sample_mask(struct pipe_context *ctx, unsigned sample_mask) { struct iris_context *ice = (struct iris_context *) ctx; /* We only support 16x MSAA, so we have 16 bits of sample maks. * st/mesa may pass us 0xffffffff though, meaning "enable all samples". */ ice->state.sample_mask = sample_mask & 0xffff; ice->state.dirty |= IRIS_DIRTY_SAMPLE_MASK; } /** * The pipe->set_scissor_states() driver hook. * * This corresponds to our SCISSOR_RECT state structures. It's an * exact match, so we just store them, and memcpy them out later. */ static void iris_set_scissor_states(struct pipe_context *ctx, unsigned start_slot, unsigned num_scissors, const struct pipe_scissor_state *rects) { struct iris_context *ice = (struct iris_context *) ctx; for (unsigned i = 0; i < num_scissors; i++) { if (rects[i].minx == rects[i].maxx || rects[i].miny == rects[i].maxy) { /* If the scissor was out of bounds and got clamped to 0 width/height * at the bounds, the subtraction of 1 from maximums could produce a * negative number and thus not clip anything. Instead, just provide * a min > max scissor inside the bounds, which produces the expected * no rendering. */ ice->state.scissors[start_slot + i] = (struct pipe_scissor_state) { .minx = 1, .maxx = 0, .miny = 1, .maxy = 0, }; } else { ice->state.scissors[start_slot + i] = (struct pipe_scissor_state) { .minx = rects[i].minx, .miny = rects[i].miny, .maxx = rects[i].maxx - 1, .maxy = rects[i].maxy - 1, }; } } ice->state.dirty |= IRIS_DIRTY_SCISSOR_RECT; } /** * The pipe->set_stencil_ref() driver hook. * * This is added to 3DSTATE_WM_DEPTH_STENCIL dynamically at draw time. */ static void iris_set_stencil_ref(struct pipe_context *ctx, const struct pipe_stencil_ref *state) { struct iris_context *ice = (struct iris_context *) ctx; memcpy(&ice->state.stencil_ref, state, sizeof(*state)); if (GEN_GEN == 8) ice->state.dirty |= IRIS_DIRTY_COLOR_CALC_STATE; else ice->state.dirty |= IRIS_DIRTY_WM_DEPTH_STENCIL; } static float viewport_extent(const struct pipe_viewport_state *state, int axis, float sign) { return copysignf(state->scale[axis], sign) + state->translate[axis]; } static void calculate_guardband_size(uint32_t fb_width, uint32_t fb_height, float m00, float m11, float m30, float m31, float *xmin, float *xmax, float *ymin, float *ymax) { /* According to the "Vertex X,Y Clamping and Quantization" section of the * Strips and Fans documentation: * * "The vertex X and Y screen-space coordinates are also /clamped/ to the * fixed-point "guardband" range supported by the rasterization hardware" * * and * * "In almost all circumstances, if an object’s vertices are actually * modified by this clamping (i.e., had X or Y coordinates outside of * the guardband extent the rendered object will not match the intended * result. Therefore software should take steps to ensure that this does * not happen - e.g., by clipping objects such that they do not exceed * these limits after the Drawing Rectangle is applied." * * I believe the fundamental restriction is that the rasterizer (in * the SF/WM stages) have a limit on the number of pixels that can be * rasterized. We need to ensure any coordinates beyond the rasterizer * limit are handled by the clipper. So effectively that limit becomes * the clipper's guardband size. * * It goes on to say: * * "In addition, in order to be correctly rendered, objects must have a * screenspace bounding box not exceeding 8K in the X or Y direction. * This additional restriction must also be comprehended by software, * i.e., enforced by use of clipping." * * This makes no sense. Gen7+ hardware supports 16K render targets, * and you definitely need to be able to draw polygons that fill the * surface. Our assumption is that the rasterizer was limited to 8K * on Sandybridge, which only supports 8K surfaces, and it was actually * increased to 16K on Ivybridge and later. * * So, limit the guardband to 16K on Gen7+ and 8K on Sandybridge. */ const float gb_size = GEN_GEN >= 7 ? 16384.0f : 8192.0f; if (m00 != 0 && m11 != 0) { /* First, we compute the screen-space render area */ const float ss_ra_xmin = MIN3( 0, m30 + m00, m30 - m00); const float ss_ra_xmax = MAX3( fb_width, m30 + m00, m30 - m00); const float ss_ra_ymin = MIN3( 0, m31 + m11, m31 - m11); const float ss_ra_ymax = MAX3(fb_height, m31 + m11, m31 - m11); /* We want the guardband to be centered on that */ const float ss_gb_xmin = (ss_ra_xmin + ss_ra_xmax) / 2 - gb_size; const float ss_gb_xmax = (ss_ra_xmin + ss_ra_xmax) / 2 + gb_size; const float ss_gb_ymin = (ss_ra_ymin + ss_ra_ymax) / 2 - gb_size; const float ss_gb_ymax = (ss_ra_ymin + ss_ra_ymax) / 2 + gb_size; /* Now we need it in native device coordinates */ const float ndc_gb_xmin = (ss_gb_xmin - m30) / m00; const float ndc_gb_xmax = (ss_gb_xmax - m30) / m00; const float ndc_gb_ymin = (ss_gb_ymin - m31) / m11; const float ndc_gb_ymax = (ss_gb_ymax - m31) / m11; /* Thanks to Y-flipping and ORIGIN_UPPER_LEFT, the Y coordinates may be * flipped upside-down. X should be fine though. */ assert(ndc_gb_xmin <= ndc_gb_xmax); *xmin = ndc_gb_xmin; *xmax = ndc_gb_xmax; *ymin = MIN2(ndc_gb_ymin, ndc_gb_ymax); *ymax = MAX2(ndc_gb_ymin, ndc_gb_ymax); } else { /* The viewport scales to 0, so nothing will be rendered. */ *xmin = 0.0f; *xmax = 0.0f; *ymin = 0.0f; *ymax = 0.0f; } } /** * The pipe->set_viewport_states() driver hook. * * This corresponds to our SF_CLIP_VIEWPORT states. We can't calculate * the guardband yet, as we need the framebuffer dimensions, but we can * at least fill out the rest. */ static void iris_set_viewport_states(struct pipe_context *ctx, unsigned start_slot, unsigned count, const struct pipe_viewport_state *states) { struct iris_context *ice = (struct iris_context *) ctx; memcpy(&ice->state.viewports[start_slot], states, sizeof(*states) * count); ice->state.dirty |= IRIS_DIRTY_SF_CL_VIEWPORT; if (ice->state.cso_rast && (!ice->state.cso_rast->depth_clip_near || !ice->state.cso_rast->depth_clip_far)) ice->state.dirty |= IRIS_DIRTY_CC_VIEWPORT; } /** * The pipe->set_framebuffer_state() driver hook. * * Sets the current draw FBO, including color render targets, depth, * and stencil buffers. */ static void iris_set_framebuffer_state(struct pipe_context *ctx, const struct pipe_framebuffer_state *state) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_screen *screen = (struct iris_screen *)ctx->screen; struct isl_device *isl_dev = &screen->isl_dev; struct pipe_framebuffer_state *cso = &ice->state.framebuffer; struct iris_resource *zres; struct iris_resource *stencil_res; unsigned samples = util_framebuffer_get_num_samples(state); unsigned layers = util_framebuffer_get_num_layers(state); if (cso->samples != samples) { ice->state.dirty |= IRIS_DIRTY_MULTISAMPLE; } if (cso->nr_cbufs != state->nr_cbufs) { ice->state.dirty |= IRIS_DIRTY_BLEND_STATE; } if ((cso->layers == 0) != (layers == 0)) { ice->state.dirty |= IRIS_DIRTY_CLIP; } if (cso->width != state->width || cso->height != state->height) { ice->state.dirty |= IRIS_DIRTY_SF_CL_VIEWPORT; } util_copy_framebuffer_state(cso, state); cso->samples = samples; cso->layers = layers; struct iris_depth_buffer_state *cso_z = &ice->state.genx->depth_buffer; struct isl_view view = { .base_level = 0, .levels = 1, .base_array_layer = 0, .array_len = 1, .swizzle = ISL_SWIZZLE_IDENTITY, }; struct isl_depth_stencil_hiz_emit_info info = { .view = &view }; if (cso->zsbuf) { iris_get_depth_stencil_resources(cso->zsbuf->texture, &zres, &stencil_res); view.base_level = cso->zsbuf->u.tex.level; view.base_array_layer = cso->zsbuf->u.tex.first_layer; view.array_len = cso->zsbuf->u.tex.last_layer - cso->zsbuf->u.tex.first_layer + 1; if (zres) { view.usage |= ISL_SURF_USAGE_DEPTH_BIT; info.depth_surf = &zres->surf; info.depth_address = zres->bo->gtt_offset + zres->offset; info.mocs = mocs(zres->bo); view.format = zres->surf.format; if (iris_resource_level_has_hiz(zres, view.base_level)) { info.hiz_usage = ISL_AUX_USAGE_HIZ; info.hiz_surf = &zres->aux.surf; info.hiz_address = zres->aux.bo->gtt_offset; } } if (stencil_res) { view.usage |= ISL_SURF_USAGE_STENCIL_BIT; info.stencil_surf = &stencil_res->surf; info.stencil_address = stencil_res->bo->gtt_offset + stencil_res->offset; if (!zres) { view.format = stencil_res->surf.format; info.mocs = mocs(stencil_res->bo); } } } isl_emit_depth_stencil_hiz_s(isl_dev, cso_z->packets, &info); /* Make a null surface for unbound buffers */ void *null_surf_map = upload_state(ice->state.surface_uploader, &ice->state.null_fb, 4 * GENX(RENDER_SURFACE_STATE_length), 64); isl_null_fill_state(&screen->isl_dev, null_surf_map, isl_extent3d(MAX2(cso->width, 1), MAX2(cso->height, 1), cso->layers ? cso->layers : 1)); ice->state.null_fb.offset += iris_bo_offset_from_base_address(iris_resource_bo(ice->state.null_fb.res)); ice->state.dirty |= IRIS_DIRTY_DEPTH_BUFFER; /* Render target change */ ice->state.dirty |= IRIS_DIRTY_BINDINGS_FS; ice->state.dirty |= IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES; ice->state.dirty |= ice->state.dirty_for_nos[IRIS_NOS_FRAMEBUFFER]; #if GEN_GEN == 11 // XXX: we may want to flag IRIS_DIRTY_MULTISAMPLE (or SAMPLE_MASK?) // XXX: see commit 979fc1bc9bcc64027ff2cfafd285676f31b930a6 /* The PIPE_CONTROL command description says: * * "Whenever a Binding Table Index (BTI) used by a Render Target Message * points to a different RENDER_SURFACE_STATE, SW must issue a Render * Target Cache Flush by enabling this bit. When render target flush * is set due to new association of BTI, PS Scoreboard Stall bit must * be set in this packet." */ // XXX: does this need to happen at 3DSTATE_BTP_PS time? iris_emit_pipe_control_flush(&ice->batches[IRIS_BATCH_RENDER], PIPE_CONTROL_RENDER_TARGET_FLUSH | PIPE_CONTROL_STALL_AT_SCOREBOARD); #endif } /** * The pipe->set_constant_buffer() driver hook. * * This uploads any constant data in user buffers, and references * any UBO resources containing constant data. */ static void iris_set_constant_buffer(struct pipe_context *ctx, enum pipe_shader_type p_stage, unsigned index, const struct pipe_constant_buffer *input) { struct iris_context *ice = (struct iris_context *) ctx; gl_shader_stage stage = stage_from_pipe(p_stage); struct iris_shader_state *shs = &ice->state.shaders[stage]; struct pipe_shader_buffer *cbuf = &shs->constbuf[index]; if (input && input->buffer) { shs->bound_cbufs |= 1u << index; assert(index > 0); pipe_resource_reference(&cbuf->buffer, input->buffer); cbuf->buffer_offset = input->buffer_offset; cbuf->buffer_size = MIN2(input->buffer_size, iris_resource_bo(input->buffer)->size - cbuf->buffer_offset); struct iris_resource *res = (void *) cbuf->buffer; res->bind_history |= PIPE_BIND_CONSTANT_BUFFER; iris_upload_ubo_ssbo_surf_state(ice, cbuf, &shs->constbuf_surf_state[index], false); } else { shs->bound_cbufs &= ~(1u << index); pipe_resource_reference(&cbuf->buffer, NULL); pipe_resource_reference(&shs->constbuf_surf_state[index].res, NULL); } if (index == 0) { if (input) memcpy(&shs->cbuf0, input, sizeof(shs->cbuf0)); else memset(&shs->cbuf0, 0, sizeof(shs->cbuf0)); shs->cbuf0_needs_upload = true; } ice->state.dirty |= IRIS_DIRTY_CONSTANTS_VS << stage; // XXX: maybe not necessary all the time...? // XXX: we need 3DS_BTP to commit these changes, and if we fell back to // XXX: pull model we may need actual new bindings... ice->state.dirty |= IRIS_DIRTY_BINDINGS_VS << stage; } static void upload_uniforms(struct iris_context *ice, gl_shader_stage stage) { UNUSED struct iris_genx_state *genx = ice->state.genx; struct iris_shader_state *shs = &ice->state.shaders[stage]; struct pipe_shader_buffer *cbuf = &shs->constbuf[0]; struct iris_compiled_shader *shader = ice->shaders.prog[stage]; unsigned upload_size = shader->num_system_values * sizeof(uint32_t) + shs->cbuf0.buffer_size; if (upload_size == 0) return; uint32_t *map = NULL; u_upload_alloc(ice->ctx.const_uploader, 0, upload_size, 64, &cbuf->buffer_offset, &cbuf->buffer, (void **) &map); for (int i = 0; i < shader->num_system_values; i++) { uint32_t sysval = shader->system_values[i]; uint32_t value = 0; if (BRW_PARAM_DOMAIN(sysval) == BRW_PARAM_DOMAIN_IMAGE) { #if GEN_GEN == 8 unsigned img = BRW_PARAM_IMAGE_IDX(sysval); unsigned offset = BRW_PARAM_IMAGE_OFFSET(sysval); struct brw_image_param *param = &genx->shaders[stage].image_param[img]; assert(offset < sizeof(struct brw_image_param)); value = ((uint32_t *) param)[offset]; #endif } else if (sysval == BRW_PARAM_BUILTIN_ZERO) { value = 0; } else if (BRW_PARAM_BUILTIN_IS_CLIP_PLANE(sysval)) { int plane = BRW_PARAM_BUILTIN_CLIP_PLANE_IDX(sysval); int comp = BRW_PARAM_BUILTIN_CLIP_PLANE_COMP(sysval); value = fui(ice->state.clip_planes.ucp[plane][comp]); } else if (sysval == BRW_PARAM_BUILTIN_PATCH_VERTICES_IN) { if (stage == MESA_SHADER_TESS_CTRL) { value = ice->state.vertices_per_patch; } else { assert(stage == MESA_SHADER_TESS_EVAL); const struct shader_info *tcs_info = iris_get_shader_info(ice, MESA_SHADER_TESS_CTRL); if (tcs_info) value = tcs_info->tess.tcs_vertices_out; else value = ice->state.vertices_per_patch; } } else if (sysval >= BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X && sysval <= BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_W) { unsigned i = sysval - BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X; value = fui(ice->state.default_outer_level[i]); } else if (sysval == BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_X) { value = fui(ice->state.default_inner_level[0]); } else if (sysval == BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_Y) { value = fui(ice->state.default_inner_level[1]); } else { assert(!"unhandled system value"); } *map++ = value; } if (shs->cbuf0.user_buffer) { memcpy(map, shs->cbuf0.user_buffer, shs->cbuf0.buffer_size); } cbuf->buffer_size = upload_size; iris_upload_ubo_ssbo_surf_state(ice, cbuf, &shs->constbuf_surf_state[0], false); } /** * The pipe->set_shader_buffers() driver hook. * * This binds SSBOs and ABOs. Unfortunately, we need to stream out * SURFACE_STATE here, as the buffer offset may change each time. */ static void iris_set_shader_buffers(struct pipe_context *ctx, enum pipe_shader_type p_stage, unsigned start_slot, unsigned count, const struct pipe_shader_buffer *buffers, unsigned writable_bitmask) { struct iris_context *ice = (struct iris_context *) ctx; gl_shader_stage stage = stage_from_pipe(p_stage); struct iris_shader_state *shs = &ice->state.shaders[stage]; unsigned modified_bits = u_bit_consecutive(start_slot, count); shs->bound_ssbos &= ~modified_bits; shs->writable_ssbos &= ~modified_bits; shs->writable_ssbos |= writable_bitmask << start_slot; for (unsigned i = 0; i < count; i++) { if (buffers && buffers[i].buffer) { struct iris_resource *res = (void *) buffers[i].buffer; struct pipe_shader_buffer *ssbo = &shs->ssbo[start_slot + i]; struct iris_state_ref *surf_state = &shs->ssbo_surf_state[start_slot + i]; pipe_resource_reference(&ssbo->buffer, &res->base); ssbo->buffer_offset = buffers[i].buffer_offset; ssbo->buffer_size = MIN2(buffers[i].buffer_size, res->bo->size - ssbo->buffer_offset); shs->bound_ssbos |= 1 << (start_slot + i); iris_upload_ubo_ssbo_surf_state(ice, ssbo, surf_state, true); res->bind_history |= PIPE_BIND_SHADER_BUFFER; util_range_add(&res->valid_buffer_range, ssbo->buffer_offset, ssbo->buffer_offset + ssbo->buffer_size); } else { pipe_resource_reference(&shs->ssbo[start_slot + i].buffer, NULL); pipe_resource_reference(&shs->ssbo_surf_state[start_slot + i].res, NULL); } } ice->state.dirty |= IRIS_DIRTY_BINDINGS_VS << stage; } static void iris_delete_state(struct pipe_context *ctx, void *state) { free(state); } /** * The pipe->set_vertex_buffers() driver hook. * * This translates pipe_vertex_buffer to our 3DSTATE_VERTEX_BUFFERS packet. */ static void iris_set_vertex_buffers(struct pipe_context *ctx, unsigned start_slot, unsigned count, const struct pipe_vertex_buffer *buffers) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_genx_state *genx = ice->state.genx; ice->state.bound_vertex_buffers &= ~u_bit_consecutive64(start_slot, count); for (unsigned i = 0; i < count; i++) { const struct pipe_vertex_buffer *buffer = buffers ? &buffers[i] : NULL; struct iris_vertex_buffer_state *state = &genx->vertex_buffers[start_slot + i]; if (!buffer) { pipe_resource_reference(&state->resource, NULL); continue; } /* We may see user buffers that are NULL bindings. */ assert(!(buffer->is_user_buffer && buffer->buffer.user != NULL)); pipe_resource_reference(&state->resource, buffer->buffer.resource); struct iris_resource *res = (void *) state->resource; if (res) { ice->state.bound_vertex_buffers |= 1ull << (start_slot + i); res->bind_history |= PIPE_BIND_VERTEX_BUFFER; } iris_pack_state(GENX(VERTEX_BUFFER_STATE), state->state, vb) { vb.VertexBufferIndex = start_slot + i; vb.AddressModifyEnable = true; vb.BufferPitch = buffer->stride; if (res) { vb.BufferSize = res->bo->size - (int) buffer->buffer_offset; vb.BufferStartingAddress = ro_bo(NULL, res->bo->gtt_offset + (int) buffer->buffer_offset); vb.MOCS = mocs(res->bo); } else { vb.NullVertexBuffer = true; } } } ice->state.dirty |= IRIS_DIRTY_VERTEX_BUFFERS; } /** * Gallium CSO for vertex elements. */ struct iris_vertex_element_state { uint32_t vertex_elements[1 + 33 * GENX(VERTEX_ELEMENT_STATE_length)]; uint32_t vf_instancing[33 * GENX(3DSTATE_VF_INSTANCING_length)]; uint32_t edgeflag_ve[GENX(VERTEX_ELEMENT_STATE_length)]; uint32_t edgeflag_vfi[GENX(3DSTATE_VF_INSTANCING_length)]; unsigned count; }; /** * The pipe->create_vertex_elements() driver hook. * * This translates pipe_vertex_element to our 3DSTATE_VERTEX_ELEMENTS * and 3DSTATE_VF_INSTANCING commands. The vertex_elements and vf_instancing * arrays are ready to be emitted at draw time if no EdgeFlag or SGVs are * needed. In these cases we will need information available at draw time. * We setup edgeflag_ve and edgeflag_vfi as alternatives last * 3DSTATE_VERTEX_ELEMENT and 3DSTATE_VF_INSTANCING that can be used at * draw time if we detect that EdgeFlag is needed by the Vertex Shader. */ static void * iris_create_vertex_elements(struct pipe_context *ctx, unsigned count, const struct pipe_vertex_element *state) { struct iris_screen *screen = (struct iris_screen *)ctx->screen; const struct gen_device_info *devinfo = &screen->devinfo; struct iris_vertex_element_state *cso = malloc(sizeof(struct iris_vertex_element_state)); cso->count = count; iris_pack_command(GENX(3DSTATE_VERTEX_ELEMENTS), cso->vertex_elements, ve) { ve.DWordLength = 1 + GENX(VERTEX_ELEMENT_STATE_length) * MAX2(count, 1) - 2; } uint32_t *ve_pack_dest = &cso->vertex_elements[1]; uint32_t *vfi_pack_dest = cso->vf_instancing; if (count == 0) { iris_pack_state(GENX(VERTEX_ELEMENT_STATE), ve_pack_dest, ve) { ve.Valid = true; ve.SourceElementFormat = ISL_FORMAT_R32G32B32A32_FLOAT; ve.Component0Control = VFCOMP_STORE_0; ve.Component1Control = VFCOMP_STORE_0; ve.Component2Control = VFCOMP_STORE_0; ve.Component3Control = VFCOMP_STORE_1_FP; } iris_pack_command(GENX(3DSTATE_VF_INSTANCING), vfi_pack_dest, vi) { } } for (int i = 0; i < count; i++) { const struct iris_format_info fmt = iris_format_for_usage(devinfo, state[i].src_format, 0); unsigned comp[4] = { VFCOMP_STORE_SRC, VFCOMP_STORE_SRC, VFCOMP_STORE_SRC, VFCOMP_STORE_SRC }; switch (isl_format_get_num_channels(fmt.fmt)) { case 0: comp[0] = VFCOMP_STORE_0; /* fallthrough */ case 1: comp[1] = VFCOMP_STORE_0; /* fallthrough */ case 2: comp[2] = VFCOMP_STORE_0; /* fallthrough */ case 3: comp[3] = isl_format_has_int_channel(fmt.fmt) ? VFCOMP_STORE_1_INT : VFCOMP_STORE_1_FP; break; } iris_pack_state(GENX(VERTEX_ELEMENT_STATE), ve_pack_dest, ve) { ve.EdgeFlagEnable = false; ve.VertexBufferIndex = state[i].vertex_buffer_index; ve.Valid = true; ve.SourceElementOffset = state[i].src_offset; ve.SourceElementFormat = fmt.fmt; ve.Component0Control = comp[0]; ve.Component1Control = comp[1]; ve.Component2Control = comp[2]; ve.Component3Control = comp[3]; } iris_pack_command(GENX(3DSTATE_VF_INSTANCING), vfi_pack_dest, vi) { vi.VertexElementIndex = i; vi.InstancingEnable = state[i].instance_divisor > 0; vi.InstanceDataStepRate = state[i].instance_divisor; } ve_pack_dest += GENX(VERTEX_ELEMENT_STATE_length); vfi_pack_dest += GENX(3DSTATE_VF_INSTANCING_length); } /* An alternative version of the last VE and VFI is stored so it * can be used at draw time in case Vertex Shader uses EdgeFlag */ if (count) { const unsigned edgeflag_index = count - 1; const struct iris_format_info fmt = iris_format_for_usage(devinfo, state[edgeflag_index].src_format, 0); iris_pack_state(GENX(VERTEX_ELEMENT_STATE), cso->edgeflag_ve, ve) { ve.EdgeFlagEnable = true ; ve.VertexBufferIndex = state[edgeflag_index].vertex_buffer_index; ve.Valid = true; ve.SourceElementOffset = state[edgeflag_index].src_offset; ve.SourceElementFormat = fmt.fmt; ve.Component0Control = VFCOMP_STORE_SRC; ve.Component1Control = VFCOMP_STORE_0; ve.Component2Control = VFCOMP_STORE_0; ve.Component3Control = VFCOMP_STORE_0; } iris_pack_command(GENX(3DSTATE_VF_INSTANCING), cso->edgeflag_vfi, vi) { /* The vi.VertexElementIndex of the EdgeFlag Vertex Element is filled * at draw time, as it should change if SGVs are emitted. */ vi.InstancingEnable = state[edgeflag_index].instance_divisor > 0; vi.InstanceDataStepRate = state[edgeflag_index].instance_divisor; } } return cso; } /** * The pipe->bind_vertex_elements_state() driver hook. */ static void iris_bind_vertex_elements_state(struct pipe_context *ctx, void *state) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_vertex_element_state *old_cso = ice->state.cso_vertex_elements; struct iris_vertex_element_state *new_cso = state; /* 3DSTATE_VF_SGVs overrides the last VE, so if the count is changing, * we need to re-emit it to ensure we're overriding the right one. */ if (new_cso && cso_changed(count)) ice->state.dirty |= IRIS_DIRTY_VF_SGVS; ice->state.cso_vertex_elements = state; ice->state.dirty |= IRIS_DIRTY_VERTEX_ELEMENTS; } /** * The pipe->create_stream_output_target() driver hook. * * "Target" here refers to a destination buffer. We translate this into * a 3DSTATE_SO_BUFFER packet. We can handle most fields, but don't yet * know which buffer this represents, or whether we ought to zero the * write-offsets, or append. Those are handled in the set() hook. */ static struct pipe_stream_output_target * iris_create_stream_output_target(struct pipe_context *ctx, struct pipe_resource *p_res, unsigned buffer_offset, unsigned buffer_size) { struct iris_resource *res = (void *) p_res; struct iris_stream_output_target *cso = calloc(1, sizeof(*cso)); if (!cso) return NULL; res->bind_history |= PIPE_BIND_STREAM_OUTPUT; pipe_reference_init(&cso->base.reference, 1); pipe_resource_reference(&cso->base.buffer, p_res); cso->base.buffer_offset = buffer_offset; cso->base.buffer_size = buffer_size; cso->base.context = ctx; util_range_add(&res->valid_buffer_range, buffer_offset, buffer_offset + buffer_size); upload_state(ctx->stream_uploader, &cso->offset, sizeof(uint32_t), 4); return &cso->base; } static void iris_stream_output_target_destroy(struct pipe_context *ctx, struct pipe_stream_output_target *state) { struct iris_stream_output_target *cso = (void *) state; pipe_resource_reference(&cso->base.buffer, NULL); pipe_resource_reference(&cso->offset.res, NULL); free(cso); } /** * The pipe->set_stream_output_targets() driver hook. * * At this point, we know which targets are bound to a particular index, * and also whether we want to append or start over. We can finish the * 3DSTATE_SO_BUFFER packets we started earlier. */ static void iris_set_stream_output_targets(struct pipe_context *ctx, unsigned num_targets, struct pipe_stream_output_target **targets, const unsigned *offsets) { struct iris_context *ice = (struct iris_context *) ctx; struct iris_genx_state *genx = ice->state.genx; uint32_t *so_buffers = genx->so_buffers; const bool active = num_targets > 0; if (ice->state.streamout_active != active) { ice->state.streamout_active = active; ice->state.dirty |= IRIS_DIRTY_STREAMOUT; /* We only emit 3DSTATE_SO_DECL_LIST when streamout is active, because * it's a non-pipelined command. If we're switching streamout on, we * may have missed emitting it earlier, so do so now. (We're already * taking a stall to update 3DSTATE_SO_BUFFERS anyway...) */ if (active) { ice->state.dirty |= IRIS_DIRTY_SO_DECL_LIST; } else { uint32_t flush = 0; for (int i = 0; i < PIPE_MAX_SO_BUFFERS; i++) { struct iris_stream_output_target *tgt = (void *) ice->state.so_target[i]; if (tgt) { struct iris_resource *res = (void *) tgt->base.buffer; flush |= iris_flush_bits_for_history(res); iris_dirty_for_history(ice, res); } } iris_emit_pipe_control_flush(&ice->batches[IRIS_BATCH_RENDER], flush); } } for (int i = 0; i < 4; i++) { pipe_so_target_reference(&ice->state.so_target[i], i < num_targets ? targets[i] : NULL); } /* No need to update 3DSTATE_SO_BUFFER unless SOL is active. */ if (!active) return; for (unsigned i = 0; i < 4; i++, so_buffers += GENX(3DSTATE_SO_BUFFER_length)) { struct iris_stream_output_target *tgt = (void *) ice->state.so_target[i]; unsigned offset = offsets[i]; if (!tgt) { iris_pack_command(GENX(3DSTATE_SO_BUFFER), so_buffers, sob) sob.SOBufferIndex = i; continue; } struct iris_resource *res = (void *) tgt->base.buffer; /* Note that offsets[i] will either be 0, causing us to zero * the value in the buffer, or 0xFFFFFFFF, which happens to mean * "continue appending at the existing offset." */ assert(offset == 0 || offset == 0xFFFFFFFF); /* We might be called by Begin (offset = 0), Pause, then Resume * (offset = 0xFFFFFFFF) before ever drawing (where these commands * will actually be sent to the GPU). In this case, we don't want * to append - we still want to do our initial zeroing. */ if (!tgt->zeroed) offset = 0; iris_pack_command(GENX(3DSTATE_SO_BUFFER), so_buffers, sob) { sob.SurfaceBaseAddress = rw_bo(NULL, res->bo->gtt_offset + tgt->base.buffer_offset); sob.SOBufferEnable = true; sob.StreamOffsetWriteEnable = true; sob.StreamOutputBufferOffsetAddressEnable = true; sob.MOCS = mocs(res->bo); sob.SurfaceSize = MAX2(tgt->base.buffer_size / 4, 1) - 1; sob.SOBufferIndex = i; sob.StreamOffset = offset; sob.StreamOutputBufferOffsetAddress = rw_bo(NULL, iris_resource_bo(tgt->offset.res)->gtt_offset + tgt->offset.offset); } } ice->state.dirty |= IRIS_DIRTY_SO_BUFFERS; } /** * An iris-vtable helper for encoding the 3DSTATE_SO_DECL_LIST and * 3DSTATE_STREAMOUT packets. * * 3DSTATE_SO_DECL_LIST is a list of shader outputs we want the streamout * hardware to record. We can create it entirely based on the shader, with * no dynamic state dependencies. * * 3DSTATE_STREAMOUT is an annoying mix of shader-based information and * state-based settings. We capture the shader-related ones here, and merge * the rest in at draw time. */ static uint32_t * iris_create_so_decl_list(const struct pipe_stream_output_info *info, const struct brw_vue_map *vue_map) { struct GENX(SO_DECL) so_decl[MAX_VERTEX_STREAMS][128]; int buffer_mask[MAX_VERTEX_STREAMS] = {0, 0, 0, 0}; int next_offset[MAX_VERTEX_STREAMS] = {0, 0, 0, 0}; int decls[MAX_VERTEX_STREAMS] = {0, 0, 0, 0}; int max_decls = 0; STATIC_ASSERT(ARRAY_SIZE(so_decl[0]) >= MAX_PROGRAM_OUTPUTS); memset(so_decl, 0, sizeof(so_decl)); /* Construct the list of SO_DECLs to be emitted. The formatting of the * command feels strange -- each dword pair contains a SO_DECL per stream. */ for (unsigned i = 0; i < info->num_outputs; i++) { const struct pipe_stream_output *output = &info->output[i]; const int buffer = output->output_buffer; const int varying = output->register_index; const unsigned stream_id = output->stream; assert(stream_id < MAX_VERTEX_STREAMS); buffer_mask[stream_id] |= 1 << buffer; assert(vue_map->varying_to_slot[varying] >= 0); /* Mesa doesn't store entries for gl_SkipComponents in the Outputs[] * array. Instead, it simply increments DstOffset for the following * input by the number of components that should be skipped. * * Our hardware is unusual in that it requires us to program SO_DECLs * for fake "hole" components, rather than simply taking the offset * for each real varying. Each hole can have size 1, 2, 3, or 4; we * program as many size = 4 holes as we can, then a final hole to * accommodate the final 1, 2, or 3 remaining. */ int skip_components = output->dst_offset - next_offset[buffer]; while (skip_components > 0) { so_decl[stream_id][decls[stream_id]++] = (struct GENX(SO_DECL)) { .HoleFlag = 1, .OutputBufferSlot = output->output_buffer, .ComponentMask = (1 << MIN2(skip_components, 4)) - 1, }; skip_components -= 4; } next_offset[buffer] = output->dst_offset + output->num_components; so_decl[stream_id][decls[stream_id]++] = (struct GENX(SO_DECL)) { .OutputBufferSlot = output->output_buffer, .RegisterIndex = vue_map->varying_to_slot[varying], .ComponentMask = ((1 << output->num_components) - 1) << output->start_component, }; if (decls[stream_id] > max_decls) max_decls = decls[stream_id]; } unsigned dwords = GENX(3DSTATE_STREAMOUT_length) + (3 + 2 * max_decls); uint32_t *map = ralloc_size(NULL, sizeof(uint32_t) * dwords); uint32_t *so_decl_map = map + GENX(3DSTATE_STREAMOUT_length); iris_pack_command(GENX(3DSTATE_STREAMOUT), map, sol) { int urb_entry_read_offset = 0; int urb_entry_read_length = (vue_map->num_slots + 1) / 2 - urb_entry_read_offset; /* We always read the whole vertex. This could be reduced at some * point by reading less and offsetting the register index in the * SO_DECLs. */ sol.Stream0VertexReadOffset = urb_entry_read_offset; sol.Stream0VertexReadLength = urb_entry_read_length - 1; sol.Stream1VertexReadOffset = urb_entry_read_offset; sol.Stream1VertexReadLength = urb_entry_read_length - 1; sol.Stream2VertexReadOffset = urb_entry_read_offset; sol.Stream2VertexReadLength = urb_entry_read_length - 1; sol.Stream3VertexReadOffset = urb_entry_read_offset; sol.Stream3VertexReadLength = urb_entry_read_length - 1; /* Set buffer pitches; 0 means unbound. */ sol.Buffer0SurfacePitch = 4 * info->stride[0]; sol.Buffer1SurfacePitch = 4 * info->stride[1]; sol.Buffer2SurfacePitch = 4 * info->stride[2]; sol.Buffer3SurfacePitch = 4 * info->stride[3]; } iris_pack_command(GENX(3DSTATE_SO_DECL_LIST), so_decl_map, list) { list.DWordLength = 3 + 2 * max_decls - 2; list.StreamtoBufferSelects0 = buffer_mask[0]; list.StreamtoBufferSelects1 = buffer_mask[1]; list.StreamtoBufferSelects2 = buffer_mask[2]; list.StreamtoBufferSelects3 = buffer_mask[3]; list.NumEntries0 = decls[0]; list.NumEntries1 = decls[1]; list.NumEntries2 = decls[2]; list.NumEntries3 = decls[3]; } for (int i = 0; i < max_decls; i++) { iris_pack_state(GENX(SO_DECL_ENTRY), so_decl_map + 3 + i * 2, entry) { entry.Stream0Decl = so_decl[0][i]; entry.Stream1Decl = so_decl[1][i]; entry.Stream2Decl = so_decl[2][i]; entry.Stream3Decl = so_decl[3][i]; } } return map; } static void iris_compute_sbe_urb_read_interval(uint64_t fs_input_slots, const struct brw_vue_map *last_vue_map, bool two_sided_color, unsigned *out_offset, unsigned *out_length) { /* The compiler computes the first URB slot without considering COL/BFC * swizzling (because it doesn't know whether it's enabled), so we need * to do that here too. This may result in a smaller offset, which * should be safe. */ const unsigned first_slot = brw_compute_first_urb_slot_required(fs_input_slots, last_vue_map); /* This becomes the URB read offset (counted in pairs of slots). */ assert(first_slot % 2 == 0); *out_offset = first_slot / 2; /* We need to adjust the inputs read to account for front/back color * swizzling, as it can make the URB length longer. */ for (int c = 0; c <= 1; c++) { if (fs_input_slots & (VARYING_BIT_COL0 << c)) { /* If two sided color is enabled, the fragment shader's gl_Color * (COL0) input comes from either the gl_FrontColor (COL0) or * gl_BackColor (BFC0) input varyings. Mark BFC as used, too. */ if (two_sided_color) fs_input_slots |= (VARYING_BIT_BFC0 << c); /* If front color isn't written, we opt to give them back color * instead of an undefined value. Switch from COL to BFC. */ if (last_vue_map->varying_to_slot[VARYING_SLOT_COL0 + c] == -1) { fs_input_slots &= ~(VARYING_BIT_COL0 << c); fs_input_slots |= (VARYING_BIT_BFC0 << c); } } } /* Compute the minimum URB Read Length necessary for the FS inputs. * * From the Sandy Bridge PRM, Volume 2, Part 1, documentation for * 3DSTATE_SF DWord 1 bits 15:11, "Vertex URB Entry Read Length": * * "This field should be set to the minimum length required to read the * maximum source attribute. The maximum source attribute is indicated * by the maximum value of the enabled Attribute # Source Attribute if * Attribute Swizzle Enable is set, Number of Output Attributes-1 if * enable is not set. * read_length = ceiling((max_source_attr + 1) / 2) * * [errata] Corruption/Hang possible if length programmed larger than * recommended" * * Similar text exists for Ivy Bridge. * * We find the last URB slot that's actually read by the FS. */ unsigned last_read_slot = last_vue_map->num_slots - 1; while (last_read_slot > first_slot && !(fs_input_slots & (1ull << last_vue_map->slot_to_varying[last_read_slot]))) --last_read_slot; /* The URB read length is the difference of the two, counted in pairs. */ *out_length = DIV_ROUND_UP(last_read_slot - first_slot + 1, 2); } static void iris_emit_sbe_swiz(struct iris_batch *batch, const struct iris_context *ice, unsigned urb_read_offset, unsigned sprite_coord_enables) { struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL) attr_overrides[16] = {}; const struct brw_wm_prog_data *wm_prog_data = (void *) ice->shaders.prog[MESA_SHADER_FRAGMENT]->prog_data; const struct brw_vue_map *vue_map = ice->shaders.last_vue_map; const struct iris_rasterizer_state *cso_rast = ice->state.cso_rast; /* XXX: this should be generated when putting programs in place */ for (int fs_attr = 0; fs_attr < VARYING_SLOT_MAX; fs_attr++) { const int input_index = wm_prog_data->urb_setup[fs_attr]; if (input_index < 0 || input_index >= 16) continue; struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL) *attr = &attr_overrides[input_index]; int slot = vue_map->varying_to_slot[fs_attr]; /* Viewport and Layer are stored in the VUE header. We need to override * them to zero if earlier stages didn't write them, as GL requires that * they read back as zero when not explicitly set. */ switch (fs_attr) { case VARYING_SLOT_VIEWPORT: case VARYING_SLOT_LAYER: attr->ComponentOverrideX = true; attr->ComponentOverrideW = true; attr->ConstantSource = CONST_0000; if (!(vue_map->slots_valid & VARYING_BIT_LAYER)) attr->ComponentOverrideY = true; if (!(vue_map->slots_valid & VARYING_BIT_VIEWPORT)) attr->ComponentOverrideZ = true; continue; case VARYING_SLOT_PRIMITIVE_ID: /* Override if the previous shader stage didn't write gl_PrimitiveID. */ if (slot == -1) { attr->ComponentOverrideX = true; attr->ComponentOverrideY = true; attr->ComponentOverrideZ = true; attr->ComponentOverrideW = true; attr->ConstantSource = PRIM_ID; continue; } default: break; } if (sprite_coord_enables & (1 << input_index)) continue; /* If there was only a back color written but not front, use back * as the color instead of undefined. */ if (slot == -1 && fs_attr == VARYING_SLOT_COL0) slot = vue_map->varying_to_slot[VARYING_SLOT_BFC0]; if (slot == -1 && fs_attr == VARYING_SLOT_COL1) slot = vue_map->varying_to_slot[VARYING_SLOT_BFC1]; /* Not written by the previous stage - undefined. */ if (slot == -1) { attr->ComponentOverrideX = true; attr->ComponentOverrideY = true; attr->ComponentOverrideZ = true; attr->ComponentOverrideW = true; attr->ConstantSource = CONST_0001_FLOAT; continue; } /* Compute the location of the attribute relative to the read offset, * which is counted in 256-bit increments (two 128-bit VUE slots). */ const int source_attr = slot - 2 * urb_read_offset; assert(source_attr >= 0 && source_attr <= 32); attr->SourceAttribute = source_attr; /* If we are doing two-sided color, and the VUE slot following this one * represents a back-facing color, then we need to instruct the SF unit * to do back-facing swizzling. */ if (cso_rast->light_twoside && ((vue_map->slot_to_varying[slot] == VARYING_SLOT_COL0 && vue_map->slot_to_varying[slot+1] == VARYING_SLOT_BFC0) || (vue_map->slot_to_varying[slot] == VARYING_SLOT_COL1 && vue_map->slot_to_varying[slot+1] == VARYING_SLOT_BFC1))) attr->SwizzleSelect = INPUTATTR_FACING; } iris_emit_cmd(batch, GENX(3DSTATE_SBE_SWIZ), sbes) { for (int i = 0; i < 16; i++) sbes.Attribute[i] = attr_overrides[i]; } } static unsigned iris_calculate_point_sprite_overrides(const struct brw_wm_prog_data *prog_data, const struct iris_rasterizer_state *cso) { unsigned overrides = 0; if (prog_data->urb_setup[VARYING_SLOT_PNTC] != -1) overrides |= 1 << prog_data->urb_setup[VARYING_SLOT_PNTC]; for (int i = 0; i < 8; i++) { if ((cso->sprite_coord_enable & (1 << i)) && prog_data->urb_setup[VARYING_SLOT_TEX0 + i] != -1) overrides |= 1 << prog_data->urb_setup[VARYING_SLOT_TEX0 + i]; } return overrides; } static void iris_emit_sbe(struct iris_batch *batch, const struct iris_context *ice) { const struct iris_rasterizer_state *cso_rast = ice->state.cso_rast; const struct brw_wm_prog_data *wm_prog_data = (void *) ice->shaders.prog[MESA_SHADER_FRAGMENT]->prog_data; const struct shader_info *fs_info = iris_get_shader_info(ice, MESA_SHADER_FRAGMENT); unsigned urb_read_offset, urb_read_length; iris_compute_sbe_urb_read_interval(fs_info->inputs_read, ice->shaders.last_vue_map, cso_rast->light_twoside, &urb_read_offset, &urb_read_length); unsigned sprite_coord_overrides = iris_calculate_point_sprite_overrides(wm_prog_data, cso_rast); iris_emit_cmd(batch, GENX(3DSTATE_SBE), sbe) { sbe.AttributeSwizzleEnable = true; sbe.NumberofSFOutputAttributes = wm_prog_data->num_varying_inputs; sbe.PointSpriteTextureCoordinateOrigin = cso_rast->sprite_coord_mode; sbe.VertexURBEntryReadOffset = urb_read_offset; sbe.VertexURBEntryReadLength = urb_read_length; sbe.ForceVertexURBEntryReadOffset = true; sbe.ForceVertexURBEntryReadLength = true; sbe.ConstantInterpolationEnable = wm_prog_data->flat_inputs; sbe.PointSpriteTextureCoordinateEnable = sprite_coord_overrides; #if GEN_GEN >= 9 for (int i = 0; i < 32; i++) { sbe.AttributeActiveComponentFormat[i] = ACTIVE_COMPONENT_XYZW; } #endif } iris_emit_sbe_swiz(batch, ice, urb_read_offset, sprite_coord_overrides); } /* ------------------------------------------------------------------- */ /** * Populate VS program key fields based on the current state. */ static void iris_populate_vs_key(const struct iris_context *ice, const struct shader_info *info, struct brw_vs_prog_key *key) { const struct iris_rasterizer_state *cso_rast = ice->state.cso_rast; if (info->clip_distance_array_size == 0 && (info->outputs_written & (VARYING_BIT_POS | VARYING_BIT_CLIP_VERTEX))) key->nr_userclip_plane_consts = cso_rast->num_clip_plane_consts; } /** * Populate TCS program key fields based on the current state. */ static void iris_populate_tcs_key(const struct iris_context *ice, struct brw_tcs_prog_key *key) { } /** * Populate TES program key fields based on the current state. */ static void iris_populate_tes_key(const struct iris_context *ice, struct brw_tes_prog_key *key) { } /** * Populate GS program key fields based on the current state. */ static void iris_populate_gs_key(const struct iris_context *ice, struct brw_gs_prog_key *key) { } /** * Populate FS program key fields based on the current state. */ static void iris_populate_fs_key(const struct iris_context *ice, struct brw_wm_prog_key *key) { struct iris_screen *screen = (void *) ice->ctx.screen; const struct pipe_framebuffer_state *fb = &ice->state.framebuffer; const struct iris_depth_stencil_alpha_state *zsa = ice->state.cso_zsa; const struct iris_rasterizer_state *rast = ice->state.cso_rast; const struct iris_blend_state *blend = ice->state.cso_blend; key->nr_color_regions = fb->nr_cbufs; key->clamp_fragment_color = rast->clamp_fragment_color; key->alpha_to_coverage = blend->alpha_to_coverage; key->alpha_test_replicate_alpha = fb->nr_cbufs > 1 && zsa->alpha.enabled; /* XXX: only bother if COL0/1 are read */ key->flat_shade = rast->flatshade; key->persample_interp = rast->force_persample_interp; key->multisample_fbo = rast->multisample && fb->samples > 1; key->coherent_fb_fetch = true; key->force_dual_color_blend = screen->driconf.dual_color_blend_by_location && (blend->blend_enables & 1) && blend->dual_color_blending; /* TODO: support key->force_dual_color_blend for Unigine */ /* TODO: Respect glHint for key->high_quality_derivatives */ } static void iris_populate_cs_key(const struct iris_context *ice, struct brw_cs_prog_key *key) { } static uint64_t KSP(const struct iris_compiled_shader *shader) { struct iris_resource *res = (void *) shader->assembly.res; return iris_bo_offset_from_base_address(res->bo) + shader->assembly.offset; } /* Gen11 workaround table #2056 WABTPPrefetchDisable suggests to disable * prefetching of binding tables in A0 and B0 steppings. XXX: Revisit * this WA on C0 stepping. * * TODO: Fill out SamplerCount for prefetching? */ #define INIT_THREAD_DISPATCH_FIELDS(pkt, prefix, stage) \ pkt.KernelStartPointer = KSP(shader); \ pkt.BindingTableEntryCount = GEN_GEN == 11 ? 0 : \ prog_data->binding_table.size_bytes / 4; \ pkt.FloatingPointMode = prog_data->use_alt_mode; \ \ pkt.DispatchGRFStartRegisterForURBData = \ prog_data->dispatch_grf_start_reg; \ pkt.prefix##URBEntryReadLength = vue_prog_data->urb_read_length; \ pkt.prefix##URBEntryReadOffset = 0; \ \ pkt.StatisticsEnable = true; \ pkt.Enable = true; \ \ if (prog_data->total_scratch) { \ struct iris_bo *bo = \ iris_get_scratch_space(ice, prog_data->total_scratch, stage); \ uint32_t scratch_addr = bo->gtt_offset; \ pkt.PerThreadScratchSpace = ffs(prog_data->total_scratch) - 11; \ pkt.ScratchSpaceBasePointer = rw_bo(NULL, scratch_addr); \ } /** * Encode most of 3DSTATE_VS based on the compiled shader. */ static void iris_store_vs_state(struct iris_context *ice, const struct gen_device_info *devinfo, struct iris_compiled_shader *shader) { struct brw_stage_prog_data *prog_data = shader->prog_data; struct brw_vue_prog_data *vue_prog_data = (void *) prog_data; iris_pack_command(GENX(3DSTATE_VS), shader->derived_data, vs) { INIT_THREAD_DISPATCH_FIELDS(vs, Vertex, MESA_SHADER_VERTEX); vs.MaximumNumberofThreads = devinfo->max_vs_threads - 1; vs.SIMD8DispatchEnable = true; vs.UserClipDistanceCullTestEnableBitmask = vue_prog_data->cull_distance_mask; } } /** * Encode most of 3DSTATE_HS based on the compiled shader. */ static void iris_store_tcs_state(struct iris_context *ice, const struct gen_device_info *devinfo, struct iris_compiled_shader *shader) { struct brw_stage_prog_data *prog_data = shader->prog_data; struct brw_vue_prog_data *vue_prog_data = (void *) prog_data; struct brw_tcs_prog_data *tcs_prog_data = (void *) prog_data; iris_pack_command(GENX(3DSTATE_HS), shader->derived_data, hs) { INIT_THREAD_DISPATCH_FIELDS(hs, Vertex, MESA_SHADER_TESS_CTRL); hs.InstanceCount = tcs_prog_data->instances - 1; hs.MaximumNumberofThreads = devinfo->max_tcs_threads - 1; hs.IncludeVertexHandles = true; #if GEN_GEN >= 9 hs.DispatchMode = vue_prog_data->dispatch_mode; hs.IncludePrimitiveID = tcs_prog_data->include_primitive_id; #endif } } /** * Encode 3DSTATE_TE and most of 3DSTATE_DS based on the compiled shader. */ static void iris_store_tes_state(struct iris_context *ice, const struct gen_device_info *devinfo, struct iris_compiled_shader *shader) { struct brw_stage_prog_data *prog_data = shader->prog_data; struct brw_vue_prog_data *vue_prog_data = (void *) prog_data; struct brw_tes_prog_data *tes_prog_data = (void *) prog_data; uint32_t *te_state = (void *) shader->derived_data; uint32_t *ds_state = te_state + GENX(3DSTATE_TE_length); iris_pack_command(GENX(3DSTATE_TE), te_state, te) { te.Partitioning = tes_prog_data->partitioning; te.OutputTopology = tes_prog_data->output_topology; te.TEDomain = tes_prog_data->domain; te.TEEnable = true; te.MaximumTessellationFactorOdd = 63.0; te.MaximumTessellationFactorNotOdd = 64.0; } iris_pack_command(GENX(3DSTATE_DS), ds_state, ds) { INIT_THREAD_DISPATCH_FIELDS(ds, Patch, MESA_SHADER_TESS_EVAL); ds.DispatchMode = DISPATCH_MODE_SIMD8_SINGLE_PATCH; ds.MaximumNumberofThreads = devinfo->max_tes_threads - 1; ds.ComputeWCoordinateEnable = tes_prog_data->domain == BRW_TESS_DOMAIN_TRI; ds.UserClipDistanceCullTestEnableBitmask = vue_prog_data->cull_distance_mask; } } /** * Encode most of 3DSTATE_GS based on the compiled shader. */ static void iris_store_gs_state(struct iris_context *ice, const struct gen_device_info *devinfo, struct iris_compiled_shader *shader) { struct brw_stage_prog_data *prog_data = shader->prog_data; struct brw_vue_prog_data *vue_prog_data = (void *) prog_data; struct brw_gs_prog_data *gs_prog_data = (void *) prog_data; iris_pack_command(GENX(3DSTATE_GS), shader->derived_data, gs) { INIT_THREAD_DISPATCH_FIELDS(gs, Vertex, MESA_SHADER_GEOMETRY); gs.OutputVertexSize = gs_prog_data->output_vertex_size_hwords * 2 - 1; gs.OutputTopology = gs_prog_data->output_topology; gs.ControlDataHeaderSize = gs_prog_data->control_data_header_size_hwords; gs.InstanceControl = gs_prog_data->invocations - 1; gs.DispatchMode = DISPATCH_MODE_SIMD8; gs.IncludePrimitiveID = gs_prog_data->include_primitive_id; gs.ControlDataFormat = gs_prog_data->control_data_format; gs.ReorderMode = TRAILING; gs.ExpectedVertexCount = gs_prog_data->vertices_in; gs.MaximumNumberofThreads = GEN_GEN == 8 ? (devinfo->max_gs_threads / 2 - 1) : (devinfo->max_gs_threads - 1); if (gs_prog_data->static_vertex_count != -1) { gs.StaticOutput = true; gs.StaticOutputVertexCount = gs_prog_data->static_vertex_count; } gs.IncludeVertexHandles = vue_prog_data->include_vue_handles; gs.UserClipDistanceCullTestEnableBitmask = vue_prog_data->cull_distance_mask; const int urb_entry_write_offset = 1; const uint32_t urb_entry_output_length = DIV_ROUND_UP(vue_prog_data->vue_map.num_slots, 2) - urb_entry_write_offset; gs.VertexURBEntryOutputReadOffset = urb_entry_write_offset; gs.VertexURBEntryOutputLength = MAX2(urb_entry_output_length, 1); } } /** * Encode most of 3DSTATE_PS and 3DSTATE_PS_EXTRA based on the shader. */ static void iris_store_fs_state(struct iris_context *ice, const struct gen_device_info *devinfo, struct iris_compiled_shader *shader) { struct brw_stage_prog_data *prog_data = shader->prog_data; struct brw_wm_prog_data *wm_prog_data = (void *) shader->prog_data; uint32_t *ps_state = (void *) shader->derived_data; uint32_t *psx_state = ps_state + GENX(3DSTATE_PS_length); iris_pack_command(GENX(3DSTATE_PS), ps_state, ps) { ps.VectorMaskEnable = true; // XXX: WABTPPrefetchDisable, see above, drop at C0 ps.BindingTableEntryCount = GEN_GEN == 11 ? 0 : prog_data->binding_table.size_bytes / 4; ps.FloatingPointMode = prog_data->use_alt_mode; ps.MaximumNumberofThreadsPerPSD = 64 - (GEN_GEN == 8 ? 2 : 1); ps.PushConstantEnable = prog_data->ubo_ranges[0].length > 0; /* From the documentation for this packet: * "If the PS kernel does not need the Position XY Offsets to * compute a Position Value, then this field should be programmed * to POSOFFSET_NONE." * * "SW Recommendation: If the PS kernel needs the Position Offsets * to compute a Position XY value, this field should match Position * ZW Interpolation Mode to ensure a consistent position.xyzw * computation." * * We only require XY sample offsets. So, this recommendation doesn't * look useful at the moment. We might need this in future. */ ps.PositionXYOffsetSelect = wm_prog_data->uses_pos_offset ? POSOFFSET_SAMPLE : POSOFFSET_NONE; ps._8PixelDispatchEnable = wm_prog_data->dispatch_8; ps._16PixelDispatchEnable = wm_prog_data->dispatch_16; ps._32PixelDispatchEnable = wm_prog_data->dispatch_32; // XXX: Disable SIMD32 with 16x MSAA ps.DispatchGRFStartRegisterForConstantSetupData0 = brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data, ps, 0); ps.DispatchGRFStartRegisterForConstantSetupData1 = brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data, ps, 1); ps.DispatchGRFStartRegisterForConstantSetupData2 = brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data, ps, 2); ps.KernelStartPointer0 = KSP(shader) + brw_wm_prog_data_prog_offset(wm_prog_data, ps, 0); ps.KernelStartPointer1 = KSP(shader) + brw_wm_prog_data_prog_offset(wm_prog_data, ps, 1); ps.KernelStartPointer2 = KSP(shader) + brw_wm_prog_data_prog_offset(wm_prog_data, ps, 2); if (prog_data->total_scratch) { struct iris_bo *bo = iris_get_scratch_space(ice, prog_data->total_scratch, MESA_SHADER_FRAGMENT); uint32_t scratch_addr = bo->gtt_offset; ps.PerThreadScratchSpace = ffs(prog_data->total_scratch) - 11; ps.ScratchSpaceBasePointer = rw_bo(NULL, scratch_addr); } } iris_pack_command(GENX(3DSTATE_PS_EXTRA), psx_state, psx) { psx.PixelShaderValid = true; psx.PixelShaderComputedDepthMode = wm_prog_data->computed_depth_mode; psx.PixelShaderKillsPixel = wm_prog_data->uses_kill; psx.AttributeEnable = wm_prog_data->num_varying_inputs != 0; psx.PixelShaderUsesSourceDepth = wm_prog_data->uses_src_depth; psx.PixelShaderUsesSourceW = wm_prog_data->uses_src_w; psx.PixelShaderIsPerSample = wm_prog_data->persample_dispatch; psx.oMaskPresenttoRenderTarget = wm_prog_data->uses_omask; #if GEN_GEN >= 9 psx.PixelShaderPullsBary = wm_prog_data->pulls_bary; psx.PixelShaderComputesStencil = wm_prog_data->computed_stencil; #else psx.PixelShaderUsesInputCoverageMask = wm_prog_data->uses_sample_mask; #endif // XXX: UAV bit } } /** * Compute the size of the derived data (shader command packets). * * This must match the data written by the iris_store_xs_state() functions. */ static void iris_store_cs_state(struct iris_context *ice, const struct gen_device_info *devinfo, struct iris_compiled_shader *shader) { struct brw_stage_prog_data *prog_data = shader->prog_data; struct brw_cs_prog_data *cs_prog_data = (void *) shader->prog_data; void *map = shader->derived_data; iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA), map, desc) { desc.KernelStartPointer = KSP(shader); desc.ConstantURBEntryReadLength = cs_prog_data->push.per_thread.regs; desc.NumberofThreadsinGPGPUThreadGroup = cs_prog_data->threads; desc.SharedLocalMemorySize = encode_slm_size(GEN_GEN, prog_data->total_shared); desc.BarrierEnable = cs_prog_data->uses_barrier; desc.CrossThreadConstantDataReadLength = cs_prog_data->push.cross_thread.regs; } } static unsigned iris_derived_program_state_size(enum iris_program_cache_id cache_id) { assert(cache_id <= IRIS_CACHE_BLORP); static const unsigned dwords[] = { [IRIS_CACHE_VS] = GENX(3DSTATE_VS_length), [IRIS_CACHE_TCS] = GENX(3DSTATE_HS_length), [IRIS_CACHE_TES] = GENX(3DSTATE_TE_length) + GENX(3DSTATE_DS_length), [IRIS_CACHE_GS] = GENX(3DSTATE_GS_length), [IRIS_CACHE_FS] = GENX(3DSTATE_PS_length) + GENX(3DSTATE_PS_EXTRA_length), [IRIS_CACHE_CS] = GENX(INTERFACE_DESCRIPTOR_DATA_length), [IRIS_CACHE_BLORP] = 0, }; return sizeof(uint32_t) * dwords[cache_id]; } /** * Create any state packets corresponding to the given shader stage * (i.e. 3DSTATE_VS) and save them as "derived data" in the shader variant. * This means that we can look up a program in the in-memory cache and * get most of the state packet without having to reconstruct it. */ static void iris_store_derived_program_state(struct iris_context *ice, enum iris_program_cache_id cache_id, struct iris_compiled_shader *shader) { struct iris_screen *screen = (void *) ice->ctx.screen; const struct gen_device_info *devinfo = &screen->devinfo; switch (cache_id) { case IRIS_CACHE_VS: iris_store_vs_state(ice, devinfo, shader); break; case IRIS_CACHE_TCS: iris_store_tcs_state(ice, devinfo, shader); break; case IRIS_CACHE_TES: iris_store_tes_state(ice, devinfo, shader); break; case IRIS_CACHE_GS: iris_store_gs_state(ice, devinfo, shader); break; case IRIS_CACHE_FS: iris_store_fs_state(ice, devinfo, shader); break; case IRIS_CACHE_CS: iris_store_cs_state(ice, devinfo, shader); case IRIS_CACHE_BLORP: break; default: break; } } /* ------------------------------------------------------------------- */ static const uint32_t push_constant_opcodes[] = { [MESA_SHADER_VERTEX] = 21, [MESA_SHADER_TESS_CTRL] = 25, /* HS */ [MESA_SHADER_TESS_EVAL] = 26, /* DS */ [MESA_SHADER_GEOMETRY] = 22, [MESA_SHADER_FRAGMENT] = 23, [MESA_SHADER_COMPUTE] = 0, }; static uint32_t use_null_surface(struct iris_batch *batch, struct iris_context *ice) { struct iris_bo *state_bo = iris_resource_bo(ice->state.unbound_tex.res); iris_use_pinned_bo(batch, state_bo, false); return ice->state.unbound_tex.offset; } static uint32_t use_null_fb_surface(struct iris_batch *batch, struct iris_context *ice) { /* If set_framebuffer_state() was never called, fall back to 1x1x1 */ if (!ice->state.null_fb.res) return use_null_surface(batch, ice); struct iris_bo *state_bo = iris_resource_bo(ice->state.null_fb.res); iris_use_pinned_bo(batch, state_bo, false); return ice->state.null_fb.offset; } static uint32_t surf_state_offset_for_aux(struct iris_resource *res, unsigned aux_modes, enum isl_aux_usage aux_usage) { return SURFACE_STATE_ALIGNMENT * util_bitcount(res->aux.possible_usages & ((1 << aux_usage) - 1)); } static void surf_state_update_clear_value(struct iris_batch *batch, struct iris_resource *res, struct iris_state_ref *state, unsigned aux_modes, enum isl_aux_usage aux_usage) { struct isl_device *isl_dev = &batch->screen->isl_dev; struct iris_bo *state_bo = iris_resource_bo(state->res); uint64_t real_offset = state->offset + IRIS_MEMZONE_BINDER_START; uint32_t offset_into_bo = real_offset - state_bo->gtt_offset; uint32_t clear_offset = offset_into_bo + isl_dev->ss.clear_value_offset + surf_state_offset_for_aux(res, aux_modes, aux_usage); batch->vtbl->copy_mem_mem(batch, state_bo, clear_offset, res->aux.clear_color_bo, res->aux.clear_color_offset, isl_dev->ss.clear_value_size); } static void update_clear_value(struct iris_context *ice, struct iris_batch *batch, struct iris_resource *res, struct iris_state_ref *state, unsigned aux_modes, struct isl_view *view) { struct iris_screen *screen = batch->screen; const struct gen_device_info *devinfo = &screen->devinfo; /* We only need to update the clear color in the surface state for gen8 and * gen9. Newer gens can read it directly from the clear color state buffer. */ if (devinfo->gen > 9) return; if (devinfo->gen == 9) { /* Skip updating the ISL_AUX_USAGE_NONE surface state */ aux_modes &= ~(1 << ISL_AUX_USAGE_NONE); while (aux_modes) { enum isl_aux_usage aux_usage = u_bit_scan(&aux_modes); surf_state_update_clear_value(batch, res, state, aux_modes, aux_usage); } } else if (devinfo->gen == 8) { pipe_resource_reference(&state->res, NULL); void *map = alloc_surface_states(ice->state.surface_uploader, state, res->aux.possible_usages); while (aux_modes) { enum isl_aux_usage aux_usage = u_bit_scan(&aux_modes); fill_surface_state(&screen->isl_dev, map, res, view, aux_usage); map += SURFACE_STATE_ALIGNMENT; } } } /** * Add a surface to the validation list, as well as the buffer containing * the corresponding SURFACE_STATE. * * Returns the binding table entry (offset to SURFACE_STATE). */ static uint32_t use_surface(struct iris_context *ice, struct iris_batch *batch, struct pipe_surface *p_surf, bool writeable, enum isl_aux_usage aux_usage) { struct iris_surface *surf = (void *) p_surf; struct iris_resource *res = (void *) p_surf->texture; iris_use_pinned_bo(batch, iris_resource_bo(p_surf->texture), writeable); iris_use_pinned_bo(batch, iris_resource_bo(surf->surface_state.res), false); if (res->aux.bo) { iris_use_pinned_bo(batch, res->aux.bo, writeable); if (res->aux.clear_color_bo) iris_use_pinned_bo(batch, res->aux.clear_color_bo, false); if (memcmp(&res->aux.clear_color, &surf->clear_color, sizeof(surf->clear_color)) != 0) { update_clear_value(ice, batch, res, &surf->surface_state, res->aux.possible_usages, &surf->view); surf->clear_color = res->aux.clear_color; } } return surf->surface_state.offset + surf_state_offset_for_aux(res, res->aux.possible_usages, aux_usage); } static uint32_t use_sampler_view(struct iris_context *ice, struct iris_batch *batch, struct iris_sampler_view *isv) { // XXX: ASTC hacks enum isl_aux_usage aux_usage = iris_resource_texture_aux_usage(ice, isv->res, isv->view.format, 0); iris_use_pinned_bo(batch, isv->res->bo, false); iris_use_pinned_bo(batch, iris_resource_bo(isv->surface_state.res), false); if (isv->res->aux.bo) { iris_use_pinned_bo(batch, isv->res->aux.bo, false); if (isv->res->aux.clear_color_bo) iris_use_pinned_bo(batch, isv->res->aux.clear_color_bo, false); if (memcmp(&isv->res->aux.clear_color, &isv->clear_color, sizeof(isv->clear_color)) != 0) { update_clear_value(ice, batch, isv->res, &isv->surface_state, isv->res->aux.sampler_usages, &isv->view); isv->clear_color = isv->res->aux.clear_color; } } return isv->surface_state.offset + surf_state_offset_for_aux(isv->res, isv->res->aux.sampler_usages, aux_usage); } static uint32_t use_ubo_ssbo(struct iris_batch *batch, struct iris_context *ice, struct pipe_shader_buffer *buf, struct iris_state_ref *surf_state, bool writable) { if (!buf->buffer) return use_null_surface(batch, ice); iris_use_pinned_bo(batch, iris_resource_bo(buf->buffer), writable); iris_use_pinned_bo(batch, iris_resource_bo(surf_state->res), false); return surf_state->offset; } static uint32_t use_image(struct iris_batch *batch, struct iris_context *ice, struct iris_shader_state *shs, int i) { struct iris_image_view *iv = &shs->image[i]; struct iris_resource *res = (void *) iv->base.resource; if (!res) return use_null_surface(batch, ice); bool write = iv->base.shader_access & PIPE_IMAGE_ACCESS_WRITE; iris_use_pinned_bo(batch, res->bo, write); iris_use_pinned_bo(batch, iris_resource_bo(iv->surface_state.res), false); if (res->aux.bo) iris_use_pinned_bo(batch, res->aux.bo, write); return iv->surface_state.offset; } #define push_bt_entry(addr) \ assert(addr >= binder_addr); \ assert(s < prog_data->binding_table.size_bytes / sizeof(uint32_t)); \ if (!pin_only) bt_map[s++] = (addr) - binder_addr; #define bt_assert(section, exists) \ if (!pin_only) assert(prog_data->binding_table.section == \ (exists) ? s : 0xd0d0d0d0) /** * Populate the binding table for a given shader stage. * * This fills out the table of pointers to surfaces required by the shader, * and also adds those buffers to the validation list so the kernel can make * resident before running our batch. */ static void iris_populate_binding_table(struct iris_context *ice, struct iris_batch *batch, gl_shader_stage stage, bool pin_only) { const struct iris_binder *binder = &ice->state.binder; struct iris_uncompiled_shader *ish = ice->shaders.uncompiled[stage]; struct iris_compiled_shader *shader = ice->shaders.prog[stage]; if (!shader) return; UNUSED struct brw_stage_prog_data *prog_data = shader->prog_data; struct iris_shader_state *shs = &ice->state.shaders[stage]; uint32_t binder_addr = binder->bo->gtt_offset; //struct brw_stage_prog_data *prog_data = (void *) shader->prog_data; uint32_t *bt_map = binder->map + binder->bt_offset[stage]; int s = 0; const struct shader_info *info = iris_get_shader_info(ice, stage); if (!info) { /* TCS passthrough doesn't need a binding table. */ assert(stage == MESA_SHADER_TESS_CTRL); return; } if (stage == MESA_SHADER_COMPUTE) { /* surface for gl_NumWorkGroups */ struct iris_state_ref *grid_data = &ice->state.grid_size; struct iris_state_ref *grid_state = &ice->state.grid_surf_state; iris_use_pinned_bo(batch, iris_resource_bo(grid_data->res), false); iris_use_pinned_bo(batch, iris_resource_bo(grid_state->res), false); push_bt_entry(grid_state->offset); } if (stage == MESA_SHADER_FRAGMENT) { struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer; /* Note that cso_fb->nr_cbufs == fs_key->nr_color_regions. */ if (cso_fb->nr_cbufs) { for (unsigned i = 0; i < cso_fb->nr_cbufs; i++) { uint32_t addr; if (cso_fb->cbufs[i]) { addr = use_surface(ice, batch, cso_fb->cbufs[i], true, ice->state.draw_aux_usage[i]); } else { addr = use_null_fb_surface(batch, ice); } push_bt_entry(addr); } } else { uint32_t addr = use_null_fb_surface(batch, ice); push_bt_entry(addr); } } unsigned num_textures = util_last_bit(info->textures_used); bt_assert(texture_start, num_textures > 0); for (int i = 0; i < num_textures; i++) { struct iris_sampler_view *view = shs->textures[i]; uint32_t addr = view ? use_sampler_view(ice, batch, view) : use_null_surface(batch, ice); push_bt_entry(addr); } bt_assert(image_start, info->num_images > 0); for (int i = 0; i < info->num_images; i++) { uint32_t addr = use_image(batch, ice, shs, i); push_bt_entry(addr); } bt_assert(ubo_start, shader->num_cbufs > 0); for (int i = 0; i < shader->num_cbufs; i++) { uint32_t addr = use_ubo_ssbo(batch, ice, &shs->constbuf[i], &shs->constbuf_surf_state[i], false); push_bt_entry(addr); } if (ish->const_data) { iris_use_pinned_bo(batch, iris_resource_bo(ish->const_data), false); iris_use_pinned_bo(batch, iris_resource_bo(ish->const_data_state.res), false); uint32_t addr = ish->const_data_state.offset; push_bt_entry(addr); } bt_assert(ssbo_start, info->num_abos + info->num_ssbos > 0); /* XXX: st is wasting 16 binding table slots for ABOs. Should add a cap * for changing nir_lower_atomics_to_ssbos setting and buffer_base offset * in st_atom_storagebuf.c so it'll compact them into one range, with * SSBOs starting at info->num_abos. Ideally it'd reset num_abos to 0 too */ if (info->num_abos + info->num_ssbos > 0) { for (int i = 0; i < IRIS_MAX_ABOS + info->num_ssbos; i++) { uint32_t addr = use_ubo_ssbo(batch, ice, &shs->ssbo[i], &shs->ssbo_surf_state[i], shs->writable_ssbos & (1u << i)); push_bt_entry(addr); } } #if 0 /* XXX: YUV surfaces not implemented yet */ bt_assert(plane_start[1], ...); bt_assert(plane_start[2], ...); #endif } static void iris_use_optional_res(struct iris_batch *batch, struct pipe_resource *res, bool writeable) { if (res) { struct iris_bo *bo = iris_resource_bo(res); iris_use_pinned_bo(batch, bo, writeable); } } static void pin_depth_and_stencil_buffers(struct iris_batch *batch, struct pipe_surface *zsbuf, struct iris_depth_stencil_alpha_state *cso_zsa) { if (!zsbuf) return; struct iris_resource *zres, *sres; iris_get_depth_stencil_resources(zsbuf->texture, &zres, &sres); if (zres) { iris_use_pinned_bo(batch, zres->bo, cso_zsa->depth_writes_enabled); if (zres->aux.bo) { iris_use_pinned_bo(batch, zres->aux.bo, cso_zsa->depth_writes_enabled); } } if (sres) { iris_use_pinned_bo(batch, sres->bo, cso_zsa->stencil_writes_enabled); } } /* ------------------------------------------------------------------- */ /** * Pin any BOs which were installed by a previous batch, and restored * via the hardware logical context mechanism. * * We don't need to re-emit all state every batch - the hardware context * mechanism will save and restore it for us. This includes pointers to * various BOs...which won't exist unless we ask the kernel to pin them * by adding them to the validation list. * * We can skip buffers if we've re-emitted those packets, as we're * overwriting those stale pointers with new ones, and don't actually * refer to the old BOs. */ static void iris_restore_render_saved_bos(struct iris_context *ice, struct iris_batch *batch, const struct pipe_draw_info *draw) { struct iris_genx_state *genx = ice->state.genx; const uint64_t clean = ~ice->state.dirty; if (clean & IRIS_DIRTY_CC_VIEWPORT) { iris_use_optional_res(batch, ice->state.last_res.cc_vp, false); } if (clean & IRIS_DIRTY_SF_CL_VIEWPORT) { iris_use_optional_res(batch, ice->state.last_res.sf_cl_vp, false); } if (clean & IRIS_DIRTY_BLEND_STATE) { iris_use_optional_res(batch, ice->state.last_res.blend, false); } if (clean & IRIS_DIRTY_COLOR_CALC_STATE) { iris_use_optional_res(batch, ice->state.last_res.color_calc, false); } if (clean & IRIS_DIRTY_SCISSOR_RECT) { iris_use_optional_res(batch, ice->state.last_res.scissor, false); } if (ice->state.streamout_active && (clean & IRIS_DIRTY_SO_BUFFERS)) { for (int i = 0; i < 4; i++) { struct iris_stream_output_target *tgt = (void *) ice->state.so_target[i]; if (tgt) { iris_use_pinned_bo(batch, iris_resource_bo(tgt->base.buffer), true); iris_use_pinned_bo(batch, iris_resource_bo(tgt->offset.res), true); } } } for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) { if (!(clean & (IRIS_DIRTY_CONSTANTS_VS << stage))) continue; struct iris_shader_state *shs = &ice->state.shaders[stage]; struct iris_compiled_shader *shader = ice->shaders.prog[stage]; if (!shader) continue; struct brw_stage_prog_data *prog_data = (void *) shader->prog_data; for (int i = 0; i < 4; i++) { const struct brw_ubo_range *range = &prog_data->ubo_ranges[i]; if (range->length == 0) continue; struct pipe_shader_buffer *cbuf = &shs->constbuf[range->block]; struct iris_resource *res = (void *) cbuf->buffer; if (res) iris_use_pinned_bo(batch, res->bo, false); else iris_use_pinned_bo(batch, batch->screen->workaround_bo, false); } } for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) { if (clean & (IRIS_DIRTY_BINDINGS_VS << stage)) { /* Re-pin any buffers referred to by the binding table. */ iris_populate_binding_table(ice, batch, stage, true); } } for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) { struct iris_shader_state *shs = &ice->state.shaders[stage]; struct pipe_resource *res = shs->sampler_table.res; if (res) iris_use_pinned_bo(batch, iris_resource_bo(res), false); } for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) { if (clean & (IRIS_DIRTY_VS << stage)) { struct iris_compiled_shader *shader = ice->shaders.prog[stage]; if (shader) { struct iris_bo *bo = iris_resource_bo(shader->assembly.res); iris_use_pinned_bo(batch, bo, false); struct brw_stage_prog_data *prog_data = shader->prog_data; if (prog_data->total_scratch > 0) { struct iris_bo *bo = iris_get_scratch_space(ice, prog_data->total_scratch, stage); iris_use_pinned_bo(batch, bo, true); } } } } if ((clean & IRIS_DIRTY_DEPTH_BUFFER) && (clean & IRIS_DIRTY_WM_DEPTH_STENCIL)) { struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer; pin_depth_and_stencil_buffers(batch, cso_fb->zsbuf, ice->state.cso_zsa); } if (draw->index_size == 0 && ice->state.last_res.index_buffer) { /* This draw didn't emit a new index buffer, so we are inheriting the * older index buffer. This draw didn't need it, but future ones may. */ struct iris_bo *bo = iris_resource_bo(ice->state.last_res.index_buffer); iris_use_pinned_bo(batch, bo, false); } if (clean & IRIS_DIRTY_VERTEX_BUFFERS) { uint64_t bound = ice->state.bound_vertex_buffers; while (bound) { const int i = u_bit_scan64(&bound); struct pipe_resource *res = genx->vertex_buffers[i].resource; iris_use_pinned_bo(batch, iris_resource_bo(res), false); } } } static void iris_restore_compute_saved_bos(struct iris_context *ice, struct iris_batch *batch, const struct pipe_grid_info *grid) { const uint64_t clean = ~ice->state.dirty; const int stage = MESA_SHADER_COMPUTE; struct iris_shader_state *shs = &ice->state.shaders[stage]; if (clean & IRIS_DIRTY_CONSTANTS_CS) { struct iris_compiled_shader *shader = ice->shaders.prog[stage]; if (shader) { struct brw_stage_prog_data *prog_data = (void *) shader->prog_data; const struct brw_ubo_range *range = &prog_data->ubo_ranges[0]; if (range->length > 0) { struct pipe_shader_buffer *cbuf = &shs->constbuf[range->block]; struct iris_resource *res = (void *) cbuf->buffer; if (res) iris_use_pinned_bo(batch, res->bo, false); else iris_use_pinned_bo(batch, batch->screen->workaround_bo, false); } } } if (clean & IRIS_DIRTY_BINDINGS_CS) { /* Re-pin any buffers referred to by the binding table. */ iris_populate_binding_table(ice, batch, stage, true); } struct pipe_resource *sampler_res = shs->sampler_table.res; if (sampler_res) iris_use_pinned_bo(batch, iris_resource_bo(sampler_res), false); if (clean & IRIS_DIRTY_CS) { struct iris_compiled_shader *shader = ice->shaders.prog[stage]; if (shader) { struct iris_bo *bo = iris_resource_bo(shader->assembly.res); iris_use_pinned_bo(batch, bo, false); struct brw_stage_prog_data *prog_data = shader->prog_data; if (prog_data->total_scratch > 0) { struct iris_bo *bo = iris_get_scratch_space(ice, prog_data->total_scratch, stage); iris_use_pinned_bo(batch, bo, true); } } } } /** * Possibly emit STATE_BASE_ADDRESS to update Surface State Base Address. */ static void iris_update_surface_base_address(struct iris_batch *batch, struct iris_binder *binder) { if (batch->last_surface_base_address == binder->bo->gtt_offset) return; flush_for_state_base_change(batch); iris_emit_cmd(batch, GENX(STATE_BASE_ADDRESS), sba) { sba.SurfaceStateMOCS = MOCS_WB; sba.SurfaceStateBaseAddressModifyEnable = true; sba.SurfaceStateBaseAddress = ro_bo(binder->bo, 0); } batch->last_surface_base_address = binder->bo->gtt_offset; } static void iris_upload_dirty_render_state(struct iris_context *ice, struct iris_batch *batch, const struct pipe_draw_info *draw) { const uint64_t dirty = ice->state.dirty; if (!(dirty & IRIS_ALL_DIRTY_FOR_RENDER)) return; struct iris_genx_state *genx = ice->state.genx; struct iris_binder *binder = &ice->state.binder; struct brw_wm_prog_data *wm_prog_data = (void *) ice->shaders.prog[MESA_SHADER_FRAGMENT]->prog_data; if (dirty & IRIS_DIRTY_CC_VIEWPORT) { const struct iris_rasterizer_state *cso_rast = ice->state.cso_rast; uint32_t cc_vp_address; /* XXX: could avoid streaming for depth_clip [0,1] case. */ uint32_t *cc_vp_map = stream_state(batch, ice->state.dynamic_uploader, &ice->state.last_res.cc_vp, 4 * ice->state.num_viewports * GENX(CC_VIEWPORT_length), 32, &cc_vp_address); for (int i = 0; i < ice->state.num_viewports; i++) { float zmin, zmax; util_viewport_zmin_zmax(&ice->state.viewports[i], cso_rast->clip_halfz, &zmin, &zmax); if (cso_rast->depth_clip_near) zmin = 0.0; if (cso_rast->depth_clip_far) zmax = 1.0; iris_pack_state(GENX(CC_VIEWPORT), cc_vp_map, ccv) { ccv.MinimumDepth = zmin; ccv.MaximumDepth = zmax; } cc_vp_map += GENX(CC_VIEWPORT_length); } iris_emit_cmd(batch, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_CC), ptr) { ptr.CCViewportPointer = cc_vp_address; } } if (dirty & IRIS_DIRTY_SF_CL_VIEWPORT) { struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer; uint32_t sf_cl_vp_address; uint32_t *vp_map = stream_state(batch, ice->state.dynamic_uploader, &ice->state.last_res.sf_cl_vp, 4 * ice->state.num_viewports * GENX(SF_CLIP_VIEWPORT_length), 64, &sf_cl_vp_address); for (unsigned i = 0; i < ice->state.num_viewports; i++) { const struct pipe_viewport_state *state = &ice->state.viewports[i]; float gb_xmin, gb_xmax, gb_ymin, gb_ymax; float vp_xmin = viewport_extent(state, 0, -1.0f); float vp_xmax = viewport_extent(state, 0, 1.0f); float vp_ymin = viewport_extent(state, 1, -1.0f); float vp_ymax = viewport_extent(state, 1, 1.0f); calculate_guardband_size(cso_fb->width, cso_fb->height, state->scale[0], state->scale[1], state->translate[0], state->translate[1], &gb_xmin, &gb_xmax, &gb_ymin, &gb_ymax); iris_pack_state(GENX(SF_CLIP_VIEWPORT), vp_map, vp) { vp.ViewportMatrixElementm00 = state->scale[0]; vp.ViewportMatrixElementm11 = state->scale[1]; vp.ViewportMatrixElementm22 = state->scale[2]; vp.ViewportMatrixElementm30 = state->translate[0]; vp.ViewportMatrixElementm31 = state->translate[1]; vp.ViewportMatrixElementm32 = state->translate[2]; vp.XMinClipGuardband = gb_xmin; vp.XMaxClipGuardband = gb_xmax; vp.YMinClipGuardband = gb_ymin; vp.YMaxClipGuardband = gb_ymax; vp.XMinViewPort = MAX2(vp_xmin, 0); vp.XMaxViewPort = MIN2(vp_xmax, cso_fb->width) - 1; vp.YMinViewPort = MAX2(vp_ymin, 0); vp.YMaxViewPort = MIN2(vp_ymax, cso_fb->height) - 1; } vp_map += GENX(SF_CLIP_VIEWPORT_length); } iris_emit_cmd(batch, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP), ptr) { ptr.SFClipViewportPointer = sf_cl_vp_address; } } if (dirty & IRIS_DIRTY_URB) { unsigned size[4]; for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) { if (!ice->shaders.prog[i]) { size[i] = 1; } else { struct brw_vue_prog_data *vue_prog_data = (void *) ice->shaders.prog[i]->prog_data; size[i] = vue_prog_data->urb_entry_size; } assert(size[i] != 0); } genX(emit_urb_setup)(ice, batch, size, ice->shaders.prog[MESA_SHADER_TESS_EVAL] != NULL, ice->shaders.prog[MESA_SHADER_GEOMETRY] != NULL); } if (dirty & IRIS_DIRTY_BLEND_STATE) { struct iris_blend_state *cso_blend = ice->state.cso_blend; struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer; struct iris_depth_stencil_alpha_state *cso_zsa = ice->state.cso_zsa; const int header_dwords = GENX(BLEND_STATE_length); /* Always write at least one BLEND_STATE - the final RT message will * reference BLEND_STATE[0] even if there aren't color writes. There * may still be alpha testing, computed depth, and so on. */ const int rt_dwords = MAX2(cso_fb->nr_cbufs, 1) * GENX(BLEND_STATE_ENTRY_length); uint32_t blend_offset; uint32_t *blend_map = stream_state(batch, ice->state.dynamic_uploader, &ice->state.last_res.blend, 4 * (header_dwords + rt_dwords), 64, &blend_offset); uint32_t blend_state_header; iris_pack_state(GENX(BLEND_STATE), &blend_state_header, bs) { bs.AlphaTestEnable = cso_zsa->alpha.enabled; bs.AlphaTestFunction = translate_compare_func(cso_zsa->alpha.func); } blend_map[0] = blend_state_header | cso_blend->blend_state[0]; memcpy(&blend_map[1], &cso_blend->blend_state[1], 4 * rt_dwords); iris_emit_cmd(batch, GENX(3DSTATE_BLEND_STATE_POINTERS), ptr) { ptr.BlendStatePointer = blend_offset; ptr.BlendStatePointerValid = true; } } if (dirty & IRIS_DIRTY_COLOR_CALC_STATE) { struct iris_depth_stencil_alpha_state *cso = ice->state.cso_zsa; #if GEN_GEN == 8 struct pipe_stencil_ref *p_stencil_refs = &ice->state.stencil_ref; #endif uint32_t cc_offset; void *cc_map = stream_state(batch, ice->state.dynamic_uploader, &ice->state.last_res.color_calc, sizeof(uint32_t) * GENX(COLOR_CALC_STATE_length), 64, &cc_offset); iris_pack_state(GENX(COLOR_CALC_STATE), cc_map, cc) { cc.AlphaTestFormat = ALPHATEST_FLOAT32; cc.AlphaReferenceValueAsFLOAT32 = cso->alpha.ref_value; cc.BlendConstantColorRed = ice->state.blend_color.color[0]; cc.BlendConstantColorGreen = ice->state.blend_color.color[1]; cc.BlendConstantColorBlue = ice->state.blend_color.color[2]; cc.BlendConstantColorAlpha = ice->state.blend_color.color[3]; #if GEN_GEN == 8 cc.StencilReferenceValue = p_stencil_refs->ref_value[0]; cc.BackfaceStencilReferenceValue = p_stencil_refs->ref_value[1]; #endif } iris_emit_cmd(batch, GENX(3DSTATE_CC_STATE_POINTERS), ptr) { ptr.ColorCalcStatePointer = cc_offset; ptr.ColorCalcStatePointerValid = true; } } for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) { if (!(dirty & (IRIS_DIRTY_CONSTANTS_VS << stage))) continue; struct iris_shader_state *shs = &ice->state.shaders[stage]; struct iris_compiled_shader *shader = ice->shaders.prog[stage]; if (!shader) continue; if (shs->cbuf0_needs_upload) upload_uniforms(ice, stage); struct brw_stage_prog_data *prog_data = (void *) shader->prog_data; iris_emit_cmd(batch, GENX(3DSTATE_CONSTANT_VS), pkt) { pkt._3DCommandSubOpcode = push_constant_opcodes[stage]; if (prog_data) { /* The Skylake PRM contains the following restriction: * * "The driver must ensure The following case does not occur * without a flush to the 3D engine: 3DSTATE_CONSTANT_* with * buffer 3 read length equal to zero committed followed by a * 3DSTATE_CONSTANT_* with buffer 0 read length not equal to * zero committed." * * To avoid this, we program the buffers in the highest slots. * This way, slot 0 is only used if slot 3 is also used. */ int n = 3; for (int i = 3; i >= 0; i--) { const struct brw_ubo_range *range = &prog_data->ubo_ranges[i]; if (range->length == 0) continue; struct pipe_shader_buffer *cbuf = &shs->constbuf[range->block]; struct iris_resource *res = (void *) cbuf->buffer; assert(cbuf->buffer_offset % 32 == 0); pkt.ConstantBody.ReadLength[n] = range->length; pkt.ConstantBody.Buffer[n] = res ? ro_bo(res->bo, range->start * 32 + cbuf->buffer_offset) : ro_bo(batch->screen->workaround_bo, 0); n--; } } } } for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) { if (dirty & (IRIS_DIRTY_BINDINGS_VS << stage)) { iris_emit_cmd(batch, GENX(3DSTATE_BINDING_TABLE_POINTERS_VS), ptr) { ptr._3DCommandSubOpcode = 38 + stage; ptr.PointertoVSBindingTable = binder->bt_offset[stage]; } } } for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) { if (dirty & (IRIS_DIRTY_BINDINGS_VS << stage)) { iris_populate_binding_table(ice, batch, stage, false); } } for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) { if (!(dirty & (IRIS_DIRTY_SAMPLER_STATES_VS << stage)) || !ice->shaders.prog[stage]) continue; iris_upload_sampler_states(ice, stage); struct iris_shader_state *shs = &ice->state.shaders[stage]; struct pipe_resource *res = shs->sampler_table.res; if (res) iris_use_pinned_bo(batch, iris_resource_bo(res), false); iris_emit_cmd(batch, GENX(3DSTATE_SAMPLER_STATE_POINTERS_VS), ptr) { ptr._3DCommandSubOpcode = 43 + stage; ptr.PointertoVSSamplerState = shs->sampler_table.offset; } } if (ice->state.need_border_colors) iris_use_pinned_bo(batch, ice->state.border_color_pool.bo, false); if (dirty & IRIS_DIRTY_MULTISAMPLE) { iris_emit_cmd(batch, GENX(3DSTATE_MULTISAMPLE), ms) { ms.PixelLocation = ice->state.cso_rast->half_pixel_center ? CENTER : UL_CORNER; if (ice->state.framebuffer.samples > 0) ms.NumberofMultisamples = ffs(ice->state.framebuffer.samples) - 1; } } if (dirty & IRIS_DIRTY_SAMPLE_MASK) { iris_emit_cmd(batch, GENX(3DSTATE_SAMPLE_MASK), ms) { ms.SampleMask = ice->state.sample_mask; } } for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) { if (!(dirty & (IRIS_DIRTY_VS << stage))) continue; struct iris_compiled_shader *shader = ice->shaders.prog[stage]; if (shader) { struct brw_stage_prog_data *prog_data = shader->prog_data; struct iris_resource *cache = (void *) shader->assembly.res; iris_use_pinned_bo(batch, cache->bo, false); if (prog_data->total_scratch > 0) { struct iris_bo *bo = iris_get_scratch_space(ice, prog_data->total_scratch, stage); iris_use_pinned_bo(batch, bo, true); } #if GEN_GEN >= 9 if (stage == MESA_SHADER_FRAGMENT && wm_prog_data->uses_sample_mask) { uint32_t psx_state[GENX(3DSTATE_PS_EXTRA_length)] = {0}; uint32_t *shader_psx = ((uint32_t*)shader->derived_data) + GENX(3DSTATE_PS_length); struct iris_rasterizer_state *cso = ice->state.cso_rast; iris_pack_command(GENX(3DSTATE_PS_EXTRA), &psx_state, psx) { if (wm_prog_data->post_depth_coverage) psx.InputCoverageMaskState = ICMS_DEPTH_COVERAGE; else if (wm_prog_data->inner_coverage && cso->conservative_rasterization) psx.InputCoverageMaskState = ICMS_INNER_CONSERVATIVE; else psx.InputCoverageMaskState = ICMS_NORMAL; } iris_batch_emit(batch, shader->derived_data, sizeof(uint32_t) * GENX(3DSTATE_PS_length)); iris_emit_merge(batch, shader_psx, psx_state, GENX(3DSTATE_PS_EXTRA_length)); } else #endif iris_batch_emit(batch, shader->derived_data, iris_derived_program_state_size(stage)); } else { if (stage == MESA_SHADER_TESS_EVAL) { iris_emit_cmd(batch, GENX(3DSTATE_HS), hs); iris_emit_cmd(batch, GENX(3DSTATE_TE), te); iris_emit_cmd(batch, GENX(3DSTATE_DS), ds); } else if (stage == MESA_SHADER_GEOMETRY) { iris_emit_cmd(batch, GENX(3DSTATE_GS), gs); } } } if (ice->state.streamout_active) { if (dirty & IRIS_DIRTY_SO_BUFFERS) { iris_batch_emit(batch, genx->so_buffers, 4 * 4 * GENX(3DSTATE_SO_BUFFER_length)); for (int i = 0; i < 4; i++) { struct iris_stream_output_target *tgt = (void *) ice->state.so_target[i]; if (tgt) { tgt->zeroed = true; iris_use_pinned_bo(batch, iris_resource_bo(tgt->base.buffer), true); iris_use_pinned_bo(batch, iris_resource_bo(tgt->offset.res), true); } } } if ((dirty & IRIS_DIRTY_SO_DECL_LIST) && ice->state.streamout) { uint32_t *decl_list = ice->state.streamout + GENX(3DSTATE_STREAMOUT_length); iris_batch_emit(batch, decl_list, 4 * ((decl_list[0] & 0xff) + 2)); } if (dirty & IRIS_DIRTY_STREAMOUT) { const struct iris_rasterizer_state *cso_rast = ice->state.cso_rast; uint32_t dynamic_sol[GENX(3DSTATE_STREAMOUT_length)]; iris_pack_command(GENX(3DSTATE_STREAMOUT), dynamic_sol, sol) { sol.SOFunctionEnable = true; sol.SOStatisticsEnable = true; sol.RenderingDisable = cso_rast->rasterizer_discard && !ice->state.prims_generated_query_active; sol.ReorderMode = cso_rast->flatshade_first ? LEADING : TRAILING; } assert(ice->state.streamout); iris_emit_merge(batch, ice->state.streamout, dynamic_sol, GENX(3DSTATE_STREAMOUT_length)); } } else { if (dirty & IRIS_DIRTY_STREAMOUT) { iris_emit_cmd(batch, GENX(3DSTATE_STREAMOUT), sol); } } if (dirty & IRIS_DIRTY_CLIP) { struct iris_rasterizer_state *cso_rast = ice->state.cso_rast; struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer; bool gs_or_tes = ice->shaders.prog[MESA_SHADER_GEOMETRY] || ice->shaders.prog[MESA_SHADER_TESS_EVAL]; bool points_or_lines = cso_rast->fill_mode_point_or_line || (gs_or_tes ? ice->shaders.output_topology_is_points_or_lines : ice->state.prim_is_points_or_lines); uint32_t dynamic_clip[GENX(3DSTATE_CLIP_length)]; iris_pack_command(GENX(3DSTATE_CLIP), &dynamic_clip, cl) { cl.StatisticsEnable = ice->state.statistics_counters_enabled; cl.ClipMode = cso_rast->rasterizer_discard ? CLIPMODE_REJECT_ALL : CLIPMODE_NORMAL; cl.ViewportXYClipTestEnable = !points_or_lines; if (wm_prog_data->barycentric_interp_modes & BRW_BARYCENTRIC_NONPERSPECTIVE_BITS) cl.NonPerspectiveBarycentricEnable = true; cl.ForceZeroRTAIndexEnable = cso_fb->layers == 0; cl.MaximumVPIndex = ice->state.num_viewports - 1; } iris_emit_merge(batch, cso_rast->clip, dynamic_clip, ARRAY_SIZE(cso_rast->clip)); } if (dirty & IRIS_DIRTY_RASTER) { struct iris_rasterizer_state *cso = ice->state.cso_rast; iris_batch_emit(batch, cso->raster, sizeof(cso->raster)); iris_batch_emit(batch, cso->sf, sizeof(cso->sf)); } if (dirty & IRIS_DIRTY_WM) { struct iris_rasterizer_state *cso = ice->state.cso_rast; uint32_t dynamic_wm[GENX(3DSTATE_WM_length)]; iris_pack_command(GENX(3DSTATE_WM), &dynamic_wm, wm) { wm.StatisticsEnable = ice->state.statistics_counters_enabled; wm.BarycentricInterpolationMode = wm_prog_data->barycentric_interp_modes; if (wm_prog_data->early_fragment_tests) wm.EarlyDepthStencilControl = EDSC_PREPS; else if (wm_prog_data->has_side_effects) wm.EarlyDepthStencilControl = EDSC_PSEXEC; /* We could skip this bit if color writes are enabled. */ if (wm_prog_data->has_side_effects || wm_prog_data->uses_kill) wm.ForceThreadDispatchEnable = ForceON; } iris_emit_merge(batch, cso->wm, dynamic_wm, ARRAY_SIZE(cso->wm)); } if (dirty & IRIS_DIRTY_SBE) { iris_emit_sbe(batch, ice); } if (dirty & IRIS_DIRTY_PS_BLEND) { struct iris_blend_state *cso_blend = ice->state.cso_blend; struct iris_depth_stencil_alpha_state *cso_zsa = ice->state.cso_zsa; const struct shader_info *fs_info = iris_get_shader_info(ice, MESA_SHADER_FRAGMENT); uint32_t dynamic_pb[GENX(3DSTATE_PS_BLEND_length)]; iris_pack_command(GENX(3DSTATE_PS_BLEND), &dynamic_pb, pb) { pb.HasWriteableRT = has_writeable_rt(cso_blend, fs_info); pb.AlphaTestEnable = cso_zsa->alpha.enabled; /* The dual source blending docs caution against using SRC1 factors * when the shader doesn't use a dual source render target write. * Empirically, this can lead to GPU hangs, and the results are * undefined anyway, so simply disable blending to avoid the hang. */ pb.ColorBufferBlendEnable = (cso_blend->blend_enables & 1) && (!cso_blend->dual_color_blending || wm_prog_data->dual_src_blend); } iris_emit_merge(batch, cso_blend->ps_blend, dynamic_pb, ARRAY_SIZE(cso_blend->ps_blend)); } if (dirty & IRIS_DIRTY_WM_DEPTH_STENCIL) { struct iris_depth_stencil_alpha_state *cso = ice->state.cso_zsa; #if GEN_GEN >= 9 struct pipe_stencil_ref *p_stencil_refs = &ice->state.stencil_ref; uint32_t stencil_refs[GENX(3DSTATE_WM_DEPTH_STENCIL_length)]; iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL), &stencil_refs, wmds) { wmds.StencilReferenceValue = p_stencil_refs->ref_value[0]; wmds.BackfaceStencilReferenceValue = p_stencil_refs->ref_value[1]; } iris_emit_merge(batch, cso->wmds, stencil_refs, ARRAY_SIZE(cso->wmds)); #else iris_batch_emit(batch, cso->wmds, sizeof(cso->wmds)); #endif } if (dirty & IRIS_DIRTY_SCISSOR_RECT) { uint32_t scissor_offset = emit_state(batch, ice->state.dynamic_uploader, &ice->state.last_res.scissor, ice->state.scissors, sizeof(struct pipe_scissor_state) * ice->state.num_viewports, 32); iris_emit_cmd(batch, GENX(3DSTATE_SCISSOR_STATE_POINTERS), ptr) { ptr.ScissorRectPointer = scissor_offset; } } if (dirty & IRIS_DIRTY_DEPTH_BUFFER) { struct iris_depth_buffer_state *cso_z = &ice->state.genx->depth_buffer; /* Do not emit the clear params yets. We need to update the clear value * first. */ uint32_t clear_length = GENX(3DSTATE_CLEAR_PARAMS_length) * 4; uint32_t cso_z_size = sizeof(cso_z->packets) - clear_length; iris_batch_emit(batch, cso_z->packets, cso_z_size); union isl_color_value clear_value = { .f32 = { 0, } }; struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer; if (cso_fb->zsbuf) { struct iris_resource *zres, *sres; iris_get_depth_stencil_resources(cso_fb->zsbuf->texture, &zres, &sres); if (zres && zres->aux.bo) clear_value = iris_resource_get_clear_color(zres, NULL, NULL); } uint32_t clear_params[GENX(3DSTATE_CLEAR_PARAMS_length)]; iris_pack_command(GENX(3DSTATE_CLEAR_PARAMS), clear_params, clear) { clear.DepthClearValueValid = true; clear.DepthClearValue = clear_value.f32[0]; } iris_batch_emit(batch, clear_params, clear_length); } if (dirty & (IRIS_DIRTY_DEPTH_BUFFER | IRIS_DIRTY_WM_DEPTH_STENCIL)) { /* Listen for buffer changes, and also write enable changes. */ struct pipe_framebuffer_state *cso_fb = &ice->state.framebuffer; pin_depth_and_stencil_buffers(batch, cso_fb->zsbuf, ice->state.cso_zsa); } if (dirty & IRIS_DIRTY_POLYGON_STIPPLE) { iris_emit_cmd(batch, GENX(3DSTATE_POLY_STIPPLE_PATTERN), poly) { for (int i = 0; i < 32; i++) { poly.PatternRow[i] = ice->state.poly_stipple.stipple[i]; } } } if (dirty & IRIS_DIRTY_LINE_STIPPLE) { struct iris_rasterizer_state *cso = ice->state.cso_rast; iris_batch_emit(batch, cso->line_stipple, sizeof(cso->line_stipple)); } if (dirty & IRIS_DIRTY_VF_TOPOLOGY) { iris_emit_cmd(batch, GENX(3DSTATE_VF_TOPOLOGY), topo) { topo.PrimitiveTopologyType = translate_prim_type(draw->mode, draw->vertices_per_patch); } } if (dirty & IRIS_DIRTY_VERTEX_BUFFERS) { int count = util_bitcount64(ice->state.bound_vertex_buffers); int dynamic_bound = ice->state.bound_vertex_buffers; if (ice->state.vs_uses_draw_params) { if (ice->draw.draw_params_offset == 0) { u_upload_data(ice->state.dynamic_uploader, 0, sizeof(ice->draw.params), 4, &ice->draw.params, &ice->draw.draw_params_offset, &ice->draw.draw_params_res); } assert(ice->draw.draw_params_res); struct iris_vertex_buffer_state *state = &(ice->state.genx->vertex_buffers[count]); pipe_resource_reference(&state->resource, ice->draw.draw_params_res); struct iris_resource *res = (void *) state->resource; iris_pack_state(GENX(VERTEX_BUFFER_STATE), state->state, vb) { vb.VertexBufferIndex = count; vb.AddressModifyEnable = true; vb.BufferPitch = 0; vb.BufferSize = res->bo->size - ice->draw.draw_params_offset; vb.BufferStartingAddress = ro_bo(NULL, res->bo->gtt_offset + (int) ice->draw.draw_params_offset); vb.MOCS = mocs(res->bo); } dynamic_bound |= 1ull << count; count++; } if (ice->state.vs_uses_derived_draw_params) { u_upload_data(ice->state.dynamic_uploader, 0, sizeof(ice->draw.derived_params), 4, &ice->draw.derived_params, &ice->draw.derived_draw_params_offset, &ice->draw.derived_draw_params_res); struct iris_vertex_buffer_state *state = &(ice->state.genx->vertex_buffers[count]); pipe_resource_reference(&state->resource, ice->draw.derived_draw_params_res); struct iris_resource *res = (void *) ice->draw.derived_draw_params_res; iris_pack_state(GENX(VERTEX_BUFFER_STATE), state->state, vb) { vb.VertexBufferIndex = count; vb.AddressModifyEnable = true; vb.BufferPitch = 0; vb.BufferSize = res->bo->size - ice->draw.derived_draw_params_offset; vb.BufferStartingAddress = ro_bo(NULL, res->bo->gtt_offset + (int) ice->draw.derived_draw_params_offset); vb.MOCS = mocs(res->bo); } dynamic_bound |= 1ull << count; count++; } if (count) { /* The VF cache designers cut corners, and made the cache key's * tuple only consider the bottom * 32 bits of the address. If you have two vertex buffers which get * placed exactly 4 GiB apart and use them in back-to-back draw calls, * you can get collisions (even within a single batch). * * So, we need to do a VF cache invalidate if the buffer for a VB * slot slot changes [48:32] address bits from the previous time. */ unsigned flush_flags = 0; uint64_t bound = dynamic_bound; while (bound) { const int i = u_bit_scan64(&bound); uint16_t high_bits = 0; struct iris_resource *res = (void *) genx->vertex_buffers[i].resource; if (res) { iris_use_pinned_bo(batch, res->bo, false); high_bits = res->bo->gtt_offset >> 32ull; if (high_bits != ice->state.last_vbo_high_bits[i]) { flush_flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE | PIPE_CONTROL_CS_STALL; ice->state.last_vbo_high_bits[i] = high_bits; } } } if (flush_flags) iris_emit_pipe_control_flush(batch, flush_flags); const unsigned vb_dwords = GENX(VERTEX_BUFFER_STATE_length); uint32_t *map = iris_get_command_space(batch, 4 * (1 + vb_dwords * count)); _iris_pack_command(batch, GENX(3DSTATE_VERTEX_BUFFERS), map, vb) { vb.DWordLength = (vb_dwords * count + 1) - 2; } map += 1; bound = dynamic_bound; while (bound) { const int i = u_bit_scan64(&bound); memcpy(map, genx->vertex_buffers[i].state, sizeof(uint32_t) * vb_dwords); map += vb_dwords; } } } if (dirty & IRIS_DIRTY_VERTEX_ELEMENTS) { struct iris_vertex_element_state *cso = ice->state.cso_vertex_elements; const unsigned entries = MAX2(cso->count, 1); if (!(ice->state.vs_needs_sgvs_element || ice->state.vs_uses_derived_draw_params || ice->state.vs_needs_edge_flag)) { iris_batch_emit(batch, cso->vertex_elements, sizeof(uint32_t) * (1 + entries * GENX(VERTEX_ELEMENT_STATE_length))); } else { uint32_t dynamic_ves[1 + 33 * GENX(VERTEX_ELEMENT_STATE_length)]; const unsigned dyn_count = cso->count + ice->state.vs_needs_sgvs_element + ice->state.vs_uses_derived_draw_params; iris_pack_command(GENX(3DSTATE_VERTEX_ELEMENTS), &dynamic_ves, ve) { ve.DWordLength = 1 + GENX(VERTEX_ELEMENT_STATE_length) * dyn_count - 2; } memcpy(&dynamic_ves[1], &cso->vertex_elements[1], (cso->count - ice->state.vs_needs_edge_flag) * GENX(VERTEX_ELEMENT_STATE_length) * sizeof(uint32_t)); uint32_t *ve_pack_dest = &dynamic_ves[1 + (cso->count - ice->state.vs_needs_edge_flag) * GENX(VERTEX_ELEMENT_STATE_length)]; if (ice->state.vs_needs_sgvs_element) { uint32_t base_ctrl = ice->state.vs_uses_draw_params ? VFCOMP_STORE_SRC : VFCOMP_STORE_0; iris_pack_state(GENX(VERTEX_ELEMENT_STATE), ve_pack_dest, ve) { ve.Valid = true; ve.VertexBufferIndex = util_bitcount64(ice->state.bound_vertex_buffers); ve.SourceElementFormat = ISL_FORMAT_R32G32_UINT; ve.Component0Control = base_ctrl; ve.Component1Control = base_ctrl; ve.Component2Control = VFCOMP_STORE_0; ve.Component3Control = VFCOMP_STORE_0; } ve_pack_dest += GENX(VERTEX_ELEMENT_STATE_length); } if (ice->state.vs_uses_derived_draw_params) { iris_pack_state(GENX(VERTEX_ELEMENT_STATE), ve_pack_dest, ve) { ve.Valid = true; ve.VertexBufferIndex = util_bitcount64(ice->state.bound_vertex_buffers) + ice->state.vs_uses_draw_params; ve.SourceElementFormat = ISL_FORMAT_R32G32_UINT; ve.Component0Control = VFCOMP_STORE_SRC; ve.Component1Control = VFCOMP_STORE_SRC; ve.Component2Control = VFCOMP_STORE_0; ve.Component3Control = VFCOMP_STORE_0; } ve_pack_dest += GENX(VERTEX_ELEMENT_STATE_length); } if (ice->state.vs_needs_edge_flag) { for (int i = 0; i < GENX(VERTEX_ELEMENT_STATE_length); i++) ve_pack_dest[i] = cso->edgeflag_ve[i]; } iris_batch_emit(batch, &dynamic_ves, sizeof(uint32_t) * (1 + dyn_count * GENX(VERTEX_ELEMENT_STATE_length))); } if (!ice->state.vs_needs_edge_flag) { iris_batch_emit(batch, cso->vf_instancing, sizeof(uint32_t) * entries * GENX(3DSTATE_VF_INSTANCING_length)); } else { assert(cso->count > 0); const unsigned edgeflag_index = cso->count - 1; uint32_t dynamic_vfi[33 * GENX(3DSTATE_VF_INSTANCING_length)]; memcpy(&dynamic_vfi[0], cso->vf_instancing, edgeflag_index * GENX(3DSTATE_VF_INSTANCING_length) * sizeof(uint32_t)); uint32_t *vfi_pack_dest = &dynamic_vfi[0] + edgeflag_index * GENX(3DSTATE_VF_INSTANCING_length); iris_pack_command(GENX(3DSTATE_VF_INSTANCING), vfi_pack_dest, vi) { vi.VertexElementIndex = edgeflag_index + ice->state.vs_needs_sgvs_element + ice->state.vs_uses_derived_draw_params; } for (int i = 0; i < GENX(3DSTATE_VF_INSTANCING_length); i++) vfi_pack_dest[i] |= cso->edgeflag_vfi[i]; iris_batch_emit(batch, &dynamic_vfi[0], sizeof(uint32_t) * entries * GENX(3DSTATE_VF_INSTANCING_length)); } } if (dirty & IRIS_DIRTY_VF_SGVS) { const struct brw_vs_prog_data *vs_prog_data = (void *) ice->shaders.prog[MESA_SHADER_VERTEX]->prog_data; struct iris_vertex_element_state *cso = ice->state.cso_vertex_elements; iris_emit_cmd(batch, GENX(3DSTATE_VF_SGVS), sgv) { if (vs_prog_data->uses_vertexid) { sgv.VertexIDEnable = true; sgv.VertexIDComponentNumber = 2; sgv.VertexIDElementOffset = cso->count - ice->state.vs_needs_edge_flag; } if (vs_prog_data->uses_instanceid) { sgv.InstanceIDEnable = true; sgv.InstanceIDComponentNumber = 3; sgv.InstanceIDElementOffset = cso->count - ice->state.vs_needs_edge_flag; } } } if (dirty & IRIS_DIRTY_VF) { iris_emit_cmd(batch, GENX(3DSTATE_VF), vf) { if (draw->primitive_restart) { vf.IndexedDrawCutIndexEnable = true; vf.CutIndex = draw->restart_index; } } } if (dirty & IRIS_DIRTY_VF_STATISTICS) { iris_emit_cmd(batch, GENX(3DSTATE_VF_STATISTICS), vf) { vf.StatisticsEnable = true; } } /* TODO: Gen8 PMA fix */ } static void iris_upload_render_state(struct iris_context *ice, struct iris_batch *batch, const struct pipe_draw_info *draw) { bool use_predicate = ice->state.predicate == IRIS_PREDICATE_STATE_USE_BIT; /* Always pin the binder. If we're emitting new binding table pointers, * we need it. If not, we're probably inheriting old tables via the * context, and need it anyway. Since true zero-bindings cases are * practically non-existent, just pin it and avoid last_res tracking. */ iris_use_pinned_bo(batch, ice->state.binder.bo, false); if (!batch->contains_draw) { iris_restore_render_saved_bos(ice, batch, draw); batch->contains_draw = true; } iris_upload_dirty_render_state(ice, batch, draw); if (draw->index_size > 0) { unsigned offset; if (draw->has_user_indices) { u_upload_data(ice->ctx.stream_uploader, 0, draw->count * draw->index_size, 4, draw->index.user, &offset, &ice->state.last_res.index_buffer); } else { struct iris_resource *res = (void *) draw->index.resource; res->bind_history |= PIPE_BIND_INDEX_BUFFER; pipe_resource_reference(&ice->state.last_res.index_buffer, draw->index.resource); offset = 0; } struct iris_bo *bo = iris_resource_bo(ice->state.last_res.index_buffer); iris_emit_cmd(batch, GENX(3DSTATE_INDEX_BUFFER), ib) { ib.IndexFormat = draw->index_size >> 1; ib.MOCS = mocs(bo); ib.BufferSize = bo->size - offset; ib.BufferStartingAddress = ro_bo(bo, offset); } /* The VF cache key only uses 32-bits, see vertex buffer comment above */ uint16_t high_bits = bo->gtt_offset >> 32ull; if (high_bits != ice->state.last_index_bo_high_bits) { iris_emit_pipe_control_flush(batch, PIPE_CONTROL_VF_CACHE_INVALIDATE | PIPE_CONTROL_CS_STALL); ice->state.last_index_bo_high_bits = high_bits; } } #define _3DPRIM_END_OFFSET 0x2420 #define _3DPRIM_START_VERTEX 0x2430 #define _3DPRIM_VERTEX_COUNT 0x2434 #define _3DPRIM_INSTANCE_COUNT 0x2438 #define _3DPRIM_START_INSTANCE 0x243C #define _3DPRIM_BASE_VERTEX 0x2440 if (draw->indirect) { if (draw->indirect->indirect_draw_count) { use_predicate = true; struct iris_bo *draw_count_bo = iris_resource_bo(draw->indirect->indirect_draw_count); unsigned draw_count_offset = draw->indirect->indirect_draw_count_offset; iris_emit_pipe_control_flush(batch, PIPE_CONTROL_FLUSH_ENABLE); if (ice->state.predicate == IRIS_PREDICATE_STATE_USE_BIT) { static const uint32_t math[] = { MI_MATH | (9 - 2), /* Compute (draw index < draw count). * We do this by subtracting and storing the carry bit. */ MI_ALU2(LOAD, SRCA, R0), MI_ALU2(LOAD, SRCB, R1), MI_ALU0(SUB), MI_ALU2(STORE, R3, CF), /* Compute (subtracting result & MI_PREDICATE). */ MI_ALU2(LOAD, SRCA, R3), MI_ALU2(LOAD, SRCB, R2), MI_ALU0(AND), MI_ALU2(STORE, R3, ACCU), }; /* Upload the current draw count from the draw parameters * buffer to GPR1. */ ice->vtbl.load_register_mem32(batch, CS_GPR(1), draw_count_bo, draw_count_offset); /* Zero the top 32-bits of GPR1. */ ice->vtbl.load_register_imm32(batch, CS_GPR(1) + 4, 0); /* Upload the id of the current primitive to GPR0. */ ice->vtbl.load_register_imm64(batch, CS_GPR(0), draw->drawid); iris_batch_emit(batch, math, sizeof(math)); /* Store result of MI_MATH computations to MI_PREDICATE_RESULT. */ ice->vtbl.load_register_reg64(batch, MI_PREDICATE_RESULT, CS_GPR(3)); } else { uint32_t mi_predicate; /* Upload the id of the current primitive to MI_PREDICATE_SRC1. */ ice->vtbl.load_register_imm64(batch, MI_PREDICATE_SRC1, draw->drawid); /* Upload the current draw count from the draw parameters buffer * to MI_PREDICATE_SRC0. */ ice->vtbl.load_register_mem32(batch, MI_PREDICATE_SRC0, draw_count_bo, draw_count_offset); /* Zero the top 32-bits of MI_PREDICATE_SRC0 */ ice->vtbl.load_register_imm32(batch, MI_PREDICATE_SRC0 + 4, 0); if (draw->drawid == 0) { mi_predicate = MI_PREDICATE | MI_PREDICATE_LOADOP_LOADINV | MI_PREDICATE_COMBINEOP_SET | MI_PREDICATE_COMPAREOP_SRCS_EQUAL; } else { /* While draw_index < draw_count the predicate's result will be * (draw_index == draw_count) ^ TRUE = TRUE * When draw_index == draw_count the result is * (TRUE) ^ TRUE = FALSE * After this all results will be: * (FALSE) ^ FALSE = FALSE */ mi_predicate = MI_PREDICATE | MI_PREDICATE_LOADOP_LOAD | MI_PREDICATE_COMBINEOP_XOR | MI_PREDICATE_COMPAREOP_SRCS_EQUAL; } iris_batch_emit(batch, &mi_predicate, sizeof(uint32_t)); } } struct iris_bo *bo = iris_resource_bo(draw->indirect->buffer); assert(bo); iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = _3DPRIM_VERTEX_COUNT; lrm.MemoryAddress = ro_bo(bo, draw->indirect->offset + 0); } iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = _3DPRIM_INSTANCE_COUNT; lrm.MemoryAddress = ro_bo(bo, draw->indirect->offset + 4); } iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = _3DPRIM_START_VERTEX; lrm.MemoryAddress = ro_bo(bo, draw->indirect->offset + 8); } if (draw->index_size) { iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = _3DPRIM_BASE_VERTEX; lrm.MemoryAddress = ro_bo(bo, draw->indirect->offset + 12); } iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = _3DPRIM_START_INSTANCE; lrm.MemoryAddress = ro_bo(bo, draw->indirect->offset + 16); } } else { iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = _3DPRIM_START_INSTANCE; lrm.MemoryAddress = ro_bo(bo, draw->indirect->offset + 12); } iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_IMM), lri) { lri.RegisterOffset = _3DPRIM_BASE_VERTEX; lri.DataDWord = 0; } } } else if (draw->count_from_stream_output) { struct iris_stream_output_target *so = (void *) draw->count_from_stream_output; /* XXX: Replace with actual cache tracking */ iris_emit_pipe_control_flush(batch, PIPE_CONTROL_CS_STALL); iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = CS_GPR(0); lrm.MemoryAddress = ro_bo(iris_resource_bo(so->offset.res), so->offset.offset); } if (so->base.buffer_offset) iris_math_add32_gpr0(ice, batch, -so->base.buffer_offset); iris_math_div32_gpr0(ice, batch, so->stride); _iris_emit_lrr(batch, _3DPRIM_VERTEX_COUNT, CS_GPR(0)); _iris_emit_lri(batch, _3DPRIM_START_VERTEX, 0); _iris_emit_lri(batch, _3DPRIM_BASE_VERTEX, 0); _iris_emit_lri(batch, _3DPRIM_START_INSTANCE, 0); _iris_emit_lri(batch, _3DPRIM_INSTANCE_COUNT, draw->instance_count); } iris_emit_cmd(batch, GENX(3DPRIMITIVE), prim) { prim.VertexAccessType = draw->index_size > 0 ? RANDOM : SEQUENTIAL; prim.PredicateEnable = use_predicate; if (draw->indirect || draw->count_from_stream_output) { prim.IndirectParameterEnable = true; } else { prim.StartInstanceLocation = draw->start_instance; prim.InstanceCount = draw->instance_count; prim.VertexCountPerInstance = draw->count; // XXX: this is probably bonkers. prim.StartVertexLocation = draw->start; if (draw->index_size) { prim.BaseVertexLocation += draw->index_bias; } else { prim.StartVertexLocation += draw->index_bias; } //prim.BaseVertexLocation = ...; } } } static void iris_upload_compute_state(struct iris_context *ice, struct iris_batch *batch, const struct pipe_grid_info *grid) { const uint64_t dirty = ice->state.dirty; struct iris_screen *screen = batch->screen; const struct gen_device_info *devinfo = &screen->devinfo; struct iris_binder *binder = &ice->state.binder; struct iris_shader_state *shs = &ice->state.shaders[MESA_SHADER_COMPUTE]; struct iris_compiled_shader *shader = ice->shaders.prog[MESA_SHADER_COMPUTE]; struct brw_stage_prog_data *prog_data = shader->prog_data; struct brw_cs_prog_data *cs_prog_data = (void *) prog_data; /* Always pin the binder. If we're emitting new binding table pointers, * we need it. If not, we're probably inheriting old tables via the * context, and need it anyway. Since true zero-bindings cases are * practically non-existent, just pin it and avoid last_res tracking. */ iris_use_pinned_bo(batch, ice->state.binder.bo, false); if ((dirty & IRIS_DIRTY_CONSTANTS_CS) && shs->cbuf0_needs_upload) upload_uniforms(ice, MESA_SHADER_COMPUTE); if (dirty & IRIS_DIRTY_BINDINGS_CS) iris_populate_binding_table(ice, batch, MESA_SHADER_COMPUTE, false); if (dirty & IRIS_DIRTY_SAMPLER_STATES_CS) iris_upload_sampler_states(ice, MESA_SHADER_COMPUTE); iris_use_optional_res(batch, shs->sampler_table.res, false); iris_use_pinned_bo(batch, iris_resource_bo(shader->assembly.res), false); if (ice->state.need_border_colors) iris_use_pinned_bo(batch, ice->state.border_color_pool.bo, false); if (dirty & IRIS_DIRTY_CS) { /* The MEDIA_VFE_STATE documentation for Gen8+ says: * * "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless * the only bits that are changed are scoreboard related: Scoreboard * Enable, Scoreboard Type, Scoreboard Mask, Scoreboard Delta. For * these scoreboard related states, a MEDIA_STATE_FLUSH is * sufficient." */ iris_emit_pipe_control_flush(batch, PIPE_CONTROL_CS_STALL); iris_emit_cmd(batch, GENX(MEDIA_VFE_STATE), vfe) { if (prog_data->total_scratch) { struct iris_bo *bo = iris_get_scratch_space(ice, prog_data->total_scratch, MESA_SHADER_COMPUTE); vfe.PerThreadScratchSpace = ffs(prog_data->total_scratch) - 11; vfe.ScratchSpaceBasePointer = rw_bo(bo, 0); } vfe.MaximumNumberofThreads = devinfo->max_cs_threads * screen->subslice_total - 1; #if GEN_GEN < 11 vfe.ResetGatewayTimer = Resettingrelativetimerandlatchingtheglobaltimestamp; #endif #if GEN_GEN == 8 vfe.BypassGatewayControl = true; #endif vfe.NumberofURBEntries = 2; vfe.URBEntryAllocationSize = 2; vfe.CURBEAllocationSize = ALIGN(cs_prog_data->push.per_thread.regs * cs_prog_data->threads + cs_prog_data->push.cross_thread.regs, 2); } } /* TODO: Combine subgroup-id with cbuf0 so we can push regular uniforms */ uint32_t curbe_data_offset = 0; assert(cs_prog_data->push.cross_thread.dwords == 0 && cs_prog_data->push.per_thread.dwords == 1 && cs_prog_data->base.param[0] == BRW_PARAM_BUILTIN_SUBGROUP_ID); struct pipe_resource *curbe_data_res = NULL; uint32_t *curbe_data_map = stream_state(batch, ice->state.dynamic_uploader, &curbe_data_res, ALIGN(cs_prog_data->push.total.size, 64), 64, &curbe_data_offset); assert(curbe_data_map); memset(curbe_data_map, 0x5a, ALIGN(cs_prog_data->push.total.size, 64)); iris_fill_cs_push_const_buffer(cs_prog_data, curbe_data_map); if (dirty & IRIS_DIRTY_CONSTANTS_CS) { iris_emit_cmd(batch, GENX(MEDIA_CURBE_LOAD), curbe) { curbe.CURBETotalDataLength = ALIGN(cs_prog_data->push.total.size, 64); curbe.CURBEDataStartAddress = curbe_data_offset; } } if (dirty & (IRIS_DIRTY_SAMPLER_STATES_CS | IRIS_DIRTY_BINDINGS_CS | IRIS_DIRTY_CONSTANTS_CS | IRIS_DIRTY_CS)) { struct pipe_resource *desc_res = NULL; uint32_t desc[GENX(INTERFACE_DESCRIPTOR_DATA_length)]; iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA), desc, idd) { idd.SamplerStatePointer = shs->sampler_table.offset; idd.BindingTablePointer = binder->bt_offset[MESA_SHADER_COMPUTE]; } for (int i = 0; i < GENX(INTERFACE_DESCRIPTOR_DATA_length); i++) desc[i] |= ((uint32_t *) shader->derived_data)[i]; iris_emit_cmd(batch, GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD), load) { load.InterfaceDescriptorTotalLength = GENX(INTERFACE_DESCRIPTOR_DATA_length) * sizeof(uint32_t); load.InterfaceDescriptorDataStartAddress = emit_state(batch, ice->state.dynamic_uploader, &desc_res, desc, sizeof(desc), 32); } pipe_resource_reference(&desc_res, NULL); } uint32_t group_size = grid->block[0] * grid->block[1] * grid->block[2]; uint32_t remainder = group_size & (cs_prog_data->simd_size - 1); uint32_t right_mask; if (remainder > 0) right_mask = ~0u >> (32 - remainder); else right_mask = ~0u >> (32 - cs_prog_data->simd_size); #define GPGPU_DISPATCHDIMX 0x2500 #define GPGPU_DISPATCHDIMY 0x2504 #define GPGPU_DISPATCHDIMZ 0x2508 if (grid->indirect) { struct iris_state_ref *grid_size = &ice->state.grid_size; struct iris_bo *bo = iris_resource_bo(grid_size->res); iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = GPGPU_DISPATCHDIMX; lrm.MemoryAddress = ro_bo(bo, grid_size->offset + 0); } iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = GPGPU_DISPATCHDIMY; lrm.MemoryAddress = ro_bo(bo, grid_size->offset + 4); } iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = GPGPU_DISPATCHDIMZ; lrm.MemoryAddress = ro_bo(bo, grid_size->offset + 8); } } iris_emit_cmd(batch, GENX(GPGPU_WALKER), ggw) { ggw.IndirectParameterEnable = grid->indirect != NULL; ggw.SIMDSize = cs_prog_data->simd_size / 16; ggw.ThreadDepthCounterMaximum = 0; ggw.ThreadHeightCounterMaximum = 0; ggw.ThreadWidthCounterMaximum = cs_prog_data->threads - 1; ggw.ThreadGroupIDXDimension = grid->grid[0]; ggw.ThreadGroupIDYDimension = grid->grid[1]; ggw.ThreadGroupIDZDimension = grid->grid[2]; ggw.RightExecutionMask = right_mask; ggw.BottomExecutionMask = 0xffffffff; } iris_emit_cmd(batch, GENX(MEDIA_STATE_FLUSH), msf); if (!batch->contains_draw) { iris_restore_compute_saved_bos(ice, batch, grid); batch->contains_draw = true; } } /** * State module teardown. */ static void iris_destroy_state(struct iris_context *ice) { struct iris_genx_state *genx = ice->state.genx; uint64_t bound_vbs = ice->state.bound_vertex_buffers; while (bound_vbs) { const int i = u_bit_scan64(&bound_vbs); pipe_resource_reference(&genx->vertex_buffers[i].resource, NULL); } free(ice->state.genx); for (unsigned i = 0; i < ice->state.framebuffer.nr_cbufs; i++) { pipe_surface_reference(&ice->state.framebuffer.cbufs[i], NULL); } pipe_surface_reference(&ice->state.framebuffer.zsbuf, NULL); for (int stage = 0; stage < MESA_SHADER_STAGES; stage++) { struct iris_shader_state *shs = &ice->state.shaders[stage]; pipe_resource_reference(&shs->sampler_table.res, NULL); for (int i = 0; i < PIPE_MAX_CONSTANT_BUFFERS; i++) { pipe_resource_reference(&shs->constbuf[i].buffer, NULL); pipe_resource_reference(&shs->constbuf_surf_state[i].res, NULL); } for (int i = 0; i < PIPE_MAX_SHADER_IMAGES; i++) { pipe_resource_reference(&shs->image[i].base.resource, NULL); pipe_resource_reference(&shs->image[i].surface_state.res, NULL); } for (int i = 0; i < PIPE_MAX_SHADER_BUFFERS; i++) { pipe_resource_reference(&shs->ssbo[i].buffer, NULL); pipe_resource_reference(&shs->ssbo_surf_state[i].res, NULL); } for (int i = 0; i < IRIS_MAX_TEXTURE_SAMPLERS; i++) { pipe_sampler_view_reference((struct pipe_sampler_view **) &shs->textures[i], NULL); } } pipe_resource_reference(&ice->state.grid_size.res, NULL); pipe_resource_reference(&ice->state.grid_surf_state.res, NULL); pipe_resource_reference(&ice->state.null_fb.res, NULL); pipe_resource_reference(&ice->state.unbound_tex.res, NULL); pipe_resource_reference(&ice->state.last_res.cc_vp, NULL); pipe_resource_reference(&ice->state.last_res.sf_cl_vp, NULL); pipe_resource_reference(&ice->state.last_res.color_calc, NULL); pipe_resource_reference(&ice->state.last_res.scissor, NULL); pipe_resource_reference(&ice->state.last_res.blend, NULL); pipe_resource_reference(&ice->state.last_res.index_buffer, NULL); } /* ------------------------------------------------------------------- */ static void iris_rebind_buffer(struct iris_context *ice, struct iris_resource *res, uint64_t old_address) { struct pipe_context *ctx = &ice->ctx; struct iris_screen *screen = (void *) ctx->screen; struct iris_genx_state *genx = ice->state.genx; assert(res->base.target == PIPE_BUFFER); /* Buffers can't be framebuffer attachments, nor display related, * and we don't have upstream Clover support. */ assert(!(res->bind_history & (PIPE_BIND_DEPTH_STENCIL | PIPE_BIND_RENDER_TARGET | PIPE_BIND_BLENDABLE | PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_CURSOR | PIPE_BIND_COMPUTE_RESOURCE | PIPE_BIND_GLOBAL))); if (res->bind_history & PIPE_BIND_VERTEX_BUFFER) { uint64_t bound_vbs = ice->state.bound_vertex_buffers; while (bound_vbs) { const int i = u_bit_scan64(&bound_vbs); struct iris_vertex_buffer_state *state = &genx->vertex_buffers[i]; /* Update the CPU struct */ STATIC_ASSERT(GENX(VERTEX_BUFFER_STATE_BufferStartingAddress_start) == 32); STATIC_ASSERT(GENX(VERTEX_BUFFER_STATE_BufferStartingAddress_bits) == 64); uint64_t *addr = (uint64_t *) &state->state[1]; if (*addr == old_address) { *addr = res->bo->gtt_offset; ice->state.dirty |= IRIS_DIRTY_VERTEX_BUFFERS; } } } /* No need to handle these: * - PIPE_BIND_INDEX_BUFFER (emitted for every indexed draw) * - PIPE_BIND_COMMAND_ARGS_BUFFER (emitted for every indirect draw) * - PIPE_BIND_QUERY_BUFFER (no persistent state references) */ if (res->bind_history & PIPE_BIND_STREAM_OUTPUT) { /* XXX: be careful about resetting vs appending... */ assert(false); } for (int s = MESA_SHADER_VERTEX; s < MESA_SHADER_STAGES; s++) { struct iris_shader_state *shs = &ice->state.shaders[s]; enum pipe_shader_type p_stage = stage_to_pipe(s); if (res->bind_history & PIPE_BIND_CONSTANT_BUFFER) { /* Skip constant buffer 0, it's for regular uniforms, not UBOs */ uint32_t bound_cbufs = shs->bound_cbufs & ~1u; while (bound_cbufs) { const int i = u_bit_scan(&bound_cbufs); struct pipe_shader_buffer *cbuf = &shs->constbuf[i]; struct iris_state_ref *surf_state = &shs->constbuf_surf_state[i]; if (res->bo == iris_resource_bo(cbuf->buffer)) { iris_upload_ubo_ssbo_surf_state(ice, cbuf, surf_state, false); ice->state.dirty |= IRIS_DIRTY_CONSTANTS_VS << s; } } } if (res->bind_history & PIPE_BIND_SHADER_BUFFER) { uint32_t bound_ssbos = shs->bound_ssbos; while (bound_ssbos) { const int i = u_bit_scan(&bound_ssbos); struct pipe_shader_buffer *ssbo = &shs->ssbo[i]; if (res->bo == iris_resource_bo(ssbo->buffer)) { struct pipe_shader_buffer buf = { .buffer = &res->base, .buffer_offset = ssbo->buffer_offset, .buffer_size = ssbo->buffer_size, }; iris_set_shader_buffers(ctx, p_stage, i, 1, &buf, (shs->writable_ssbos >> i) & 1); } } } if (res->bind_history & PIPE_BIND_SAMPLER_VIEW) { uint32_t bound_sampler_views = shs->bound_sampler_views; while (bound_sampler_views) { const int i = u_bit_scan(&bound_sampler_views); struct iris_sampler_view *isv = shs->textures[i]; if (res->bo == iris_resource_bo(isv->base.texture)) { void *map = alloc_surface_states(ice->state.surface_uploader, &isv->surface_state, isv->res->aux.sampler_usages); assert(map); fill_buffer_surface_state(&screen->isl_dev, isv->res, map, isv->view.format, isv->view.swizzle, isv->base.u.buf.offset, isv->base.u.buf.size); ice->state.dirty |= IRIS_DIRTY_BINDINGS_VS << s; } } } if (res->bind_history & PIPE_BIND_SHADER_IMAGE) { uint32_t bound_image_views = shs->bound_image_views; while (bound_image_views) { const int i = u_bit_scan(&bound_image_views); struct iris_image_view *iv = &shs->image[i]; if (res->bo == iris_resource_bo(iv->base.resource)) { iris_set_shader_images(ctx, p_stage, i, 1, &iv->base); } } } } } /* ------------------------------------------------------------------- */ static void iris_load_register_reg32(struct iris_batch *batch, uint32_t dst, uint32_t src) { _iris_emit_lrr(batch, dst, src); } static void iris_load_register_reg64(struct iris_batch *batch, uint32_t dst, uint32_t src) { _iris_emit_lrr(batch, dst, src); _iris_emit_lrr(batch, dst + 4, src + 4); } static void iris_load_register_imm32(struct iris_batch *batch, uint32_t reg, uint32_t val) { _iris_emit_lri(batch, reg, val); } static void iris_load_register_imm64(struct iris_batch *batch, uint32_t reg, uint64_t val) { _iris_emit_lri(batch, reg + 0, val & 0xffffffff); _iris_emit_lri(batch, reg + 4, val >> 32); } /** * Emit MI_LOAD_REGISTER_MEM to load a 32-bit MMIO register from a buffer. */ static void iris_load_register_mem32(struct iris_batch *batch, uint32_t reg, struct iris_bo *bo, uint32_t offset) { iris_emit_cmd(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) { lrm.RegisterAddress = reg; lrm.MemoryAddress = ro_bo(bo, offset); } } /** * Load a 64-bit value from a buffer into a MMIO register via * two MI_LOAD_REGISTER_MEM commands. */ static void iris_load_register_mem64(struct iris_batch *batch, uint32_t reg, struct iris_bo *bo, uint32_t offset) { iris_load_register_mem32(batch, reg + 0, bo, offset + 0); iris_load_register_mem32(batch, reg + 4, bo, offset + 4); } static void iris_store_register_mem32(struct iris_batch *batch, uint32_t reg, struct iris_bo *bo, uint32_t offset, bool predicated) { iris_emit_cmd(batch, GENX(MI_STORE_REGISTER_MEM), srm) { srm.RegisterAddress = reg; srm.MemoryAddress = rw_bo(bo, offset); srm.PredicateEnable = predicated; } } static void iris_store_register_mem64(struct iris_batch *batch, uint32_t reg, struct iris_bo *bo, uint32_t offset, bool predicated) { iris_store_register_mem32(batch, reg + 0, bo, offset + 0, predicated); iris_store_register_mem32(batch, reg + 4, bo, offset + 4, predicated); } static void iris_store_data_imm32(struct iris_batch *batch, struct iris_bo *bo, uint32_t offset, uint32_t imm) { iris_emit_cmd(batch, GENX(MI_STORE_DATA_IMM), sdi) { sdi.Address = rw_bo(bo, offset); sdi.ImmediateData = imm; } } static void iris_store_data_imm64(struct iris_batch *batch, struct iris_bo *bo, uint32_t offset, uint64_t imm) { /* Can't use iris_emit_cmd because MI_STORE_DATA_IMM has a length of * 2 in genxml but it's actually variable length and we need 5 DWords. */ void *map = iris_get_command_space(batch, 4 * 5); _iris_pack_command(batch, GENX(MI_STORE_DATA_IMM), map, sdi) { sdi.DWordLength = 5 - 2; sdi.Address = rw_bo(bo, offset); sdi.ImmediateData = imm; } } static void iris_copy_mem_mem(struct iris_batch *batch, struct iris_bo *dst_bo, uint32_t dst_offset, struct iris_bo *src_bo, uint32_t src_offset, unsigned bytes) { /* MI_COPY_MEM_MEM operates on DWords. */ assert(bytes % 4 == 0); assert(dst_offset % 4 == 0); assert(src_offset % 4 == 0); for (unsigned i = 0; i < bytes; i += 4) { iris_emit_cmd(batch, GENX(MI_COPY_MEM_MEM), cp) { cp.DestinationMemoryAddress = rw_bo(dst_bo, dst_offset + i); cp.SourceMemoryAddress = ro_bo(src_bo, src_offset + i); } } } /* ------------------------------------------------------------------- */ static unsigned flags_to_post_sync_op(uint32_t flags) { if (flags & PIPE_CONTROL_WRITE_IMMEDIATE) return WriteImmediateData; if (flags & PIPE_CONTROL_WRITE_DEPTH_COUNT) return WritePSDepthCount; if (flags & PIPE_CONTROL_WRITE_TIMESTAMP) return WriteTimestamp; return 0; } /** * Do the given flags have a Post Sync or LRI Post Sync operation? */ static enum pipe_control_flags get_post_sync_flags(enum pipe_control_flags flags) { flags &= PIPE_CONTROL_WRITE_IMMEDIATE | PIPE_CONTROL_WRITE_DEPTH_COUNT | PIPE_CONTROL_WRITE_TIMESTAMP | PIPE_CONTROL_LRI_POST_SYNC_OP; /* Only one "Post Sync Op" is allowed, and it's mutually exclusive with * "LRI Post Sync Operation". So more than one bit set would be illegal. */ assert(util_bitcount(flags) <= 1); return flags; } #define IS_COMPUTE_PIPELINE(batch) (batch->name == IRIS_BATCH_COMPUTE) /** * Emit a series of PIPE_CONTROL commands, taking into account any * workarounds necessary to actually accomplish the caller's request. * * Unless otherwise noted, spec quotations in this function come from: * * Synchronization of the 3D Pipeline > PIPE_CONTROL Command > Programming * Restrictions for PIPE_CONTROL. * * You should not use this function directly. Use the helpers in * iris_pipe_control.c instead, which may split the pipe control further. */ static void iris_emit_raw_pipe_control(struct iris_batch *batch, uint32_t flags, struct iris_bo *bo, uint32_t offset, uint64_t imm) { UNUSED const struct gen_device_info *devinfo = &batch->screen->devinfo; enum pipe_control_flags post_sync_flags = get_post_sync_flags(flags); enum pipe_control_flags non_lri_post_sync_flags = post_sync_flags & ~PIPE_CONTROL_LRI_POST_SYNC_OP; /* Recursive PIPE_CONTROL workarounds -------------------------------- * (http://knowyourmeme.com/memes/xzibit-yo-dawg) * * We do these first because we want to look at the original operation, * rather than any workarounds we set. */ if (GEN_GEN == 9 && (flags & PIPE_CONTROL_VF_CACHE_INVALIDATE)) { /* The PIPE_CONTROL "VF Cache Invalidation Enable" bit description * lists several workarounds: * * "Project: SKL, KBL, BXT * * If the VF Cache Invalidation Enable is set to a 1 in a * PIPE_CONTROL, a separate Null PIPE_CONTROL, all bitfields * sets to 0, with the VF Cache Invalidation Enable set to 0 * needs to be sent prior to the PIPE_CONTROL with VF Cache * Invalidation Enable set to a 1." */ iris_emit_raw_pipe_control(batch, 0, NULL, 0, 0); } if (GEN_GEN == 9 && IS_COMPUTE_PIPELINE(batch) && post_sync_flags) { /* Project: SKL / Argument: LRI Post Sync Operation [23] * * "PIPECONTROL command with “Command Streamer Stall Enable” must be * programmed prior to programming a PIPECONTROL command with "LRI * Post Sync Operation" in GPGPU mode of operation (i.e when * PIPELINE_SELECT command is set to GPGPU mode of operation)." * * The same text exists a few rows below for Post Sync Op. */ iris_emit_raw_pipe_control(batch, PIPE_CONTROL_CS_STALL, bo, offset, imm); } if (GEN_GEN == 10 && (flags & PIPE_CONTROL_RENDER_TARGET_FLUSH)) { /* Cannonlake: * "Before sending a PIPE_CONTROL command with bit 12 set, SW must issue * another PIPE_CONTROL with Render Target Cache Flush Enable (bit 12) * = 0 and Pipe Control Flush Enable (bit 7) = 1" */ iris_emit_raw_pipe_control(batch, PIPE_CONTROL_FLUSH_ENABLE, bo, offset, imm); } /* "Flush Types" workarounds --------------------------------------------- * We do these now because they may add post-sync operations or CS stalls. */ if (GEN_GEN < 11 && flags & PIPE_CONTROL_VF_CACHE_INVALIDATE) { /* Project: BDW, SKL+ (stopping at CNL) / Argument: VF Invalidate * * "'Post Sync Operation' must be enabled to 'Write Immediate Data' or * 'Write PS Depth Count' or 'Write Timestamp'." */ if (!bo) { flags |= PIPE_CONTROL_WRITE_IMMEDIATE; post_sync_flags |= PIPE_CONTROL_WRITE_IMMEDIATE; non_lri_post_sync_flags |= PIPE_CONTROL_WRITE_IMMEDIATE; bo = batch->screen->workaround_bo; } } /* #1130 from Gen10 workarounds page: * * "Enable Depth Stall on every Post Sync Op if Render target Cache * Flush is not enabled in same PIPE CONTROL and Enable Pixel score * board stall if Render target cache flush is enabled." * * Applicable to CNL B0 and C0 steppings only. * * The wording here is unclear, and this workaround doesn't look anything * like the internal bug report recommendations, but leave it be for now... */ if (GEN_GEN == 10) { if (flags & PIPE_CONTROL_RENDER_TARGET_FLUSH) { flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD; } else if (flags & non_lri_post_sync_flags) { flags |= PIPE_CONTROL_DEPTH_STALL; } } if (flags & PIPE_CONTROL_DEPTH_STALL) { /* From the PIPE_CONTROL instruction table, bit 13 (Depth Stall Enable): * * "This bit must be DISABLED for operations other than writing * PS_DEPTH_COUNT." * * This seems like nonsense. An Ivybridge workaround requires us to * emit a PIPE_CONTROL with a depth stall and write immediate post-sync * operation. Gen8+ requires us to emit depth stalls and depth cache * flushes together. So, it's hard to imagine this means anything other * than "we originally intended this to be used for PS_DEPTH_COUNT". * * We ignore the supposed restriction and do nothing. */ } if (flags & (PIPE_CONTROL_RENDER_TARGET_FLUSH | PIPE_CONTROL_STALL_AT_SCOREBOARD)) { /* From the PIPE_CONTROL instruction table, bit 12 and bit 1: * * "This bit must be DISABLED for End-of-pipe (Read) fences, * PS_DEPTH_COUNT or TIMESTAMP queries." * * TODO: Implement end-of-pipe checking. */ assert(!(post_sync_flags & (PIPE_CONTROL_WRITE_DEPTH_COUNT | PIPE_CONTROL_WRITE_TIMESTAMP))); } if (GEN_GEN < 11 && (flags & PIPE_CONTROL_STALL_AT_SCOREBOARD)) { /* From the PIPE_CONTROL instruction table, bit 1: * * "This bit is ignored if Depth Stall Enable is set. * Further, the render cache is not flushed even if Write Cache * Flush Enable bit is set." * * We assert that the caller doesn't do this combination, to try and * prevent mistakes. It shouldn't hurt the GPU, though. * * We skip this check on Gen11+ as the "Stall at Pixel Scoreboard" * and "Render Target Flush" combo is explicitly required for BTI * update workarounds. */ assert(!(flags & (PIPE_CONTROL_DEPTH_STALL | PIPE_CONTROL_RENDER_TARGET_FLUSH))); } /* PIPE_CONTROL page workarounds ------------------------------------- */ if (GEN_GEN <= 8 && (flags & PIPE_CONTROL_STATE_CACHE_INVALIDATE)) { /* From the PIPE_CONTROL page itself: * * "IVB, HSW, BDW * Restriction: Pipe_control with CS-stall bit set must be issued * before a pipe-control command that has the State Cache * Invalidate bit set." */ flags |= PIPE_CONTROL_CS_STALL; } if (flags & PIPE_CONTROL_FLUSH_LLC) { /* From the PIPE_CONTROL instruction table, bit 26 (Flush LLC): * * "Project: ALL * SW must always program Post-Sync Operation to "Write Immediate * Data" when Flush LLC is set." * * For now, we just require the caller to do it. */ assert(flags & PIPE_CONTROL_WRITE_IMMEDIATE); } /* "Post-Sync Operation" workarounds -------------------------------- */ /* Project: All / Argument: Global Snapshot Count Reset [19] * * "This bit must not be exercised on any product. * Requires stall bit ([20] of DW1) set." * * We don't use this, so we just assert that it isn't used. The * PIPE_CONTROL instruction page indicates that they intended this * as a debug feature and don't think it is useful in production, * but it may actually be usable, should we ever want to. */ assert((flags & PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET) == 0); if (flags & (PIPE_CONTROL_MEDIA_STATE_CLEAR | PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE)) { /* Project: All / Arguments: * * - Generic Media State Clear [16] * - Indirect State Pointers Disable [16] * * "Requires stall bit ([20] of DW1) set." * * Also, the PIPE_CONTROL instruction table, bit 16 (Generic Media * State Clear) says: * * "PIPECONTROL command with “Command Streamer Stall Enable” must be * programmed prior to programming a PIPECONTROL command with "Media * State Clear" set in GPGPU mode of operation" * * This is a subset of the earlier rule, so there's nothing to do. */ flags |= PIPE_CONTROL_CS_STALL; } if (flags & PIPE_CONTROL_STORE_DATA_INDEX) { /* Project: All / Argument: Store Data Index * * "Post-Sync Operation ([15:14] of DW1) must be set to something other * than '0'." * * For now, we just assert that the caller does this. We might want to * automatically add a write to the workaround BO... */ assert(non_lri_post_sync_flags != 0); } if (flags & PIPE_CONTROL_SYNC_GFDT) { /* Project: All / Argument: Sync GFDT * * "Post-Sync Operation ([15:14] of DW1) must be set to something other * than '0' or 0x2520[13] must be set." * * For now, we just assert that the caller does this. */ assert(non_lri_post_sync_flags != 0); } if (flags & PIPE_CONTROL_TLB_INVALIDATE) { /* Project: IVB+ / Argument: TLB inv * * "Requires stall bit ([20] of DW1) set." * * Also, from the PIPE_CONTROL instruction table: * * "Project: SKL+ * Post Sync Operation or CS stall must be set to ensure a TLB * invalidation occurs. Otherwise no cycle will occur to the TLB * cache to invalidate." * * This is not a subset of the earlier rule, so there's nothing to do. */ flags |= PIPE_CONTROL_CS_STALL; } if (GEN_GEN == 9 && devinfo->gt == 4) { /* TODO: The big Skylake GT4 post sync op workaround */ } /* "GPGPU specific workarounds" (both post-sync and flush) ------------ */ if (IS_COMPUTE_PIPELINE(batch)) { if (GEN_GEN >= 9 && (flags & PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE)) { /* Project: SKL+ / Argument: Tex Invalidate * "Requires stall bit ([20] of DW) set for all GPGPU Workloads." */ flags |= PIPE_CONTROL_CS_STALL; } if (GEN_GEN == 8 && (post_sync_flags || (flags & (PIPE_CONTROL_NOTIFY_ENABLE | PIPE_CONTROL_DEPTH_STALL | PIPE_CONTROL_RENDER_TARGET_FLUSH | PIPE_CONTROL_DEPTH_CACHE_FLUSH | PIPE_CONTROL_DATA_CACHE_FLUSH)))) { /* Project: BDW / Arguments: * * - LRI Post Sync Operation [23] * - Post Sync Op [15:14] * - Notify En [8] * - Depth Stall [13] * - Render Target Cache Flush [12] * - Depth Cache Flush [0] * - DC Flush Enable [5] * * "Requires stall bit ([20] of DW) set for all GPGPU and Media * Workloads." */ flags |= PIPE_CONTROL_CS_STALL; /* Also, from the PIPE_CONTROL instruction table, bit 20: * * "Project: BDW * This bit must be always set when PIPE_CONTROL command is * programmed by GPGPU and MEDIA workloads, except for the cases * when only Read Only Cache Invalidation bits are set (State * Cache Invalidation Enable, Instruction cache Invalidation * Enable, Texture Cache Invalidation Enable, Constant Cache * Invalidation Enable). This is to WA FFDOP CG issue, this WA * need not implemented when FF_DOP_CG is disable via "Fixed * Function DOP Clock Gate Disable" bit in RC_PSMI_CTRL register." * * It sounds like we could avoid CS stalls in some cases, but we * don't currently bother. This list isn't exactly the list above, * either... */ } } /* "Stall" workarounds ---------------------------------------------- * These have to come after the earlier ones because we may have added * some additional CS stalls above. */ if (GEN_GEN < 9 && (flags & PIPE_CONTROL_CS_STALL)) { /* Project: PRE-SKL, VLV, CHV * * "[All Stepping][All SKUs]: * * One of the following must also be set: * * - Render Target Cache Flush Enable ([12] of DW1) * - Depth Cache Flush Enable ([0] of DW1) * - Stall at Pixel Scoreboard ([1] of DW1) * - Depth Stall ([13] of DW1) * - Post-Sync Operation ([13] of DW1) * - DC Flush Enable ([5] of DW1)" * * If we don't already have one of those bits set, we choose to add * "Stall at Pixel Scoreboard". Some of the other bits require a * CS stall as a workaround (see above), which would send us into * an infinite recursion of PIPE_CONTROLs. "Stall at Pixel Scoreboard" * appears to be safe, so we choose that. */ const uint32_t wa_bits = PIPE_CONTROL_RENDER_TARGET_FLUSH | PIPE_CONTROL_DEPTH_CACHE_FLUSH | PIPE_CONTROL_WRITE_IMMEDIATE | PIPE_CONTROL_WRITE_DEPTH_COUNT | PIPE_CONTROL_WRITE_TIMESTAMP | PIPE_CONTROL_STALL_AT_SCOREBOARD | PIPE_CONTROL_DEPTH_STALL | PIPE_CONTROL_DATA_CACHE_FLUSH; if (!(flags & wa_bits)) flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD; } /* Emit --------------------------------------------------------------- */ iris_emit_cmd(batch, GENX(PIPE_CONTROL), pc) { pc.LRIPostSyncOperation = NoLRIOperation; pc.PipeControlFlushEnable = flags & PIPE_CONTROL_FLUSH_ENABLE; pc.DCFlushEnable = flags & PIPE_CONTROL_DATA_CACHE_FLUSH; pc.StoreDataIndex = 0; pc.CommandStreamerStallEnable = flags & PIPE_CONTROL_CS_STALL; pc.GlobalSnapshotCountReset = flags & PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET; pc.TLBInvalidate = flags & PIPE_CONTROL_TLB_INVALIDATE; pc.GenericMediaStateClear = flags & PIPE_CONTROL_MEDIA_STATE_CLEAR; pc.StallAtPixelScoreboard = flags & PIPE_CONTROL_STALL_AT_SCOREBOARD; pc.RenderTargetCacheFlushEnable = flags & PIPE_CONTROL_RENDER_TARGET_FLUSH; pc.DepthCacheFlushEnable = flags & PIPE_CONTROL_DEPTH_CACHE_FLUSH; pc.StateCacheInvalidationEnable = flags & PIPE_CONTROL_STATE_CACHE_INVALIDATE; pc.VFCacheInvalidationEnable = flags & PIPE_CONTROL_VF_CACHE_INVALIDATE; pc.ConstantCacheInvalidationEnable = flags & PIPE_CONTROL_CONST_CACHE_INVALIDATE; pc.PostSyncOperation = flags_to_post_sync_op(flags); pc.DepthStallEnable = flags & PIPE_CONTROL_DEPTH_STALL; pc.InstructionCacheInvalidateEnable = flags & PIPE_CONTROL_INSTRUCTION_INVALIDATE; pc.NotifyEnable = flags & PIPE_CONTROL_NOTIFY_ENABLE; pc.IndirectStatePointersDisable = flags & PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE; pc.TextureCacheInvalidationEnable = flags & PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE; pc.Address = rw_bo(bo, offset); pc.ImmediateData = imm; } } void genX(emit_urb_setup)(struct iris_context *ice, struct iris_batch *batch, const unsigned size[4], bool tess_present, bool gs_present) { const struct gen_device_info *devinfo = &batch->screen->devinfo; const unsigned push_size_kB = 32; unsigned entries[4]; unsigned start[4]; ice->shaders.last_vs_entry_size = size[MESA_SHADER_VERTEX]; gen_get_urb_config(devinfo, 1024 * push_size_kB, 1024 * ice->shaders.urb_size, tess_present, gs_present, size, entries, start); for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) { iris_emit_cmd(batch, GENX(3DSTATE_URB_VS), urb) { urb._3DCommandSubOpcode += i; urb.VSURBStartingAddress = start[i]; urb.VSURBEntryAllocationSize = size[i] - 1; urb.VSNumberofURBEntries = entries[i]; } } } #if GEN_GEN == 9 /** * Preemption on Gen9 has to be enabled or disabled in various cases. * * See these workarounds for preemption: * - WaDisableMidObjectPreemptionForGSLineStripAdj * - WaDisableMidObjectPreemptionForTrifanOrPolygon * - WaDisableMidObjectPreemptionForLineLoop * - WA#0798 * * We don't put this in the vtable because it's only used on Gen9. */ void gen9_toggle_preemption(struct iris_context *ice, struct iris_batch *batch, const struct pipe_draw_info *draw) { struct iris_genx_state *genx = ice->state.genx; bool object_preemption = true; /* WaDisableMidObjectPreemptionForGSLineStripAdj * * "WA: Disable mid-draw preemption when draw-call is a linestrip_adj * and GS is enabled." */ if (draw->mode == PIPE_PRIM_LINE_STRIP_ADJACENCY && ice->shaders.prog[MESA_SHADER_GEOMETRY]) object_preemption = false; /* WaDisableMidObjectPreemptionForTrifanOrPolygon * * "TriFan miscompare in Execlist Preemption test. Cut index that is * on a previous context. End the previous, the resume another context * with a tri-fan or polygon, and the vertex count is corrupted. If we * prempt again we will cause corruption. * * WA: Disable mid-draw preemption when draw-call has a tri-fan." */ if (draw->mode == PIPE_PRIM_TRIANGLE_FAN) object_preemption = false; /* WaDisableMidObjectPreemptionForLineLoop * * "VF Stats Counters Missing a vertex when preemption enabled. * * WA: Disable mid-draw preemption when the draw uses a lineloop * topology." */ if (draw->mode == PIPE_PRIM_LINE_LOOP) object_preemption = false; /* WA#0798 * * "VF is corrupting GAFS data when preempted on an instance boundary * and replayed with instancing enabled. * * WA: Disable preemption when using instanceing." */ if (draw->instance_count > 1) object_preemption = false; if (genx->object_preemption != object_preemption) { iris_enable_obj_preemption(batch, object_preemption); genx->object_preemption = object_preemption; } } #endif void genX(init_state)(struct iris_context *ice) { struct pipe_context *ctx = &ice->ctx; struct iris_screen *screen = (struct iris_screen *)ctx->screen; ctx->create_blend_state = iris_create_blend_state; ctx->create_depth_stencil_alpha_state = iris_create_zsa_state; ctx->create_rasterizer_state = iris_create_rasterizer_state; ctx->create_sampler_state = iris_create_sampler_state; ctx->create_sampler_view = iris_create_sampler_view; ctx->create_surface = iris_create_surface; ctx->create_vertex_elements_state = iris_create_vertex_elements; ctx->bind_blend_state = iris_bind_blend_state; ctx->bind_depth_stencil_alpha_state = iris_bind_zsa_state; ctx->bind_sampler_states = iris_bind_sampler_states; ctx->bind_rasterizer_state = iris_bind_rasterizer_state; ctx->bind_vertex_elements_state = iris_bind_vertex_elements_state; ctx->delete_blend_state = iris_delete_state; ctx->delete_depth_stencil_alpha_state = iris_delete_state; ctx->delete_rasterizer_state = iris_delete_state; ctx->delete_sampler_state = iris_delete_state; ctx->delete_vertex_elements_state = iris_delete_state; ctx->set_blend_color = iris_set_blend_color; ctx->set_clip_state = iris_set_clip_state; ctx->set_constant_buffer = iris_set_constant_buffer; ctx->set_shader_buffers = iris_set_shader_buffers; ctx->set_shader_images = iris_set_shader_images; ctx->set_sampler_views = iris_set_sampler_views; ctx->set_tess_state = iris_set_tess_state; ctx->set_framebuffer_state = iris_set_framebuffer_state; ctx->set_polygon_stipple = iris_set_polygon_stipple; ctx->set_sample_mask = iris_set_sample_mask; ctx->set_scissor_states = iris_set_scissor_states; ctx->set_stencil_ref = iris_set_stencil_ref; ctx->set_vertex_buffers = iris_set_vertex_buffers; ctx->set_viewport_states = iris_set_viewport_states; ctx->sampler_view_destroy = iris_sampler_view_destroy; ctx->surface_destroy = iris_surface_destroy; ctx->draw_vbo = iris_draw_vbo; ctx->launch_grid = iris_launch_grid; ctx->create_stream_output_target = iris_create_stream_output_target; ctx->stream_output_target_destroy = iris_stream_output_target_destroy; ctx->set_stream_output_targets = iris_set_stream_output_targets; ice->vtbl.destroy_state = iris_destroy_state; ice->vtbl.init_render_context = iris_init_render_context; ice->vtbl.init_compute_context = iris_init_compute_context; ice->vtbl.upload_render_state = iris_upload_render_state; ice->vtbl.update_surface_base_address = iris_update_surface_base_address; ice->vtbl.upload_compute_state = iris_upload_compute_state; ice->vtbl.emit_raw_pipe_control = iris_emit_raw_pipe_control; ice->vtbl.rebind_buffer = iris_rebind_buffer; ice->vtbl.load_register_reg32 = iris_load_register_reg32; ice->vtbl.load_register_reg64 = iris_load_register_reg64; ice->vtbl.load_register_imm32 = iris_load_register_imm32; ice->vtbl.load_register_imm64 = iris_load_register_imm64; ice->vtbl.load_register_mem32 = iris_load_register_mem32; ice->vtbl.load_register_mem64 = iris_load_register_mem64; ice->vtbl.store_register_mem32 = iris_store_register_mem32; ice->vtbl.store_register_mem64 = iris_store_register_mem64; ice->vtbl.store_data_imm32 = iris_store_data_imm32; ice->vtbl.store_data_imm64 = iris_store_data_imm64; ice->vtbl.copy_mem_mem = iris_copy_mem_mem; ice->vtbl.derived_program_state_size = iris_derived_program_state_size; ice->vtbl.store_derived_program_state = iris_store_derived_program_state; ice->vtbl.create_so_decl_list = iris_create_so_decl_list; ice->vtbl.populate_vs_key = iris_populate_vs_key; ice->vtbl.populate_tcs_key = iris_populate_tcs_key; ice->vtbl.populate_tes_key = iris_populate_tes_key; ice->vtbl.populate_gs_key = iris_populate_gs_key; ice->vtbl.populate_fs_key = iris_populate_fs_key; ice->vtbl.populate_cs_key = iris_populate_cs_key; ice->vtbl.mocs = mocs; ice->state.dirty = ~0ull; ice->state.statistics_counters_enabled = true; ice->state.sample_mask = 0xffff; ice->state.num_viewports = 1; ice->state.genx = calloc(1, sizeof(struct iris_genx_state)); /* Make a 1x1x1 null surface for unbound textures */ void *null_surf_map = upload_state(ice->state.surface_uploader, &ice->state.unbound_tex, 4 * GENX(RENDER_SURFACE_STATE_length), 64); isl_null_fill_state(&screen->isl_dev, null_surf_map, isl_extent3d(1, 1, 1)); ice->state.unbound_tex.offset += iris_bo_offset_from_base_address(iris_resource_bo(ice->state.unbound_tex.res)); /* Default all scissor rectangles to be empty regions. */ for (int i = 0; i < IRIS_MAX_VIEWPORTS; i++) { ice->state.scissors[i] = (struct pipe_scissor_state) { .minx = 1, .maxx = 0, .miny = 1, .maxy = 0, }; } }