#include "util/ralloc.h"
#include "brw_context.h"
+#include "brw_defines.h"
#include "brw_cs.h"
-#include "brw_eu.h"
#include "brw_wm.h"
-#include "brw_shader.h"
#include "intel_mipmap_tree.h"
#include "intel_batchbuffer.h"
#include "brw_state.h"
-
-static unsigned
-get_cs_thread_count(const struct brw_cs_prog_data *cs_prog_data)
-{
- const unsigned simd_size = cs_prog_data->simd_size;
- unsigned group_size = cs_prog_data->local_size[0] *
- cs_prog_data->local_size[1] * cs_prog_data->local_size[2];
-
- return (group_size + simd_size - 1) / simd_size;
-}
-
-
-static void
-brw_upload_cs_state(struct brw_context *brw)
-{
- if (!brw->cs.prog_data)
- return;
-
- uint32_t offset;
- uint32_t *desc = (uint32_t*) brw_state_batch(brw, AUB_TRACE_SURFACE_STATE,
- 8 * 4, 64, &offset);
- struct gl_program *prog = (struct gl_program *) brw->compute_program;
- struct brw_stage_state *stage_state = &brw->cs.base;
- struct brw_cs_prog_data *cs_prog_data = brw->cs.prog_data;
- struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
-
- if (INTEL_DEBUG & DEBUG_SHADER_TIME) {
- brw->vtbl.emit_buffer_surface_state(
- brw, &stage_state->surf_offset[
- prog_data->binding_table.shader_time_start],
- brw->shader_time.bo, 0, BRW_SURFACEFORMAT_RAW,
- brw->shader_time.bo->size, 1, true);
- }
-
- uint32_t *bind = (uint32_t*) brw_state_batch(brw, AUB_TRACE_BINDING_TABLE,
- prog_data->binding_table.size_bytes,
- 32, &stage_state->bind_bo_offset);
-
- unsigned local_id_dwords = 0;
-
- if (prog->SystemValuesRead & SYSTEM_BIT_LOCAL_INVOCATION_ID) {
- local_id_dwords =
- brw_cs_prog_local_id_payload_dwords(cs_prog_data->simd_size);
- }
-
- unsigned push_constant_data_size =
- (prog_data->nr_params + local_id_dwords) * sizeof(gl_constant_value);
- unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
- unsigned push_constant_regs = reg_aligned_constant_size / 32;
- unsigned threads = get_cs_thread_count(cs_prog_data);
-
- uint32_t dwords = brw->gen < 8 ? 8 : 9;
- BEGIN_BATCH(dwords);
- OUT_BATCH(MEDIA_VFE_STATE << 16 | (dwords - 2));
-
- if (prog_data->total_scratch) {
- if (brw->gen >= 8)
- OUT_RELOC64(stage_state->scratch_bo,
- I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
- ffs(prog_data->total_scratch) - 11);
- else
- OUT_RELOC(stage_state->scratch_bo,
- I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
- ffs(prog_data->total_scratch) - 11);
- } else {
- OUT_BATCH(0);
- if (brw->gen >= 8)
- OUT_BATCH(0);
- }
-
- const uint32_t vfe_num_urb_entries = brw->gen >= 8 ? 2 : 0;
- const uint32_t vfe_gpgpu_mode =
- brw->gen == 7 ? SET_FIELD(1, GEN7_MEDIA_VFE_STATE_GPGPU_MODE) : 0;
- OUT_BATCH(SET_FIELD(brw->max_cs_threads - 1, MEDIA_VFE_STATE_MAX_THREADS) |
- SET_FIELD(vfe_num_urb_entries, MEDIA_VFE_STATE_URB_ENTRIES) |
- SET_FIELD(1, MEDIA_VFE_STATE_RESET_GTW_TIMER) |
- SET_FIELD(1, MEDIA_VFE_STATE_BYPASS_GTW) |
- vfe_gpgpu_mode);
-
- OUT_BATCH(0);
- const uint32_t vfe_urb_allocation = brw->gen >= 8 ? 2 : 0;
-
- /* We are uploading duplicated copies of push constant uniforms for each
- * thread. Although the local id data needs to vary per thread, it won't
- * change for other uniform data. Unfortunately this duplication is
- * required for gen7. As of Haswell, this duplication can be avoided, but
- * this older mechanism with duplicated data continues to work.
- *
- * FINISHME: As of Haswell, we could make use of the
- * INTERFACE_DESCRIPTOR_DATA "Cross-Thread Constant Data Read Length" field
- * to only store one copy of uniform data.
- *
- * FINISHME: Broadwell adds a new alternative "Indirect Payload Storage"
- * which is described in the GPGPU_WALKER command and in the Broadwell PRM
- * Volume 7: 3D Media GPGPU, under Media GPGPU Pipeline => Mode of
- * Operations => GPGPU Mode => Indirect Payload Storage.
- *
- * Note: The constant data is built in brw_upload_cs_push_constants below.
- */
- const uint32_t vfe_curbe_allocation = push_constant_regs * threads;
- OUT_BATCH(SET_FIELD(vfe_urb_allocation, MEDIA_VFE_STATE_URB_ALLOC) |
- SET_FIELD(vfe_curbe_allocation, MEDIA_VFE_STATE_CURBE_ALLOC));
- OUT_BATCH(0);
- OUT_BATCH(0);
- OUT_BATCH(0);
- ADVANCE_BATCH();
-
- if (reg_aligned_constant_size > 0) {
- BEGIN_BATCH(4);
- OUT_BATCH(MEDIA_CURBE_LOAD << 16 | (4 - 2));
- OUT_BATCH(0);
- OUT_BATCH(reg_aligned_constant_size * threads);
- OUT_BATCH(stage_state->push_const_offset);
- ADVANCE_BATCH();
- }
-
- /* BRW_NEW_SURFACES and BRW_NEW_*_CONSTBUF */
- memcpy(bind, stage_state->surf_offset,
- prog_data->binding_table.size_bytes);
-
- memset(desc, 0, 8 * 4);
-
- int dw = 0;
- desc[dw++] = brw->cs.base.prog_offset;
- if (brw->gen >= 8)
- desc[dw++] = 0; /* Kernel Start Pointer High */
- desc[dw++] = 0;
- desc[dw++] = stage_state->sampler_offset |
- ((stage_state->sampler_count + 3) / 4);
- desc[dw++] = stage_state->bind_bo_offset;
- desc[dw++] = SET_FIELD(push_constant_regs, MEDIA_CURBE_READ_LENGTH);
- const uint32_t media_threads =
- brw->gen >= 8 ?
- SET_FIELD(threads, GEN8_MEDIA_GPGPU_THREAD_COUNT) :
- SET_FIELD(threads, MEDIA_GPGPU_THREAD_COUNT);
- assert(threads <= brw->max_cs_threads);
- desc[dw++] =
- SET_FIELD(cs_prog_data->uses_barrier, MEDIA_BARRIER_ENABLE) |
- media_threads;
-
- BEGIN_BATCH(4);
- OUT_BATCH(MEDIA_INTERFACE_DESCRIPTOR_LOAD << 16 | (4 - 2));
- OUT_BATCH(0);
- OUT_BATCH(8 * 4);
- OUT_BATCH(offset);
- ADVANCE_BATCH();
-}
-
-const struct brw_tracked_state brw_cs_state = {
- .dirty = {
- .mesa = _NEW_PROGRAM_CONSTANTS,
- .brw = BRW_NEW_BATCH |
- BRW_NEW_CS_PROG_DATA |
- BRW_NEW_PUSH_CONSTANT_ALLOCATION |
- BRW_NEW_SURFACES,
- },
- .emit = brw_upload_cs_state
-};
-
-
-/**
- * We are building the local ID push constant data using the simplest possible
- * method. We simply push the local IDs directly as they should appear in the
- * registers for the uvec3 gl_LocalInvocationID variable.
- *
- * Therefore, for SIMD8, we use 3 full registers, and for SIMD16 we use 6
- * registers worth of push constant space.
- *
- * Note: Any updates to brw_cs_prog_local_id_payload_dwords,
- * fill_local_id_payload or fs_visitor::emit_cs_local_invocation_id_setup need
- * to coordinated.
- *
- * FINISHME: There are a few easy optimizations to consider.
- *
- * 1. If gl_WorkGroupSize x, y or z is 1, we can just use zero, and there is
- * no need for using push constant space for that dimension.
- *
- * 2. Since GL_MAX_COMPUTE_WORK_GROUP_SIZE is currently 1024 or less, we can
- * easily use 16-bit words rather than 32-bit dwords in the push constant
- * data.
- *
- * 3. If gl_WorkGroupSize x, y or z is small, then we can use bytes for
- * conveying the data, and thereby reduce push constant usage.
- *
- */
-unsigned
-brw_cs_prog_local_id_payload_dwords(unsigned dispatch_width)
-{
- return 3 * dispatch_width;
-}
-
-
-static void
-fill_local_id_payload(const struct brw_cs_prog_data *cs_prog_data,
- void *buffer, unsigned *x, unsigned *y, unsigned *z)
-{
- uint32_t *param = (uint32_t *)buffer;
- for (unsigned i = 0; i < cs_prog_data->simd_size; i++) {
- param[0 * cs_prog_data->simd_size + i] = *x;
- param[1 * cs_prog_data->simd_size + i] = *y;
- param[2 * cs_prog_data->simd_size + i] = *z;
-
- (*x)++;
- if (*x == cs_prog_data->local_size[0]) {
- *x = 0;
- (*y)++;
- if (*y == cs_prog_data->local_size[1]) {
- *y = 0;
- (*z)++;
- if (*z == cs_prog_data->local_size[2])
- *z = 0;
- }
- }
- }
-}
-
+#include "program/prog_statevars.h"
+#include "compiler/glsl/ir_uniform.h"
+#include "main/shaderapi.h"
/**
* Creates a region containing the push constants for the CS on gen7+.
brw_upload_cs_push_constants(struct brw_context *brw,
const struct gl_program *prog,
const struct brw_cs_prog_data *cs_prog_data,
- struct brw_stage_state *stage_state,
- enum aub_state_struct_type type)
+ struct brw_stage_state *stage_state)
{
struct gl_context *ctx = &brw->ctx;
const struct brw_stage_prog_data *prog_data =
(struct brw_stage_prog_data*) cs_prog_data;
- unsigned local_id_dwords = 0;
-
- if (prog->SystemValuesRead & SYSTEM_BIT_LOCAL_INVOCATION_ID) {
- local_id_dwords =
- brw_cs_prog_local_id_payload_dwords(cs_prog_data->simd_size);
- }
/* Updates the ParamaterValues[i] pointers for all parameters of the
* basic type of PROGRAM_STATE_VAR.
/* XXX: Should this happen somewhere before to get our state flag set? */
_mesa_load_state_parameters(ctx, prog->Parameters);
- if (prog_data->nr_params == 0 && local_id_dwords == 0) {
+ if (cs_prog_data->push.total.size == 0) {
stage_state->push_const_size = 0;
- } else {
- gl_constant_value *param;
- unsigned i, t;
+ return;
+ }
- const unsigned push_constant_data_size =
- (local_id_dwords + prog_data->nr_params) * sizeof(gl_constant_value);
- const unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
- const unsigned param_aligned_count =
- reg_aligned_constant_size / sizeof(*param);
- unsigned threads = get_cs_thread_count(cs_prog_data);
+ gl_constant_value *param = (gl_constant_value*)
+ brw_state_batch(brw, ALIGN(cs_prog_data->push.total.size, 64),
+ 64, &stage_state->push_const_offset);
+ assert(param);
- param = (gl_constant_value*)
- brw_state_batch(brw, type,
- reg_aligned_constant_size * threads,
- 32, &stage_state->push_const_offset);
- assert(param);
+ STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
- STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
+ if (cs_prog_data->push.cross_thread.size > 0) {
+ gl_constant_value *param_copy = param;
+ assert(cs_prog_data->thread_local_id_index < 0 ||
+ cs_prog_data->thread_local_id_index >=
+ cs_prog_data->push.cross_thread.dwords);
+ for (unsigned i = 0;
+ i < cs_prog_data->push.cross_thread.dwords;
+ i++) {
+ param_copy[i] = *prog_data->param[i];
+ }
+ }
- /* _NEW_PROGRAM_CONSTANTS */
- unsigned x = 0, y = 0, z = 0;
- for (t = 0; t < threads; t++) {
- gl_constant_value *next_param = ¶m[t * param_aligned_count];
- if (local_id_dwords > 0) {
- fill_local_id_payload(cs_prog_data, (void*)next_param, &x, &y, &z);
- next_param += local_id_dwords;
- }
- for (i = 0; i < prog_data->nr_params; i++) {
- next_param[i] = *prog_data->param[i];
+ gl_constant_value thread_id;
+ if (cs_prog_data->push.per_thread.size > 0) {
+ for (unsigned t = 0; t < cs_prog_data->threads; t++) {
+ unsigned dst =
+ 8 * (cs_prog_data->push.per_thread.regs * t +
+ cs_prog_data->push.cross_thread.regs);
+ unsigned src = cs_prog_data->push.cross_thread.dwords;
+ for ( ; src < prog_data->nr_params; src++, dst++) {
+ if (src != cs_prog_data->thread_local_id_index)
+ param[dst] = *prog_data->param[src];
+ else {
+ thread_id.u = t * cs_prog_data->simd_size;
+ param[dst] = thread_id;
+ }
}
}
-
- stage_state->push_const_size = ALIGN(prog_data->nr_params, 8) / 8;
}
+
+ stage_state->push_const_size =
+ cs_prog_data->push.cross_thread.regs +
+ cs_prog_data->push.per_thread.regs;
}
struct brw_stage_state *stage_state = &brw->cs.base;
/* BRW_NEW_COMPUTE_PROGRAM */
- const struct brw_compute_program *cp =
- (struct brw_compute_program *) brw->compute_program;
+ const struct brw_program *cp = (struct brw_program *) brw->compute_program;
if (cp) {
- /* CACHE_NEW_CS_PROG */
- struct brw_cs_prog_data *cs_prog_data = brw->cs.prog_data;
+ /* BRW_NEW_CS_PROG_DATA */
+ struct brw_cs_prog_data *cs_prog_data =
+ brw_cs_prog_data(brw->cs.base.prog_data);
- brw_upload_cs_push_constants(brw, &cp->program.Base, cs_prog_data,
- stage_state, AUB_TRACE_WM_CONSTANTS);
+ _mesa_shader_write_subroutine_indices(&brw->ctx, MESA_SHADER_COMPUTE);
+ brw_upload_cs_push_constants(brw, &cp->program, cs_prog_data,
+ stage_state);
}
}
.dirty = {
.mesa = _NEW_PROGRAM_CONSTANTS,
.brw = BRW_NEW_BATCH |
+ BRW_NEW_BLORP |
BRW_NEW_COMPUTE_PROGRAM |
- BRW_NEW_PUSH_CONSTANT_ALLOCATION,
+ BRW_NEW_CS_PROG_DATA,
},
.emit = gen7_upload_cs_push_constants,
};
+
+/**
+ * Creates a new CS constant buffer reflecting the current CS program's
+ * constants, if needed by the CS program.
+ */
+static void
+brw_upload_cs_pull_constants(struct brw_context *brw)
+{
+ struct brw_stage_state *stage_state = &brw->cs.base;
+
+ /* BRW_NEW_COMPUTE_PROGRAM */
+ struct brw_program *cp = (struct brw_program *) brw->compute_program;
+
+ /* BRW_NEW_CS_PROG_DATA */
+ const struct brw_stage_prog_data *prog_data = brw->cs.base.prog_data;
+
+ _mesa_shader_write_subroutine_indices(&brw->ctx, MESA_SHADER_COMPUTE);
+ /* _NEW_PROGRAM_CONSTANTS */
+ brw_upload_pull_constants(brw, BRW_NEW_SURFACES, &cp->program,
+ stage_state, prog_data);
+}
+
+const struct brw_tracked_state brw_cs_pull_constants = {
+ .dirty = {
+ .mesa = _NEW_PROGRAM_CONSTANTS,
+ .brw = BRW_NEW_BATCH |
+ BRW_NEW_BLORP |
+ BRW_NEW_COMPUTE_PROGRAM |
+ BRW_NEW_CS_PROG_DATA,
+ },
+ .emit = brw_upload_cs_pull_constants,
+};
projects / mesa.git / blobdiff