gallium: add a new interface for pipe_context::launch_grid()

This introduces pipe_grid_info which contains all information to
describe a launch_grid call. This will be used to implement indirect
compute in the same fashion as indirect draw.

Changes from v2:
 - correctly initialize pipe_grid_info for nv50/nvc0

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>

src/gallium/drivers/ilo/ilo_gpgpu.c

@@ -79,9 +79,7 @@ launch_grid(struct ilo_context *ilo,
 }
 
 static void
-ilo_launch_grid(struct pipe_context *pipe,
+ilo_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
-                const uint *block_layout, const uint *grid_layout,
-                uint32_t pc, const void *input)
 {
    struct ilo_context *ilo = ilo_context(pipe);
    struct ilo_shader_state *cs = ilo->state_vector.cs;
@@ -92,13 +90,13 @@ ilo_launch_grid(struct pipe_context *pipe,
    input_buf.buffer_size =
       ilo_shader_get_kernel_param(cs, ILO_KERNEL_CS_INPUT_SIZE);
    if (input_buf.buffer_size) {
-      u_upload_data(ilo->uploader, 0, input_buf.buffer_size, 16, input,
+      u_upload_data(ilo->uploader, 0, input_buf.buffer_size, 16, info->input,
             &input_buf.buffer_offset, &input_buf.buffer);
    }
 
    ilo_shader_cache_upload(ilo->shader_cache, &ilo->cp->builder);
 
-   launch_grid(ilo, block_layout, grid_layout, &input_buf, pc);
+   launch_grid(ilo, info->block, info->grid, &input_buf, info->pc);
 
    ilo_render_invalidate_hw(ilo->render);
 

src/gallium/drivers/nouveau/nv50/nv50_compute.c

@@ -270,13 +270,11 @@ nv50_compute_find_symbol(struct nv50_context *nv50, uint32_t label)
 }
 
 void
-nv50_launch_grid(struct pipe_context *pipe,
+nv50_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
-                 const uint *block_layout, const uint *grid_layout,
-                 uint32_t label, const void *input)
 {
    struct nv50_context *nv50 = nv50_context(pipe);
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
-   unsigned block_size = block_layout[0] * block_layout[1] * block_layout[2];
+   unsigned block_size = info->block[0] * info->block[1] * info->block[2];
    struct nv50_program *cp = nv50->compprog;
    bool ret;
 
@@ -286,10 +284,10 @@ nv50_launch_grid(struct pipe_context *pipe,
       return;
    }
 
-   nv50_compute_upload_input(nv50, input);
+   nv50_compute_upload_input(nv50, info->input);
 
    BEGIN_NV04(push, NV50_COMPUTE(CP_START_ID), 1);
-   PUSH_DATA (push, nv50_compute_find_symbol(nv50, label));
+   PUSH_DATA (push, nv50_compute_find_symbol(nv50, info->pc));
 
    BEGIN_NV04(push, NV50_COMPUTE(SHARED_SIZE), 1);
    PUSH_DATA (push, align(cp->cp.smem_size + cp->parm_size + 0x10, 0x40));
@@ -298,14 +296,14 @@ nv50_launch_grid(struct pipe_context *pipe,
 
    /* grid/block setup */
    BEGIN_NV04(push, NV50_COMPUTE(BLOCKDIM_XY), 2);
-   PUSH_DATA (push, block_layout[1] << 16 | block_layout[0]);
+   PUSH_DATA (push, info->block[1] << 16 | info->block[0]);
-   PUSH_DATA (push, block_layout[2]);
+   PUSH_DATA (push, info->block[2]);
    BEGIN_NV04(push, NV50_COMPUTE(BLOCK_ALLOC), 1);
    PUSH_DATA (push, 1 << 16 | block_size);
    BEGIN_NV04(push, NV50_COMPUTE(BLOCKDIM_LATCH), 1);
    PUSH_DATA (push, 1);
    BEGIN_NV04(push, NV50_COMPUTE(GRIDDIM), 1);
-   PUSH_DATA (push, grid_layout[1] << 16 | grid_layout[0]);
+   PUSH_DATA (push, info->grid[1] << 16 | info->grid[0]);
    BEGIN_NV04(push, NV50_COMPUTE(GRIDID), 1);
    PUSH_DATA (push, 1);
 

src/gallium/drivers/nouveau/nv50/nv50_context.h

@@ -322,7 +322,6 @@ nv98_video_buffer_create(struct pipe_context *pipe,
 
 /* nv50_compute.c */
 void
-nv50_launch_grid(struct pipe_context *, const uint *, const uint *,
+nv50_launch_grid(struct pipe_context *, const struct pipe_grid_info *);
-                 uint32_t, const void *);
 
 #endif

src/gallium/drivers/nouveau/nv50/nv50_query_hw_sm.c

@@ -218,11 +218,12 @@ nv50_hw_sm_end_query(struct nv50_context *nv50, struct nv50_hw_query *hq)
    struct pipe_context *pipe = &nv50->base.pipe;
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
    struct nv50_hw_sm_query *hsq = nv50_hw_sm_query(hq);
+   struct pipe_grid_info info = {};
    uint32_t mask;
    uint32_t input[3];
    const uint block[3] = { 32, 1, 1 };
    const uint grid[3] = { screen->MPsInTP, screen->TPs, 1 };
-   int c;
+   int c, i;
 
    if (unlikely(!screen->pm.prog)) {
       struct nv50_program *prog = CALLOC_STRUCT(nv50_program);
@@ -262,7 +263,14 @@ nv50_hw_sm_end_query(struct nv50_context *nv50, struct nv50_hw_query *hq)
    pipe->bind_compute_state(pipe, screen->pm.prog);
    input[0] = hq->bo->offset + hq->base_offset;
    input[1] = hq->sequence;
-   pipe->launch_grid(pipe, block, grid, 0, input);
+
+   for (i = 0; i < 3; i++) {
+      info.block[i] = block[i];
+      info.grid[i] = grid[i];
+   }
+   info.pc = 0;
+   info.input = input;
+   pipe->launch_grid(pipe, &info);
 
    nouveau_bufctx_reset(nv50->bufctx_cp, NV50_BIND_CP_QUERY);
 

src/gallium/drivers/nouveau/nvc0/nvc0_compute.c

@@ -183,10 +183,7 @@ nvc0_compute_upload_input(struct nvc0_context *nvc0, const void *input)
 }
 
 void
-nvc0_launch_grid(struct pipe_context *pipe,
+nvc0_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
-                 const uint *block_layout, const uint *grid_layout,
-                 uint32_t label,
-                 const void *input)
 {
    struct nvc0_context *nvc0 = nvc0_context(pipe);
    struct nouveau_pushbuf *push = nvc0->base.pushbuf;
@@ -200,10 +197,10 @@ nvc0_launch_grid(struct pipe_context *pipe,
       return;
    }
 
-   nvc0_compute_upload_input(nvc0, input);
+   nvc0_compute_upload_input(nvc0, info->input);
 
    BEGIN_NVC0(push, NVC0_COMPUTE(CP_START_ID), 1);
-   PUSH_DATA (push, nvc0_program_symbol_offset(cp, label));
+   PUSH_DATA (push, nvc0_program_symbol_offset(cp, info->pc));
 
    BEGIN_NVC0(push, NVC0_COMPUTE(LOCAL_POS_ALLOC), 3);
    PUSH_DATA (push, align(cp->cp.lmem_size, 0x10));
@@ -212,18 +209,18 @@ nvc0_launch_grid(struct pipe_context *pipe,
 
    BEGIN_NVC0(push, NVC0_COMPUTE(SHARED_SIZE), 3);
    PUSH_DATA (push, align(cp->cp.smem_size, 0x100));
-   PUSH_DATA (push, block_layout[0] * block_layout[1] * block_layout[2]);
+   PUSH_DATA (push, info->block[0] * info->block[1] * info->block[2]);
    PUSH_DATA (push, cp->num_barriers);
    BEGIN_NVC0(push, NVC0_COMPUTE(CP_GPR_ALLOC), 1);
    PUSH_DATA (push, cp->num_gprs);
 
    /* grid/block setup */
    BEGIN_NVC0(push, NVC0_COMPUTE(GRIDDIM_YX), 2);
-   PUSH_DATA (push, (grid_layout[1] << 16) | grid_layout[0]);
+   PUSH_DATA (push, (info->grid[1] << 16) | info->grid[0]);
-   PUSH_DATA (push, grid_layout[2]);
+   PUSH_DATA (push, info->grid[2]);
    BEGIN_NVC0(push, NVC0_COMPUTE(BLOCKDIM_YX), 2);
-   PUSH_DATA (push, (block_layout[1] << 16) | block_layout[0]);
+   PUSH_DATA (push, (info->block[1] << 16) | info->block[0]);
-   PUSH_DATA (push, block_layout[2]);
+   PUSH_DATA (push, info->block[2]);
 
    /* launch preliminary setup */
    BEGIN_NVC0(push, NVC0_COMPUTE(GRIDID), 1);

src/gallium/drivers/nouveau/nvc0/nvc0_context.h

@@ -331,11 +331,9 @@ nvc0_video_buffer_create(struct pipe_context *pipe,
 void nvc0_push_vbo(struct nvc0_context *, const struct pipe_draw_info *);
 
 /* nve4_compute.c */
-void nve4_launch_grid(struct pipe_context *,
+void nve4_launch_grid(struct pipe_context *, const struct pipe_grid_info *);
-                      const uint *, const uint *, uint32_t, const void *);
 
 /* nvc0_compute.c */
-void nvc0_launch_grid(struct pipe_context *,
+void nvc0_launch_grid(struct pipe_context *, const struct pipe_grid_info *);
-                      const uint *, const uint *, uint32_t, const void *);
 
 #endif

src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_sm.c

@@ -937,11 +937,12 @@ nvc0_hw_sm_end_query(struct nvc0_context *nvc0, struct nvc0_hw_query *hq)
    struct nouveau_pushbuf *push = nvc0->base.pushbuf;
    const bool is_nve4 = screen->base.class_3d >= NVE4_3D_CLASS;
    struct nvc0_hw_sm_query *hsq = nvc0_hw_sm_query(hq);
+   struct pipe_grid_info info = {};
    uint32_t mask;
    uint32_t input[3];
    const uint block[3] = { 32, is_nve4 ? 4 : 1, 1 };
    const uint grid[3] = { screen->mp_count, screen->gpc_count, 1 };
-   unsigned c;
+   unsigned c, i;
 
    if (unlikely(!screen->pm.prog)) {
       struct nvc0_program *prog = CALLOC_STRUCT(nvc0_program);
@@ -989,7 +990,14 @@ nvc0_hw_sm_end_query(struct nvc0_context *nvc0, struct nvc0_hw_query *hq)
    input[0] = (hq->bo->offset + hq->base_offset);
    input[1] = (hq->bo->offset + hq->base_offset) >> 32;
    input[2] = hq->sequence;
-   pipe->launch_grid(pipe, block, grid, 0, input);
+
+   for (i = 0; i < 3; i++) {
+      info.block[i] = block[i];
+      info.grid[i] = grid[i];
+   }
+   info.pc = 0;
+   info.input = input;
+   pipe->launch_grid(pipe, &info);
 
    nouveau_bufctx_reset(nvc0->bufctx_cp, NVC0_BIND_CP_QUERY);
 

src/gallium/drivers/nouveau/nvc0/nve4_compute.c

@@ -429,10 +429,7 @@ nve4_compute_alloc_launch_desc(struct nouveau_context *nv,
 }
 
 void
-nve4_launch_grid(struct pipe_context *pipe,
+nve4_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
-                 const uint *block_layout, const uint *grid_layout,
-                 uint32_t label,
-                 const void *input)
 {
    struct nvc0_context *nvc0 = nvc0_context(pipe);
    struct nouveau_pushbuf *push = nvc0->base.pushbuf;
@@ -453,13 +450,14 @@ nve4_launch_grid(struct pipe_context *pipe,
    if (ret)
       goto out;
 
-   nve4_compute_setup_launch_desc(nvc0, desc, label, block_layout, grid_layout);
+   nve4_compute_setup_launch_desc(nvc0, desc, info->pc,
+                                  info->block, info->grid);
 #ifdef DEBUG
    if (debug_get_num_option("NV50_PROG_DEBUG", 0))
       nve4_compute_dump_launch_desc(desc);
 #endif
 
-   nve4_compute_upload_input(nvc0, input, block_layout, grid_layout);
+   nve4_compute_upload_input(nvc0, info->input, info->block, info->grid);
 
    /* upload descriptor and flush */
 #if 0

src/gallium/drivers/r600/evergreen_compute.c

@@ -553,25 +553,24 @@ void evergreen_emit_cs_shader(
 }
 
 static void evergreen_launch_grid(
-		struct pipe_context *ctx_,
+		struct pipe_context *ctx_, const struct pipe_grid_info *info)
-		const uint *block_layout, const uint *grid_layout,
-		uint32_t pc, const void *input)
 {
 	struct r600_context *ctx = (struct r600_context *)ctx_;
 #ifdef HAVE_OPENCL
 	struct r600_pipe_compute *shader = ctx->cs_shader_state.shader;
 	boolean use_kill;
 
-	ctx->cs_shader_state.pc = pc;
+	ctx->cs_shader_state.pc = info->pc;
 	/* Get the config information for this kernel. */
-	r600_shader_binary_read_config(&shader->binary, &shader->bc, pc, &use_kill);
+	r600_shader_binary_read_config(&shader->binary, &shader->bc,
+                                  info->pc, &use_kill);
 #endif
 
-	COMPUTE_DBG(ctx->screen, "*** evergreen_launch_grid: pc = %u\n", pc);
+	COMPUTE_DBG(ctx->screen, "*** evergreen_launch_grid: pc = %u\n", info->pc);
 
 
-	evergreen_compute_upload_input(ctx_, block_layout, grid_layout, input);
+	evergreen_compute_upload_input(ctx_, info->block, info->grid, info->input);
-	compute_emit_cs(ctx, block_layout, grid_layout);
+	compute_emit_cs(ctx, info->block, info->grid);
 }
 
 static void evergreen_set_compute_resources(struct pipe_context * ctx_,

src/gallium/drivers/radeonsi/si_compute.c

@@ -196,9 +196,7 @@ static unsigned compute_num_waves_for_scratch(
 }
 
 static void si_launch_grid(
-		struct pipe_context *ctx,
+		struct pipe_context *ctx, const struct pipe_grid_info *info)
-		const uint *block_layout, const uint *grid_layout,
-		uint32_t pc, const void *input)
 {
 	struct si_context *sctx = (struct si_context*)ctx;
 	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
@@ -232,7 +230,7 @@ static void si_launch_grid(
 	pm4->compute_pkt = true;
 
 	/* Read the config information */
-	si_shader_binary_read_config(&shader->binary, &shader->config, pc);
+	si_shader_binary_read_config(&shader->binary, &shader->config, info->pc);
 
 	/* Upload the kernel arguments */
 
@@ -242,15 +240,16 @@ static void si_launch_grid(
 	kernel_args = sctx->b.ws->buffer_map(input_buffer->buf,
 			sctx->b.gfx.cs, PIPE_TRANSFER_WRITE);
 	for (i = 0; i < 3; i++) {
-		kernel_args[i] = grid_layout[i];
+		kernel_args[i] = info->grid[i];
-		kernel_args[i + 3] = grid_layout[i] * block_layout[i];
+		kernel_args[i + 3] = info->grid[i] * info->block[i];
-		kernel_args[i + 6] = block_layout[i];
+		kernel_args[i + 6] = info->block[i];
 	}
 
 	num_waves_for_scratch =	compute_num_waves_for_scratch(
-		&sctx->screen->b.info, block_layout, grid_layout);
+		&sctx->screen->b.info, info->block, info->grid);
 
-	memcpy(kernel_args + (num_work_size_bytes / 4), input, program->input_size);
+	memcpy(kernel_args + (num_work_size_bytes / 4), info->input,
+          program->input_size);
 
 	if (shader->config.scratch_bytes_per_wave > 0) {
 
@@ -291,11 +290,11 @@ static void si_launch_grid(
 	si_pm4_set_reg(pm4, R_00B818_COMPUTE_START_Z, 0);
 
 	si_pm4_set_reg(pm4, R_00B81C_COMPUTE_NUM_THREAD_X,
-				S_00B81C_NUM_THREAD_FULL(block_layout[0]));
+				S_00B81C_NUM_THREAD_FULL(info->block[0]));
 	si_pm4_set_reg(pm4, R_00B820_COMPUTE_NUM_THREAD_Y,
-				S_00B820_NUM_THREAD_FULL(block_layout[1]));
+				S_00B820_NUM_THREAD_FULL(info->block[1]));
 	si_pm4_set_reg(pm4, R_00B824_COMPUTE_NUM_THREAD_Z,
-				S_00B824_NUM_THREAD_FULL(block_layout[2]));
+				S_00B824_NUM_THREAD_FULL(info->block[2]));
 
 	/* Global buffers */
 	for (i = 0; i < MAX_GLOBAL_BUFFERS; i++) {
@@ -323,7 +322,7 @@ static void si_launch_grid(
 	}
 
 	shader_va = shader->bo->gpu_address;
-	shader_va += pc;
+	shader_va += info->pc;
 
 	radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, shader->bo,
 				  RADEON_USAGE_READ, RADEON_PRIO_USER_SHADER);
@@ -375,9 +374,9 @@ static void si_launch_grid(
 		;
 
 	si_pm4_cmd_begin(pm4, PKT3_DISPATCH_DIRECT);
-	si_pm4_cmd_add(pm4, grid_layout[0]); /* Thread groups DIM_X */
+	si_pm4_cmd_add(pm4, info->grid[0]); /* Thread groups DIM_X */
-	si_pm4_cmd_add(pm4, grid_layout[1]); /* Thread groups DIM_Y */
+	si_pm4_cmd_add(pm4, info->grid[1]); /* Thread groups DIM_Y */
-	si_pm4_cmd_add(pm4, grid_layout[2]); /* Thread gropus DIM_Z */
+	si_pm4_cmd_add(pm4, info->grid[2]); /* Thread gropus DIM_Z */
 	si_pm4_cmd_add(pm4, 1); /* DISPATCH_INITIATOR */
         si_pm4_cmd_end(pm4, false);
 

src/gallium/include/pipe/p_context.h

@@ -48,6 +48,7 @@ struct pipe_constant_buffer;
 struct pipe_debug_callback;
 struct pipe_depth_stencil_alpha_state;
 struct pipe_draw_info;
+struct pipe_grid_info;
 struct pipe_fence_handle;
 struct pipe_framebuffer_state;
 struct pipe_image_view;
@@ -618,23 +619,9 @@ struct pipe_context {
    /**
     * Launch the compute kernel starting from instruction \a pc of the
     * currently bound compute program.
-    *
-    * \a grid_layout and \a block_layout are arrays of size \a
-    * PIPE_COMPUTE_CAP_GRID_DIMENSION that determine the layout of the
-    * grid (in block units) and working block (in thread units) to be
-    * used, respectively.
-    *
-    * \a pc For drivers that use PIPE_SHADER_IR_LLVM as their prefered IR,
-    * this value will be the index of the kernel in the opencl.kernels
-    * metadata list.
-    *
-    * \a input will be used to initialize the INPUT resource, and it
-    * should point to a buffer of at least
-    * pipe_compute_state::req_input_mem bytes.
     */
    void (*launch_grid)(struct pipe_context *context,
-                       const uint *block_layout, const uint *grid_layout,
+                       const struct pipe_grid_info *info);
-                       uint32_t pc, const void *input);
    /*@}*/
 
    /**

src/gallium/include/pipe/p_state.h

@@ -678,6 +678,33 @@ struct pipe_blit_info
    boolean alpha_blend; /* dst.rgb = src.rgb * src.a + dst.rgb * (1 - src.a) */
 };
 
+/**
+ * Information to describe a launch_grid call.
+ */
+struct pipe_grid_info
+{
+   /**
+    * For drivers that use PIPE_SHADER_IR_LLVM as their prefered IR, this value
+    * will be the index of the kernel in the opencl.kernels metadata list.
+    */
+   uint32_t pc;
+
+   /**
+    * Will be used to initialize the INPUT resource, and it should point to a
+    * buffer of at least pipe_compute_state::req_input_mem bytes.
+    */
+   void *input;
+
+   /**
+    * Determine the layout of the working block (in thread units) to be used.
+    */
+   uint block[3];
+
+   /**
+    * Determine the layout of the grid (in block units) to be used.
+    */
+   uint grid[3];
+};
 
 /**
  * Structure used as a header for serialized LLVM programs.

src/gallium/state_trackers/clover/core/kernel.cpp

@@ -55,6 +55,7 @@ kernel::launch(command_queue &q,
    const auto reduced_grid_size =
       map(divides(), grid_size, block_size);
    void *st = exec.bind(&q, grid_offset);
+   struct pipe_grid_info info;
 
    // The handles are created during exec_context::bind(), so we need make
    // sure to call exec_context::bind() before retrieving them.
@@ -74,11 +75,13 @@ kernel::launch(command_queue &q,
    q.pipe->set_global_binding(q.pipe, 0, exec.g_buffers.size(),
                               exec.g_buffers.data(), g_handles.data());
 
-   q.pipe->launch_grid(q.pipe,
+   // Fill information for the launch_grid() call.
-                       pad_vector(q, block_size, 1).data(),
+   info.block = pad_vector(q, block_size, 1).data(),
-                       pad_vector(q, reduced_grid_size, 1).data(),
+   info.grid = pad_vector(q, reduced_grid_size, 1).data(),
-                       find(name_equals(_name), m.syms).offset,
+   info.pc = find(name_equals(_name), m.sysm).offset;
-                       exec.input.data());
+   info.input = exec.input.data();
+
+   q.pipe->launch_grid(q.pipe, &info);
 
    q.pipe->set_global_binding(q.pipe, 0, exec.g_buffers.size(), NULL, NULL);
    q.pipe->set_compute_resources(q.pipe, 0, exec.resources.size(), NULL);

src/gallium/tests/trivial/compute.c

@@ -424,8 +424,17 @@ static void launch_grid(struct context *ctx, const uint *block_layout,
                         const void *input)
 {
         struct pipe_context *pipe = ctx->pipe;
+        struct pipe_grid_info info;
+        int i;
 
-        pipe->launch_grid(pipe, block_layout, grid_layout, pc, input);
+        for (i = 0; i < 3; i++) {
+                info.block[i] = block_layout[i];
+                info.grid[i] = grid_layout[i];
+        }
+        info.pc = pc;
+        info.input = input;
+
+        pipe->launch_grid(pipe, &info);
 }
 
 static void test_default_init(void *p, int s, int x, int y)