vc4: Add simulator kernel validation for multithreaded fragment shaders.

This is Jonas Pfeil's code from the kernel, brought back to Mesa by
anholt.

3 files changed, 76 insertions, 5 deletions

diff --git a/src/gallium/drivers/vc4/kernel/vc4_drv.h b/src/gallium/drivers/vc4/kernel/vc4_drv.h
index 90f45397d83..8f5ed00d96f 100644
--- a/src/gallium/drivers/vc4/kernel/vc4_drv.h
+++ b/src/gallium/drivers/vc4/kernel/vc4_drv.h
@@ -150,6 +150,8 @@ struct vc4_validated_shader_info
 
 	uint32_t num_uniform_addr_offsets;
 	uint32_t *uniform_addr_offsets;
+
+	bool is_threaded;
 };
 
 /* vc4_validate.c */
diff --git a/src/gallium/drivers/vc4/kernel/vc4_validate.c b/src/gallium/drivers/vc4/kernel/vc4_validate.c
index a9dce1fa379..bd193b993b3 100644
--- a/src/gallium/drivers/vc4/kernel/vc4_validate.c
+++ b/src/gallium/drivers/vc4/kernel/vc4_validate.c
@@ -780,11 +780,6 @@ validate_gl_shader_rec(struct drm_device *dev,
 	exec->shader_rec_v += roundup(packet_size, 16);
 	exec->shader_rec_size -= packet_size;
 
-	if (!(*(uint16_t *)pkt_u & VC4_SHADER_FLAG_FS_SINGLE_THREAD)) {
-		DRM_ERROR("Multi-threaded fragment shaders not supported.\n");
-		return -EINVAL;
-	}
-
 	for (i = 0; i < shader_reloc_count; i++) {
 		if (src_handles[i] > exec->bo_count) {
 			DRM_ERROR("Shader handle %d too big\n", src_handles[i]);
@@ -801,6 +796,18 @@ validate_gl_shader_rec(struct drm_device *dev,
 			return -EINVAL;
 	}
 
+	if (((*(uint16_t *)pkt_u & VC4_SHADER_FLAG_FS_SINGLE_THREAD) == 0) !=
+	    to_vc4_bo(&bo[0]->base)->validated_shader->is_threaded) {
+		DRM_ERROR("Thread mode of CL and FS do not match\n");
+		return -EINVAL;
+	}
+
+	if (to_vc4_bo(&bo[1]->base)->validated_shader->is_threaded ||
+	    to_vc4_bo(&bo[2]->base)->validated_shader->is_threaded) {
+		DRM_ERROR("cs and vs cannot be threaded\n");
+		return -EINVAL;
+	}
+
 	for (i = 0; i < shader_reloc_count; i++) {
 		struct vc4_validated_shader_info *validated_shader;
 		uint32_t o = shader_reloc_offsets[i];
diff --git a/src/gallium/drivers/vc4/kernel/vc4_validate_shaders.c b/src/gallium/drivers/vc4/kernel/vc4_validate_shaders.c
index 0ff3d01f3f2..d93f5239d7d 100644
--- a/src/gallium/drivers/vc4/kernel/vc4_validate_shaders.c
+++ b/src/gallium/drivers/vc4/kernel/vc4_validate_shaders.c
@@ -84,6 +84,14 @@ struct vc4_shader_validation_state {
 	 * basic blocks.
 	 */
 	bool needs_uniform_address_for_loop;
+
+	/* Set when we find an instruction which violates the criterion for a
+	 * threaded shader. These are:
+	 * 	- only write the lower half of the register space
+	 * 	- last thread switch signaled at the end
+	 * So track the usage of the thread switches and the register usage.
+	 */
+	bool all_registers_used;
 };
 
 static uint32_t
@@ -119,6 +127,12 @@ raddr_add_a_to_live_reg_index(uint64_t inst)
 		return ~0;
 }
 
+static bool live_reg_is_upper_half(uint32_t lri)
+{
+	return	(lri >=16 && lri < 32) ||
+		(lri >=32 + 16 && lri < 32 + 32);
+}
+
 static bool
 is_tmu_submit(uint32_t waddr)
 {
@@ -385,6 +399,9 @@ check_reg_write(struct vc4_validated_shader_info *validated_shader,
 		} else {
 			validation_state->live_immediates[lri] = ~0;
 		}
+
+		if (live_reg_is_upper_half(lri))
+			validation_state->all_registers_used = true;
 	}
 
 	switch (waddr) {
@@ -593,6 +610,11 @@ check_instruction_reads(struct vc4_validated_shader_info *validated_shader,
 		}
 	}
 
+	if ((raddr_a >= 16 && raddr_a < 32) ||
+	    (raddr_b >= 16 && raddr_b < 32 && sig != QPU_SIG_SMALL_IMM)) {
+		validation_state->all_registers_used = true;
+	}
+
 	return true;
 }
 
@@ -748,6 +770,7 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 {
 	bool found_shader_end = false;
 	int shader_end_ip = 0;
+	uint32_t last_thread_switch_ip = -3;
 	uint32_t ip;
 	struct vc4_validated_shader_info *validated_shader = NULL;
 	struct vc4_shader_validation_state validation_state;
@@ -780,6 +803,16 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 		if (!vc4_handle_branch_target(&validation_state))
 			goto fail;
 
+		if (ip == last_thread_switch_ip + 3) {
+			/* Reset r0-r3 live clamp data */
+			int i;
+			for (i = 64; i < LIVE_REG_COUNT; i++) {
+				validation_state.live_min_clamp_offsets[i] = ~0;
+				validation_state.live_max_clamp_regs[i] = false;
+				validation_state.live_immediates[i] = ~0;
+			}
+		}
+
 		switch (sig) {
 		case QPU_SIG_NONE:
 		case QPU_SIG_WAIT_FOR_SCOREBOARD:
@@ -789,6 +822,8 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 		case QPU_SIG_LOAD_TMU1:
 		case QPU_SIG_PROG_END:
 		case QPU_SIG_SMALL_IMM:
+		case QPU_SIG_THREAD_SWITCH:
+		case QPU_SIG_LAST_THREAD_SWITCH:
 			if (!check_instruction_writes(validated_shader,
 						      &validation_state)) {
 				DRM_ERROR("Bad write at ip %d\n", ip);
@@ -804,6 +839,18 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 				shader_end_ip = ip;
 			}
 
+			if (sig == QPU_SIG_THREAD_SWITCH ||
+			    sig == QPU_SIG_LAST_THREAD_SWITCH) {
+				validated_shader->is_threaded = true;
+
+				if (ip < last_thread_switch_ip + 3) {
+					DRM_ERROR("Thread switch too soon after "
+						  "last switch at ip %d\n", ip);
+					goto fail;
+				}
+				last_thread_switch_ip = ip;
+			}
+
 			break;
 
 		case QPU_SIG_LOAD_IMM:
@@ -818,6 +865,13 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 			if (!check_branch(inst, validated_shader,
 					  &validation_state, ip))
 				goto fail;
+
+			if (ip < last_thread_switch_ip + 3) {
+				DRM_ERROR("Branch in thread switch at ip %d",
+					  ip);
+				goto fail;
+			}
+
 			break;
 		default:
 			DRM_ERROR("Unsupported QPU signal %d at "
@@ -839,6 +893,14 @@ vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
 		goto fail;
 	}
 
+	/* Might corrupt other thread */
+	if (validated_shader->is_threaded &&
+	    validation_state.all_registers_used) {
+		DRM_ERROR("Shader uses threading, but uses the upper "
+			  "half of the registers, too\n");
+		goto fail;
+	}
+
 	/* If we did a backwards branch and we haven't emitted a uniforms
 	 * reset since then, we still need the uniforms stream to have the
 	 * uniforms address available so that the backwards branch can do its