panfrost: Implement a disk cache

Wire up the Mesa shader disk cache into Panfrost. Coupled with the
precompiles from the previous patch, this should greatly reduce shader
recompile jank.

This is a bare bones implementation. Obvious future work includes:

- Caching internal (outside of Gallium) shaders
- Implement finalize_nir to reduce on disk size of shaders

That doesn't need to come in this patch.

This patch does shuffle some allocation patterns around to avoid extra
nir_shader_clones, but the result should be pretty clean.

---

Consider dEQP-GLES31.functional.ssbo.layout.basic_unsized_array.* in the CTS.
With a cold cache:

   44.11user 0.66system 0:45.44elapsed 98%CPU (0avgtext+0avgdata 267804maxresident)
   k 0inputs+0outputs (130major+74725minor)pagefaults 0swaps

But with this commit and a warm cache:

   4.07user 0.35system 0:04.56elapsed 96%CPU (0avgtext+0avgdata 211012maxresident)
   k0inputs+0outputs (1major+49489minor)pagefaults 0swaps

That's an 11x improvement!

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19363>

src/gallium/drivers/panfrost/meson.build

@@ -20,6 +20,7 @@
 # SOFTWARE.
 
 files_panfrost = files(
+  'pan_disk_cache.c',
   'pan_helpers.c',
   'pan_public.h',
   'pan_screen.c',

src/gallium/drivers/panfrost/pan_context.c

@@ -51,6 +51,7 @@
 #include "pan_util.h"
 #include "decode.h"
 #include "util/pan_lower_framebuffer.h"
+#include "compiler/nir/nir_serialize.h"
 
 static void
 panfrost_clear(

src/gallium/drivers/panfrost/pan_context.h

@@ -279,8 +279,18 @@ struct panfrost_fs_key {
 };
 
 struct panfrost_shader_key {
-        /* If we need vertex shader keys, union it in */
-        struct panfrost_fs_key fs;
+        union {
+                /* Vertex shaders do not use shader keys. However, we have a
+                 * special "transform feedback" vertex program derived from a
+                 * vertex shader. If vs_is_xfb is set on a vertex shader, this
+                 * is a transform feedback shader, else it is a regular
+                 * (unkeyed) vertex shader.
+                 */
+                bool vs_is_xfb;
+
+                /* Fragment shaders use regular shader keys */
+                struct panfrost_fs_key fs;
+        };
 };
 
 struct panfrost_compiled_shader {
@@ -308,7 +318,14 @@ struct panfrost_compiled_shader {
 
 /* Shader CSO */
 struct panfrost_uncompiled_shader {
-        nir_shader *nir;
+        /* NIR for the shader. For graphics, this will be non-NULL even for
+         * TGSI. For compute, this will be NULL after the shader is compiled,
+         * as we don't need any compute variants.
+         */
+        const nir_shader *nir;
+
+        /* A SHA1 of the serialized NIR for the disk cache. */
+        unsigned char nir_sha1[20];
 
         /* Stream output information */
         struct pipe_stream_output_info stream_output;
@@ -329,6 +346,35 @@ struct panfrost_uncompiled_shader {
         uint32_t fixed_varying_mask;
 };
 
+/* The binary artefacts of compiling a shader. This differs from
+ * panfrost_compiled_shader, which adds extra metadata beyond compiling but
+ * throws away information not needed after the initial compile.
+ *
+ * This structure is serialized for the shader disk cache.
+ */
+struct panfrost_shader_binary {
+        /* Collected information about the compiled shader */
+        struct pan_shader_info info;
+
+        /* The binary itself */
+        struct util_dynarray binary;
+};
+
+void
+panfrost_disk_cache_store(struct disk_cache *cache,
+                          const struct panfrost_uncompiled_shader *uncompiled,
+                          const struct panfrost_shader_key *key,
+                          const struct panfrost_shader_binary *binary);
+
+bool
+panfrost_disk_cache_retrieve(struct disk_cache *cache,
+                             const struct panfrost_uncompiled_shader *uncompiled,
+                             const struct panfrost_shader_key *key,
+                             struct panfrost_shader_binary *binary);
+
+void
+panfrost_disk_cache_init(struct panfrost_screen *screen);
+
 /** (Vertex buffer index, divisor) tuple that will become an Attribute Buffer
  * Descriptor at draw-time on Midgard
  */

src/gallium/drivers/panfrost/pan_disk_cache.c

@@ -0,0 +1,175 @@
+/*
+ * Copyright (c) 2022 Amazon.com, Inc. or its affiliates.
+ * Copyright © 2018 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.
+ */
+
+#include <stdio.h>
+#include <stdint.h>
+#include <assert.h>
+#include <string.h>
+
+#include "compiler/nir/nir.h"
+#include "util/blob.h"
+#include "util/build_id.h"
+#include "util/disk_cache.h"
+#include "util/mesa-sha1.h"
+
+#include "pan_context.h"
+
+static bool debug = false;
+
+extern int midgard_debug;
+extern int bifrost_debug;
+
+/**
+ * Compute a disk cache key for the given uncompiled shader and shader key.
+ */
+static void
+panfrost_disk_cache_compute_key(struct disk_cache *cache,
+                                const struct panfrost_uncompiled_shader *uncompiled,
+                                const struct panfrost_shader_key *shader_key,
+                                cache_key cache_key)
+{
+        uint8_t data[sizeof(uncompiled->nir_sha1) + sizeof(*shader_key)];
+
+        memcpy(data, uncompiled->nir_sha1, sizeof(uncompiled->nir_sha1));
+        memcpy(data + sizeof(uncompiled->nir_sha1), shader_key, sizeof(*shader_key));
+
+        disk_cache_compute_key(cache, data, sizeof(data), cache_key);
+}
+
+/**
+ * Store the given compiled shader in the disk cache.
+ *
+ * This should only be called on newly compiled shaders.  No checking is
+ * done to prevent repeated stores of the same shader.
+ */
+void
+panfrost_disk_cache_store(struct disk_cache *cache,
+                          const struct panfrost_uncompiled_shader *uncompiled,
+                          const struct panfrost_shader_key *key,
+                          const struct panfrost_shader_binary *binary)
+{
+#ifdef ENABLE_SHADER_CACHE
+        if (!cache)
+                return;
+
+        cache_key cache_key;
+        panfrost_disk_cache_compute_key(cache, uncompiled, key, cache_key);
+
+        if (debug) {
+                char sha1[41];
+                _mesa_sha1_format(sha1, cache_key);
+                fprintf(stderr, "[mesa disk cache] storing %s\n", sha1);
+        }
+
+        struct blob blob;
+        blob_init(&blob);
+
+        /* We write the following data to the cache blob:
+         *
+         * 1. Size of program binary
+         * 2. Program binary
+         * 3. Shader info
+         */
+        blob_write_uint32(&blob, binary->binary.size);
+        blob_write_bytes(&blob, binary->binary.data, binary->binary.size);
+        blob_write_bytes(&blob, &binary->info, sizeof(binary->info));
+
+        disk_cache_put(cache, cache_key, blob.data, blob.size, NULL);
+        blob_finish(&blob);
+#endif
+}
+
+/**
+ * Search for a compiled shader in the disk cache.
+ */
+bool
+panfrost_disk_cache_retrieve(struct disk_cache *cache,
+                             const struct panfrost_uncompiled_shader *uncompiled,
+                             const struct panfrost_shader_key *key,
+                             struct panfrost_shader_binary *binary)
+{
+#ifdef ENABLE_SHADER_CACHE
+        if (!cache)
+                return false;
+
+        cache_key cache_key;
+        panfrost_disk_cache_compute_key(cache, uncompiled, key, cache_key);
+
+        if (debug) {
+                char sha1[41];
+                _mesa_sha1_format(sha1, cache_key);
+                fprintf(stderr, "[mesa disk cache] retrieving %s: ", sha1);
+        }
+
+        size_t size;
+        void *buffer = disk_cache_get(cache, cache_key, &size);
+
+        if (debug)
+                fprintf(stderr, "%s\n", buffer ? "found" : "missing");
+
+        if (!buffer)
+                return false;
+
+        struct blob_reader blob;
+        blob_reader_init(&blob, buffer, size);
+
+        util_dynarray_init(&binary->binary, NULL);
+
+        uint32_t binary_size = blob_read_uint32(&blob);
+        void *ptr = util_dynarray_resize_bytes(&binary->binary, binary_size, 1);
+
+        blob_copy_bytes(&blob, ptr, binary_size);
+        blob_copy_bytes(&blob, &binary->info, sizeof(binary->info));
+
+        return true;
+#else
+        return false;
+#endif
+}
+
+/**
+ * Initialize the on-disk shader cache.
+ */
+void
+panfrost_disk_cache_init(struct panfrost_screen *screen)
+{
+#ifdef ENABLE_SHADER_CACHE
+        const char *renderer = screen->base.get_name(&screen->base);
+
+        const struct build_id_note *note =
+                build_id_find_nhdr_for_addr(panfrost_disk_cache_init);
+        assert(note && build_id_length(note) == 20); /* sha1 */
+
+        const uint8_t *id_sha1 = build_id_data(note);
+        assert(id_sha1);
+
+        char timestamp[41];
+        _mesa_sha1_format(timestamp, id_sha1);
+
+        /* Consider any flags affecting the compile when caching */
+        uint64_t driver_flags = screen->dev.debug;
+        driver_flags |= ((uint64_t) (midgard_debug | bifrost_debug) << 32);
+
+        screen->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
+#endif
+}

src/gallium/drivers/panfrost/pan_screen.c

@@ -753,6 +753,8 @@ panfrost_destroy_screen(struct pipe_screen *pscreen)
         if (dev->ro)
                 dev->ro->destroy(dev->ro);
         panfrost_close_device(dev);
+
+        disk_cache_destroy(screen->disk_cache);
         ralloc_free(pscreen);
 }
 
@@ -853,6 +855,12 @@ panfrost_screen_get_compiler_options(struct pipe_screen *pscreen,
         return pan_screen(pscreen)->vtbl.get_compiler_options();
 }
 
+static struct disk_cache *
+panfrost_get_disk_shader_cache(struct pipe_screen *pscreen)
+{
+        return pan_screen(pscreen)->disk_cache;
+}
+
 struct pipe_screen *
 panfrost_create_screen(int fd, struct renderonly *ro)
 {
@@ -896,12 +904,16 @@ panfrost_create_screen(int fd, struct renderonly *ro)
                panfrost_is_dmabuf_modifier_supported;
         screen->base.context_create = panfrost_create_context;
         screen->base.get_compiler_options = panfrost_screen_get_compiler_options;
+        screen->base.get_disk_shader_cache = panfrost_get_disk_shader_cache;
         screen->base.fence_reference = panfrost_fence_reference;
         screen->base.fence_finish = panfrost_fence_finish;
         screen->base.set_damage_region = panfrost_resource_set_damage_region;
 
         panfrost_resource_screen_init(&screen->base);
         pan_blend_shaders_init(dev);
+
+        panfrost_disk_cache_init(screen);
+
         panfrost_pool_init(&screen->indirect_draw.bin_pool, NULL, dev,
                            PAN_BO_EXECUTE, 65536, "Indirect draw shaders",
                            false, true);

src/gallium/drivers/panfrost/pan_screen.h

@@ -37,6 +37,7 @@
 #include "util/bitset.h"
 #include "util/set.h"
 #include "util/log.h"
+#include "util/disk_cache.h"
 
 #include "pan_device.h"
 #include "pan_mempool.h"
@@ -107,6 +108,7 @@ struct panfrost_screen {
         } indirect_draw;
 
         struct panfrost_vtable vtbl;
+        struct disk_cache *disk_cache;
 };
 
 static inline struct panfrost_screen *

src/gallium/drivers/panfrost/pan_shader.c

@@ -37,12 +37,26 @@
 #include "nir_serialize.h"
 
 static struct panfrost_uncompiled_shader *
-panfrost_alloc_shader(void)
+panfrost_alloc_shader(const nir_shader *nir)
 {
-        struct panfrost_uncompiled_shader *so = CALLOC_STRUCT(panfrost_uncompiled_shader);
+        struct panfrost_uncompiled_shader *so =
+                rzalloc(NULL, struct panfrost_uncompiled_shader);
 
         simple_mtx_init(&so->lock, mtx_plain);
-        util_dynarray_init(&so->variants, NULL);
+        util_dynarray_init(&so->variants, so);
+
+        so->nir = nir;
+
+        /* Serialize the NIR to a binary blob that we can hash for the disk
+         * cache. Drop unnecessary information (like variable names) so the
+         * serialized NIR is smaller, and also to let us detect more isomorphic
+         * shaders when hashing, increasing cache hits.
+         */
+        struct blob blob;
+        blob_init(&blob);
+        nir_serialize(&blob, nir, true);
+        _mesa_sha1_compute(blob.data, blob.size, so->nir_sha1);
+        blob_finish(&blob);
 
         return so;
 }
@@ -54,17 +68,15 @@ panfrost_alloc_variant(struct panfrost_uncompiled_shader *so)
 }
 
 static void
-panfrost_shader_compile(struct pipe_screen *pscreen,
-                        struct panfrost_pool *shader_pool,
-                        struct panfrost_pool *desc_pool,
+panfrost_shader_compile(struct panfrost_screen *screen,
                         const nir_shader *ir,
                         struct util_debug_callback *dbg,
-                        struct panfrost_compiled_shader *state,
+                        struct panfrost_shader_key *key,
                         unsigned req_local_mem,
-                        unsigned fixed_varying_mask)
+                        unsigned fixed_varying_mask,
+                        struct panfrost_shader_binary *out)
 {
-        struct panfrost_screen *screen = pan_screen(pscreen);
-        struct panfrost_device *dev = pan_device(pscreen);
+        struct panfrost_device *dev = pan_device(&screen->base);
 
         nir_shader *s = nir_shader_clone(NULL, ir);
 
@@ -76,27 +88,27 @@ panfrost_shader_compile(struct pipe_screen *pscreen,
 
         /* Lower this early so the backends don't have to worry about it */
         if (s->info.stage == MESA_SHADER_FRAGMENT) {
-                inputs.fixed_varying_mask = state->key.fs.fixed_varying_mask;
+                inputs.fixed_varying_mask = key->fs.fixed_varying_mask;
 
                 if (s->info.outputs_written & BITFIELD_BIT(FRAG_RESULT_COLOR)) {
                         NIR_PASS_V(s, nir_lower_fragcolor,
-                                   state->key.fs.nr_cbufs_for_fragcolor);
+                                   key->fs.nr_cbufs_for_fragcolor);
                 }
 
-                if (state->key.fs.sprite_coord_enable) {
+                if (key->fs.sprite_coord_enable) {
                         NIR_PASS_V(s, nir_lower_texcoord_replace,
-                                   state->key.fs.sprite_coord_enable,
+                                   key->fs.sprite_coord_enable,
                                    true /* point coord is sysval */,
                                    false /* Y-invert */);
                 }
 
-                if (state->key.fs.clip_plane_enable) {
+                if (key->fs.clip_plane_enable) {
                         NIR_PASS_V(s, nir_lower_clip_fs,
-                                   state->key.fs.clip_plane_enable,
+                                   key->fs.clip_plane_enable,
                                    false);
                 }
 
-                memcpy(inputs.rt_formats, state->key.fs.rt_formats, sizeof(inputs.rt_formats));
+                memcpy(inputs.rt_formats, key->fs.rt_formats, sizeof(inputs.rt_formats));
         } else if (s->info.stage == MESA_SHADER_VERTEX) {
                 inputs.fixed_varying_mask = fixed_varying_mask;
 
@@ -104,41 +116,67 @@ panfrost_shader_compile(struct pipe_screen *pscreen,
                 inputs.no_idvs = s->info.has_transform_feedback_varyings;
         }
 
-        struct util_dynarray binary;
+        util_dynarray_init(&out->binary, NULL);
+        screen->vtbl.compile_shader(s, &inputs, &out->binary, &out->info);
 
-        util_dynarray_init(&binary, NULL);
-        screen->vtbl.compile_shader(s, &inputs, &binary, &state->info);
+        assert(req_local_mem >= out->info.wls_size);
+        out->info.wls_size = req_local_mem;
 
-        assert(req_local_mem >= state->info.wls_size);
-        state->info.wls_size = req_local_mem;
+        /* In both clone and tgsi_to_nir paths, the shader is ralloc'd against
+         * a NULL context
+         */
+        ralloc_free(s);
+}
 
-        if (binary.size) {
-                state->bin = panfrost_pool_take_ref(shader_pool,
-                        pan_pool_upload_aligned(&shader_pool->base,
-                                binary.data, binary.size, 128));
+static void
+panfrost_shader_get(struct pipe_screen *pscreen,
+                    struct panfrost_pool *shader_pool,
+                    struct panfrost_pool *desc_pool,
+                    struct panfrost_uncompiled_shader *uncompiled,
+                    struct util_debug_callback *dbg,
+                    struct panfrost_compiled_shader *state,
+                    unsigned req_local_mem)
+{
+        struct panfrost_screen *screen = pan_screen(pscreen);
+        struct panfrost_device *dev = pan_device(pscreen);
+
+        struct panfrost_shader_binary res = { 0 };
+
+        /* Try to retrieve the variant from the disk cache. If that fails,
+         * compile a new variant and store in the disk cache for later reuse.
+         */
+        if (!panfrost_disk_cache_retrieve(screen->disk_cache, uncompiled, &state->key, &res)) {
+                panfrost_shader_compile(screen, uncompiled->nir, dbg, &state->key,
+                                        req_local_mem,
+                                        uncompiled->fixed_varying_mask, &res);
+
+                panfrost_disk_cache_store(screen->disk_cache, uncompiled, &state->key, &res);
         }
 
+        state->info = res.info;
+
+        if (res.binary.size) {
+                state->bin = panfrost_pool_take_ref(shader_pool,
+                        pan_pool_upload_aligned(&shader_pool->base,
+                                res.binary.data, res.binary.size, 128));
+        }
+
+        util_dynarray_fini(&res.binary);
 
         /* Don't upload RSD for fragment shaders since they need draw-time
          * merging for e.g. depth/stencil/alpha. RSDs are replaced by simpler
          * shader program descriptors on Valhall, which can be preuploaded even
          * for fragment shaders. */
-        bool upload = !(s->info.stage == MESA_SHADER_FRAGMENT && dev->arch <= 7);
+        bool upload = !(uncompiled->nir->info.stage == MESA_SHADER_FRAGMENT && dev->arch <= 7);
         screen->vtbl.prepare_shader(state, desc_pool, upload);
 
         panfrost_analyze_sysvals(state);
-
-        util_dynarray_fini(&binary);
-
-        /* In both clone and tgsi_to_nir paths, the shader is ralloc'd against
-         * a NULL context */
-        ralloc_free(s);
 }
 
 static void
 panfrost_build_key(struct panfrost_context *ctx,
                    struct panfrost_shader_key *key,
-                   nir_shader *nir)
+                   const nir_shader *nir)
 {
         /* We don't currently have vertex shader variants */
         if (nir->info.stage != MESA_SHADER_FRAGMENT)
@@ -237,10 +275,8 @@ panfrost_new_variant_locked(
                 .stream_output = uncompiled->stream_output,
         };
 
-        panfrost_shader_compile(ctx->base.screen,
-                                &ctx->shaders, &ctx->descs, uncompiled->nir,
-                                &ctx->base.debug, prog, 0,
-                                uncompiled->fixed_varying_mask);
+        panfrost_shader_get(ctx->base.screen, &ctx->shaders, &ctx->descs,
+                            uncompiled, &ctx->base.debug, prog, 0);
 
         /* Fixup the stream out information */
         prog->so_mask =
@@ -333,14 +369,19 @@ panfrost_create_shader_state(
         struct pipe_context *pctx,
         const struct pipe_shader_state *cso)
 {
-        struct panfrost_uncompiled_shader *so = panfrost_alloc_shader();
+        nir_shader *nir = (cso->type == PIPE_SHADER_IR_TGSI) ?
+                          tgsi_to_nir(cso->tokens, pctx->screen, false) :
+                          cso->ir.nir;
+
+        struct panfrost_uncompiled_shader *so = panfrost_alloc_shader(nir);
+
+        /* The driver gets ownership of the nir_shader for graphics. The NIR is
+         * ralloc'd. Free the NIR when we free the uncompiled shader.
+         */
+        ralloc_steal(so, nir);
 
         so->stream_output = cso->stream_output;
-
-        if (cso->type == PIPE_SHADER_IR_TGSI)
-                so->nir = tgsi_to_nir(cso->tokens, pctx->screen, false);
-        else
-                so->nir = cso->ir.nir;
+        so->nir = nir;
 
         /* Fix linkage early */
         if (so->nir->info.stage == MESA_SHADER_VERTEX) {
@@ -353,7 +394,6 @@ panfrost_create_shader_state(
          * feedback program. This is a special shader variant.
          */
         struct panfrost_context *ctx = pan_context(pctx);
-        struct util_debug_callback *dbg = &ctx->base.debug;
 
         if (so->nir->xfb_info) {
                 nir_shader *xfb = nir_shader_clone(NULL, so->nir);
@@ -361,14 +401,15 @@ panfrost_create_shader_state(
                 xfb->info.internal = true;
 
                 so->xfb = calloc(1, sizeof(struct panfrost_compiled_shader));
-                panfrost_shader_compile(pctx->screen, &ctx->shaders,
-                                        &ctx->descs, xfb, dbg, so->xfb, 0,
-                                        so->fixed_varying_mask);
+                so->xfb->key.vs_is_xfb = true;
+
+                panfrost_shader_get(ctx->base.screen, &ctx->shaders, &ctx->descs,
+                                    so, &ctx->base.debug, so->xfb, 0);
 
                 /* Since transform feedback is handled via the transform
                  * feedback program, the original program no longer uses XFB
                  */
-                so->nir->info.has_transform_feedback_varyings = false;
+                nir->info.has_transform_feedback_varyings = false;
         }
 
         /* Compile the program. We don't use vertex shader keys, so there will
@@ -401,14 +442,10 @@ panfrost_create_shader_state(
 }
 
 static void
-panfrost_delete_shader_state(
-        struct pipe_context *pctx,
-        void *so)
+panfrost_delete_shader_state(struct pipe_context *pctx, void *so)
 {
         struct panfrost_uncompiled_shader *cso = (struct panfrost_uncompiled_shader *) so;
 
-        ralloc_free(cso->nir);
-
         util_dynarray_foreach(&cso->variants, struct panfrost_compiled_shader, so) {
                 panfrost_bo_unreference(so->bin.bo);
                 panfrost_bo_unreference(so->state.bo);
@@ -424,8 +461,7 @@ panfrost_delete_shader_state(
 
         simple_mtx_destroy(&cso->lock);
 
-        util_dynarray_fini(&cso->variants);
-        free(so);
+        ralloc_free(so);
 }
 
 /*
@@ -438,15 +474,19 @@ panfrost_create_compute_state(
         const struct pipe_compute_state *cso)
 {
         struct panfrost_context *ctx = pan_context(pctx);
-        struct panfrost_uncompiled_shader *so = panfrost_alloc_shader();
+        struct panfrost_uncompiled_shader *so = panfrost_alloc_shader(cso->prog);
         struct panfrost_compiled_shader *v = panfrost_alloc_variant(so);
         memset(v, 0, sizeof *v);
 
         assert(cso->ir_type == PIPE_SHADER_IR_NIR && "TGSI kernels unsupported");
 
-        panfrost_shader_compile(pctx->screen, &ctx->shaders, &ctx->descs,
-                                cso->prog, &ctx->base.debug, v,
-                                cso->req_local_mem, 0);
+        panfrost_shader_get(pctx->screen, &ctx->shaders, &ctx->descs,
+                            so, &ctx->base.debug, v, cso->req_local_mem);
+
+        /* The NIR becomes invalid after this. For compute kernels, we never
+         * need to access it again. Don't keep a dangling pointer around.
+         */
+        so->nir = NULL;
 
         return so;
 }
@@ -463,16 +503,6 @@ panfrost_bind_compute_state(struct pipe_context *pipe, void *cso)
                 uncompiled ? util_dynarray_begin(&uncompiled->variants) : NULL;
 }
 
-static void
-panfrost_delete_compute_state(struct pipe_context *pipe, void *cso)
-{
-        struct panfrost_uncompiled_shader *so =
-                (struct panfrost_uncompiled_shader *)cso;
-
-        util_dynarray_fini(&so->variants);
-        free(cso);
-}
-
 void
 panfrost_shader_context_init(struct pipe_context *pctx)
 {
@@ -486,5 +516,5 @@ panfrost_shader_context_init(struct pipe_context *pctx)
 
         pctx->create_compute_state = panfrost_create_compute_state;
         pctx->bind_compute_state = panfrost_bind_compute_state;
-        pctx->delete_compute_state = panfrost_delete_compute_state;
+        pctx->delete_compute_state = panfrost_delete_shader_state;
 }