swr: simd16 shaders work in progress

Start building vertex shaders as simd16.

Disabled by default, set USE_SIMD16_SHADERS in knobs.h to experiment.

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

3 files changed, 21 insertions, 2 deletions

diff --git a/src/gallium/drivers/swr/swr_screen.cpp b/src/gallium/drivers/swr/swr_screen.cpp
index 639b18f930..46b3a003c6 100644
--- a/src/gallium/drivers/swr/swr_screen.cpp
+++ b/src/gallium/drivers/swr/swr_screen.cpp
@@ -1058,6 +1058,9 @@ swr_destroy_screen(struct pipe_screen *p_screen)
    swr_fence_reference(p_screen, &screen->flush_fence, NULL);
 
    JitDestroyContext(screen->hJitMgr);
+#if USE_SIMD16_SHADERS
+   JitDestroyContext(screen->hJitMgr16);
+#endif
 
    if (winsys->destroy)
       winsys->destroy(winsys);
@@ -1141,6 +1144,9 @@ swr_create_screen_internal(struct sw_winsys *winsys)
 
    // Pass in "" for architecture for run-time determination
    screen->hJitMgr = JitCreateContext(KNOB_SIMD_WIDTH, "", "swr");
+#if USE_SIMD16_SHADERS
+   screen->hJitMgr16 = JitCreateContext(16, "", "swr");
+#endif
 
    swr_fence_init(&screen->base);
 
diff --git a/src/gallium/drivers/swr/swr_screen.h b/src/gallium/drivers/swr/swr_screen.h
index a11ea9f41d..1c4e331583 100644
--- a/src/gallium/drivers/swr/swr_screen.h
+++ b/src/gallium/drivers/swr/swr_screen.h
@@ -49,6 +49,9 @@ struct swr_screen {
    uint32_t client_copy_limit;
 
    HANDLE hJitMgr;
+#if USE_SIMD16_SHADERS
+   HANDLE hJitMgr16;
+#endif
 
    PFNSwrGetInterface pfnSwrGetInterface;
 };
diff --git a/src/gallium/drivers/swr/swr_shader.cpp b/src/gallium/drivers/swr/swr_shader.cpp
index 510bc0e457..732e08dae7 100644
--- a/src/gallium/drivers/swr/swr_shader.cpp
+++ b/src/gallium/drivers/swr/swr_shader.cpp
@@ -693,7 +693,7 @@ swr_compile_gs(struct swr_context *ctx, swr_jit_gs_key &key)
 void
 BuilderSWR::WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput, unsigned slot, unsigned channel)
 {
-#if USE_SIMD16_FRONTEND
+#if USE_SIMD16_FRONTEND && !USE_SIMD16_SHADERS
    // interleave the simdvertex components into the dest simd16vertex
    //   slot16offset = slot8offset * 2
    //   comp16offset = comp8offset * 2 + alternateOffset
@@ -756,6 +756,9 @@ BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
    const_sizes_ptr->setName("num_vs_constants");
 
    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
+#if USE_SIMD16_SHADERS
+   vtxInput = BITCAST(vtxInput, PointerType::get(Gen_simd16vertex(JM()), 0));
+#endif
 
    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
@@ -777,7 +780,7 @@ BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 
    lp_build_tgsi_soa(gallivm,
                      swr_vs->pipe.tokens,
-                     lp_type_float_vec(32, 32 * 8),
+                     lp_type_float_vec(32, 32 * mVWidth),
                      NULL, // mask
                      wrap(consts_ptr),
                      wrap(const_sizes_ptr),
@@ -795,6 +798,9 @@ BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 
    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
+#if USE_SIMD16_SHADERS
+   vtxOutput = BITCAST(vtxOutput, PointerType::get(Gen_simd16vertex(JM()), 0));
+#endif
 
    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
@@ -905,7 +911,11 @@ swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
       return NULL;
 
    BuilderSWR builder(
+#if USE_SIMD16_SHADERS
+      reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr16),
+#else
       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
+#endif
       "VS");
    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);